nn.py 527.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
Y
Yu Yang 已提交
14
"""
15
All layers just related to the neural network.
Y
Yu Yang 已提交
16 17
"""

18 19
from __future__ import print_function

20
import numpy as np
21
import warnings
S
sneaxiy 已提交
22
import six
P
peizhilin 已提交
23
import os
S
sneaxiy 已提交
24
import inspect
Y
Yu Yang 已提交
25
from ..layer_helper import LayerHelper
26
from ..initializer import Normal, Constant, NumpyArrayInitializer
L
lujun 已提交
27
from ..framework import Variable, OpProtoHolder, in_dygraph_mode
L
lujun 已提交
28
from ..dygraph import base
Y
yangyaming 已提交
29
from ..param_attr import ParamAttr
S
sneaxiy 已提交
30
from .layer_function_generator import autodoc, templatedoc, _generate_doc_string_
31
from .tensor import concat, assign, fill_constant, zeros
32
from . import utils
F
fengjiayi 已提交
33
from .. import unique_name
34
from functools import reduce
35
from .. import core
L
lujun 已提交
36
from ..dygraph import layers
37
from ..data_feeder import convert_dtype, check_type_and_dtype, check_type, check_dtype
Y
Yu Yang 已提交
38 39

__all__ = [
X
Xin Pan 已提交
40 41 42 43 44 45 46 47 48 49 50
    'fc',
    'embedding',
    'linear_chain_crf',
    'crf_decoding',
    'cos_sim',
    'chunk_eval',
    'conv2d',
    'conv3d',
    'softmax',
    'pool2d',
    'pool3d',
51 52
    'adaptive_pool2d',
    'adaptive_pool3d',
X
Xin Pan 已提交
53
    'batch_norm',
L
lvmengsi 已提交
54
    'instance_norm',
H
heqiaozhi 已提交
55
    'data_norm',
X
Xin Pan 已提交
56 57 58 59 60 61 62
    'conv2d_transpose',
    'conv3d_transpose',
    'reduce_sum',
    'reduce_mean',
    'reduce_max',
    'reduce_min',
    'reduce_prod',
Z
zhoukunsheng 已提交
63 64
    'reduce_all',
    'reduce_any',
X
Xin Pan 已提交
65 66 67 68 69 70 71 72 73 74 75
    'dropout',
    'split',
    'ctc_greedy_decoder',
    'l2_normalize',
    'matmul',
    'topk',
    'transpose',
    'im2sequence',
    'row_conv',
    'multiplex',
    'layer_norm',
D
Dun 已提交
76
    'group_norm',
D
dengkaipeng 已提交
77
    'spectral_norm',
X
Xin Pan 已提交
78 79 80 81 82 83 84
    'smooth_l1',
    'one_hot',
    'autoincreased_step_counter',
    'reshape',
    'squeeze',
    'unsqueeze',
    'lod_reset',
85
    'lod_append',
X
Xin Pan 已提交
86 87 88 89 90
    'lrn',
    'pad',
    'pad_constant_like',
    'label_smooth',
    'roi_pool',
J
jerrywgz 已提交
91
    'roi_align',
X
Xin Pan 已提交
92 93 94 95
    'dice_loss',
    'image_resize',
    'image_resize_short',
    'resize_bilinear',
K
Kaipeng Deng 已提交
96
    'resize_trilinear',
97
    'resize_nearest',
X
Xin Pan 已提交
98
    'gather',
99
    'gather_nd',
X
Xin Pan 已提交
100
    'scatter',
101 102
    'scatter_nd_add',
    'scatter_nd',
X
Xin Pan 已提交
103 104 105
    'random_crop',
    'mean_iou',
    'relu',
C
chengduo 已提交
106
    'selu',
X
Xin Pan 已提交
107 108
    'log',
    'crop',
109
    'crop_tensor',
X
Xin Pan 已提交
110 111 112 113 114 115 116 117 118 119 120 121 122 123
    'elu',
    'relu6',
    'pow',
    'stanh',
    'hard_sigmoid',
    'swish',
    'prelu',
    'brelu',
    'leaky_relu',
    'soft_relu',
    'flatten',
    'stack',
    'pad2d',
    'unstack',
Z
zhoukunsheng 已提交
124
    'unique',
125
    'unique_with_counts',
X
Xin Pan 已提交
126
    'expand',
127
    'expand_as',
X
Xin Pan 已提交
128 129 130 131 132 133 134 135
    'scale',
    'elementwise_add',
    'elementwise_div',
    'elementwise_sub',
    'elementwise_mul',
    'elementwise_max',
    'elementwise_min',
    'elementwise_pow',
Z
zhoukunsheng 已提交
136 137
    'elementwise_mod',
    'elementwise_floordiv',
X
Xin Pan 已提交
138 139 140 141 142 143
    'uniform_random_batch_size_like',
    'gaussian_random',
    'sampling_id',
    'gaussian_random_batch_size_like',
    'sum',
    'slice',
W
wangchaochaohu 已提交
144
    'strided_slice',
X
Xin Pan 已提交
145
    'shape',
Z
zhoukunsheng 已提交
146
    'rank',
Z
zhoukunsheng 已提交
147
    'size',
X
Xin Pan 已提交
148 149 150 151 152 153 154 155 156
    'logical_and',
    'logical_or',
    'logical_xor',
    'logical_not',
    'clip',
    'clip_by_norm',
    'mean',
    'mul',
    'maxout',
J
JiabinYang 已提交
157
    'space_to_depth',
W
whs 已提交
158
    'affine_grid',
159
    'affine_channel',
B
barrierye 已提交
160
    'similarity_focus',
M
minqiyang 已提交
161
    'hash',
D
dengkaipeng 已提交
162
    'grid_sampler',
G
gmcather 已提交
163 164
    'log_loss',
    'add_position_encoding',
Q
Qiao Longfei 已提交
165
    'bilinear_tensor_product',
C
chengduo 已提交
166 167
    'merge_selected_rows',
    'get_tensor_from_selected_rows',
S
shippingwang 已提交
168
    'shuffle_channel',
169
    'temporal_shift',
S
sneaxiy 已提交
170
    'py_func',
171
    'psroi_pool',
172
    'prroi_pool',
R
ruri 已提交
173
    'pixel_shuffle',
174
    'fsp_matrix',
H
heqiaozhi 已提交
175
    'continuous_value_model',
Z
zhoukunsheng 已提交
176
    'where',
Z
zhoukunsheng 已提交
177
    'sign',
178
    'deformable_conv',
179
    'unfold',
C
cjt222 已提交
180
    'deformable_roi_pooling',
J
Jiawei Wang 已提交
181
    'filter_by_instag',
182
    'shard_index',
H
huangjun12 已提交
183
    'hard_swish',
G
Guo Sheng 已提交
184
    'gather_tree',
185
    'uniform_random',
Y
Yu Yang 已提交
186 187 188 189 190 191 192 193 194
]


def fc(input,
       size,
       num_flatten_dims=1,
       param_attr=None,
       bias_attr=None,
       act=None,
195
       name=None):
Y
Yu Yang 已提交
196
    """
197
    **Fully Connected Layer**
Y
Yu Yang 已提交
198

199 200 201
    This operator creates a fully connected layer in the network. It can take
    a Tensor(or LoDTensor) or a list of Tensor(or LoDTensor) as its inputs(see
    Args in detail). It creates a variable called weight for each input Tensor,
202
    which represents a fully connected weight matrix from each input unit to
203 204 205 206
    each output unit. The fully connected layer multiplies each input Tensor
    with its corresponding weight to produce an output Tensor with shape :math:`[M, size]` ,
    where M is batch size. If a list of Tensor is given, the results of
    multiple output Tensors with shape :math:`[M, size]` will be summed up. If :attr:`bias_attr`
207
    is not None, a bias variable will be created and added to the output.
208
    Finally, if :attr:`act` is not None, it will be applied to the output as well.
C
caoying03 已提交
209

210
    When the input is a single Tensor(or LoDTensor):
C
caoying03 已提交
211

212 213 214 215
    .. math::

        Out = Act({XW + b})

216
    When the input is a list of Tensor(or LoDTensor):
217 218 219

    .. math::

220
        Out = Act({\sum_{i=0}^{N-1}X_iW_i + b})
221 222 223

    In the above equation:

224 225 226
    * :math:`N`: Number of the input. N equals to len(input) if input is list of Variable.
    * :math:`X_i`: The i-th input tensor.
    * :math:`W_i`: The i-th weights matrix corresponding i-th input tensor.
C
caoying03 已提交
227
    * :math:`b`: The bias parameter created by this layer (if needed).
228
    * :math:`Act`: The activation function.
229
    * :math:`Out`: The output Tensor.
230 231 232

    .. code-block:: text

233 234 235 236 237 238 239 240 241 242 243 244 245 246
        Case 1:
        Given a single Tensor data_1, and num_flatten_dims = 2:
            data_1.data = [[[0.1, 0.2],
                            [0.3, 0.4]]]
            data_1.shape = (1, 2, 2) # 1 is batch_size

            out = fluid.layers.fc(input=data_1, size=1, num_flatten_dims=2)

        Then output is:
            out.data = [[0.83234344], [0.34936576]]
            out.shape = (1, 2, 1)

        Case 2:
        Given a list of Tensor:
247 248 249 250 251 252 253 254 255 256 257 258 259
            data_1.data = [[[0.1, 0.2],
                           [0.3, 0.4]]]
            data_1.shape = (1, 2, 2) # 1 is batch_size

            data_2 = [[[0.1, 0.2, 0.3]]]
            data_2.shape = (1, 1, 3)

            out = fluid.layers.fc(input=[data_1, data_2], size=2)

        Then:
            out.data = [[0.18669507, 0.1893476]]
            out.shape = (1, 2)

Y
Yu Yang 已提交
260
    Args:
261 262 263 264 265 266
        input (Variable|list of Variable): A Tensor(or LoDTensor) with shape :math:`[N_1, N_2,..., N_k]` or
            a list of Tensor(or LoDTensor). The dimensions of the input Tensor is at least 2 and the data
            type should be float32 or float64.
        size(int): The number of output units in this layer, which also means the feature size of ouput
            Tensor(or LoDTensor).
        num_flatten_dims (int): The fc layer can accept an input Tensor with more than
R
ranqiu 已提交
267
            two dimensions. If this happens, the multidimensional tensor will first be flattened
268 269
            into a 2-D matrix. The parameter :attr:`num_flatten_dims` determines how the input
            Tensor is flattened: the first :attr:`num_flatten_dims` (inclusive, index starts from 1)
R
ranqiu 已提交
270
            dimensions will be flatten to form the first dimension of the final matrix (height of
271 272 273 274 275 276 277 278 279 280 281 282 283 284 285
            the matrix), and the rest :math:`rank(X) - num\_flatten\_dims` dimensions are flattened to
            form the second dimension of the final matrix (width of the matrix). For example, assuming that
            X is a 5-dimensional Tensor with a shape [2, 3, 4, 5, 6], and :attr:`num_flatten_dims` = 3.
            Then, the flattened matrix will have a shape [2 x 3 x 4, 5 x 6] = [24, 30]. Default: 1.
        param_attr (ParamAttr): To specify the weight parameter property. Default: None, which means the
            default weight parameter property is used. See usage for details in :ref:`api_fluid_ParamAttr` .
        bias_attr (ParamAttr): To specify the bias parameter property. Default: None, which means the
            default bias parameter property is used. See usage for details in :ref:`api_fluid_ParamAttr` .
        act (str): Activation to be applied to the output of this layer, such as tanh, softmax,
            sigmoid, relu. For more information, please refer to :ref:`api_guide_activations_en` . Default: None.
        name (str, optional): The default value is None.  Normally there is no need for user to set this property.
            For more information, please refer to :ref:`api_guide_Name` .

    Returns:
        Variable: Tensor or LoDTensor calculated by fc layer. The data type is same with input.
286 287

    Raises:
288
        ValueError: If dimensions of the input Tensor is less than 2.
289 290 291 292

    Examples:
        .. code-block:: python

293
          import paddle.fluid as fluid
294
          # when input is single tensor
295
          data = fluid.data(name="data", shape=[-1, 32], dtype="float32")
296
          fc = fluid.layers.fc(input=data, size=1000, act="tanh")
297 298

          # when input are multiple tensors
299 300
          data_1 = fluid.data(name="data_1", shape=[-1, 32], dtype="float32")
          data_2 = fluid.data(name="data_2", shape=[-1, 36], dtype="float32")
301
          fc = fluid.layers.fc(input=[data_1, data_2], size=1000, act="tanh")
Y
Yu Yang 已提交
302
    """
C
caoying03 已提交
303
    helper = LayerHelper("fc", **locals())
304
    check_type(input, 'input', (list, tuple, Variable), 'fc')
305 306
    if isinstance(input, (list, tuple)):
        for i, input_x in enumerate(input):
307
            check_type(input_x, 'input[' + str(i) + ']', Variable, 'fc')
Y
Yu Yang 已提交
308
    dtype = helper.input_dtype()
309
    check_dtype(dtype, 'input', ['float16', 'float32', 'float64'], 'fc')
Y
Yu Yang 已提交
310
    mul_results = []
311 312
    for input_var, param_attr in helper.iter_inputs_and_params():
        input_shape = input_var.shape
Y
Yu Yang 已提交
313 314 315
        param_shape = [
            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
        ] + [size]
Y
ying 已提交
316

Y
Yu Yang 已提交
317
        w = helper.create_parameter(
318
            attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False)
X
Xin Pan 已提交
319
        tmp = helper.create_variable_for_type_inference(dtype)
320
        helper.append_op(
321 322 323
            type="mul",
            inputs={"X": input_var,
                    "Y": w},
324
            outputs={"Out": tmp},
M
mozga-intel 已提交
325 326
            attrs={"x_num_col_dims": num_flatten_dims,
                   "y_num_col_dims": 1})
327 328 329 330
        mul_results.append(tmp)

    if len(mul_results) == 1:
        pre_bias = mul_results[0]
331
    else:
X
Xin Pan 已提交
332
        pre_bias = helper.create_variable_for_type_inference(dtype)
333
        helper.append_op(
334 335 336
            type="sum",
            inputs={"X": mul_results},
            outputs={"Out": pre_bias},
X
Xin Pan 已提交
337
            attrs={"use_mkldnn": False})
338 339 340 341
    # add bias
    pre_activation = helper.append_bias_op(pre_bias, dim_start=num_flatten_dims)
    # add activation
    return helper.append_activation(pre_activation)
Y
Yu Yang 已提交
342 343


344 345 346
def embedding(input,
              size,
              is_sparse=False,
347
              is_distributed=False,
348 349 350
              padding_idx=None,
              param_attr=None,
              dtype='float32'):
Y
Yu Yang 已提交
351
    """
352

353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389
    **WARING:** This OP will be deprecated in a future release. This OP requires the
    last dimension of Tensor shape must be equal to 1. It is recommended to use
    fluid. :ref:`api_fluid_embedding` .

    The operator is used to lookup embeddings vector of ids provided by :attr:`input` .
    It automatically constructs a 2D embedding matrix based on the
    input :attr:`size` (vocab_size, emb_size) and :attr:`dtype` .

    This OP requires the last dimension of Tensor shape must be equal to 1. The shape
    of output Tensor is generated by replacing the last dimension of the input Tensor shape
    with emb_size.

    **Note:** The id in :attr:`input` must satisfy :math:`0 =< id < size[0]` , 
    otherwise the program will throw an exception and exit.

    .. code-block:: text

        Case 1:

        input is a Tensor. padding_idx = -1
            input.data = [[[1], [3]], [[2], [4]], [[4], [127]]]
            input.shape = [3, 2, 1]
        Given size = [128, 16]
        output is a Tensor:
            out.shape = [3, 2, 16]
            out.data = [[[0.129435295, 0.244512452, ..., 0.436322452],
                        [0.345421456, 0.524563927, ..., 0.144534654]],

                        [[0.345249859, 0.124939536, ..., 0.194353745],
                        [0.945345345, 0.435394634, ..., 0.435345365]],
                        
                        [[0.945345345, 0.435394634, ..., 0.435345365],
                        [0.0,         0.0,         ..., 0.0        ]]]  # padding data
        The input padding_idx is less than 0, it is automatically converted to padding_idx = -1 + 128 = 127
        It will pad all-zero data when ids is 127.
        
        Case 2:
390

391 392 393 394 395 396 397 398 399 400 401 402 403 404
        input is a LoDTensor with 1-level LoD. padding_idx = 0
            input.lod = [[2, 3]]
            input.data = [[1], [3], [2], [4], [0]]
            input.shape = [5, 1]
        Given size = [128, 16]
        output is a LoDTensor:
            out.lod = [[2, 3]]
            out.shape = [5, 16]
            out.data = [[0.129435295, 0.244512452, ..., 0.436322452],
                        [0.345421456, 0.524563927, ..., 0.144534654],
                        [0.345249859, 0.124939536, ..., 0.194353745],
                        [0.945345345, 0.435394634, ..., 0.435345365],
                        [0.0,         0.0,         ..., 0.0        ]]  # padding data
        It will pad all-zero data when ids is 0.
Y
Yu Yang 已提交
405 406

    Args:
407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433
        input(Variable): A Tensor or LoDTensor with type int64, which contains the id information.
            The last dimension of Tensor shape must be equal to 1. The value of the input id should
            satisfy :math:`0<= id < size[0]` .
        size(tuple|list): The shape of lookup table parameter. It should have two elements which
            indicates the size of the dictionary of embeddings and the size of each embedding vector respectively.
        is_sparse(bool): The flag indicating whether to use sparse update. This parameter only
            affects the performance of the backwards gradient update. It is recommended to set 
            True because sparse update is faster. But some optimizer does not support sparse update,
            such as :ref:`api_fluid_optimizer_AdadeltaOptimizer` , :ref:`api_fluid_optimizer_AdamaxOptimizer` , 
            :ref:`api_fluid_optimizer_DecayedAdagradOptimizer` , :ref:`api_fluid_optimizer_FtrlOptimizer` ,
            :ref:`api_fluid_optimizer_LambOptimizer` and :ref:`api_fluid_optimizer_LarsMomentumOptimizer` .
            In these case, is_sparse must be False. Default: False.
        is_distributed(bool): Whether to store the embedding matrix in a distributed manner. Only used
            in multi-machine distributed CPU training. Default: False.
        padding_idx(int|long|None): padding_idx needs to be in the interval [-vocab_size, vocab_size). 
            If :math:`padding\_idx < 0`, the :math:`padding\_idx` will automatically be converted
            to :math:`vocab\_size + padding\_idx` . It will output all-zero padding data whenever lookup
            encounters :math:`padding\_idx` in id. And the padding data will not be updated while training.
            If set None, it makes no effect to output. Default: None.
        param_attr(ParamAttr): To specify the weight parameter property. Default: None, which means the
            default weight parameter property is used. See usage for details in :ref:`api_fluid_ParamAttr` . In addition,
            user-defined or pre-trained word vectors can be loaded with the :attr:`param_attr` parameter. 
            The local word vector needs to be transformed into numpy format, and the shape of local word
            vector shoud be consistent with :attr:`size` . Then :ref:`api_fluid_initializer_NumpyArrayInitializer`
            is used to load custom or pre-trained word vectors. See code example 2 for details.
        dtype(str|core.VarDesc.VarType): It refers to the data type of output Tensor.
            It must be float32 or float64. Default: float32.
Y
Yu Yang 已提交
434

435
    Returns:
436
        Variable: Embedding Tensor or LoDTensor mapped by input. The data type is the same as :attr:`dtype` .
Y
Yu Yang 已提交
437

438 439
    Examples:
        .. code-block:: python
Y
Yu Yang 已提交
440

B
bdzhuxiaoning 已提交
441
          import paddle.fluid as fluid
442 443 444 445 446 447 448 449 450 451 452 453 454 455
          import numpy as np
          data = fluid.data(name='x', shape=[None, 1], dtype='int64')

          # exampel 1
          emb_1 = fluid.embedding(input=data, size=[128, 64])

          # example 2: load custom or pre-trained word vectors
          weight_data = np.random.random(size=(128, 100))  # word vectors with numpy format
          w_param_attrs = fluid.ParamAttr(
              name="emb_weight",
              learning_rate=0.5,
              initializer=fluid.initializer.NumpyArrayInitializer(weight_data),
              trainable=True)
          emb_2 = fluid.layers.embedding(input=data, size=(128, 100), param_attr=w_param_attrs, dtype='float32')   
Y
Yu Yang 已提交
456 457 458
    """

    helper = LayerHelper('embedding', **locals())
459 460 461 462
    check_type_and_dtype(input, 'input', Variable, ['int64'],
                         'fluid.layers.embedding')
    check_dtype(dtype, 'dtype', ['float16', 'float32', 'float64'],
                'fluid.layers.embedding')
463
    remote_prefetch = is_sparse and (not is_distributed)
Q
Qiao Longfei 已提交
464 465
    if remote_prefetch:
        assert is_sparse is True and is_distributed is False
Y
Yu Yang 已提交
466 467
    w = helper.create_parameter(
        attr=helper.param_attr, shape=size, dtype=dtype, is_bias=False)
X
Xin Pan 已提交
468
    tmp = helper.create_variable_for_type_inference(dtype)
469 470
    padding_idx = -1 if padding_idx is None else padding_idx if padding_idx >= 0 else (
        size[0] + padding_idx)
Y
Yu Yang 已提交
471 472 473 474 475
    helper.append_op(
        type='lookup_table',
        inputs={'Ids': input,
                'W': w},
        outputs={'Out': tmp},
476 477 478
        attrs={
            'is_sparse': is_sparse,
            'is_distributed': is_distributed,
Q
Qiao Longfei 已提交
479
            'remote_prefetch': remote_prefetch,
480 481
            'padding_idx': padding_idx
        })
Y
Yu Yang 已提交
482 483 484
    return tmp


H
hutuxian 已提交
485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532
def _pull_box_sparse(input, size, dtype='float32'):
    """
    **Pull Box Sparse Layer**

    This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in
    BoxPS lookup table. The result of this lookup is the embedding of each ID in the
    :attr:`input`.

    Args:
        input(Variable|list of Variable): Input is a Tensor<int64> Variable, which 
            contains the IDs information.
        size(int): The embedding size parameter, which indicates the size of 
            each embedding vector respectively.
        dtype(str): The dtype refers to the data type of output tensor. Only supports 
	    float32 now.

    Returns:
        Variable|list of Variable: The tensor variable storing the embeddings of the \
                  supplied inputs.

    Examples:
        .. code-block:: python

          import paddle.fluid as fluid
          data = fluid.layers.data(name='sequence', shape=[1], dtype='int64', lod_level=1)
          emb = fluid.layers.pull_box_sparse(input=data, size=[11])    
    """
    helper = LayerHelper('pull_box_sparse', **locals())
    if dtype != 'float32':
        raise ValueError(
            "BoxPS only support float type embedding now, and your type is: " +
            dtype)
    helper.input_dtype()
    inputs = helper.multiple_input()
    outs = [
        helper.create_variable_for_type_inference(dtype)
        for i in range(len(inputs))
    ]
    helper.append_op(
        type='pull_box_sparse',
        inputs={'Ids': inputs},
        outputs={'Out': outs},
        attrs={'size': size})
    if len(outs) == 1:
        return outs[0]
    return outs


Y
yuyang18 已提交
533
@templatedoc()
534
def linear_chain_crf(input, label, param_attr=None, length=None):
Y
yuyang18 已提交
535 536 537 538 539 540
    """
    Linear Chain CRF.

    ${comment}

    Args:
541
        input(${emission_type}): ${emission_comment} 
Y
yuyang18 已提交
542
        label(${label_type}): ${label_comment}
543
        Length(${length_type}): ${length_comment}
544
        param_attr(ParamAttr): The attribute of the learnable parameter for transition parameter.
Y
yuyang18 已提交
545 546

    Returns:
D
dzhwinter 已提交
547 548
        output(${emission_exps_type}): ${emission_exps_comment} \n
        output(${transition_exps_type}): ${transition_exps_comment} \n
549
        output(${log_likelihood_type}): ${log_likelihood_comment} \n
Y
yuyang18 已提交
550

J
JesseyXujin 已提交
551 552 553
    Examples:
        .. code-block:: python

554 555 556 557 558 559 560
            import paddle.fluid as fluid
            import numpy as np

            #define net structure, using LodTensor
            train_program = fluid.Program()
            startup_program = fluid.Program()
            with fluid.program_guard(train_program, startup_program):
561 562
                input_data = fluid.data(name='input_data', shape=[-1,10], dtype='float32')
                label = fluid.data(name='label', shape=[-1,1], dtype='int')
563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584
                emission= fluid.layers.fc(input=input_data, size=10, act="tanh")
                crf_cost = fluid.layers.linear_chain_crf(
                    input=emission,
                    label=label,
                    param_attr=fluid.ParamAttr(
                    name='crfw',
                    learning_rate=0.01)) 
            use_cuda = False
            place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
            exe = fluid.Executor(place)
            exe.run(startup_program)    
            #define data, using LoDTensor
            a = fluid.create_lod_tensor(np.random.rand(12,10).astype('float32'), [[3,3,4,2]], place)
            b = fluid.create_lod_tensor(np.array([[1],[1],[2],[3],[1],[1],[1],[3],[1],[1],[1],[1]]),[[3,3,4,2]] , place)
            feed1 = {'input_data':a,'label':b}
            loss= exe.run(train_program,feed=feed1, fetch_list=[crf_cost])
            print(loss) 

            #define net structure, using padding
            train_program = fluid.Program()
            startup_program = fluid.Program()
            with fluid.program_guard(train_program, startup_program):
585 586 587
                input_data2 = fluid.data(name='input_data2', shape=[-1,10,10], dtype='float32')
                label2 = fluid.data(name='label2', shape=[-1,10,1], dtype='int')
                label_length = fluid.data(name='length', shape=[-1,1], dtype='int')
588 589 590 591 592 593
                emission2= fluid.layers.fc(input=input_data2, size=10, act="tanh", num_flatten_dims=2)
                crf_cost2 = fluid.layers.linear_chain_crf(
                    input=emission2,
                    label=label2,
                    length=label_length,
                    param_attr=fluid.ParamAttr(
J
JesseyXujin 已提交
594
                     name='crfw',
595 596 597 598 599 600
                     learning_rate=0.01))

            use_cuda = False
            place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
            exe = fluid.Executor(place)
            exe.run(startup_program)
J
JesseyXujin 已提交
601

602 603 604
            #define data, using padding
            cc=np.random.rand(4,10,10).astype('float32')
            dd=np.random.rand(4,10,1).astype('int64')
605
            ll=np.array([[3],[3],[4],[2]])
606 607 608
            feed2 = {'input_data2':cc,'label2':dd,'length':ll}
            loss2= exe.run(train_program,feed=feed2, fetch_list=[crf_cost2])
            print(loss2) 
609 610 611 612 613
            #[array([[ 7.8902354],
            #        [ 7.3602567],
            #        [ 10.004011],
            #        [ 5.86721  ]], dtype=float32)]

614 615 616
            #you can use find_var to get transition parameter.
            transition=np.array(fluid.global_scope().find_var('crfw').get_tensor())
            print(transition)
617
            
Y
yuyang18 已提交
618
    """
Y
Yu Yang 已提交
619
    helper = LayerHelper('linear_chain_crf', **locals())
620
    size = input.shape[2] if length else input.shape[1]
Y
Yu Yang 已提交
621 622 623 624
    transition = helper.create_parameter(
        attr=helper.param_attr,
        shape=[size + 2, size],
        dtype=helper.input_dtype())
X
Xin Pan 已提交
625 626 627 628 629 630 631 632
    alpha = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype())
    emission_exps = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype())
    transition_exps = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype())
    log_likelihood = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype())
633 634 635 636 637 638
    this_inputs = {
        "Emission": [input],
        "Transition": transition,
        "Label": [label]
    }
    if length:
639
        this_inputs['Length'] = [length]
Y
Yu Yang 已提交
640 641
    helper.append_op(
        type='linear_chain_crf',
642
        inputs=this_inputs,
Y
Yu Yang 已提交
643 644 645 646 647 648 649 650 651 652
        outputs={
            "Alpha": [alpha],
            "EmissionExps": [emission_exps],
            "TransitionExps": transition_exps,
            "LogLikelihood": log_likelihood
        })

    return log_likelihood


W
wopeizl 已提交
653
@templatedoc()
654
def crf_decoding(input, param_attr, label=None, length=None):
W
wopeizl 已提交
655 656
    """
    ${comment}
Y
yi.wu 已提交
657

W
wopeizl 已提交
658 659
    Args:
        input(${emission_type}): ${emission_comment}
Y
yi.wu 已提交
660

Y
Yibing Liu 已提交
661 662 663
        param_attr (ParamAttr|None): To specify the weight parameter attribute. 
            Default: None, which means the default weight parameter property is 
            used. See usage for details in :ref:`api_fluid_ParamAttr` .
Y
yuyang18 已提交
664

Y
Yibing Liu 已提交
665
        label(${label_type}, optional): ${label_comment}
666
        
Y
Yibing Liu 已提交
667
        length(${length_type}, optional): ${length_comment}
668

W
wopeizl 已提交
669 670
    Returns:
        Variable: ${viterbi_path_comment}
Y
yi.wu 已提交
671

W
wopeizl 已提交
672 673
    Examples:
        .. code-block:: python
Y
yi.wu 已提交
674

675
           import paddle.fluid as fluid
676 677 678

           # LoDTensor-based example
           num_labels = 10
Y
Yibing Liu 已提交
679 680
           feature = fluid.data(name='word_emb', shape=[-1, 784], dtype='float32', lod_level=1)
           label = fluid.data(name='label', shape=[-1, 1], dtype='int64', lod_level=1)
681 682 683
           emission = fluid.layers.fc(input=feature, size=num_labels)
           
           crf_cost = fluid.layers.linear_chain_crf(input=emission, label=label, 
Y
Yibing Liu 已提交
684
                     param_attr=fluid.ParamAttr(name="crfw"))
685
           crf_decode = fluid.layers.crf_decoding(input=emission, 
Y
Yibing Liu 已提交
686
                     param_attr=fluid.ParamAttr(name="crfw"))
687 688 689

           # Common tensor example
           num_labels, max_len = 10, 20
Y
Yibing Liu 已提交
690 691 692
           feature = fluid.data(name='word_emb_pad', shape=[-1, max_len, 784], dtype='float32')
           label = fluid.data(name='label_pad', shape=[-1, max_len, 1], dtype='int64')
           length = fluid.data(name='length', shape=[-1, 1], dtype='int64')
693 694 695 696 697 698 699
           emission = fluid.layers.fc(input=feature, size=num_labels,
                                      num_flatten_dims=2)
           
           crf_cost = fluid.layers.linear_chain_crf(input=emission, label=label, length=length, 
                     param_attr=fluid.ParamAttr(name="crfw_pad"))
           crf_decode = fluid.layers.crf_decoding(input=emission, length=length,
                     param_attr=fluid.ParamAttr(name="crfw_pad"))
W
wopeizl 已提交
700 701 702 703 704
    """
    helper = LayerHelper('crf_decoding', **locals())
    transition = helper.get_parameter(param_attr.name)
    viterbi_path = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype())
705 706 707
    inputs = {"Emission": [input], "Transition": transition, "Label": label}
    if length:
        inputs['Length'] = length
W
wopeizl 已提交
708 709
    helper.append_op(
        type='crf_decoding',
710
        inputs=inputs,
W
wopeizl 已提交
711
        outputs={"ViterbiPath": [viterbi_path]})
Y
Yu Yang 已提交
712

W
wopeizl 已提交
713
    return viterbi_path
Y
Yu Yang 已提交
714 715


Y
yi.wu 已提交
716
@templatedoc()
F
fengjiayi 已提交
717
def cos_sim(X, Y):
Y
Yu Yang 已提交
718
    """
Y
yi.wu 已提交
719 720 721
    ${comment}

    Args:
722 723
        X (Variable): ${x_comment}.
        Y (Variable): ${y_comment}.
F
fengjiayi 已提交
724

Y
yi.wu 已提交
725
    Returns:
L
lvmengsi 已提交
726
        A Variable holding LoDTensor representing the output of cosine(X, Y).
L
lvmengsi 已提交
727 728 729 730

    Examples:
        .. code-block:: python

731
            import paddle.fluid as fluid
L
lvmengsi 已提交
732 733
            x = fluid.data(name='x', shape=[3, 7], dtype='float32')
            y = fluid.data(name='y', shape=[1, 7], dtype='float32')
L
lvmengsi 已提交
734
            out = fluid.layers.cos_sim(x, y)
Y
Yu Yang 已提交
735
    """
F
fengjiayi 已提交
736
    helper = LayerHelper('cos_sim', **locals())
X
Xin Pan 已提交
737 738 739
    out = helper.create_variable_for_type_inference(dtype=X.dtype)
    xnorm = helper.create_variable_for_type_inference(dtype=X.dtype)
    ynorm = helper.create_variable_for_type_inference(dtype=X.dtype)
Y
Yu Yang 已提交
740 741 742 743 744 745 746 747 748 749
    helper.append_op(
        type='cos_sim',
        inputs={'X': [X],
                'Y': [Y]},
        outputs={'Out': [out],
                 'XNorm': [xnorm],
                 'YNorm': [ynorm]})
    return out


P
phlrain 已提交
750 751 752 753 754
def dropout(x,
            dropout_prob,
            is_test=False,
            seed=None,
            name=None,
P
phlrain 已提交
755
            dropout_implementation="downgrade_in_infer"):
756 757 758 759 760
    """
    Computes dropout.

    Drop or keep each element of `x` independently. Dropout is a regularization
    technique for reducing overfitting by preventing neuron co-adaption during
761
    training. The dropout operator randomly sets (according to the given dropout
762 763 764
    probability) the outputs of some units to zero, while others are remain
    unchanged.

H
haowang101779990 已提交
765 766
    dropout op can be removed from the program to make the program more efficient.

767
    Args:
L
lvmengsi 已提交
768
        x (Variable): The input tensor variable. The data type is float16 or float32 or float64.
769
        dropout_prob (float): Probability of setting units to zero.
770 771 772 773
        is_test (bool): A flag indicating whether it is in test phrase or not.
        seed (int): A Python integer used to create random seeds. If this
                    parameter is set to None, a random seed is used.
                    NOTE: If an integer seed is given, always the same output
L
lvmengsi 已提交
774
                    units will be dropped. DO NOT use a fixed seed in training.Default: None.
775 776
        name (str|None): A name for this layer(optional). If set None, the layer
                         will be named automatically.
H
haowang101779990 已提交
777 778
        dropout_implementation(string): ['downgrade_in_infer'(default)|'upscale_in_train']

P
phlrain 已提交
779
                                        1. downgrade_in_infer(default), downgrade the outcome at inference
H
haowang101779990 已提交
780 781

                                           - train: out = input * mask
C
ceci3 已提交
782
                                           - inference: out = input * (1.0 - dropout_prob)
H
haowang101779990 已提交
783 784 785

                                           (mask is a tensor same shape with input, value is 0 or 1
                                           ratio of 0 is dropout_prob)
P
phlrain 已提交
786
                                        2. upscale_in_train, upscale the outcome at training time
787

H
haowang101779990 已提交
788 789
                                           - train: out = input * mask / ( 1.0 - dropout_prob )
                                           - inference: out = input
P
phlrain 已提交
790

H
haowang101779990 已提交
791 792
                                           (mask is a tensor same shape with input, value is 0 or 1
                                           ratio of 0 is dropout_prob)
793

M
minqiyang 已提交
794

795
    Returns:
L
lvmengsi 已提交
796
        A Variable holding Tensor representing the dropout, has same shape and data type with `x`.
797 798

    Examples:
799

800 801
        .. code-block:: python

802
            import paddle.fluid as fluid
L
lvmengsi 已提交
803
            x = fluid.data(name="data", shape=[None, 32, 32], dtype="float32")
804
            droped = fluid.layers.dropout(x, dropout_prob=0.5)
805 806
    """

F
fengjiayi 已提交
807
    helper = LayerHelper('dropout', **locals())
808 809
    check_type_and_dtype(x, 'x', Variable, ['float16', 'float32', 'float64'],
                         'dropout')
X
Xin Pan 已提交
810 811
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    mask = helper.create_variable_for_type_inference(
Z
Zeng Jinle 已提交
812
        dtype=core.VarDesc.VarType.UINT8, stop_gradient=True)
C
chengduo 已提交
813 814 815 816

    if (seed is None or seed == 0) and helper.main_program.random_seed != 0:
        seed = helper.main_program.random_seed

817 818 819 820 821
    helper.append_op(
        type='dropout',
        inputs={'X': [x]},
        outputs={'Out': [out],
                 'Mask': [mask]},
822 823 824 825
        attrs={
            'dropout_prob': dropout_prob,
            'is_test': is_test,
            'fix_seed': seed is not None,
L
lvmengsi 已提交
826
            'seed': seed if seed is not None else 0,
P
phlrain 已提交
827
            'dropout_implementation': dropout_implementation,
828
        })
829 830 831
    return out


Y
yi.wu 已提交
832
@templatedoc()
Y
Yu Yang 已提交
833 834 835 836
def chunk_eval(input,
               label,
               chunk_scheme,
               num_chunk_types,
837 838
               excluded_chunk_types=None,
               seq_length=None):
Y
Yu Yang 已提交
839
    """
G
Guo Sheng 已提交
840 841
    This operator computes the precision, recall and F1-score for chunk detection.
    It is often used in sequence tagging tasks, such as Named Entity Recognition(NER).
Y
yi.wu 已提交
842

M
minqiyang 已提交
843
    For some basics of chunking, please refer to
H
haowang101779990 已提交
844
    `Chunking with Support Vector Machines <https://aclanthology.info/pdf/N/N01/N01-1025.pdf>`_ .
Y
yi.wu 已提交
845

G
Guo Sheng 已提交
846 847
    This operator supports IOB, IOE, IOBES and IO (also known as plain) tagging schemes.
    Here is a NER example for the usage of these tagging schemes:
Y
yi.wu 已提交
848 849

    .. code-block:: python
850

Y
yi.wu 已提交
851 852 853 854 855 856 857 858 859 860
       ====== ====== ======  =====  ==  ============   =====  ===== =====  ==  =========
              Li     Ming    works  at  Agricultural   Bank   of    China  in  Beijing.
       ====== ====== ======  =====  ==  ============   =====  ===== =====  ==  =========
       IO     I-PER  I-PER   O      O   I-ORG          I-ORG  I-ORG I-ORG  O   I-LOC
       IOB    B-PER  I-PER   O      O   B-ORG          I-ORG  I-ORG I-ORG  O   B-LOC
       IOE    I-PER  E-PER   O      O   I-ORG          I-ORG  I-ORG E-ORG  O   E-LOC
       IOBES  B-PER  E-PER   O      O   I-ORG          I-ORG  I-ORG E-ORG  O   S-LOC
       ====== ====== ======  =====  ==  ============   =====  ===== =====  ==  =========

    There are three chunk types(named entity types) including PER(person), ORG(organization)
G
Guo Sheng 已提交
861
    and LOC(location), and we can see that the labels have the form `<tag type>-<chunk type>` .
Y
yi.wu 已提交
862

G
Guo Sheng 已提交
863 864 865
    Since the implementation of this operator actually uses label ids rather than
    label strings, to make it work, there should be a way to map label ids to
    tag types and chunk types. This operator uses the following way to do mapping:
Y
yi.wu 已提交
866 867 868 869 870 871 872 873 874 875

    .. code-block:: python

       tag_type = label % num_tag_type
       chunk_type = label / num_tag_type

    where `num_tag_type` is the num of tag types in the tagging scheme, `num_chunk_type`
    is the num of chunk types, and `tag_type` get its value from the following table.

    .. code-block:: python
876

Y
yi.wu 已提交
877 878 879 880 881 882
       Scheme Begin Inside End   Single
        plain   0     -      -     -
        IOB     0     1      -     -
        IOE     -     0      1     -
        IOBES   0     1      2     3

G
Guo Sheng 已提交
883 884
    Accordingly, in the above NER example, if the tagging scheme is IOB and chunk
    types are ORG, PER and LOC, then the label ids would be as follows:
Y
yi.wu 已提交
885 886 887 888 889 890 891 892 893 894 895

    .. code-block:: python

       B-ORG  0
       I-ORG  1
       B-PER  2
       I-PER  3
       B-LOC  4
       I-LOC  5
       O      6

G
Guo Sheng 已提交
896 897
    With which we can map each label id to the corresponding tag type and chunk
    type correctly.
Y
yi.wu 已提交
898

Y
yi.wu 已提交
899
    Args:
G
Guo Sheng 已提交
900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915
        input (Variable): A Tensor or LoDTensor, representing the predicted labels
            from the network. When it is a Tensor, its shape would be `[N, M, 1]`,
            where `N` stands for batch size, `M` for sequence length; When it is
            a LoDTensor, its shape would be `[N, 1]` where `N` stands for the total
            sequence lengths in this mini-batch. The data type should be int64.
        label (Variable): A Tensor or LoDTensor representing the ground-truth labels.
            It shoud have the same shape, lod and data type as ``input`` .
        chunk_scheme (str): Indicate the tagging schemes used here. The value must
            be IOB, IOE, IOBES or plain.
        num_chunk_types (int): The number of chunk types.
        excluded_chunk_types (list, optional): Indicate the chunk types shouldn't
            be taken into account. It should be a list of chunk type ids(integer).
            Default None.
        seq_length(Variable, optional): A 1D Tensor containing the length of each
            sequence when ``input`` and ``label`` are Tensor. It needn't be
            provided if ``input`` and ``label`` are LoDTensor. Default None.
F
fengjiayi 已提交
916

Y
yi.wu 已提交
917
    Returns:
G
Guo Sheng 已提交
918 919 920 921
        tuple: A tuple including precision, recall, F1-score, chunk number detected, \
            chunk number in ground-truth, chunk number correctly detected. Each \
            is a Tensor with shape `[1]`. The data type of precision, recall and \
            F1-score all is float32, and the others' data type all is int64.
922

Y
yi.wu 已提交
923 924 925
    Examples:
        .. code-block:: python

926 927 928 929
            import paddle.fluid as fluid

            dict_size = 10000
            label_dict_len = 7
G
Guo Sheng 已提交
930 931 932
            sequence = fluid.data(
                name='id', shape=[-1, 1], lod_level=1, dtype='int64')
            embedding = fluid.embedding(
933 934 935 936
                input=sequence, size=[dict_size, 512])
            hidden = fluid.layers.fc(input=embedding, size=512)
            label = fluid.layers.data(
                name='label', shape=[1], lod_level=1, dtype='int32')
Y
yi.wu 已提交
937
            crf = fluid.layers.linear_chain_crf(
938
                input=hidden, label=label, param_attr=fluid.ParamAttr(name="crfw"))
Y
yi.wu 已提交
939
            crf_decode = fluid.layers.crf_decoding(
940
                input=hidden, param_attr=fluid.ParamAttr(name="crfw"))
Y
yi.wu 已提交
941 942 943 944 945
            fluid.layers.chunk_eval(
                input=crf_decode,
                label=label,
                chunk_scheme="IOB",
                num_chunk_types=(label_dict_len - 1) / 2)
Y
Yu Yang 已提交
946
    """
F
fengjiayi 已提交
947
    helper = LayerHelper("chunk_eval", **locals())
Y
Yu Yang 已提交
948 949

    # prepare output
X
Xin Pan 已提交
950 951 952 953 954 955 956
    precision = helper.create_variable_for_type_inference(dtype="float32")
    recall = helper.create_variable_for_type_inference(dtype="float32")
    f1_score = helper.create_variable_for_type_inference(dtype="float32")
    num_infer_chunks = helper.create_variable_for_type_inference(dtype="int64")
    num_label_chunks = helper.create_variable_for_type_inference(dtype="int64")
    num_correct_chunks = helper.create_variable_for_type_inference(
        dtype="int64")
Y
Yu Yang 已提交
957

958 959 960 961 962
    this_input = {"Inference": [input], "Label": [label]}

    if seq_length:
        this_input["SeqLength"] = [seq_length]

Y
Yu Yang 已提交
963 964
    helper.append_op(
        type="chunk_eval",
965
        inputs=this_input,
Y
Yu Yang 已提交
966 967 968
        outputs={
            "Precision": [precision],
            "Recall": [recall],
969 970 971 972
            "F1-Score": [f1_score],
            "NumInferChunks": [num_infer_chunks],
            "NumLabelChunks": [num_label_chunks],
            "NumCorrectChunks": [num_correct_chunks]
Y
Yu Yang 已提交
973 974 975
        },
        attrs={
            "num_chunk_types": num_chunk_types,
G
guosheng 已提交
976 977
            "chunk_scheme": chunk_scheme,
            "excluded_chunk_types": excluded_chunk_types or []
Y
Yu Yang 已提交
978
        })
979 980
    return (precision, recall, f1_score, num_infer_chunks, num_label_chunks,
            num_correct_chunks)
Y
Yu Yang 已提交
981 982


983
def softmax(input, use_cudnn=False, name=None, axis=-1):
Y
Yu Yang 已提交
984
    """
985
    This operator implements the softmax layer. The calculation process is as follows:
986

987
    1. The dimension :attr:`axis` of the ``input`` will be permuted to the last.
988
    
989 990 991 992 993 994 995
    2. Then the input tensor will be logically flattened to a 2-D matrix. The matrix's
    second dimension(row length) is the same as the dimension :attr:`axis` of the input
    tensor, and the first dimension(column length) is the product of all other
    dimensions of the input tensor. For each row of the matrix, the softmax operator
    squashes the K-dimensional(K is the width of the matrix, which is also the size
    of the input tensor's dimension :attr:`axis`) vector of arbitrary real values to a
    K-dimensional vector of real values in the range [0, 1] that add up to 1.
996

997 998
    3. After the softmax operation is completed, the inverse operations of steps 1 and 2 
    are performed to restore the two-dimensional matrix to the same dimension as the ``input``.
999

1000 1001 1002 1003 1004
    It computes the exponential of the given dimension and the sum of exponential
    values of all the other dimensions in the K-dimensional vector input.
    Then the ratio of the exponential of the given dimension and the sum of
    exponential values of all the other dimensions is the output of the softmax
    operator.
1005

1006
    For each row :math:`i` and each column :math:`j` in the matrix, we have:
1007

1008
    .. math::
1009

1010
        Out[i, j] = \\frac{\exp(X[i, j])}{\sum_j(exp(X[i, j])}
1011

1012
    Example:
1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058

    .. code-block:: text

        Case 1:
          Input:
            X.shape = [2, 3, 4]
            X.data = [[[2.0, 3.0, 4.0, 5.0],
                       [3.0, 4.0, 5.0, 6.0],
                       [7.0, 8.0, 8.0, 9.0]],
                      [[1.0, 2.0, 3.0, 4.0],
                       [5.0, 6.0, 7.0, 8.0],
                       [6.0, 7.0, 8.0, 9.0]]]

          Attrs:
            axis = -1

          Output:
            Out.shape = [2, 3, 4]
            Out.data = [[[0.0320586 , 0.08714432, 0.23688282, 0.64391426],
                         [0.0320586 , 0.08714432, 0.23688282, 0.64391426],
                         [0.07232949, 0.19661193, 0.19661193, 0.53444665]],
                        [[0.0320586 , 0.08714432, 0.23688282, 0.64391426],
                         [0.0320586 , 0.08714432, 0.23688282, 0.64391426],
                         [0.0320586 , 0.08714432, 0.23688282, 0.64391426]]]

        Case 2:
          Input:
            X.shape = [2, 3, 4]
            X.data = [[[2.0, 3.0, 4.0, 5.0],
                       [3.0, 4.0, 5.0, 6.0],
                       [7.0, 8.0, 8.0, 9.0]],
                      [[1.0, 2.0, 3.0, 4.0],
                       [5.0, 6.0, 7.0, 8.0],
                       [6.0, 7.0, 8.0, 9.0]]]
          Attrs:
            axis = 1

          Output:
            Out.shape = [2, 3, 4]
            Out.data = [[[0.00657326, 0.00657326, 0.01714783, 0.01714783],
                         [0.01786798, 0.01786798, 0.04661262, 0.04661262],
                         [0.97555875, 0.97555875, 0.93623955, 0.93623955]],
                        [[0.00490169, 0.00490169, 0.00490169, 0.00490169],
                         [0.26762315, 0.26762315, 0.26762315, 0.26762315],
                         [0.72747516, 0.72747516, 0.72747516, 0.72747516]]] 

Q
qiaolongfei 已提交
1059
    Args:
1060 1061
        input (Variable): The input variable. A multi-dimension ``Tensor`` with type float32 or float64.
        use_cudnn (bool, optional): Use cudnn kernel or not, it is valid only when the cudnn \
J
jerrywgz 已提交
1062
            library is installed. To improve numerical stablity, set use_cudnn to \
1063 1064
            False by default.
        name (str, optional): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name` . Default: None.
C
chengduo 已提交
1065
            will be named automatically. Default: None.
1066
        axis (int, optional): The index of dimension to perform softmax calculations, it should
D
dengkaipeng 已提交
1067
            be in range :math:`[-1, rank - 1]`, while :math:`rank` is the rank of
1068
            input variable. Default: -1. -1 means the last dimension.
Q
qiaolongfei 已提交
1069 1070

    Returns:
1071
        Variable: ``Tensor`` indicates the output of softmax. The data type and shape are the same as ``input`` .
Q
qiaolongfei 已提交
1072 1073 1074 1075 1076

    Examples:

        .. code-block:: python

1077 1078
            import paddle.fluid as fluid
            import numpy as np
Q
qiaolongfei 已提交
1079

1080 1081 1082 1083 1084 1085 1086 1087 1088
            data = fluid.data(name="input", shape=[-1, 3],dtype="float32")
            result = fluid.layers.softmax(data,axis=1)
            place = fluid.CPUPlace()
            exe = fluid.Executor(place)
            exe.run(fluid.default_startup_program())
            x = np.random.rand(3, 3).astype("float32")
            output= exe.run(feed={"input": x},
                             fetch_list=[result[0]])
            print(output)
Q
qiaolongfei 已提交
1089
    """
1090
    helper = LayerHelper('softmax', **locals())
1091 1092
    check_type_and_dtype(input, 'input', Variable,
                         ['float16', 'float32', 'float64'], 'softmax')
1093

1094
    dtype = helper.input_dtype()
X
Xin Pan 已提交
1095
    softmax_out = helper.create_variable_for_type_inference(dtype)
1096 1097 1098 1099
    helper.append_op(
        type="softmax",
        inputs={"X": input},
        outputs={"Out": softmax_out},
D
dengkaipeng 已提交
1100 1101
        attrs={"axis": axis,
               "use_cudnn": use_cudnn})
1102 1103 1104
    return softmax_out


Y
Yu Yang 已提交
1105 1106 1107
def conv2d(input,
           num_filters,
           filter_size,
C
chengduoZH 已提交
1108 1109
           stride=1,
           padding=0,
1110
           dilation=1,
Y
Yu Yang 已提交
1111 1112 1113
           groups=None,
           param_attr=None,
           bias_attr=None,
C
chengduoZH 已提交
1114
           use_cudnn=True,
1115
           act=None,
L
liym27 已提交
1116 1117
           name=None,
           data_format="NCHW"):
Y
Yu Yang 已提交
1118
    """
C
chengduoZH 已提交
1119
    The convolution2D layer calculates the output based on the input, filter
T
tensor-tang 已提交
1120
    and strides, paddings, dilations, groups parameters. Input and
L
liym27 已提交
1121
    Output are in NCHW or NHWC format, where N is batch size, C is the number of
1122
    channels, H is the height of the feature, and W is the width of the feature.
T
tensor-tang 已提交
1123 1124 1125 1126 1127 1128
    Filter is in MCHW format, where M is the number of output image channels,
    C is the number of input image channels, H is the height of the filter,
    and W is the width of the filter. If the groups is greater than 1,
    C will equal the number of input image channels divided by the groups.
    Please refer to UFLDL's `convolution
    <http://ufldl.stanford.edu/tutorial/supervised/FeatureExtractionUsingConvolution/>`_
1129
    for more details.
1130 1131 1132
    If bias attribution and activation type are provided, bias is added to the
    output of the convolution, and the corresponding activation function is
    applied to the final result.
C
chengduoZH 已提交
1133

1134
    For each input :math:`X`, the equation is:
C
refine  
chengduoZH 已提交
1135

C
chengduoZH 已提交
1136 1137
    .. math::

C
refine  
chengduoZH 已提交
1138
        Out = \sigma (W \\ast X + b)
C
chengduoZH 已提交
1139

T
tensor-tang 已提交
1140
    Where:
C
chengduoZH 已提交
1141

L
liym27 已提交
1142
    * :math:`X`: Input value, a tensor with NCHW or NHWC format.
1143 1144 1145 1146
    * :math:`W`: Filter value, a tensor with MCHW format.
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
T
tensor-tang 已提交
1147
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
1148 1149 1150

    Example:

1151 1152
        - Input:

W
weixing02 已提交
1153
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
C
refine  
chengduoZH 已提交
1154

W
weixing02 已提交
1155
          Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)`
C
refine  
chengduoZH 已提交
1156

1157
        - Output:
T
tensor-tang 已提交
1158

W
weixing02 已提交
1159
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
C
refine  
chengduoZH 已提交
1160

C
chengduoZH 已提交
1161
        Where
1162 1163

        .. math::
C
chengduoZH 已提交
1164

W
weixing02 已提交
1165 1166
            H_{out}&= \\frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]} + 1 \\\\
            W_{out}&= \\frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]} + 1
C
chengduoZH 已提交
1167 1168

    Args:
L
lvmengsi 已提交
1169 1170
        input (Variable): The input is 4-D Tensor with shape [N, C, H, W], the data type 
            of input is float16 or float32 or float64.
T
tensor-tang 已提交
1171
        num_filters(int): The number of filter. It is as same as the output
1172
            image channel.
1173 1174
        filter_size (int|tuple): The filter size. If filter_size 
            is a tuple, it must contain two integers, (filter_size_height, 
L
lvmengsi 已提交
1175 1176 1177 1178 1179 1180 1181
            filter_size_width). Otherwise, filter_size_height = filter_size_width =\
            filter_size.
        stride (int|tuple): The stride size. It means the stride in convolution. 
            If stride is a tuple, it must contain two integers, (stride_height, stride_width). 
            Otherwise, stride_height = stride_width = stride. Default: stride = 1.
        padding (string|int|list|tuple): The padding size. It means the number of zero-paddings
            on both sides for each dimention.If `padding` is a string, either 'VALID' or
L
liym27 已提交
1182 1183
            'SAME' which is the padding algorithm. If padding size is a tuple or list,
            it could be in three forms: `[pad_height, pad_width]` or
L
lvmengsi 已提交
1184 1185 1186
            `[pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`, and when 
            `data_format` is `"NCHW"`, `padding` can be in the form `[[0,0], [0,0], 
            [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`.
L
liym27 已提交
1187 1188 1189
            when `data_format` is `"NHWC"`, `pool_padding` can be in the form
            `[[0,0], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`.
            Default: padding = 0.
L
lvmengsi 已提交
1190 1191 1192 1193
        dilation (int|tuple): The dilation size. It means the spacing between the kernel
            points. If dilation is a tuple, it must contain two integers, (dilation_height, 
            dilation_width). Otherwise, dilation_height = dilation_width = dilation. 
            Default: dilation = 1.
1194 1195 1196 1197
        groups (int): The groups number of the Conv2d Layer. According to grouped
            convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
            the first half of the filters is only connected to the first half
            of the input channels, while the second half of the filters is only
C
chengduo 已提交
1198 1199 1200 1201 1202
            connected to the second half of the input channels. Default: groups=1.
        param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
            of conv2d. If it is set to None or one attribute of ParamAttr, conv2d
            will create ParamAttr as param_attr. If the Initializer of the param_attr
            is not set, the parameter is initialized with :math:`Normal(0.0, std)`,
H
haowang101779990 已提交
1203
            and the :math:`std` is :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None.
C
chengduo 已提交
1204 1205 1206 1207 1208
        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv2d.
            If it is set to False, no bias will be added to the output units.
            If it is set to None or one attribute of ParamAttr, conv2d
            will create ParamAttr as bias_attr. If the Initializer of the bias_attr
            is not set, the bias is initialized zero. Default: None.
1209 1210
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
C
chengduo 已提交
1211 1212
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None
L
lvmengsi 已提交
1213 1214 1215
        name(str|None): For detailed information, please refer 
           to :ref:`api_guide_Name`. Usually name is no need to set and 
           None by default.
1216 1217
        data_format (str, optional): Specify the data format of the input, and the data format of the output 
            will be consistent with that of the input. An optional string from: `"NCHW"`, `"NHWC"`.
L
liym27 已提交
1218 1219
            The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`.
C
chengduoZH 已提交
1220 1221

    Returns:
L
lvmengsi 已提交
1222 1223 1224 1225
        A Variable holding Tensor representing the conv2d, whose data type is the 
        same with input. If act is None, the tensor variable storing the convolution 
        result, and if act is not None, the tensor variable storing convolution 
        and non-linearity activation result.
C
refine  
chengduoZH 已提交
1226

1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239
    Raises:
        ValueError: If the type of `use_cudnn` is not bool.
        ValueError: If `data_format` is not "NCHW" or "NHWC".
        ValueError: If the channel dimmention of the input is less than or equal to zero.
        ValueError: If `padding` is a string, but not "SAME" or "VALID".
        ValueError: If `padding` is a tuple, but the element corresponding to the input's batch size is not 0 
            or the element corresponding to the input's channel is not 0.
        ShapeError: If the input is not 4-D Tensor.
        ShapeError: If the input's dimension size and filter's dimension size not equal.
        ShapeError: If the dimension size of input minus the size of `stride` is not 2.
        ShapeError: If the number of input channels is not equal to filter's channels * groups.
        ShapeError: If the number of output channels is not be divided by groups.

C
chengduoZH 已提交
1240 1241 1242
    Examples:
        .. code-block:: python

1243
          import paddle.fluid as fluid
L
lvmengsi 已提交
1244
          data = fluid.data(name='data', shape=[None, 3, 32, 32], dtype='float32')
1245
          conv2d = fluid.layers.conv2d(input=data, num_filters=2, filter_size=3, act="relu")
Y
Yu Yang 已提交
1246 1247
    """

1248 1249
    check_type_and_dtype(input, 'input', Variable,
                         ['float16', 'float32', 'float64'], 'conv2d')
1250
    num_channels = input.shape[1]
L
liym27 已提交
1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265
    if not isinstance(use_cudnn, bool):
        raise ValueError("Attr(use_cudnn) should be True or False. Received "
                         "Attr(use_cudnn): %s. " % str(use_cudnn))

    if data_format not in ["NCHW", "NHWC"]:
        raise ValueError(
            "Attr(data_format) should be 'NCHW' or 'NHWC'. Received "
            "Attr(data_format): %s." % str(data_format))

    channel_last = (data_format == "NHWC")
    num_channels = input.shape[3] if channel_last else input.shape[1]
    if num_channels < 0:
        raise ValueError(
            "The channel dimmention of the input(%s) should be defined. "
            "Received: %s." % (str(input.shape), str(num_channels)))
C
chengduo 已提交
1266
    assert param_attr is not False, "param_attr should not be False here."
L
liym27 已提交
1267

1268
    l_type = 'conv2d'
X
xzl 已提交
1269 1270
    if (num_channels == groups and num_filters % num_channels == 0 and
            not use_cudnn):
1271
        l_type = 'depthwise_conv2d'
1272 1273 1274 1275

    helper = LayerHelper(l_type, **locals())
    dtype = helper.input_dtype()

Y
Yu Yang 已提交
1276 1277 1278 1279
    if groups is None:
        num_filter_channels = num_channels
    else:
        if num_channels % groups != 0:
L
liym27 已提交
1280
            raise ValueError(
1281 1282 1283
                "the channel of input must be divisible by groups,"
                "received: the channel of input is {}, the shape of input is {}"
                ", the groups is {}".format(num_channels, input.shape, groups))
M
minqiyang 已提交
1284
        num_filter_channels = num_channels // groups
Y
Yu Yang 已提交
1285

C
chengduoZH 已提交
1286 1287
    filter_size = utils.convert_to_list(filter_size, 2, 'filter_size')
    stride = utils.convert_to_list(stride, 2, 'stride')
1288
    dilation = utils.convert_to_list(dilation, 2, 'dilation')
C
chengduoZH 已提交
1289

L
liym27 已提交
1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312
    # padding
    def _update_padding(padding, data_format):
        def is_list_or_tuple(ele):
            if isinstance(ele, list) or isinstance(ele, tuple):
                return True
            return False

        if is_list_or_tuple(padding) and len(padding) == 4:
            if is_list_or_tuple(padding[0]) and (data_format == "NCHW"):
                if not (padding[0] == [0, 0] and padding[1] == [0, 0]):
                    raise ValueError(
                        "Non-zero padding(%s) in the batch or channel dimensions "
                        "is not supported." % str(padding))
                padding = padding[2:4]
                padding = [ele for a_list in padding for ele in a_list]
            elif is_list_or_tuple(padding[0]) and (data_format == "NHWC"):
                if not (padding[0] == [0, 0] and padding[3] == [0, 0]):
                    raise ValueError(
                        "Non-zero padding(%s) in the batch or channel dimensions "
                        "is not supported." % str(padding))
                padding = padding[1:3]
                padding = [ele for a_list in padding for ele in a_list]
            padding = utils.convert_to_list(padding, 4, 'padding')
1313 1314 1315
            if utils._is_symmetric_padding(padding, 2):
                padding = [padding[0], padding[2]]

L
liym27 已提交
1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329
        else:
            padding = utils.convert_to_list(padding, 2, 'padding')

        return padding

    padding_algorithm = "EXPLICIT"
    if isinstance(padding, str):
        padding = padding.upper()
        if padding not in ["SAME", "VALID"]:
            raise ValueError(
                "Unknown padding: '%s'. It can only be 'SAME' or 'VALID'." %
                str(padding))
        if padding == "VALID":
            padding_algorithm = "VALID"
1330
            padding = [0, 0]
L
liym27 已提交
1331 1332
        elif padding == "SAME":
            padding_algorithm = "SAME"
1333
            padding = [0, 0]
L
liym27 已提交
1334 1335

    padding = _update_padding(padding, data_format)
Y
Yu Yang 已提交
1336

M
minqiyang 已提交
1337
    filter_shape = [num_filters, int(num_filter_channels)] + filter_size
Y
Yu Yang 已提交
1338 1339

    def _get_default_param_initializer():
C
chengduo 已提交
1340 1341
        filter_elem_num = filter_size[0] * filter_size[1] * num_channels
        std = (2.0 / filter_elem_num)**0.5
Y
Yu Yang 已提交
1342 1343 1344 1345 1346 1347 1348 1349
        return Normal(0.0, std, 0)

    filter_param = helper.create_parameter(
        attr=helper.param_attr,
        shape=filter_shape,
        dtype=dtype,
        default_initializer=_get_default_param_initializer())

X
Xin Pan 已提交
1350
    pre_bias = helper.create_variable_for_type_inference(dtype)
Y
Yu Yang 已提交
1351 1352

    helper.append_op(
1353
        type=l_type,
Y
Yu Yang 已提交
1354 1355 1356 1357 1358
        inputs={
            'Input': input,
            'Filter': filter_param,
        },
        outputs={"Output": pre_bias},
C
chengduoZH 已提交
1359 1360 1361
        attrs={
            'strides': stride,
            'paddings': padding,
1362
            'dilations': dilation,
C
chengduoZH 已提交
1363
            'groups': groups,
1364
            'use_cudnn': use_cudnn,
1365
            'use_mkldnn': False,
L
liym27 已提交
1366 1367 1368
            'fuse_relu_before_depthwise_conv': False,
            "padding_algorithm": padding_algorithm,
            "data_format": data_format,
C
chengduoZH 已提交
1369
        })
Y
Yu Yang 已提交
1370

1371 1372 1373 1374
    if data_format == 'NCHW':
        pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
    else:
        pre_act = helper.append_bias_op(pre_bias, dim_start=3, dim_end=4)
Y
Yu Yang 已提交
1375 1376 1377 1378

    return helper.append_activation(pre_act)


C
chengduoZH 已提交
1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389
def conv3d(input,
           num_filters,
           filter_size,
           stride=1,
           padding=0,
           dilation=1,
           groups=None,
           param_attr=None,
           bias_attr=None,
           use_cudnn=True,
           act=None,
L
liym27 已提交
1390 1391
           name=None,
           data_format="NCDHW"):
C
chengduoZH 已提交
1392 1393 1394
    """
    The convolution3D layer calculates the output based on the input, filter
    and strides, paddings, dilations, groups parameters. Input(Input) and
L
liym27 已提交
1395
    Output(Output) are in NCDHW or NDHWC format. Where N is batch size C is the number of
1396 1397 1398 1399 1400
    channels, D is the depth of the feature, H is the height of the feature,
    and W is the width of the feature. Convlution3D is similar with Convlution2D
    but adds one dimension(depth). If bias attribution and activation type are
    provided, bias is added to the output of the convolution, and the
    corresponding activation function is applied to the final result.
C
chengduoZH 已提交
1401 1402 1403 1404 1405 1406 1407 1408 1409

    For each input :math:`X`, the equation is:

    .. math::

        Out = \sigma (W \\ast X + b)

    In the above equation:

L
liym27 已提交
1410
    * :math:`X`: Input value, a tensor with NCDHW or NDHWC format.
1411
    * :math:`W`: Filter value, a tensor with MCDHW format.
C
chengduoZH 已提交
1412 1413 1414
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
1415
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436

    Example:

        - Input:

          Input shape: :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})`

          Filter shape: :math:`(C_{out}, C_{in}, D_f, H_f, W_f)`

        - Output:
          Output shape: :math:`(N, C_{out}, D_{out}, H_{out}, W_{out})`

        Where

        .. math::

            D_{out}&= \\frac{(D_{in} + 2 * paddings[0] - (dilations[0] * (D_f - 1) + 1))}{strides[0]} + 1 \\\\
            H_{out}&= \\frac{(H_{in} + 2 * paddings[1] - (dilations[1] * (H_f - 1) + 1))}{strides[1]} + 1 \\\\
            W_{out}&= \\frac{(W_{in} + 2 * paddings[2] - (dilations[2] * (W_f - 1) + 1))}{strides[2]} + 1

    Args:
L
lvmengsi 已提交
1437 1438
        input (Variable): The input is 5-D Tensor with shape [N, C, D, H, W], the data 
            type of input is float16 or float32 or float64.
1439
        num_filters(int): The number of filter. It is as same as the output
C
chengduoZH 已提交
1440
            image channel.
1441 1442 1443 1444
        filter_size (int|tuple): The filter size. If filter_size is a tuple,
            it must contain three integers, (filter_size_depth, filter_size_height, 
            filter_size_width). Otherwise, filter_size_depth = filter_size_height = \
            filter_size_width = filter_size.
L
lvmengsi 已提交
1445 1446 1447 1448 1449
        stride (int|tuple): The stride size. It means the stride in convolution. If stride is a 
            tuple, it must contain three integers, (stride_depth, stride_height, stride_width). 
            Otherwise, stride_depth = stride_height = stride_width = stride. Default: stride = 1.
        padding (string|int|list|tuple): The padding size. It means the number of zero-paddings 
            on both sides for each dimention. If `padding` is a string, either 'VALID' or
L
liym27 已提交
1450 1451 1452 1453 1454 1455 1456 1457
            'SAME' which is the padding algorithm. If padding size is a tuple or list,
            it could be in three forms: `[pad_depth, pad_height, pad_width]` or
            `[pad_depth_front, pad_depth_back, pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`,
            and when `data_format` is `"NCDHW"`, `pool_padding` can be in the form
            `[[0,0], [0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`.
            when `data_format` is `"NDHWC"`, `pool_padding` can be in the form
            `[[0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`.
            Default: padding = 0.
L
lvmengsi 已提交
1458 1459 1460 1461
        dilation (int|tuple): The dilation size. It means the spacing between the kernel points. 
            If dilation is a tuple, it must contain three integers, (dilation_depth, dilation_height,
            dilation_width). Otherwise, dilation_depth = dilation_height = dilation_width = dilation. 
            Default: dilation = 1.
C
chengduoZH 已提交
1462 1463 1464 1465 1466
        groups (int): The groups number of the Conv3d Layer. According to grouped
            convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
            the first half of the filters is only connected to the first half
            of the input channels, while the second half of the filters is only
            connected to the second half of the input channels. Default: groups=1
C
chengduo 已提交
1467 1468 1469 1470 1471 1472 1473 1474 1475 1476
        param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
            of conv3d. If it is set to None or one attribute of ParamAttr, conv3d
            will create ParamAttr as param_attr. If it is set to None, the parameter
            is initialized with :math:`Normal(0.0, std)`, and the :math:`std` is
            :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None.
        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv3d.
            If it is set to False, no bias will be added to the output units.
            If it is set to None or one attribute of ParamAttr, conv3d
            will create ParamAttr as bias_attr. If the Initializer of the bias_attr
            is not set, the bias is initialized zero. Default: None.
C
chengduoZH 已提交
1477 1478
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
C
chengduo 已提交
1479 1480
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None.
L
lvmengsi 已提交
1481 1482 1483
        name(str|None): For detailed information, please refer 
           to :ref:`api_guide_Name`. Usually name is no need to set and 
           None by default.
1484 1485 1486 1487
        data_format (str, optional): Specify the data format of the input, and the data format of the output 
            will be consistent with that of the input. An optional string from: `"NCHW"`, `"NHWC"`.
            The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`.
C
chengduoZH 已提交
1488 1489

    Returns:
L
lvmengsi 已提交
1490 1491 1492 1493
        A Variable holding Tensor representing the conv3d, whose data type is 
        the same with input. If act is None, the tensor variable storing the 
        convolution result, and if act is not None, the tensor variable storing 
        convolution and non-linearity activation result.
C
chengduoZH 已提交
1494

1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507
    Raises:
        ValueError: If the type of `use_cudnn` is not bool.
        ValueError: If `data_format` is not "NCDHW" or "NDHWC".
        ValueError: If the channel dimmention of the input is less than or equal to zero.
        ValueError: If `padding` is a string, but not "SAME" or "VALID".
        ValueError: If `padding` is a tuple, but the element corresponding to the input's batch size is not 0 
            or the element corresponding to the input's channel is not 0.
        ShapeError: If the input is not 5-D Tensor.
        ShapeError: If the input's dimension size and filter's dimension size not equal.
        ShapeError: If the dimension size of input minus the size of `stride` is not 2.
        ShapeError: If the number of input channels is not equal to filter's channels * groups.
        ShapeError: If the number of output channels is not be divided by groups.

C
chengduoZH 已提交
1508 1509 1510
    Examples:
        .. code-block:: python

1511
          import paddle.fluid as fluid
L
lvmengsi 已提交
1512
          data = fluid.data(name='data', shape=[None, 3, 12, 32, 32], dtype='float32')
1513
          conv3d = fluid.layers.conv3d(input=data, num_filters=2, filter_size=3, act="relu")
C
chengduoZH 已提交
1514 1515 1516
    """

    l_type = 'conv3d'
C
chengduo 已提交
1517
    assert param_attr is not False, "param_attr should not be False here."
C
chengduoZH 已提交
1518 1519 1520
    helper = LayerHelper(l_type, **locals())
    dtype = helper.input_dtype()

L
liym27 已提交
1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535
    if not isinstance(use_cudnn, bool):
        raise ValueError("Attr(use_cudnn) should be True or False. Received "
                         "Attr(use_cudnn): %s. " % str(use_cudnn))

    if data_format not in ["NCDHW", "NDHWC"]:
        raise ValueError(
            "Attr(data_format) should be 'NCDHW' or 'NDHWC'. Received "
            "Attr(data_format): %s." % str(data_format))

    channel_last = (data_format == "NDHWC")
    num_channels = input.shape[4] if channel_last else input.shape[1]
    if num_channels < 0:
        raise ValueError(
            "The channel dimmention of the input(%s) should be defined. "
            "Received: %s." % (str(input.shape), str(num_channels)))
C
chengduoZH 已提交
1536 1537 1538 1539 1540

    if groups is None:
        num_filter_channels = num_channels
    else:
        if num_channels % groups != 0:
L
liym27 已提交
1541 1542 1543 1544
            raise ValueError(
                "The number of input channels must be divisible by Attr(groups). "
                "Received: number of channels(%s), groups(%s)." %
                (str(num_channels), str(groups)))
M
minqiyang 已提交
1545
        num_filter_channels = num_channels // groups
C
chengduoZH 已提交
1546 1547 1548 1549 1550

    filter_size = utils.convert_to_list(filter_size, 3, 'filter_size')
    stride = utils.convert_to_list(stride, 3, 'stride')
    dilation = utils.convert_to_list(dilation, 3, 'dilation')

L
liym27 已提交
1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572
    def _update_padding(padding, data_format):
        def is_list_or_tuple(ele):
            if isinstance(ele, list) or isinstance(ele, tuple):
                return True
            return False

        if is_list_or_tuple(padding) and len(padding) == 5:
            if is_list_or_tuple(padding[0]) and (data_format == "NCDHW"):
                if not (padding[0] == [0, 0] and padding[1] == [0, 0]):
                    raise ValueError(
                        "Non-zero padding(%s) in the batch or channel dimensions "
                        "is not supported." % str(padding))
                padding = padding[2:5]
                padding = [ele for a_list in padding for ele in a_list]
            elif is_list_or_tuple(padding[0]) and (data_format == "NDHWC"):
                if not (padding[0] == [0, 0] and padding[4] == [0, 0]):
                    raise ValueError(
                        "Non-zero padding(%s) in the batch or channel dimensions "
                        "is not supported." % str(padding))
                padding = padding[1:4]
                padding = [ele for a_list in padding for ele in a_list]
            padding = utils.convert_to_list(padding, 6, 'padding')
1573 1574
            if utils._is_symmetric_padding(padding, 3):
                padding = [padding[0], padding[2], padding[4]]
L
liym27 已提交
1575 1576
        elif is_list_or_tuple(padding) and len(padding) == 6:
            padding = utils.convert_to_list(padding, 6, 'padding')
1577 1578
            if utils._is_symmetric_padding(padding, 3):
                padding = [padding[0], padding[2], padding[4]]
L
liym27 已提交
1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592
        else:
            padding = utils.convert_to_list(padding, 3, 'padding')

        return padding

    padding_algorithm = "EXPLICIT"
    if isinstance(padding, str):
        padding = padding.upper()
        if padding not in ["SAME", "VALID"]:
            raise ValueError(
                "Unknown padding: '%s'. It can only be 'SAME' or 'VALID'." %
                str(padding))
        if padding == "VALID":
            padding_algorithm = "VALID"
1593
            padding = [0, 0, 0]
L
liym27 已提交
1594 1595
        elif padding == "SAME":
            padding_algorithm = "SAME"
1596
            padding = [0, 0, 0]
L
liym27 已提交
1597 1598

    padding = _update_padding(padding, data_format)
C
chengduoZH 已提交
1599 1600 1601 1602 1603

    input_shape = input.shape
    filter_shape = [num_filters, num_filter_channels] + filter_size

    def _get_default_param_initializer():
C
chengduo 已提交
1604 1605 1606
        filter_elem_num = filter_size[0] * filter_size[1] * filter_size[
            2] * num_channels
        std = (2.0 / filter_elem_num)**0.5
C
chengduoZH 已提交
1607 1608 1609 1610 1611 1612 1613 1614
        return Normal(0.0, std, 0)

    filter_param = helper.create_parameter(
        attr=helper.param_attr,
        shape=filter_shape,
        dtype=dtype,
        default_initializer=_get_default_param_initializer())

X
Xin Pan 已提交
1615
    pre_bias = helper.create_variable_for_type_inference(dtype)
C
chengduoZH 已提交
1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629

    helper.append_op(
        type=l_type,
        inputs={
            'Input': input,
            'Filter': filter_param,
        },
        outputs={"Output": pre_bias},
        attrs={
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn,
L
liym27 已提交
1630 1631 1632
            'use_mkldnn': False,
            "padding_algorithm": padding_algorithm,
            "data_format": data_format,
C
chengduoZH 已提交
1633 1634
        })

1635 1636 1637 1638
    if data_format == 'NCDHW':
        pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
    else:
        pre_act = helper.append_bias_op(pre_bias, dim_start=4, dim_end=5)
C
chengduoZH 已提交
1639 1640 1641 1642

    return helper.append_activation(pre_act)


F
fengjiayi 已提交
1643
@templatedoc()
Y
Yu Yang 已提交
1644
def pool2d(input,
C
chengduoZH 已提交
1645 1646
           pool_size=-1,
           pool_type="max",
C
chengduoZH 已提交
1647 1648
           pool_stride=1,
           pool_padding=0,
C
caoying03 已提交
1649
           global_pooling=False,
C
chengduoZH 已提交
1650
           use_cudnn=True,
1651
           ceil_mode=False,
1652
           name=None,
1653 1654
           exclusive=True,
           data_format="NCHW"):
Y
Yu Yang 已提交
1655
    """
F
fengjiayi 已提交
1656
    ${comment}
1657 1658

    Args:
K
Kaipeng Deng 已提交
1659 1660 1661 1662 1663
        input (Variable): The input tensor of pooling operator which is a 4-D tensor with
                          shape [N, C, H, W]. The format of input tensor is `"NCHW"` or
                          `"NHWC"`, where `N` is batch size, `C` is the number of channels,
                          `H` is the height of the feature, and `W` is the width of the
                          feature. The data type if float32 or float64.
J
JiabinYang 已提交
1664
        pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list,
J
JiabinYang 已提交
1665 1666
            it must contain two integers, (pool_size_Height, pool_size_Width).
            Otherwise, the pool kernel size will be a square of an int.
F
fengjiayi 已提交
1667
        pool_type: ${pooling_type_comment}
J
JiabinYang 已提交
1668 1669 1670
        pool_stride (int|list|tuple): The pool stride size. If pool stride size is a tuple or list,
            it must contain two integers, (pool_stride_Height, pool_stride_Width).
            Otherwise, the pool stride size will be a square of an int.
1671 1672 1673 1674 1675 1676 1677
        pool_padding (string|int|list|tuple): The pool padding. If `pool_padding` is a string, either 'VALID' or
            'SAME' which is the padding algorithm. If pool padding size is a tuple or list,
            it could be in three forms: `[pad_height, pad_width]` or
            `[pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`, and when `data_format` is `"NCHW"`,
            `pool_padding` can be in the form `[[0,0], [0,0], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`.
            when `data_format` is `"NHWC"`, `pool_padding` can be in the form
            `[[0,0], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`.
J
JiabinYang 已提交
1678
            Otherwise, the pool padding size will be a square of an int.
1679 1680 1681
        global_pooling (bool): ${global_pooling_comment}
        use_cudnn (bool): ${use_cudnn_comment}
        ceil_mode (bool): ${ceil_mode_comment}
K
Kaipeng Deng 已提交
1682 1683 1684
        name(str, optional): For detailed information, please refer
                             to :ref:`api_guide_Name`. Usually name is no need to set and
                             None by default.
1685
        exclusive (bool): Whether to exclude padding points in average pooling
1686 1687 1688 1689
                          mode, default is `true`.
        data_format (string): The data format of the input and output data. An optional string from: `"NCHW"`, `"NDHW"`.
                The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
                `[batch_size, input_channels, input_height, input_width]`.
F
fengjiayi 已提交
1690

1691
    Returns:
K
Kaipeng Deng 已提交
1692
        Variable: The output tensor of pooling result. The data type is same as input tensor.
F
fengjiayi 已提交
1693 1694

    Raises:
1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706
        ValueError: If `pool_type` is not "max" nor "avg".
        ValueError: If `global_pooling` is False and `pool_size` is -1.
        TypeError: If `use_cudnn` is not a bool value.
        ValueError: If `data_format` is not "NCHW" or "NHWC".
        ValueError: If `pool_padding` is a string, but not "SAME" or "VALID".
        ValueError: If `pool_padding` is "VALID", but `ceil_mode` is True.
        ValueError: If `pool_padding` is a list or tuple, but the elements in the batch or channel dimensions are non-zero.
        ShapeError: If the input is not a 4-D or 5-D Tensor.
        ShapeError: If the dimension of input minus the size of `pool_stride` is not 2.
        ShapeError: If the size of `pool_size` and `pool_stride` is not equal.
        ShapeError: If the output's shape calculated is not greater than 0.

F
fengjiayi 已提交
1707 1708 1709 1710 1711

    Examples:

        .. code-block:: python

1712
          import paddle.fluid as fluid
1713

K
Kaipeng Deng 已提交
1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738
          data = fluid.data(name='data', shape=[None, 3, 32, 32], dtype='float32')

          # max pool2d
          pool2d = fluid.layers.pool2d(
            input = data,
            pool_size = 2,
            pool_type = "max",
            pool_stride = 1,
            global_pooling=False)

          # average pool2d
          pool2d = fluid.layers.pool2d(
            input = data,
            pool_size = 2,
            pool_type = "avg",
            pool_stride = 1,
            global_pooling=False)

          # global average pool2d
          pool2d = fluid.layers.pool2d(
            input = data,
            pool_size = 2,
            pool_type = "avg",
            pool_stride = 1,
            global_pooling=True)
1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756

          # Attr(pool_padding) is a list with 4 elements, Attr(data_format) is "NCHW".
          out_1 = fluid.layers.pool2d(
            input = data,
            pool_size = 3,
            pool_type = "avg",
            pool_stride = 1,
            pool_padding = [1, 2, 1, 0],
            data_format = "NCHW")

          # Attr(pool_padding) is a string, Attr(data_format) is "NCHW".
          out_2 = fluid.layers.pool2d(
            input = data,
            pool_size = 3,
            pool_type = "avg",
            pool_stride = 1,
            pool_padding = "VALID",
            data_format = "NCHW")
Y
Yu Yang 已提交
1757 1758 1759
    """
    if pool_type not in ["max", "avg"]:
        raise ValueError(
1760
            "Unknown Attr(pool_type): '%s'. It can only be 'max' or 'avg'.",
Y
Yu Yang 已提交
1761
            str(pool_type))
C
chengduoZH 已提交
1762

C
chengduoZH 已提交
1763 1764
    if global_pooling is False and pool_size == -1:
        raise ValueError(
1765 1766 1767 1768
            "When Attr(global_pooling) is False, Attr(pool_size) must be passed "
            "and be a valid value. Received pool_size: %s." % str(pool_size))

    if not isinstance(use_cudnn, bool):
1769 1770
        raise TypeError("Attr(use_cudnn) should be True or False. Received "
                        "Attr(use_cudnn): %s." % str(use_cudnn))
1771 1772 1773 1774 1775

    if data_format not in ["NCHW", "NHWC"]:
        raise ValueError(
            "Attr(data_format) should be 'NCHW' or 'NHWC'. Received "
            "Attr(data_format): %s." % str(data_format))
C
chengduoZH 已提交
1776

C
chengduoZH 已提交
1777 1778 1779
    pool_size = utils.convert_to_list(pool_size, 2, 'pool_size')
    pool_stride = utils.convert_to_list(pool_stride, 2, 'pool_stride')

1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801
    def update_padding(padding, data_format):
        def is_list_or_tuple(ele):
            if isinstance(ele, list) or isinstance(ele, tuple):
                return True
            return False

        if is_list_or_tuple(padding) and len(padding) == 4:
            if is_list_or_tuple(padding[0]) and (data_format == "NCHW"):
                if not (padding[0] == [0, 0] and padding[1] == [0, 0]):
                    raise ValueError(
                        "Non-zero pool_padding(%s) in the batch or channel dimensions "
                        "is not supported." % str(padding))
                padding = padding[2:4]
                padding = [ele for a_list in padding for ele in a_list]
            elif is_list_or_tuple(padding[0]) and (data_format == "NHWC"):
                if not (padding[0] == [0, 0] and padding[3] == [0, 0]):
                    raise ValueError(
                        "Non-zero pool_padding(%s) in the batch or channel dimensions "
                        "is not supported." % str(padding))
                padding = padding[1:3]
                padding = [ele for a_list in padding for ele in a_list]
            padding = utils.convert_to_list(padding, 4, 'padding')
1802

1803 1804
            if utils._is_symmetric_padding(padding, 2):
                padding = [padding[0], padding[2]]
1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818
        else:
            padding = utils.convert_to_list(padding, 2, 'padding')

        return padding

    padding_algorithm = "EXPLICIT"
    if isinstance(pool_padding, str):
        pool_padding = pool_padding.upper()
        if pool_padding not in ["SAME", "VALID"]:
            raise ValueError(
                "Unknown Attr(pool_padding): '%s'. It can only be 'SAME' or 'VALID'."
                % str(pool_padding))
        if pool_padding == "VALID":
            padding_algorithm = "VALID"
1819
            pool_padding = [0, 0]
1820 1821 1822 1823 1824 1825
            if ceil_mode != False:
                raise ValueError(
                    "When Attr(pool_padding) is \"VALID\", Attr(ceil_mode) must be False. "
                    "Received ceil_mode: True.")
        elif pool_padding == "SAME":
            padding_algorithm = "SAME"
1826
            pool_padding = [0, 0]
1827 1828 1829 1830 1831

    pool_padding = update_padding(pool_padding, data_format)

    op_type = 'pool2d'
    helper = LayerHelper(op_type, **locals())
Y
Yu Yang 已提交
1832
    dtype = helper.input_dtype()
X
Xin Pan 已提交
1833
    pool_out = helper.create_variable_for_type_inference(dtype)
Y
Yu Yang 已提交
1834 1835

    helper.append_op(
1836
        type=op_type,
1837 1838 1839 1840 1841 1842 1843 1844
        inputs={"X": input},
        outputs={"Out": pool_out},
        attrs={
            "pooling_type": pool_type,
            "ksize": pool_size,
            "global_pooling": global_pooling,
            "strides": pool_stride,
            "paddings": pool_padding,
1845
            "padding_algorithm": padding_algorithm,
1846 1847
            "use_cudnn": use_cudnn,
            "ceil_mode": ceil_mode,
1848 1849
            "use_mkldnn": False,
            "exclusive": exclusive,
1850
            "data_format": data_format,
1851 1852 1853 1854 1855
        })

    return pool_out


D
dengkaipeng 已提交
1856
@templatedoc()
1857 1858 1859 1860 1861 1862 1863 1864
def pool3d(input,
           pool_size=-1,
           pool_type="max",
           pool_stride=1,
           pool_padding=0,
           global_pooling=False,
           use_cudnn=True,
           ceil_mode=False,
1865
           name=None,
1866 1867
           exclusive=True,
           data_format="NCDHW"):
1868
    """
1869
    ${comment}
1870 1871

    Args:
K
Kaipeng Deng 已提交
1872 1873
        input (Variable): The input tensor of pooling operator, which is a 5-D tensor with
                          shape [N, C, D, H, W]. The format of
1874 1875 1876
                          input tensor is `"NCDHW"` or `"NDHWC"`, where `N` is batch size, `C` is
                          the number of channels, `D` is the depth of the feature,
                          `H` is the height of the feature, and `W` is the width
D
dengkaipeng 已提交
1877
                          of the feature.
D
dengkaipeng 已提交
1878 1879 1880 1881 1882
        pool_size (int|list|tuple): The pool kernel size. If pool kernel size 
            is a tuple or list, it must contain three integers, 
            (pool_size_Depth, pool_size_Height, pool_size_Width).
            Otherwise, the pool kernel size will be the cube of an int.
        pool_type (string): ${pooling_type_comment}
1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893
        pool_stride (string|int|list|tuple)): The pool padding. If `pool_padding` is a string, either 'VALID' or
            'SAME' which is the padding algorithm. If pool stride size is a tuple or list,
            it must contain three integers, `[stride_Depth, stride_Height, stride_Width]`.
            Otherwise, the pool stride size will be a cube of an int.
        pool_padding (int|list|tuple): The pool padding size. If pool padding size is a tuple or list,
            it could be in three forms: `[pad_depth, pad_height, pad_width]` or
            `[pad_depth_front, pad_depth_back, pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`,
            and when `data_format` is `"NCDHW"`, `pool_padding` can be in the form
            `[[0,0], [0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`.
            when `data_format` is `"NDHWC"`, `pool_padding` can be in the form
            `[[0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`.
1894 1895 1896
        global_pooling (bool): ${global_pooling_comment}
        use_cudnn (bool): ${use_cudnn_comment}
        ceil_mode (bool): ${ceil_mode_comment}
K
Kaipeng Deng 已提交
1897 1898 1899
        name(str, optional): For detailed information, please refer
                             to :ref:`api_guide_Name`. Usually name is no need to set and
                             None by default.
1900
        exclusive (bool): Whether to exclude padding points in average pooling
1901 1902 1903 1904
                          mode, default is true.
        data_format (string): The data format of the input and output data. An optional string from: `"NCDHW"`, `"NDHWC"`.
                The default is `"NCDHW"`. When it is `"NCDHW"`, the data is stored in the order of:
                `[batch_size, input_channels, input_depth, input_height, input_width]`.
1905

1906
    Returns:
K
Kaipeng Deng 已提交
1907
        Variable: The output tensor of pooling result. The data type is same as input tensor.
D
dengkaipeng 已提交
1908

1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921
    Raises:
        ValueError: If `pool_type` is not "max" nor "avg".
        ValueError: If `global_pooling` is False and `pool_size` is -1.
        TypeError: If `use_cudnn` is not a bool value.
        ValueError: If `data_format` is not "NCDHW" or "NDHWC".
        ValueError: If `pool_padding` is a string, but not "SAME" or "VALID".
        ValueError: If `pool_padding` is "VALID", but `ceil_mode` is True.
        ValueError: If `pool_padding` is a list or tuple, but the elements in the batch or channel dimensions are non-zero.
        ShapeError: If the input is not a 4-D or 5-D Tensor.
        ShapeError: If the dimension of input minus the size of `pool_stride` is not 2.
        ShapeError: If the size of `pool_size` and `pool_stride` is not equal.
        ShapeError: If the output's shape calculated is not greater than 0.

D
dengkaipeng 已提交
1922 1923 1924 1925
    Examples:

        .. code-block:: python

1926
          import paddle.fluid as fluid
1927

K
Kaipeng Deng 已提交
1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952
          data = fluid.data(name='data', shape=[None, 3, 32, 32, 32], dtype='float32')

          # max pool3d
          pool3d = fluid.layers.pool3d(
            input = data,
            pool_size = 2,
            pool_type = "max",
            pool_stride = 1,
            global_pooling=False)

          # average pool3d
          pool3d = fluid.layers.pool3d(
            input = data,
            pool_size = 2,
            pool_type = "avg",
            pool_stride = 1,
            global_pooling=False)

          # global average pool3d
          pool3d = fluid.layers.pool3d(
            input = data,
            pool_size = 2,
            pool_type = "avg",
            pool_stride = 1,
            global_pooling=True)
1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975

          # example 1:
          # Attr(pool_padding) is a list with 6 elements, Attr(data_format) is "NCDHW".
          out_1 = fluid.layers.pool3d(
            input = data,
            pool_size = 2,
            pool_type = "avg",
            pool_stride = 1,
            pool_padding = [1, 2, 1, 0, 1, 2],
            global_pooling = False,
            data_format = "NCDHW")

          # example 2:
          # Attr(pool_padding) is a string, Attr(data_format) is "NCDHW".
          out_2 = fluid.layers.pool3d(
            input = data,
            pool_size = 3,
            pool_type = "avg",
            pool_stride = 1,
            pool_padding = "VALID",
            global_pooling = False,
            data_format = "NCDHW")

Y
Yu Yang 已提交
1976 1977 1978
    """
    if pool_type not in ["max", "avg"]:
        raise ValueError(
1979
            "Unknown Attr(pool_type): '%s'. It can only be 'max' or 'avg'.",
Y
Yu Yang 已提交
1980
            str(pool_type))
C
chengduoZH 已提交
1981

C
chengduoZH 已提交
1982 1983
    if global_pooling is False and pool_size == -1:
        raise ValueError(
1984 1985 1986 1987 1988
            "When Attr(global_pooling) is False, Attr(pool_size) must be passed "
            "and be a valid value. Received Attr(pool_size): %s." %
            str(pool_size))

    if not isinstance(use_cudnn, bool):
1989 1990
        raise TypeError("Attr(use_cudnn) should be True or False. Received "
                        "Attr(use_cudnn): %s. " % str(use_cudnn))
1991 1992 1993 1994 1995

    if data_format not in ["NCDHW", "NDHWC"]:
        raise ValueError(
            "Attr(data_format) should be 'NCDHW' or 'NDHWC'. Received "
            "Attr(data_format): %s" % str(data_format))
C
chengduoZH 已提交
1996

1997 1998
    pool_size = utils.convert_to_list(pool_size, 3, 'pool_size')
    pool_stride = utils.convert_to_list(pool_stride, 3, 'pool_stride')
C
chengduoZH 已提交
1999

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
    def update_padding(padding, data_format):
        def is_list_or_tuple(ele):
            if isinstance(ele, (list, tuple)):
                return True
            return False

        if is_list_or_tuple(padding) and len(padding) == 5:
            if is_list_or_tuple(padding[0]) and (data_format == "NCDHW"):
                if not (padding[0] == [0, 0] and padding[1] == [0, 0]):
                    raise ValueError(
                        "Non-zero pool_padding(%s) in the batch or channel dimensions "
                        "is not supported." % str(padding))
                padding = padding[2:5]
                padding = [ele for a_list in padding for ele in a_list]
            elif is_list_or_tuple(padding[0]) and (data_format == "NDHWC"):
                if not (padding[0] == [0, 0] and padding[4] == [0, 0]):
                    raise ValueError(
                        "Non-zero pool_padding(%s) in the batch or channel dimensions "
                        "is not supported." % str(padding))
                padding = padding[1:4]
                padding = [ele for a_list in padding for ele in a_list]
            padding = utils.convert_to_list(padding, 6, 'padding')
2022 2023
            if utils._is_symmetric_padding(padding, 3):
                padding = [padding[0], padding[2], padding[4]]
2024 2025 2026

        elif is_list_or_tuple(padding) and len(padding) == 6:
            padding = utils.convert_to_list(padding, 6, 'padding')
2027 2028
            if utils._is_symmetric_padding(padding, 3):
                padding = [padding[0], padding[2], padding[4]]
2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042
        else:
            padding = utils.convert_to_list(padding, 3, 'padding')

        return padding

    padding_algorithm = "EXPLICIT"
    if isinstance(pool_padding, str):
        pool_padding = pool_padding.upper()
        if pool_padding not in ["SAME", "VALID"]:
            raise ValueError(
                "Unknown Attr(pool_padding): '%s'. It can only be 'SAME' or 'VALID'."
                % str(pool_padding))
        if pool_padding == "VALID":
            padding_algorithm = "VALID"
2043
            pool_padding = [0, 0, 0]
2044 2045 2046 2047 2048 2049
            if ceil_mode != False:
                raise ValueError(
                    "When Attr(pool_padding) is \"VALID\", ceil_mode must be False. "
                    "Received ceil_mode: True.")
        elif pool_padding == "SAME":
            padding_algorithm = "SAME"
2050
            pool_padding = [0, 0, 0]
2051 2052 2053 2054 2055

    pool_padding = update_padding(pool_padding, data_format)

    op_type = "pool3d"
    helper = LayerHelper(op_type, **locals())
Y
Yu Yang 已提交
2056
    dtype = helper.input_dtype()
X
Xin Pan 已提交
2057
    pool_out = helper.create_variable_for_type_inference(dtype)
Y
Yu Yang 已提交
2058 2059

    helper.append_op(
2060
        type=op_type,
Y
Yu Yang 已提交
2061 2062 2063 2064 2065 2066 2067
        inputs={"X": input},
        outputs={"Out": pool_out},
        attrs={
            "pooling_type": pool_type,
            "ksize": pool_size,
            "global_pooling": global_pooling,
            "strides": pool_stride,
C
chengduoZH 已提交
2068
            "paddings": pool_padding,
2069
            "padding_algorithm": padding_algorithm,
2070
            "use_cudnn": use_cudnn,
2071
            "ceil_mode": ceil_mode,
2072 2073
            "use_mkldnn": False,
            "exclusive": exclusive,
2074
            "data_format": data_format,
Y
Yu Yang 已提交
2075 2076 2077 2078 2079
        })

    return pool_out


2080 2081 2082 2083 2084 2085 2086
@templatedoc(op_type="pool2d")
def adaptive_pool2d(input,
                    pool_size,
                    pool_type="max",
                    require_index=False,
                    name=None):
    """
K
Kaipeng Deng 已提交
2087
    This operation calculates the output based on the input, pool_size,
D
dengkaipeng 已提交
2088 2089 2090 2091
    pool_type parameters. Input(X) and output(Out) are in NCHW format, where N is batch
    size, C is the number of channels, H is the height of the feature, and W is
    the width of the feature. Parameters(pool_size) should contain two elements which
    represent height and width, respectively. Also the H and W dimensions of output(Out)
K
Kaipeng Deng 已提交
2092
    is same as Parameter(pool_size). The output tensor shape will be [N, C, pool_size[0], pool_size[1]]
2093

2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106
    For average adaptive pool2d:

    ..  math::

       hstart &= floor(i * H_{in} / H_{out})

       hend &= ceil((i + 1) * H_{in} / H_{out})

       wstart &= floor(j * W_{in} / W_{out})

       wend &= ceil((j + 1) * W_{in} / W_{out})

       Output(i ,j) &= \\frac{sum(Input[hstart:hend, wstart:wend])}{(hend - hstart) * (wend - wstart)}
2107 2108

    Args:
K
Kaipeng Deng 已提交
2109 2110 2111 2112 2113
        input (Variable): The input tensor of pooling operator, which is a 4-D tensor
                          with shape [N, C, H, W].  The format of input tensor is NCHW,
                          where N is batch size, C is the number of channels, H is the
                          height of the feature, and W is the width of the feature.
                          The data type is float32 or float64.
2114 2115 2116
        pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list,
            it must contain two integers, (pool_size_Height, pool_size_Width).
        pool_type: ${pooling_type_comment}
D
dengkaipeng 已提交
2117
        require_index (bool): If true, the index of max pooling point will be returned along
K
Kaipeng Deng 已提交
2118 2119 2120 2121
            with outputs. It cannot be set in average pooling type. Default False.
        name(str, optional): For detailed information, please refer
                             to :ref:`api_guide_Name`. Usually name is no need to set and
                             None by default.
2122 2123

    Returns:
K
Kaipeng Deng 已提交
2124 2125
        Variable: The output tensor of adaptive pooling result. The data type is same 
                  as input tensor.
2126 2127 2128 2129 2130 2131 2132 2133 2134

    Raises:
        ValueError: 'pool_type' is not 'max' nor 'avg'.
        ValueError: invalid setting 'require_index' true when 'pool_type' is 'avg'.
        ValueError: 'pool_size' should be a list or tuple with length as 2.

    Examples:
        .. code-block:: python

K
Kaipeng Deng 已提交
2135
          # average adaptive pool2d
M
minqiyang 已提交
2136
          # suppose input data in shape of [N, C, H, W], `pool_size` is [m, n],
2137
          # output shape is [N, C, m, n], adaptive pool divide H and W dimentions
M
minqiyang 已提交
2138
          # of input data into m * n grids averagely and performs poolings in each
2139 2140
          # grid to get output.
          # adaptive average pool performs calculations as follow:
M
minqiyang 已提交
2141
          #
2142 2143 2144 2145 2146 2147 2148 2149
          #     for i in range(m):
          #         for j in range(n):
          #             hstart = floor(i * H / m)
          #             hend = ceil((i + 1) * H / m)
          #             wstart = floor(i * W / n)
          #             wend = ceil((i + 1) * W / n)
          #             output[:, :, i, j] = avg(input[:, :, hstart: hend, wstart: wend])
          #
2150
          import paddle.fluid as fluid
K
Kaipeng Deng 已提交
2151
          data = fluid.data(name='data', shape=[None, 3, 32, 32], dtype='float32')
D
dengkaipeng 已提交
2152
          pool_out = fluid.layers.adaptive_pool2d(
2153 2154
                            input=data,
                            pool_size=[3, 3],
2155
                            pool_type='avg')
K
Kaipeng Deng 已提交
2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177

          # max adaptive pool2d
          # suppose input data in shape of [N, C, H, W], `pool_size` is [m, n],
          # output shape is [N, C, m, n], adaptive pool divide H and W dimentions
          # of input data into m * n grids averagely and performs poolings in each
          # grid to get output.
          # adaptive average pool performs calculations as follow:
          #
          #     for i in range(m):
          #         for j in range(n):
          #             hstart = floor(i * H / m)
          #             hend = ceil((i + 1) * H / m)
          #             wstart = floor(i * W / n)
          #             wend = ceil((i + 1) * W / n)
          #             output[:, :, i, j] = max(input[:, :, hstart: hend, wstart: wend])
          #
          import paddle.fluid as fluid
          data = fluid.data(name='data', shape=[None, 3, 32, 32], dtype='float32')
          pool_out = fluid.layers.adaptive_pool2d(
                            input=data,
                            pool_size=[3, 3],
                            pool_type='max')
2178 2179 2180 2181 2182 2183 2184 2185 2186 2187
    """
    if pool_type not in ["max", "avg"]:
        raise ValueError(
            "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",
            str(pool_type))

    if pool_type == "avg" and require_index:
        raise ValueError(
            "invalid setting 'require_index' true when 'pool_type' is 'avg'.")

2188
    pool_size = utils.convert_to_list(pool_size, 2, 'pool_size')
2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213

    if pool_type == "max":
        l_type = 'max_pool2d_with_index'
    else:
        l_type = "pool2d"

    helper = LayerHelper(l_type, **locals())
    dtype = helper.input_dtype()
    pool_out = helper.create_variable_for_type_inference(dtype)

    outputs = {"Out": pool_out}
    if pool_type == "max":
        mask = helper.create_variable_for_type_inference(dtype)
        outputs["Mask"] = mask

    helper.append_op(
        type=l_type,
        inputs={"X": input},
        outputs=outputs,
        attrs={
            "pooling_type": pool_type,
            "ksize": pool_size,
            "adaptive": True,
        })

D
dengkaipeng 已提交
2214
    return (pool_out, mask) if require_index else pool_out
2215 2216 2217 2218 2219 2220 2221 2222 2223


@templatedoc(op_type="pool3d")
def adaptive_pool3d(input,
                    pool_size,
                    pool_type="max",
                    require_index=False,
                    name=None):
    """
K
Kaipeng Deng 已提交
2224
    This operation calculates the output based on the input, pool_size,
D
dengkaipeng 已提交
2225 2226 2227 2228
    pool_type parameters. Input(X) and output(Out) are in NCDHW format, where N is batch
    size, C is the number of channels, D is the depth of the feature, H is the height of
    the feature, and W is the width of the feature. Parameters(pool_size) should contain
    three elements which represent height and width, respectively. Also the D, H and W
K
Kaipeng Deng 已提交
2229 2230
    dimensions of output(Out) is same as Parameter(pool_size). The output tensor shape
    will be [N, C, pool_size[0], pool_size[1], pool_size[2]]
2231

2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248
    For average adaptive pool3d:

    ..  math::

      dstart &= floor(i * D_{in} / D_{out})

      dend &= ceil((i + 1) * D_{in} / D_{out})

      hstart &= floor(j * H_{in} / H_{out})

      hend &= ceil((j + 1) * H_{in} / H_{out})

      wstart &= floor(k * W_{in} / W_{out})

      wend &= ceil((k + 1) * W_{in} / W_{out})

      Output(i ,j, k) &= \\frac{sum(Input[dstart:dend, hstart:hend, wstart:wend])}{(dend - dstart) * (hend - hstart) * (wend - wstart)}
2249 2250

    Args:
K
Kaipeng Deng 已提交
2251 2252 2253
        input (Variable): The input tensor of pooling operator, which is a 5-D tensor with 
                          shape [N, C, D, H, W]. The format of input tensor is NCDHW, where
                          N is batch size, C is the number of channels, D is the depth of the feature,
D
dengkaipeng 已提交
2254
                          H is the height of the feature, and W is the width of the feature.
K
Kaipeng Deng 已提交
2255
                          The data type is float32 or float64.
2256
        pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list,
D
dengkaipeng 已提交
2257
            it must contain three integers, (Depth, Height, Width).
2258
        pool_type: ${pooling_type_comment}
D
dengkaipeng 已提交
2259
        require_index (bool): If true, the index of max pooling point will be returned along
K
Kaipeng Deng 已提交
2260 2261 2262 2263
            with outputs. It cannot be set in average pooling type. Default False.
        name(str, optional): For detailed information, please refer
                             to :ref:`api_guide_Name`. Usually name is no need to set and
                             None by default.
2264 2265

    Returns:
K
Kaipeng Deng 已提交
2266
        Variable: The output tensor of adaptive pooling result. The data type is same as input tensor.
2267 2268 2269 2270 2271 2272 2273 2274 2275

    Raises:
        ValueError: 'pool_type' is not 'max' nor 'avg'.
        ValueError: invalid setting 'require_index' true when 'pool_type' is 'avg'.
        ValueError: 'pool_size' should be a list or tuple with length as 2.

    Examples:
        .. code-block:: python

K
Kaipeng Deng 已提交
2276
          # average adaptive pool3d
2277 2278
          # suppose input data in shape of [N, C, D, H, W], `pool_size` is [l, m, n],
          # output shape is [N, C, l, m, n], adaptive pool divide D, H and W dimentions
M
minqiyang 已提交
2279
          # of input data into l * m * n grids averagely and performs poolings in each
2280 2281
          # grid to get output.
          # adaptive average pool performs calculations as follow:
M
minqiyang 已提交
2282
          #
2283 2284 2285 2286 2287 2288 2289 2290 2291
          #     for i in range(l):
          #         for j in range(m):
          #             for k in range(n):
          #                 dstart = floor(i * D / l)
          #                 dend = ceil((i + 1) * D / l)
          #                 hstart = floor(j * H / m)
          #                 hend = ceil((j + 1) * H / m)
          #                 wstart = floor(k * W / n)
          #                 wend = ceil((k + 1) * W / n)
M
minqiyang 已提交
2292
          #                 output[:, :, i, j, k] =
2293 2294
          #                     avg(input[:, :, dstart:dend, hstart: hend, wstart: wend])
          #
K
Kaipeng Deng 已提交
2295 2296 2297

          import paddle.fluid as fluid

K
Kaipeng Deng 已提交
2298 2299
          data = fluid.data(
              name='data', shape=[None, 3, 32, 32, 32], dtype='float32')
K
Kaipeng Deng 已提交
2300
          pool_out = fluid.layers.adaptive_pool3d(
2301
                            input=data,
D
dengkaipeng 已提交
2302
                            pool_size=[3, 3, 3],
2303
                            pool_type='avg')
K
Kaipeng Deng 已提交
2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332

          # max adaptive pool3d
          # suppose input data in shape of [N, C, D, H, W], `pool_size` is [l, m, n],
          # output shape is [N, C, l, m, n], adaptive pool divide D, H and W dimentions
          # of input data into l * m * n grids averagely and performs poolings in each
          # grid to get output.
          # adaptive average pool performs calculations as follow:
          #
          #     for i in range(l):
          #         for j in range(m):
          #             for k in range(n):
          #                 dstart = floor(i * D / l)
          #                 dend = ceil((i + 1) * D / l)
          #                 hstart = floor(j * H / m)
          #                 hend = ceil((j + 1) * H / m)
          #                 wstart = floor(k * W / n)
          #                 wend = ceil((k + 1) * W / n)
          #                 output[:, :, i, j, k] =
          #                     avg(input[:, :, dstart:dend, hstart: hend, wstart: wend])
          #

          import paddle.fluid as fluid

          data = fluid.data(
              name='data', shape=[None, 3, 32, 32, 32], dtype='float32')
          pool_out = fluid.layers.adaptive_pool3d(
                            input=data,
                            pool_size=[3, 3, 3],
                            pool_type='max')
2333 2334 2335 2336 2337 2338 2339 2340 2341 2342
    """
    if pool_type not in ["max", "avg"]:
        raise ValueError(
            "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",
            str(pool_type))

    if pool_type == "avg" and require_index:
        raise ValueError(
            "invalid setting 'require_index' true when 'pool_type' is 'avg'.")

2343
    pool_size = utils.convert_to_list(pool_size, 3, 'pool_size')
2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368

    if pool_type == "max":
        l_type = 'max_pool3d_with_index'
    else:
        l_type = "pool3d"

    helper = LayerHelper(l_type, **locals())
    dtype = helper.input_dtype()
    pool_out = helper.create_variable_for_type_inference(dtype)

    outputs = {"Out": pool_out}
    if pool_type == "max":
        mask = helper.create_variable_for_type_inference(dtype)
        outputs["Mask"] = mask

    helper.append_op(
        type=l_type,
        inputs={"X": input},
        outputs=outputs,
        attrs={
            "pooling_type": pool_type,
            "ksize": pool_size,
            "adaptive": True,
        })

D
dengkaipeng 已提交
2369
    return (pool_out, mask) if require_index else pool_out
2370 2371


Y
Yu Yang 已提交
2372 2373 2374 2375 2376 2377 2378
def batch_norm(input,
               act=None,
               is_test=False,
               momentum=0.9,
               epsilon=1e-05,
               param_attr=None,
               bias_attr=None,
C
caoying03 已提交
2379
               data_layout='NCHW',
Y
Yang Yang 已提交
2380
               in_place=False,
2381 2382
               name=None,
               moving_mean_name=None,
W
wanghaoshuang 已提交
2383
               moving_variance_name=None,
2384
               do_model_average_for_mean_and_var=True,
2385
               use_global_stats=False):
Y
Yu Yang 已提交
2386
    """
Q
qiaolongfei 已提交
2387 2388
    **Batch Normalization Layer**

L
lvmengsi 已提交
2389
    Can be used as a normalizer function for convolution or fully_connected operations.
Q
qiaolongfei 已提交
2390
    The required data format for this layer is one of the following:
Q
qiaolongfei 已提交
2391

Q
qiaolongfei 已提交
2392
    1. NHWC `[batch, in_height, in_width, in_channels]`
Q
qiaolongfei 已提交
2393

Q
qiaolongfei 已提交
2394 2395
    2. NCHW `[batch, in_channels, in_height, in_width]`

Q
qiaolongfei 已提交
2396 2397 2398
    Refer to `Batch Normalization: Accelerating Deep Network Training by Reducing
    Internal Covariate Shift <https://arxiv.org/pdf/1502.03167.pdf>`_
    for more details.
Q
qiaolongfei 已提交
2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410

    :math:`input` is the input features over a mini-batch.

    ..  math::

        \\mu_{\\beta} &\\gets \\frac{1}{m} \\sum_{i=1}^{m} x_i \\qquad &//\\
        \ mini-batch\ mean \\\\
        \\sigma_{\\beta}^{2} &\\gets \\frac{1}{m} \\sum_{i=1}^{m}(x_i - \\
        \\mu_{\\beta})^2 \\qquad &//\ mini-batch\ variance \\\\
        \\hat{x_i} &\\gets \\frac{x_i - \\mu_\\beta} {\\sqrt{\\
        \\sigma_{\\beta}^{2} + \\epsilon}} \\qquad &//\ normalize \\\\
        y_i &\\gets \\gamma \\hat{x_i} + \\beta \\qquad &//\ scale\ and\ shift
2411

L
lvmengsi 已提交
2412 2413 2414
        moving\_mean = moving\_mean * momentum + mini-batch\_mean * (1. - momentum) \\\\
        moving\_var = moving\_var * momentum + mini-batch\_var * (1. - momentum) 

2415

L
lvmengsi 已提交
2416
    moving_mean is global mean and moving_var is global variance.
2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429

    When use_global_stats = True, the :math:`\\mu_{\\beta}`
    and :math:`\\sigma_{\\beta}^{2}` are not the statistics of one mini-batch.
    They are global (or running) statistics. (It usually got from the
    pre-trained model.)
    The training and testing (or inference) have the same behavior:

    ..  math::

        \\hat{x_i} &\\gets \\frac{x_i - \\mu_\\beta} {\\sqrt{\\
        \\sigma_{\\beta}^{2} + \\epsilon}}  \\\\
        y_i &\\gets \\gamma \\hat{x_i} + \\beta

L
lvmengsi 已提交
2430 2431 2432 2433
    Note:
        if build_strategy.sync_batch_norm=True, the batch_norm in network will use 
        sync_batch_norm automatically.

2434
    Args:
2435
        input(Variable): The rank of input variable can be 2, 3, 4, 5. The data type 
L
lvmengsi 已提交
2436
            is float16 or float32 or float64.
Q
qiaolongfei 已提交
2437
        act(string, Default None): Activation type, linear|relu|prelu|...
Q
qingqing01 已提交
2438 2439
        is_test (bool, Default False): A flag indicating whether it is in
            test phrase or not.
2440 2441 2442
        momentum(float|Variable, Default 0.9): The value used for the moving_mean and
            moving_var computation. This should be a float number or a Variable with
            shape [1] and data type as float32. The updated formula is:
Q
qingqing01 已提交
2443 2444 2445 2446 2447
            :math:`moving\_mean = moving\_mean * momentum + new\_mean * (1. - momentum)`
            :math:`moving\_var = moving\_var * momentum + new\_var * (1. - momentum)`
            Default is 0.9.
        epsilon(float, Default 1e-05): A value added to the denominator for
            numerical stability. Default is 1e-5.
C
chengduo 已提交
2448 2449
        param_attr(ParamAttr|None): The parameter attribute for Parameter `scale`
             of batch_norm. If it is set to None or one attribute of ParamAttr, batch_norm
2450 2451 2452
	     will create ParamAttr as param_attr, the name of scale can be set in ParamAttr.
	     If the Initializer of the param_attr is not set, the parameter is initialized 
	     with Xavier. Default: None.
C
chengduo 已提交
2453 2454
        bias_attr(ParamAttr|None): The parameter attribute for the bias of batch_norm.
             If it is set to None or one attribute of ParamAttr, batch_norm
2455 2456 2457
	     will create ParamAttr as bias_attr, the name of bias can be set in ParamAttr. 
	     If the Initializer of the bias_attr is not set, the bias is initialized zero. 
	     Default: None.
2458 2459 2460 2461
        data_layout (str, optional): Specify the data format of the input, and the data format of the output 
            will be consistent with that of the input. An optional string from: `"NCHW"`, `"NHWC"`.
            The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`.
2462
        in_place(bool, Default False): Make the input and output of batch norm reuse memory.
L
lvmengsi 已提交
2463 2464 2465
        name(str|None): For detailed information, please refer to :ref:`api_guide_Name`. 
            Usually name is no need to set and None by default. 
        moving_mean_name(str, Default None): The name of moving_mean which store the global Mean. If it 
2466 2467
            is set to None, batch_norm will save global mean with a random name, otherwise, batch_norm 
            will save global mean with the string.
L
lvmengsi 已提交
2468
        moving_variance_name(str, Default None): The name of the moving_variance which store the global Variance.
2469 2470
            If it is set to None, batch_norm will save global variance with a random name, otherwise, batch_norm 
            will save global variance with the string.
2471 2472
        do_model_average_for_mean_and_var(bool, Default True): Whether parameter mean and variance should do model
            average when model average is enabled.
2473 2474 2475 2476 2477
        use_global_stats(bool, Default False): Whether to use global mean and
            variance. In inference or test mode, set use_global_stats to true
            or is_test to true, and the behavior is equivalent.
            In train mode, when setting use_global_stats True, the global mean
            and variance are also used during train period.
2478 2479

    Returns:
L
lvmengsi 已提交
2480 2481
        A Variable holding Tensor which is the result after applying batch normalization on the input, 
        has same shape and data type with input. 
Q
qiaolongfei 已提交
2482 2483 2484 2485 2486

    Examples:

        .. code-block:: python

2487
            import paddle.fluid as fluid
L
lvmengsi 已提交
2488
            x = fluid.data(name='x', shape=[3, 7, 3, 7], dtype='float32')
Q
qiaolongfei 已提交
2489 2490
            hidden1 = fluid.layers.fc(input=x, size=200, param_attr='fc1.w')
            hidden2 = fluid.layers.batch_norm(input=hidden1)
2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517

        .. code-block:: python

            # batch_norm with momentum as Variable
            import paddle.fluid as fluid
            import paddle.fluid.layers.learning_rate_scheduler as lr_scheduler

            def get_decay_momentum(momentum_init, decay_steps, decay_rate):
                global_step = lr_scheduler._decay_step_counter()
                momentum = fluid.layers.create_global_var(
		    shape=[1],
		    value=float(momentum_init),
		    dtype='float32',
		    # set persistable for save checkpoints and resume
		    persistable=True,
		    name="momentum")
                div_res = global_step / decay_steps
                decayed_momentum = momentum_init * (decay_rate**div_res)
                fluid.layers.assign(decayed_momentum, momentum)

                return momentum

            x = fluid.data(name='x', shape=[3, 7, 3, 7], dtype='float32')
            hidden1 = fluid.layers.fc(input=x, size=200, param_attr='fc1.w')
            momentum = get_decay_momentum(0.9, 1e5, 0.9)
            hidden2 = fluid.layers.batch_norm(input=hidden1, momentum=momentum)

Y
Yu Yang 已提交
2518
    """
C
chengduo 已提交
2519
    assert bias_attr is not False, "bias_attr should not be False in batch_norm."
Y
Yu Yang 已提交
2520 2521
    helper = LayerHelper('batch_norm', **locals())

2522 2523
    check_type_and_dtype(input, 'input', Variable,
                         ['float16', 'float32', 'float64'], 'batch_norm')
2524
    dtype = helper.input_dtype()
2525 2526 2527 2528 2529 2530 2531

    has_reserve_space = False
    if data_layout == 'NHWC':
        flag = os.environ.get('FLAGS_cudnn_batchnorm_spatial_persistent')
        if flag is not None and flag.lower() in ['true', '1']:
            has_reserve_space = True

W
Wu Yi 已提交
2532 2533 2534 2535
    # use fp32 for bn parameter
    if dtype == core.VarDesc.VarType.FP16:
        dtype = core.VarDesc.VarType.FP32

Y
Yu Yang 已提交
2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553
    input_shape = input.shape
    if data_layout == 'NCHW':
        channel_num = input_shape[1]
    else:
        if data_layout == 'NHWC':
            channel_num = input_shape[-1]
        else:
            raise ValueError("unsupported data layout:" + data_layout)

    param_shape = [channel_num]

    # create parameter
    scale = helper.create_parameter(
        attr=helper.param_attr,
        shape=param_shape,
        dtype=dtype,
        default_initializer=Constant(1.0))
    bias = helper.create_parameter(
2554
        attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
Y
Yu Yang 已提交
2555

2556 2557
    mean = helper.create_parameter(
        attr=ParamAttr(
W
wanghaoshuang 已提交
2558 2559 2560
            name=moving_mean_name,
            initializer=Constant(0.0),
            trainable=False,
W
wanghaoshuang 已提交
2561
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
2562
        shape=param_shape,
W
Wu Yi 已提交
2563
        dtype=dtype)
2564 2565 2566 2567 2568 2569
    mean.stop_gradient = True

    variance = helper.create_parameter(
        attr=ParamAttr(
            name=moving_variance_name,
            initializer=Constant(1.0),
W
wanghaoshuang 已提交
2570
            trainable=False,
W
wanghaoshuang 已提交
2571
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
2572
        shape=param_shape,
W
Wu Yi 已提交
2573
        dtype=dtype)
2574
    variance.stop_gradient = True
Y
Yu Yang 已提交
2575 2576 2577 2578 2579 2580

    # create output
    # mean and mean_out share the same memory
    mean_out = mean
    # variance and variance out share the same memory
    variance_out = variance
X
Xin Pan 已提交
2581 2582 2583 2584
    saved_mean = helper.create_variable_for_type_inference(
        dtype=dtype, stop_gradient=True)
    saved_variance = helper.create_variable_for_type_inference(
        dtype=dtype, stop_gradient=True)
Y
Yu Yang 已提交
2585

2586 2587 2588 2589 2590
    reserve_space = None
    if has_reserve_space:
        reserve_space = helper.create_variable_for_type_inference(
            dtype=core.VarDesc.VarType.FP16, stop_gradient=True)

X
Xin Pan 已提交
2591 2592
    batch_norm_out = input if in_place else helper.create_variable_for_type_inference(
        dtype)
Y
Yu Yang 已提交
2593

2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612
    inputs = {
        "X": input,
        "Scale": scale,
        "Bias": bias,
        "Mean": mean,
        "Variance": variance
    }
    attrs = {
        "epsilon": epsilon,
        "is_test": is_test,
        "data_layout": data_layout,
        "use_mkldnn": False,
        "fuse_with_relu": False,
        "use_global_stats": use_global_stats
    }
    if isinstance(momentum, Variable):
        inputs['MomemtumTensor'] = momentum
    else:
        attrs['momentum'] = momentum
2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623

    outputs = {
        "Y": batch_norm_out,
        "MeanOut": mean_out,
        "VarianceOut": variance_out,
        "SavedMean": saved_mean,
        "SavedVariance": saved_variance
    }
    if reserve_space is not None:
        outputs["ReserveSpace"] = reserve_space

Y
Yu Yang 已提交
2624
    helper.append_op(
2625
        type="batch_norm", inputs=inputs, outputs=outputs, attrs=attrs)
Y
Yu Yang 已提交
2626 2627 2628 2629

    return helper.append_activation(batch_norm_out)


L
lvmengsi 已提交
2630 2631 2632 2633 2634 2635 2636 2637
def instance_norm(input,
                  epsilon=1e-05,
                  param_attr=None,
                  bias_attr=None,
                  name=None):
    """
    **Instance Normalization Layer**

L
lvmengsi 已提交
2638
    Can be used as a normalizer function for convolution or fully_connected operations.
L
lvmengsi 已提交
2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651
    The required data format for this layer is one of the following:

    DataLayout: NCHW `[batch, in_channels, in_height, in_width]`

    Refer to `Instance Normalization: The Missing Ingredient for 
    Fast Stylization <https://arxiv.org/pdf/1607.08022.pdf>`_
    for more details.

    :math:`input` is the input features over a mini-batch.

    ..  math::

        \\mu_{\\beta} &\\gets \\frac{1}{HW} \\sum_{i=1}^{HW} x_i \\qquad &//\\
L
lvmengsi 已提交
2652
        \\ mean\ of\ one\  feature\ map\ in\ mini-batch \\\\
L
lvmengsi 已提交
2653
        \\sigma_{\\beta}^{2} &\\gets \\frac{1}{HW} \\sum_{i=1}^{HW}(x_i - \\
L
lvmengsi 已提交
2654
        \\mu_{\\beta})^2 \\qquad &//\ variance\ of\ one\ feature\ map\ in\ mini-batch \\\\
L
lvmengsi 已提交
2655 2656 2657 2658
        \\hat{x_i} &\\gets \\frac{x_i - \\mu_\\beta} {\\sqrt{\\
        \\sigma_{\\beta}^{2} + \\epsilon}} \\qquad &//\ normalize \\\\
        y_i &\\gets \\gamma \\hat{x_i} + \\beta \\qquad &//\ scale\ and\ shift

L
lvmengsi 已提交
2659 2660
    Note:
        `H` means height of feature map, `W` means width of feature map.
L
lvmengsi 已提交
2661 2662

    Args:
L
lvmengsi 已提交
2663 2664
        input(variable): The rank of input variable can be 2, 3, 4, 5. 
            The data type is float32 or float64.
L
lvmengsi 已提交
2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680
        epsilon(float, Default 1e-05): A value added to the denominator for
            numerical stability. Default is 1e-5.
        param_attr(ParamAttr|None): The parameter attribute for Parameter `scale`
             of instance_norm. If it is set to None or one attribute of ParamAttr, instance_norm
	     will create ParamAttr as param_attr, the name of scale can be set in ParamAttr.
	     If the Initializer of the param_attr is not set, the parameter is initialized 
	     with Xavier. Default: None.
        bias_attr(ParamAttr|None): The parameter attribute for the bias of instance_norm.
             If it is set to None or one attribute of ParamAttr, instance_norm
	     will create ParamAttr as bias_attr, the name of bias can be set in ParamAttr. 
	     If the Initializer of the bias_attr is not set, the bias is initialized zero. 
	     Default: None.
        name(string, Default None): A name for this layer(optional). If set None, the layer
            will be named automatically.

    Returns:
L
lvmengsi 已提交
2681 2682
        A Variable holding Tensor which is the result after applying instance normalization on the input, 
        has same shape and data type with input. 
L
lvmengsi 已提交
2683 2684 2685 2686 2687 2688

    Examples:

        .. code-block:: python

            import paddle.fluid as fluid
L
lvmengsi 已提交
2689
            x = fluid.data(name='x', shape=[3, 7, 3, 7], dtype='float32')
L
lvmengsi 已提交
2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743
            hidden1 = fluid.layers.fc(input=x, size=200, param_attr='fc1.w')
            hidden2 = fluid.layers.instance_norm(input=hidden1)
    """
    assert bias_attr is not False, "bias_attr should not be False in instance_norm."
    helper = LayerHelper('instance_norm', **locals())
    dtype = helper.input_dtype()

    # use fp32 for in parameter
    if dtype == core.VarDesc.VarType.FP16:
        dtype = core.VarDesc.VarType.FP32

    input_shape = input.shape
    channel_num = input_shape[1]

    param_shape = [channel_num]

    # create parameter
    scale = helper.create_parameter(
        attr=helper.param_attr,
        shape=param_shape,
        dtype=dtype,
        default_initializer=Constant(1.0))
    bias = helper.create_parameter(
        attr=helper.bias_attr,
        shape=param_shape,
        dtype=dtype,
        is_bias=True,
        default_initializer=Constant(0.0))

    # create output
    saved_mean = helper.create_variable_for_type_inference(
        dtype=dtype, stop_gradient=True)
    saved_variance = helper.create_variable_for_type_inference(
        dtype=dtype, stop_gradient=True)

    instance_norm_out = helper.create_variable_for_type_inference(dtype)

    helper.append_op(
        type="instance_norm",
        inputs={
            "X": input,
            "Scale": scale,
            "Bias": bias,
        },
        outputs={
            "Y": instance_norm_out,
            "SavedMean": saved_mean,
            "SavedVariance": saved_variance
        },
        attrs={"epsilon": epsilon, })

    return instance_norm_out


H
heqiaozhi 已提交
2744 2745 2746 2747 2748 2749 2750 2751 2752
def data_norm(input,
              act=None,
              epsilon=1e-05,
              param_attr=None,
              data_layout='NCHW',
              in_place=False,
              name=None,
              moving_mean_name=None,
              moving_variance_name=None,
2753
              do_model_average_for_mean_and_var=True,
H
hutuxian 已提交
2754 2755 2756
              slot_dim=-1,
              sync_stats=False,
              summary_decay_rate=0.9999999):
H
heqiaozhi 已提交
2757 2758 2759
    """
    **Data Normalization Layer**

2760
    This op can be used as a normalizer function for conv2d and fully_connected operations.
H
heqiaozhi 已提交
2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783
    The required data format for this layer is one of the following:

    1. NHWC `[batch, in_height, in_width, in_channels]`

    2. NCHW `[batch, in_channels, in_height, in_width]`

    :math:`input` is the input features over a mini-batch.

    ..  math::

        \\mu_{\\beta} &\\gets \\frac{1}{m} \\sum_{i=1}^{m} x_i \\qquad &//\\
        \ mini-batch\ mean \\\\
        \\sigma_{\\beta}^{2} &\\gets \\frac{1}{m} \\sum_{i=1}^{m}(x_i - \\
        \\mu_{\\beta})^2 \\qquad &//\ mini-batch\ variance \\\\
        \\hat{x_i} &\\gets \\frac{x_i - \\mu_\\beta} {\\sqrt{\\
        \\sigma_{\\beta}^{2} + \\epsilon}} \\qquad &//\ normalize \\\\
        y_i &\\gets \\gamma \\hat{x_i} + \\beta \\qquad &//\ scale\ and\ shift

    Args:
        input(variable): The input variable which is a LoDTensor.
        act(string, Default None): Activation type, linear|relu|prelu|...
        epsilon(float, Default 1e-05):
        param_attr(ParamAttr): The parameter attribute for Parameter `scale`.
2784 2785 2786 2787
        data_layout (str, optional): Specify the data format of the input, and the data format of the output 
            will be consistent with that of the input. An optional string from: `"NCHW"`, `"NHWC"`.
            The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`.
H
heqiaozhi 已提交
2788 2789 2790 2791 2792
        in_place(bool, Default False): Make the input and output of batch norm reuse memory.
        name(string, Default None): A name for this layer(optional). If set None, the layer
            will be named automatically.
        moving_mean_name(string, Default None): The name of moving_mean which store the global Mean.
        moving_variance_name(string, Default None): The name of the moving_variance which store the global Variance.
2793 2794
        do_model_average_for_mean_and_var(bool, Default True): Whether parameter mean and variance
            should do model average when model average is enabled.
2795 2796 2797 2798 2799 2800 2801
        slot_dim(int): The embedding dimension of one slot. Slot is a set of one specific feature. In pslib mode, we 
            distinguish feature ids by slot and pull their embeddings from parameter server (pslib). The first
            place of the embedding is the historical show number (occurence time of this feature id with a label 0).
            If the input of this op is concated by slot-wise embeddings, and the show number is zero when this slot 
            is new or empty, the normalization result may be impractical. To avoid this, we add slot_dim to locate 
            the show number and judge if the show number is zero. If so, we choose to skip normalization on this
            embedding.
H
hutuxian 已提交
2802 2803 2804
        sync_stats(bool, Default False): When running with multiple GPU cards, using allreduce to sync the
            summary messages.
        summary_decay_rate(float, Default 0.9999999): The decay rate when updating summary.
H
heqiaozhi 已提交
2805 2806 2807 2808 2809 2810 2811

    Returns:
        Variable: A tensor variable which is the result after applying data normalization on the input.

    Examples:

        .. code-block:: python
2812 2813
            
            import paddle.fluid as fluid
H
heqiaozhi 已提交
2814

2815
            hidden1 = fluid.data(name="hidden1", shape=[64, 200])
2816
            hidden2 = fluid.layers.data_norm(name="hidden2", input=hidden1)
H
heqiaozhi 已提交
2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878
    """
    helper = LayerHelper('data_norm', **locals())
    dtype = helper.input_dtype()

    input_shape = input.shape
    if data_layout == 'NCHW':
        channel_num = input_shape[1]
    else:
        if data_layout == 'NHWC':
            channel_num = input_shape[-1]
        else:
            raise ValueError("unsupported data layout:" + data_layout)

    param_shape = [channel_num]

    batch_size_default = 1e4
    batch_sum_default = 0.0
    batch_square_sum_default = 1e4

    if param_attr and isinstance(param_attr, dict):
        batch_size_default = param_attr.get("batch_size", 1e4)
        batch_sum_default = param_attr.get("batch_sum", 0.0)
        batch_square_sum_default = param_attr.get("batch_square", 1e4)

    # create parameter
    batch_size = helper.create_parameter(
        attr=ParamAttr(
            name=name + '.batch_size',
            initializer=Constant(value=float(batch_size_default)),
            trainable=True),
        shape=param_shape,
        dtype=input.dtype)

    batch_sum = helper.create_parameter(
        attr=ParamAttr(
            name=name + '.batch_sum',
            initializer=Constant(value=float(batch_sum_default)),
            trainable=True),
        shape=param_shape,
        dtype=input.dtype)

    batch_square_sum = helper.create_parameter(
        attr=ParamAttr(
            name=name + '.batch_square_sum',
            initializer=Constant(value=float(batch_square_sum_default)),
            trainable=True),
        shape=param_shape,
        dtype=input.dtype)

    means = helper.create_variable(dtype=dtype, stop_gradient=True)
    scales = helper.create_variable(dtype=dtype, stop_gradient=True)

    data_norm_out = input if in_place else helper.create_variable(dtype=dtype)

    helper.append_op(
        type="data_norm",
        inputs={
            "X": input,
            "BatchSize": batch_size,
            "BatchSum": batch_sum,
            "BatchSquareSum": batch_square_sum
        },
H
hutuxian 已提交
2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892
        outputs={
            "Y": data_norm_out,
            "Means": means,
            "Scales": scales,
            "BatchSize": batch_size,
            "BatchSum": batch_sum,
            "BatchSquareSum": batch_square_sum
        },
        attrs={
            "epsilon": epsilon,
            "slot_dim": slot_dim,
            "sync_stats": sync_stats,
            "summary_decay_rate": summary_decay_rate
        })
H
heqiaozhi 已提交
2893 2894 2895 2896

    return helper.append_activation(data_norm_out)


Y
yuyang18 已提交
2897
@templatedoc()
G
guosheng 已提交
2898 2899 2900 2901 2902 2903 2904 2905 2906 2907
def layer_norm(input,
               scale=True,
               shift=True,
               begin_norm_axis=1,
               epsilon=1e-05,
               param_attr=None,
               bias_attr=None,
               act=None,
               name=None):
    """
2908 2909 2910 2911
    **Layer Normalization Layer**

    The API implements the function of the Layer Normalization Layer and can be applied to mini-batch input data.
    Refer to `Layer Normalization <https://arxiv.org/pdf/1607.06450v1.pdf>`_
G
guosheng 已提交
2912 2913 2914

    The formula is as follows:

Y
yuyang18 已提交
2915
    ..  math::
G
guosheng 已提交
2916

2917
        \\mu & = \\frac{1}{H}\\sum_{i=1}^{H} x_i
G
guosheng 已提交
2918

2919
        \\sigma & = \\sqrt{\\frac{1}{H}\sum_{i=1}^{H}{(x_i - \\mu)^2} + \\epsilon}
Y
yuyang18 已提交
2920

2921
        y & = f(\\frac{g}{\\sigma}(x - \\mu) + b)
Y
yuyang18 已提交
2922

2923 2924 2925 2926 2927
    - :math:`x`: the vector representation of the summed inputs to the neurons in that layer.
    - :math:`H`: the number of hidden units in a layers
    - :math:`\\epsilon`: the small value added to the variance to prevent division by zero.
    - :math:`g`: the trainable scale parameter.
    - :math:`b`: the trainable bias parameter.
Y
yuyang18 已提交
2928

G
guosheng 已提交
2929
    Args:
2930 2931 2932 2933 2934 2935
        input(Variable): A multi-dimension ``Tensor`` , and the data type is float32 or float64.
        scale(bool, optional): Whether to learn the adaptive gain :math:`g` after
            normalization. Default: True.
        shift(bool, optional): Whether to learn the adaptive bias :math:`b` after
            normalization. Default: True.
        begin_norm_axis(int, optional): The normalization will be performed along
G
guosheng 已提交
2936
            dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`.
2937 2938 2939 2940
            Default: 1.
        epsilon(float, optional): The small value added to the variance to prevent
            division by zero. Default: 1e-05.
        param_attr(ParamAttr, optional): The parameter attribute for the learnable
S
sneaxiy 已提交
2941 2942
            gain :math:`g`. If :attr:`scale` is False, :attr:`param_attr` is
            omitted. If :attr:`scale` is True and :attr:`param_attr` is None,
2943
            a default :code:`ParamAttr` would be added as scale. The
2944 2945
            :attr:`param_attr` is initialized as 1 if it is added. Default: None.
        bias_attr(ParamAttr, optional): The parameter attribute for the learnable
S
sneaxiy 已提交
2946 2947
            bias :math:`b`. If :attr:`shift` is False, :attr:`bias_attr` is
            omitted. If :attr:`shift` is True and :attr:`param_attr` is None,
2948
            a default :code:`ParamAttr` would be added as bias. The
2949 2950 2951 2952
            :attr:`bias_attr` is initialized as 0 if it is added. Default: None.
        act(str, optional): Activation to be applied to the output of layer normalizaiton.
                  Default: None.
        name(str): The default value is None.  Normally there is no need for user to set this property.  For more information, please refer to :ref:`api_guide_Name` .
G
guosheng 已提交
2953 2954

    Returns:
2955
        Variable: ``Tensor``  indicating the normalized result, the data type is the same as  ``input`` , and the return dimension is the same as  ``input`` .
G
guosheng 已提交
2956 2957 2958

    Examples:

2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970
        .. code-block:: python

            import paddle.fluid as fluid
            import numpy as np
            x = fluid.data(name='x', shape=[-1, 32, 32], dtype='float32')
            hidden1 = fluid.layers.layer_norm(input=x, begin_norm_axis=1)
            place = fluid.CPUPlace()
            exe = fluid.Executor(place)
            exe.run(fluid.default_startup_program())
            np_x = np.random.random(size=(8, 3, 32, 32)).astype('float32')
            output = exe.run(feed={"x": np_x}, fetch_list = [hidden1])
            print(output)
G
guosheng 已提交
2971
    """
L
lujun 已提交
2972
    assert in_dygraph_mode(
L
lujun 已提交
2973
    ) is not True, "please use FC instead of fc in dygraph mode!"
G
guosheng 已提交
2974 2975 2976 2977 2978 2979 2980 2981
    helper = LayerHelper('layer_norm', **locals())
    dtype = helper.input_dtype()

    # create intput and parameters
    inputs = {'X': input}
    input_shape = input.shape
    param_shape = [reduce(lambda x, y: x * y, input_shape[begin_norm_axis:])]
    if scale:
2982
        assert param_attr is not False, "param_attr should not be False when using scale."
G
guosheng 已提交
2983 2984 2985 2986 2987 2988
        scale = helper.create_parameter(
            attr=helper.param_attr,
            shape=param_shape,
            dtype=dtype,
            default_initializer=Constant(1.0))
        inputs['Scale'] = scale
2989 2990 2991
    else:
        if param_attr:
            warnings.warn("param_attr is only avaliable with scale is True.")
G
guosheng 已提交
2992
    if shift:
2993
        assert bias_attr is not False, "bias_attr should not be False when using shift."
G
guosheng 已提交
2994 2995 2996
        bias = helper.create_parameter(
            attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
        inputs['Bias'] = bias
2997 2998 2999
    else:
        if bias_attr:
            warnings.warn("bias_attr is only avaliable with shift is True.")
G
guosheng 已提交
3000 3001

    # create output
X
Xin Pan 已提交
3002 3003 3004 3005 3006
    mean_out = helper.create_variable_for_type_inference(
        dtype=dtype, stop_gradient=True)
    variance_out = helper.create_variable_for_type_inference(
        dtype=dtype, stop_gradient=True)
    layer_norm_out = helper.create_variable_for_type_inference(dtype)
G
guosheng 已提交
3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021

    helper.append_op(
        type="layer_norm",
        inputs=inputs,
        outputs={
            "Y": layer_norm_out,
            "Mean": mean_out,
            "Variance": variance_out,
        },
        attrs={"epsilon": epsilon,
               "begin_norm_axis": begin_norm_axis})

    return helper.append_activation(layer_norm_out)


D
Dun 已提交
3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033
@templatedoc()
def group_norm(input,
               groups,
               epsilon=1e-05,
               param_attr=None,
               bias_attr=None,
               act=None,
               data_layout='NCHW',
               name=None):
    """
    **Group Normalization Layer**

H
haowang101779990 已提交
3034
    Refer to `Group Normalization <https://arxiv.org/abs/1803.08494>`_ .
D
Dun 已提交
3035

3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050
    Parameters:
        input(Variable): 4-D Tensor, the data type is float32 or float64.
        groups(int): The number of groups that divided from channels, the data type
            is int32.
        epsilon(float, optional): The small value added to the variance to prevent
            division by zero, the data type is float32. Default: 1e-05.
        param_attr(ParamAttr|bool, optional): ParamAttr object that specifies weight parameter
            attribute. If a bool type, only False is supported, which means there is no weight parameter.
            Default: None, the default weight parameter attribute is used. For more information, please
            refer to :ref:`api_guide_ParamAttr` .
        bias_attr(ParamAttr|bool, optional): ParamAttr object that specifies bias parameter
            attribute. If a bool type, only False is supported, which means there is no bias parameter.
            Default: None, the default bias parameter attribute is used. For more information, please
            refer to :ref:`api_guide_ParamAttr` .
        act(str, optional): Activation to be applied to the output of group normalizaiton.
3051 3052 3053 3054
        data_layout(str, optional): Specify the data format of the input, and the data format of the output 
            will be consistent with that of the input. An optional string from: `"NCHW"`, `"NHWC"`.
            The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`.
3055 3056
        name (str, optional): The default value is None. Normally there is no need for user to set this
            property. For more information, please refer to :ref:`api_guide_Name` .
D
Dun 已提交
3057 3058

    Returns:
3059 3060 3061 3062
        Variable: A 4-D Tensor has same data type and data format with `input`.

    Raises:
        ValueError: If `data_layout` is neither 'NCHW' nor 'NHWC'.
3063 3064 3065 3066 3067 3068
        ValueError: If `groups` is greater than the number of input channels.
        ValueError: If `groups` is less than 1.
        ShapeError: If the param_attr(Scale) is not 1-D Tensor.
        ShapeError: If the param_attr(Scale)'s first dimension size is not equal to the input channels.
        ShapeError: If the bias_attr(Bias) is not 1-D Tensor.
        ShapeError: If the bias_attr(Bias)'s first dimension size is not equal to the input channels.
D
Dun 已提交
3069 3070

    Examples:
3071
       .. code-block:: python
D
Dun 已提交
3072

3073 3074 3075
            import paddle.fluid as fluid
            data = fluid.data(name='data', shape=[None, 8, 32, 32], dtype='float32')
            x = fluid.layers.group_norm(input=data, groups=4)
D
Dun 已提交
3076 3077 3078 3079 3080 3081 3082
    """
    helper = LayerHelper('group_norm', **locals())
    dtype = helper.input_dtype()

    # create intput and parameters
    inputs = {'X': input}
    input_shape = input.shape
3083 3084 3085 3086 3087 3088
    if data_layout != 'NCHW' and data_layout != 'NHWC':
        raise ValueError(
            "Param(data_layout) of Op(fluid.layers.group_norm) got wrong value: received "
            + data_layout + " but only NCHW or NHWC supported.")
    channel_num = input_shape[1] if data_layout == 'NCHW' else input_shape[-1]
    param_shape = [channel_num]
D
Dun 已提交
3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101
    if param_attr:
        scale = helper.create_parameter(
            attr=helper.param_attr,
            shape=param_shape,
            dtype=dtype,
            default_initializer=Constant(1.0))
        inputs['Scale'] = scale
    if bias_attr:
        bias = helper.create_parameter(
            attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
        inputs['Bias'] = bias

    # create output
H
heqiaozhi 已提交
3102 3103
    mean_out = helper.create_variable(dtype=dtype, stop_gradient=True)
    variance_out = helper.create_variable(dtype=dtype, stop_gradient=True)
D
dengkaipeng 已提交
3104 3105 3106 3107 3108 3109 3110 3111 3112 3113
    group_norm_out = helper.create_variable(dtype=dtype)

    helper.append_op(
        type="group_norm",
        inputs=inputs,
        outputs={
            "Y": group_norm_out,
            "Mean": mean_out,
            "Variance": variance_out,
        },
3114 3115 3116 3117 3118
        attrs={
            "epsilon": epsilon,
            "groups": groups,
            "data_layout": data_layout
        })
D
dengkaipeng 已提交
3119 3120 3121 3122 3123

    return helper.append_activation(group_norm_out)


@templatedoc()
3124
def spectral_norm(weight, dim=0, power_iters=1, eps=1e-12, name=None):
D
dengkaipeng 已提交
3125 3126 3127
    """
    **Spectral Normalization Layer**

K
Kaipeng Deng 已提交
3128
    This operation calculates the spectral normalization value of weight parameters of
3129
    fc, conv1d, conv2d, conv3d layers which should be 2-D, 3-D, 4-D, 5-D
K
Kaipeng Deng 已提交
3130 3131
    Parameters. Output tensor will be in same shape with input tensor.
    Calculations are showed as follows.
3132

D
dengkaipeng 已提交
3133 3134 3135
    Step 1:
    Generate vector U in shape of [H], and V in shape of [W].
    While H is the :attr:`dim` th dimension of the input weights,
D
dengkaipeng 已提交
3136
    and W is the product result of remaining dimensions.
D
dengkaipeng 已提交
3137 3138 3139

    Step 2:
    :attr:`power_iters` shoule be a positive interger, do following
K
Kaipeng Deng 已提交
3140 3141
    calculations with U and V for :attr:`power_iters` rounds. Calculations
    as follows:
D
dengkaipeng 已提交
3142 3143 3144 3145 3146 3147 3148 3149

    .. math:: 

        \mathbf{v} := \\frac{\mathbf{W}^{T} \mathbf{u}}{\|\mathbf{W}^{T} \mathbf{u}\|_2}

        \mathbf{u} := \\frac{\mathbf{W}^{T} \mathbf{v}}{\|\mathbf{W}^{T} \mathbf{v}\|_2}

    Step 3:
D
dengkaipeng 已提交
3150
    Calculate :math:`\sigma(\mathbf{W})` and normalize weight values.
D
dengkaipeng 已提交
3151 3152 3153 3154

    .. math::

        \sigma(\mathbf{W}) = \mathbf{u}^{T} \mathbf{W} \mathbf{v}
3155

D
dengkaipeng 已提交
3156
        \mathbf{W} = \\frac{\mathbf{W}}{\sigma(\mathbf{W})}
3157 3158
                

D
dengkaipeng 已提交
3159 3160 3161 3162
    Refer to `Spectral Normalization <https://arxiv.org/abs/1802.05957>`_ .

    Args:
        weight(${weight_type}): ${weight_comment}
D
dengkaipeng 已提交
3163 3164 3165
        dim(int): ${dim_comment}
        power_iters(int): ${power_iters_comment}
        eps(float): ${eps_comment}
K
Kaipeng Deng 已提交
3166 3167 3168
        name(str, optional): For detailed information, please refer
                             to :ref:`api_guide_Name`. Usually name is no need to set and
                             None by default.
D
dengkaipeng 已提交
3169 3170

    Returns:
D
dengkaipeng 已提交
3171
        Variable: A tensor variable of weight parameters after spectral normalization.
K
Kaipeng Deng 已提交
3172
                  The data type and shape is same as input tensor.
D
dengkaipeng 已提交
3173 3174

    Examples:
K
Kaipeng Deng 已提交
3175
       .. code-block:: python
D
dengkaipeng 已提交
3176

K
Kaipeng Deng 已提交
3177 3178
            import paddle.fluid as fluid

K
Kaipeng Deng 已提交
3179
            weight = fluid.data(name='weight', shape=[2, 8, 32, 32], dtype='float32')
K
Kaipeng Deng 已提交
3180
            x = fluid.layers.spectral_norm(weight=weight, dim=1, power_iters=2)
D
dengkaipeng 已提交
3181 3182
    """
    helper = LayerHelper('spectral_norm', **locals())
3183
    dtype = weight.dtype
D
dengkaipeng 已提交
3184 3185 3186

    # create intput and parameters
    inputs = {'Weight': weight}
3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204
    input_shape = weight.shape
    h = input_shape[dim]
    w = np.prod(input_shape) // h

    u = helper.create_parameter(
        attr=ParamAttr(),
        shape=[h],
        dtype=dtype,
        default_initializer=Normal(0., 1.))
    u.stop_gradient = True
    inputs['U'] = u
    v = helper.create_parameter(
        attr=ParamAttr(),
        shape=[w],
        dtype=dtype,
        default_initializer=Normal(0., 1.))
    inputs['V'] = v
    v.stop_gradient = True
D
dengkaipeng 已提交
3205 3206

    # create output
3207
    out = helper.create_variable(dtype=dtype)
D
Dun 已提交
3208 3209

    helper.append_op(
3210
        type="spectral_norm",
D
Dun 已提交
3211
        inputs=inputs,
3212 3213 3214 3215 3216 3217
        outputs={"Out": out, },
        attrs={
            "dim": dim,
            "power_iters": power_iters,
            "eps": eps,
        })
D
Dun 已提交
3218

3219
    return out
D
Dun 已提交
3220 3221


Y
Yu Yang 已提交
3222 3223 3224 3225
def conv2d_transpose(input,
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
3226 3227 3228
                     padding=0,
                     stride=1,
                     dilation=1,
3229
                     groups=None,
C
caoying03 已提交
3230
                     param_attr=None,
3231
                     bias_attr=None,
C
chengduoZH 已提交
3232
                     use_cudnn=True,
3233
                     act=None,
3234 3235
                     name=None,
                     data_format='NCHW'):
Y
Yu Yang 已提交
3236
    """
3237 3238
    The convolution2D transpose layer calculates the output based on the input,
    filter, and dilations, strides, paddings. Input(Input) and output(Output)
3239
    are in NCHW or NHWC format. Where N is batch size, C is the number of channels,
3240 3241 3242
    H is the height of the feature, and W is the width of the feature.
    Parameters(dilations, strides, paddings) are two elements. These two elements
    represent height and width, respectively. The details of convolution transpose
3243
    layer, please refer to the following explanation and references
L
lvmengsi 已提交
3244
    `therein <https://arxiv.org/pdf/1603.07285.pdf>`_.
3245 3246 3247
    If bias attribution and activation type are provided, bias is added to
    the output of the convolution, and the corresponding activation function
    is applied to the final result.
3248 3249 3250 3251 3252

    For each input :math:`X`, the equation is:

    .. math::

3253
        Out = \sigma (W \\ast X + b)
3254

3255
    Where:
3256

3257 3258
    * :math:`X`: Input value, a 4-D Tensor with NCHW or NHWC format.
    * :math:`W`: Filter value, a 4-D Tensor with MCHW format.
3259
    * :math:`\\ast`: Convolution operation.
3260
    * :math:`b`: Bias value, a 2-D Tensor with shape [M, 1].
3261
    * :math:`\\sigma`: Activation function.
3262
    * :math:`Out`: Output value, a 4-D Tensor with data format 'NCHW' or 'NHWC', the shape of :math:`Out` and :math:`X` may be different.
Y
Yu Yang 已提交
3263

3264 3265 3266 3267
    Example:

        - Input:

3268
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
3269

3270
          Filter shape: :math:`(C_{in}, C_{out}, H_f, W_f)`
3271 3272 3273

        - Output:

3274
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
3275 3276

        Where
Y
Yu Yang 已提交
3277

3278 3279
        .. math::

3280 3281
           H^\prime_{out} &= (H_{in} - 1) * strides[0] - pad_height_top - pad_height_bottom + dilations[0] * (H_f - 1) + 1 \\\\
           W^\prime_{out} &= (W_{in} - 1) * strides[1] - pad_width_left - pad_width_right + dilations[1] * (W_f - 1) + 1 \\\\
L
lvmengsi 已提交
3282
           H_{out} &\in [ H^\prime_{out}, H^\prime_{out} + strides[0] ] \\\\
3283 3284
           W_{out} &\in [ W^\prime_{out}, W^\prime_{out} + strides[1] ]

L
lvmengsi 已提交
3285
    Note:
L
lvmengsi 已提交
3286 3287 3288 3289
          The conv2d_transpose can be seen as the backward of the conv2d. For conv2d, 
          when stride > 1, conv2d maps multiple input shape to the same output shape, 
          so for conv2d_transpose, when stride > 1, input shape maps multiple output shape.
          If output_size is None, :math:`H_{out} = H^\prime_{out}, W_{out} = W^\prime_{out}`; 
L
lvmengsi 已提交
3290 3291 3292 3293
          else, the :math:`H_{out}` of the output size must between :math:`H^\prime_{out}` 
          and :math:`H^\prime_{out} + strides[0]`, and the :math:`W_{out}` of the output size must 
          between :math:`W^\prime_{out}` and :math:`W^\prime_{out} + strides[1]`, 
          conv2d_transpose can compute the kernel size automatically.
Y
Yu Yang 已提交
3294 3295

    Args:
3296 3297
        input(Variable): 4-D Tensor with [N, C, H, W] or [N, H, W, C] format,
                         its data type is float32 or float64.
3298 3299
        num_filters(int): The number of the filter. It is as same as the output
            image channel.
3300
        output_size(int|tuple, optional): The output image size. If output size is a
3301
            tuple, it must contain two integers, (image_height, image_width). None if use
3302
            filter_size, padding, and stride to calculate output_size.
L
lvmengsi 已提交
3303 3304 3305
            If output_size and filter_size are specified at the same time, They
            should follow the formula above. Default: None. output_size and filter_size 
            should not be None at the same time.
3306
        filter_size(int|tuple, optional): The filter size. If filter_size is a tuple,
3307 3308
            it must contain two integers, (filter_size_height, filter_size_width).
            Otherwise, filter_size_height = filter_size_width = filter_size. None if 
L
lvmengsi 已提交
3309 3310 3311 3312 3313 3314 3315
            use output size to calculate filter_size. Default: None. filter_size and 
            output_size should not be None at the same time.
        stride(int|tuple, optional): The stride size. It means the stride in transposed convolution. 
            If stride is a tuple, it must contain two integers, (stride_height, stride_width). 
            Otherwise, stride_height = stride_width = stride. Default: stride = 1.
        padding(int|list|str|tuple, optional): The padding size. The padding argument effectively adds
             `dilation * (kernel - 1)` amount of zero-padding on both sides of input. If `padding` is a
3316 3317 3318 3319 3320 3321 3322 3323 3324
             string, either 'VALID' or 'SAME' supported, which is the padding algorithm.
             If `padding` is a tuple or list, it could be in three forms:
             `[pad_height, pad_width]` or
            `[pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`, and
            when `data_format` is `'NCHW'`,
            `padding` can be in the form `[[0,0], [0,0], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`.
            when `data_format` is `'NHWC'`, `padding` can be in the form
            `[[0,0], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`.
            Default: padding = 0.
L
lvmengsi 已提交
3325 3326 3327 3328 3329 3330 3331
        dilation(int|tuple, optional): The dilation size. It means the spacing between the kernel points. 
            If dilation is a tuple, it must contain two integers, (dilation_height, dilation_width). 
            Otherwise, dilation_height = dilation_width = dilation. Default: dilation = 1.
        filter_size(int|tuple, optional): The filter size. If filter_size is a tuple,
            it must contain two integers, (filter_size_height, filter_size_width).
            Otherwise, filter_size_height = filter_size_width = filter_size. None if 
            use output size to calculate filter_size. Default: None.
3332
        groups(int, optional): The groups number of the Conv2d transpose layer. Inspired by
3333 3334 3335 3336
            grouped convolution in Alex Krizhevsky's Deep CNN paper, in which
            when group=2, the first half of the filters is only connected to the
            first half of the input channels, while the second half of the
            filters is only connected to the second half of the input channels.
C
chengduo 已提交
3337
            Default: groups = 1.
3338
        param_attr (ParamAttr, optional): The parameter attribute for learnable parameters/weights
C
chengduo 已提交
3339 3340 3341
            of conv2d_transpose. If it is set to None or one attribute of ParamAttr, conv2d_transpose
            will create ParamAttr as param_attr. If the Initializer of the param_attr
            is not set, the parameter is initialized with Xavier. Default: None.
3342
        bias_attr (ParamAttr|bool, optional): The parameter attribute for the bias of conv2d_transpose.
C
chengduo 已提交
3343 3344 3345 3346
            If it is set to False, no bias will be added to the output units.
            If it is set to None or one attribute of ParamAttr, conv2d_transpose
            will create ParamAttr as bias_attr. If the Initializer of the bias_attr
            is not set, the bias is initialized zero. Default: None.
3347
        use_cudnn(bool, optional): Use cudnn kernel or not, it is valid only when the cudnn
C
chengduo 已提交
3348
            library is installed. Default: True.
3349
        act (str, optional): Activation type, if it is set to None, activation is not appended.
C
chengduo 已提交
3350
            Default: None.
L
lvmengsi 已提交
3351 3352 3353
        name(str, optional): For detailed information, please refer 
           to :ref:`api_guide_Name`. Usually name is no need to set and 
           None by default.
3354 3355 3356 3357
        data_format (str, optional): Specify the data format of the input, and the data format of the output 
            will be consistent with that of the input. An optional string from: `"NCHW"`, `"NHWC"`.
            The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`.
Y
Yu Yang 已提交
3358 3359

    Returns:
L
lvmengsi 已提交
3360 3361 3362 3363 3364 3365
        A Variable holding Tensor representing the conv2d_transpose, whose 
        data type is the same with input and shape is (num_batches, channels, out_h, 
        out_w) or (num_batches, out_h, out_w, channels). If act is None, the tensor variable 
        storing the transposed convolution result, and if act is not None, the 
        tensor variable storing transposed convolution and non-linearity activation 
        result.
3366 3367

    Raises:
3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378
        ValueError: If the type of `use_cudnn` is not bool.
        ValueError: If `data_format` is not "NCHW" or "NHWC".
        ValueError: If `padding` is a string, but not "SAME" or "VALID".
        ValueError: If `padding` is a tuple, but the element corresponding to the input's batch size is not 0 
            or the element corresponding to the input's channel is not 0.
        ValueError: If `output_size` and filter_size are None at the same time.
        ShapeError: If the input is not 4-D Tensor.
        ShapeError: If the input's dimension size and filter's dimension size not equal.
        ShapeError: If the dimension size of input minus the size of `stride` is not 2.
        ShapeError: If the number of input channels is not equal to filter's channels.
        ShapeError: If the size of `output_size` is not equal to that of `stride`.
3379 3380 3381 3382

    Examples:
       .. code-block:: python

3383
          import paddle.fluid as fluid
L
lvmengsi 已提交
3384
          data = fluid.data(name='data', shape=[None, 3, 32, 32], dtype='float32')
3385
          conv2d_transpose = fluid.layers.conv2d_transpose(input=data, num_filters=2, filter_size=3)
Y
Yu Yang 已提交
3386
    """
C
chengduo 已提交
3387
    assert param_attr is not False, "param_attr should not be False in conv2d_transpose."
3388 3389 3390 3391
    if data_format not in ['NCHW', 'NHWC']:
        raise ValueError(
            "Attr(data_format) of Op(fluid.layers.conv2d_transpose) got wrong value: received "
            + data_format + " but only NCHW or NHWC supported.")
3392

3393
    input_channel = input.shape[1] if data_format == 'NCHW' else input.shape[-1]
3394 3395 3396 3397 3398 3399
    op_type = 'conv2d_transpose'
    if (input_channel == groups and num_filters == input_channel and
            not use_cudnn):
        op_type = 'depthwise_conv2d_transpose'

    helper = LayerHelper(op_type, **locals())
Y
Yu Yang 已提交
3400 3401 3402
    if not isinstance(input, Variable):
        raise TypeError("Input of conv2d_transpose must be Variable")

C
chengduoZH 已提交
3403 3404
    stride = utils.convert_to_list(stride, 2, 'stride')
    dilation = utils.convert_to_list(dilation, 2, 'dilation')
G
guosheng 已提交
3405

C
chengduoZH 已提交
3406 3407
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
G
guosheng 已提交
3408

3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451
    def _update_padding(padding, data_format):
        def is_list_or_tuple(ele):
            if isinstance(ele, list) or isinstance(ele, tuple):
                return True
            return False

        if is_list_or_tuple(padding) and len(padding) == 4:
            if is_list_or_tuple(padding[0]) and (data_format == "NCHW"):
                if not (padding[0] == [0, 0] and padding[1] == [0, 0]):
                    raise ValueError(
                        "Non-zero padding(%s) in the batch or channel dimensions "
                        "is not supported." % str(padding))
                padding = padding[2:4]
                padding = [ele for a_list in padding for ele in a_list]
            elif is_list_or_tuple(padding[0]) and (data_format == "NHWC"):
                if not (padding[0] == [0, 0] and padding[3] == [0, 0]):
                    raise ValueError(
                        "Non-zero padding(%s) in the batch or channel dimensions "
                        "is not supported." % str(padding))
                padding = padding[1:3]
                padding = [ele for a_list in padding for ele in a_list]
            padding = utils.convert_to_list(padding, 4, 'padding')
        else:
            padding = utils.convert_to_list(padding, 2, 'padding')
            padding = [padding[0], padding[0], padding[1], padding[1]]
        return padding

    padding_algorithm = "EXPLICIT"
    if isinstance(padding, str):
        padding = padding.upper()
        if padding not in ["SAME", "VALID"]:
            raise ValueError(
                "Unknown padding: '%s'. It can only be 'SAME' or 'VALID'." %
                str(padding))
        if padding == "VALID":
            padding_algorithm = "VALID"
            padding = [0, 0, 0, 0]
        elif padding == "SAME":
            padding_algorithm = "SAME"
            padding = [0, 0, 0, 0]

    padding = _update_padding(padding, data_format)

Y
Yu Yang 已提交
3452 3453 3454 3455 3456
    if filter_size is None:
        if output_size is None:
            raise ValueError("output_size must be set when filter_size is None")
        if isinstance(output_size, int):
            output_size = [output_size, output_size]
G
guosheng 已提交
3457

3458 3459
        h_in = input.shape[2] if data_format == 'NCHW' else input.shape[1]
        w_in = input.shape[3] if data_format == 'NCHW' else input.shape[2]
G
guosheng 已提交
3460

3461 3462 3463 3464
        filter_size_h = (output_size[0] - (h_in - 1) * stride[0] + padding[0] +
                         padding[1] - 1) // dilation[0] + 1
        filter_size_w = (output_size[1] - (w_in - 1) * stride[1] + padding[2] +
                         padding[3] - 1) // dilation[1] + 1
Y
Yu Yang 已提交
3465
        filter_size = [filter_size_h, filter_size_w]
C
chengduoZH 已提交
3466 3467 3468
    else:
        filter_size = utils.convert_to_list(filter_size, 2,
                                            'conv2d_transpose.filter_size')
C
chengduo 已提交
3469

3470 3471 3472
    if len(padding) == 4 and utils._is_symmetric_padding(padding, 2):
        padding = [padding[0], padding[2]]

3473 3474 3475 3476 3477 3478
    if output_size is None:
        output_size = []
    elif isinstance(output_size, list) or isinstance(output_size, int):
        output_size = utils.convert_to_list(output_size, 2, 'output_size')
    else:
        raise ValueError("output_size should be list or int")
3479
    groups = 1 if groups is None else groups
M
minqiyang 已提交
3480
    filter_shape = [input_channel, num_filters // groups] + filter_size
C
chengduo 已提交
3481

Y
Yu Yang 已提交
3482 3483 3484
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

X
Xin Pan 已提交
3485
    pre_bias = helper.create_variable_for_type_inference(dtype=input.dtype)
Y
Yu Yang 已提交
3486
    helper.append_op(
3487
        type=op_type,
Y
Yu Yang 已提交
3488 3489
        inputs={'Input': [input],
                'Filter': [img_filter]},
3490
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
3491
        attrs={
3492
            'output_size': output_size,
3493 3494
            'strides': stride,
            'paddings': padding,
3495
            'padding_algorithm': padding_algorithm,
3496 3497
            'dilations': dilation,
            'groups': groups,
3498 3499
            'use_cudnn': use_cudnn,
            'data_format': data_format
Y
Yu Yang 已提交
3500 3501
        })

3502 3503 3504 3505
    if data_format == 'NCHW':
        pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
    else:
        pre_act = helper.append_bias_op(pre_bias, dim_start=3, dim_end=4)
3506 3507
    out = helper.append_activation(pre_act)
    return out
Y
Yu Yang 已提交
3508 3509


3510
def conv3d_transpose(input,
Y
Yu Yang 已提交
3511 3512 3513
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
3514 3515 3516
                     padding=0,
                     stride=1,
                     dilation=1,
3517
                     groups=None,
C
caoying03 已提交
3518
                     param_attr=None,
3519
                     bias_attr=None,
C
chengduoZH 已提交
3520
                     use_cudnn=True,
3521
                     act=None,
3522 3523
                     name=None,
                     data_format='NCDHW'):
Y
Yu Yang 已提交
3524
    """
3525
    The convolution3D transpose layer calculates the output based on the input,
3526
    filter, and dilations, strides, paddings. Input(Input) and output(Output)
3527
    are in NCDHW or NDHWC format. Where N is batch size, C is the number of channels,
3528 3529 3530 3531
    D is the depth of the feature, H is the height of the feature, and W
    is the width of the feature. Parameters(dilations, strides, paddings) are
    two elements. These two elements represent height and width, respectively.
    The details of convolution transpose layer, please refer to the following
L
lvmengsi 已提交
3532
    explanation and references `therein <https://arxiv.org/pdf/1603.07285.pdf>`_.
3533 3534 3535
    If bias attribution and activation type are provided, bias is added to
    the output of the convolution, and the corresponding activation function
    is applied to the final result.
3536 3537 3538 3539 3540

    For each input :math:`X`, the equation is:

    .. math::

3541
        Out = \sigma (W \\ast X + b)
3542 3543 3544

    In the above equation:

3545 3546
    * :math:`X`: Input value, a Tensor with NCDHW or NDHWC format.
    * :math:`W`: Filter value, a Tensor with MCDHW format.
3547
    * :math:`\\ast`: Convolution operation.
3548
    * :math:`b`: Bias value, a 2-D Tensor with shape [M, 1].
3549 3550
    * :math:`\\sigma`: Activation function.
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
Y
Yu Yang 已提交
3551

3552 3553 3554 3555
    Example:

        - Input:

3556
          Input shape: :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})`
3557

3558
          Filter shape: :math:`(C_{in}, C_{out}, D_f, H_f, W_f)`
3559 3560 3561

        - Output:

3562
          Output shape: :math:`(N, C_{out}, D_{out}, H_{out}, W_{out})`
3563 3564

        Where
Y
Yu Yang 已提交
3565

3566 3567
        .. math::

L
lvmengsi 已提交
3568 3569 3570 3571 3572 3573
           D^\prime_{out} &= (D_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (D_f - 1) + 1 \\\\
           H^\prime_{out} &= (H_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (H_f - 1) + 1 \\\\
           W^\prime_{out} &= (W_{in} - 1) * strides[2] - 2 * paddings[2] + dilations[2] * (W_f - 1) + 1 \\\\
           D_{out} &\in [ D^\prime_{out}, D^\prime_{out} + strides[0] ] \\\\
           H_{out} &\in [ H^\prime_{out}, H^\prime_{out} + strides[1] ] \\\\
           W_{out} &\in [ W^\prime_{out}, W^\prime_{out} + strides[2] ]
Y
Yu Yang 已提交
3574

L
lvmengsi 已提交
3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589
    Note:
          The conv3d_transpose can be seen as the backward of the conv3d. For conv3d, 
          when stride > 1, conv3d maps multiple input shape to the same output shape, 
          so for conv3d_transpose, when stride > 1, input shape maps multiple output shape.
          If output_size is None, :math:`H_{out} = H^\prime_{out}, :math:`H_{out} = \
          H^\prime_{out}, W_{out} = W^\prime_{out}`; else, the :math:`D_{out}` of the output 
          size must between :math:`D^\prime_{out}` and :math:`D^\prime_{out} + strides[0]`, 
          the :math:`H_{out}` of the output size must between :math:`H^\prime_{out}` 
          and :math:`H^\prime_{out} + strides[1]`, and the :math:`W_{out}` of the output size must 
          between :math:`W^\prime_{out}` and :math:`W^\prime_{out} + strides[2]`, 
          conv3d_transpose can compute the kernel size automatically.

    Args:
        input(Variable): The input is 5-D Tensor with shape [N, C, D, H, W] or [N, D, H, W, C], the data type 
            of input is float32 or float64.
3590 3591
        num_filters(int): The number of the filter. It is as same as the output
            image channel.
3592
        output_size(int|tuple, optional): The output image size. If output size is a
L
lvmengsi 已提交
3593 3594 3595 3596
            tuple, it must contain three integers, (image_depth, image_height, image_width). This
            parameter only works when filter_size is None. If output_size and filter_size are 
            specified at the same time, They should follow the formula above. Default: None. 
            Output_size and filter_size should not be None at the same time.
3597
        filter_size(int|tuple, optional): The filter size. If filter_size is a tuple,
L
lvmengsi 已提交
3598
            it must contain three integers, (filter_size_depth, filter_size_height,
3599 3600
            filter_size_width). Otherwise, filter_size_depth = filter_size_height = \
            filter_size_width = filter_size. None if use output size to
L
lvmengsi 已提交
3601 3602 3603 3604
            calculate filter_size. Default: None. filter_size and output_size should not be 
            None at the same time.
        padding(int|list|str|tuple, optional): The padding size. The padding argument effectively
             adds `dilation * (kernel - 1)` amount of zero-padding on both sides of input. If `padding` is a string,
3605 3606 3607 3608 3609 3610 3611 3612
             either 'VALID' or 'SAME' supported, which is the padding algorithm. If `padding`
             is a tuple or list, it could be in three forms: `[pad_depth, pad_height, pad_width]` or
            `[pad_depth_front, pad_depth_back, pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`,
            and when `data_format` is `'NCDHW'`, `padding` can be in the form
            `[[0,0], [0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`.
            when `data_format` is `'NDHWC'`, `padding` can be in the form
            `[[0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`.
            Default: padding = 0.
L
lvmengsi 已提交
3613 3614 3615 3616 3617 3618 3619 3620
        stride(int|tuple, optional): The stride size. It means the stride in transposed convolution. 
            If stride is a tuple, it must contain three integers, (stride_depth, stride_height, 
            stride_width). Otherwise, stride_depth = stride_height = stride_width = stride. 
            Default: stride = 1.
        dilation(int|tuple, optional): The dilation size. It means the spacing between the kernel points. 
            If dilation is a tuple, it must contain three integers, (dilation_depth, dilation_height, 
            dilation_width). Otherwise, dilation_depth = dilation_height = dilation_width = dilation. 
            Default: dilation = 1.
3621
        groups(int, optional): The groups number of the Conv3d transpose layer. Inspired by
3622 3623 3624 3625 3626
            grouped convolution in Alex Krizhevsky's Deep CNN paper, in which
            when group=2, the first half of the filters is only connected to the
            first half of the input channels, while the second half of the
            filters is only connected to the second half of the input channels.
            Default: groups=1
3627
        param_attr (ParamAttr, optional): The parameter attribute for learnable parameters/weights
C
chengduo 已提交
3628 3629 3630
            of conv3d_transpose. If it is set to None or one attribute of ParamAttr, conv3d_transpose
            will create ParamAttr as param_attr. If the Initializer of the param_attr
            is not set, the parameter is initialized with Xavier. Default: None.
3631
        bias_attr (ParamAttr|bool, optional): The parameter attribute for the bias of conv3d_transpose.
C
chengduo 已提交
3632 3633 3634 3635
            If it is set to False, no bias will be added to the output units.
            If it is set to None or one attribute of ParamAttr, conv3d_transpose
            will create ParamAttr as bias_attr. If the Initializer of the bias_attr
            is not set, the bias is initialized zero. Default: None.
3636
        use_cudnn(bool, optional): Use cudnn kernel or not, it is valid only when the cudnn
3637
            library is installed. Default: True
3638
        act (str, optional): Activation type, if it is set to None, activation is not appended.
C
chengduo 已提交
3639
            Default: None.
L
lvmengsi 已提交
3640 3641 3642
        name(str, optional): For detailed information, please refer 
           to :ref:`api_guide_Name`. Usually name is no need to set and 
           None by default.
3643 3644 3645 3646
        data_format (str, optional): Specify the data format of the input, and the data format of the output 
            will be consistent with that of the input. An optional string from: `"NCHW"`, `"NHWC"`.
            The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`.
Y
Yu Yang 已提交
3647 3648

    Returns:
L
lvmengsi 已提交
3649 3650 3651 3652 3653
        A Variable holding Tensor representing the conv3d_transpose, whose data 
        type is the same with input and shape is (num_batches, channels, out_d, out_h, 
        out_w) or (num_batches, out_d, out_h, out_w, channels). If act is None, the tensor 
        variable storing the transposed convolution result, and if act is not None, the tensor 
        variable storing transposed convolution and non-linearity activation result.
3654 3655

    Raises:
3656 3657 3658 3659 3660 3661 3662 3663 3664 3665 3666
        ValueError: If the type of `use_cudnn` is not bool.
        ValueError: If `data_format` is not "NCDHW" or "NDHWC".
        ValueError: If `padding` is a string, but not "SAME" or "VALID".
        ValueError: If `padding` is a tuple, but the element corresponding to the input's batch size is not 0 
            or the element corresponding to the input's channel is not 0.
        ValueError: If `output_size` and filter_size are None at the same time.
        ShapeError: If the input is not 5-D Tensor.
        ShapeError: If the input's dimension size and filter's dimension size not equal.
        ShapeError: If the dimension size of input minus the size of `stride` is not 2.
        ShapeError: If the number of input channels is not equal to filter's channels.
        ShapeError: If the size of `output_size` is not equal to that of `stride`.
3667 3668 3669 3670

    Examples:
       .. code-block:: python

3671
          import paddle.fluid as fluid
L
lvmengsi 已提交
3672
          data = fluid.data(name='data', shape=[None, 3, 12, 32, 32], dtype='float32')
3673
          conv3d_transpose = fluid.layers.conv3d_transpose(input=data, num_filters=2, filter_size=3)
Y
Yu Yang 已提交
3674
    """
C
chengduo 已提交
3675
    assert param_attr is not False, "param_attr should not be False in conv3d_transpose."
3676 3677 3678 3679
    if data_format not in ['NCDHW', 'NDHWC']:
        raise ValueError(
            "Param(data_format) of Op(fluid.layers.conv3d_transpose) got wrong value: received "
            + data_format + " but only NCDHW or NDHWC supported.")
3680 3681
    l_type = "conv3d_transpose"
    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
3682
    if not isinstance(input, Variable):
3683
        raise TypeError("Input of conv3d_transpose must be Variable")
3684 3685
    input_channel = input.shape[1] if data_format == 'NCDHW' else input.shape[
        -1]
Y
Yu Yang 已提交
3686

3687 3688
    stride = utils.convert_to_list(stride, 3, 'stride')
    dilation = utils.convert_to_list(dilation, 3, 'dilation')
C
chengduoZH 已提交
3689

C
chengduoZH 已提交
3690 3691 3692
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")

3693 3694 3695 3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706
    def _update_padding(padding, data_format):
        def is_list_or_tuple(ele):
            if isinstance(ele, list) or isinstance(ele, tuple):
                return True
            return False

        if is_list_or_tuple(padding) and len(padding) == 5:
            if is_list_or_tuple(padding[0]) and (data_format == "NCDHW"):
                if not (padding[0] == [0, 0] and padding[1] == [0, 0]):
                    raise ValueError(
                        "Non-zero padding(%s) in the batch or channel dimensions "
                        "is not supported." % str(padding))
                padding = padding[2:5]
                padding = [ele for a_list in padding for ele in a_list]
3707 3708 3709 3710 3711 3712 3713 3714
            elif is_list_or_tuple(padding[0]) and (data_format == "NDHWC"):
                if not (padding[0] == [0, 0] and padding[4] == [0, 0]):
                    raise ValueError(
                        "Non-zero padding(%s) in the batch or channel dimensions "
                        "is not supported." % str(padding))
                padding = padding[1:4]
                padding = [ele for a_list in padding for ele in a_list]
            padding = utils.convert_to_list(padding, 6, 'padding')
G
Guo Sheng 已提交
3715

3716 3717
        elif is_list_or_tuple(padding) and len(padding) == 6:
            padding = utils.convert_to_list(padding, 6, 'padding')
G
Guo Sheng 已提交
3718

3719 3720 3721 3722 3723 3724 3725
        else:
            padding = utils.convert_to_list(padding, 3, 'padding')
            padding = [
                padding[0], padding[0], padding[1], padding[1], padding[2],
                padding[2]
            ]
        return padding
G
Guo Sheng 已提交
3726

3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739
    padding_algorithm = "EXPLICIT"
    if isinstance(padding, str):
        padding = padding.upper()
        if padding not in ["SAME", "VALID"]:
            raise ValueError(
                "Unknown padding: '%s'. It can only be 'SAME' or 'VALID'." %
                str(padding))
        if padding == "VALID":
            padding_algorithm = "VALID"
            padding = [0, 0, 0, 0, 0, 0]
        elif padding == "SAME":
            padding_algorithm = "SAME"
            padding = [0, 0, 0, 0, 0, 0]
G
Guo Sheng 已提交
3740

3741
    padding = _update_padding(padding, data_format)
Y
yangyaming 已提交
3742

3743 3744 3745 3746 3747
    if filter_size is None:
        if output_size is None:
            raise ValueError("output_size must be set when filter_size is None")
        if isinstance(output_size, int):
            output_size = [output_size, output_size]
Y
yangyaming 已提交
3748

3749 3750 3751
        d_in = input.shape[2] if data_format == 'NCDHW' else input.shape[1]
        h_in = input.shape[3] if data_format == 'NCDHW' else input.shape[2]
        w_in = input.shape[4] if data_format == 'NCDHW' else input.shape[3]
Y
yangyaming 已提交
3752

3753 3754 3755 3756 3757 3758 3759 3760 3761 3762
        filter_size_d = (output_size[0] - (d_in - 1) * stride[0] + padding[0] +
                         padding[1] - 1) // dilation[0] + 1
        filter_size_h = (output_size[1] - (h_in - 1) * stride[1] + padding[2] +
                         padding[3] - 1) // dilation[1] + 1
        filter_size_w = (output_size[2] - (w_in - 1) * stride[2] + padding[4] +
                         padding[5] - 1) // dilation[2] + 1
        filter_size = [filter_size_d, filter_size_h, filter_size_w]
    else:
        filter_size = utils.convert_to_list(filter_size, 3,
                                            'conv3d_transpose.filter_size')
Y
yangyaming 已提交
3763

3764 3765
    if len(padding) == 6 and utils._is_symmetric_padding(padding, 3):
        padding = [padding[0], padding[2], padding[4]]
Y
yangyaming 已提交
3766

3767 3768 3769 3770
    groups = 1 if groups is None else groups
    filter_shape = [input_channel, num_filters // groups] + filter_size
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)
3771

3772 3773 3774 3775
    if data_format == 'NCDHW':
        data_format = 'NCHW'
    if data_format == 'NDHWC':
        data_format = 'NHWC'
Y
yangyaming 已提交
3776

3777
    pre_bias = helper.create_variable_for_type_inference(dtype=input.dtype)
Y
yangyaming 已提交
3778
    helper.append_op(
3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791
        type=l_type,
        inputs={'Input': [input],
                'Filter': [img_filter]},
        outputs={'Output': pre_bias},
        attrs={
            'strides': stride,
            'paddings': padding,
            'padding_algorithm': padding_algorithm,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn,
            'data_format': data_format
        })
Y
yangyaming 已提交
3792

3793 3794 3795 3796 3797 3798
    if data_format == 'NCHW':
        pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
    else:
        pre_act = helper.append_bias_op(pre_bias, dim_start=4, dim_end=5)
    out = helper.append_activation(pre_act)
    return out
G
guosheng 已提交
3799 3800


C
caoying03 已提交
3801
def reduce_sum(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
3802
    """
Y
yangyaming 已提交
3803
    Computes the sum of tensor elements over the given dimension.
G
guosheng 已提交
3804 3805

    Args:
3806 3807 3808
        input (Variable): The input variable which is a Tensor, the data type is float32,
            float64, int32, int64.
        dim (list|int, optional): The dimensions along which the sum is performed. If
Y
yangyaming 已提交
3809 3810
            :attr:`None`, sum all elements of :attr:`input` and return a
            Tensor variable with a single element, otherwise must be in the
W
whs 已提交
3811 3812
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
3813
        keep_dim (bool, optional): Whether to reserve the reduced dimension in the
Y
yangyaming 已提交
3814
            output Tensor. The result tensor will have one fewer dimension
3815 3816 3817 3818
            than the :attr:`input` unless :attr:`keep_dim` is true, default
            value is False.
        name(str, optional): The default value is None.  Normally there is no need for
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`
G
guosheng 已提交
3819 3820

    Returns:
3821 3822
        Variable: Tensor, results of summation operation on the specified dim of input tensor,
        it's data type is the same as input's Tensor.
F
fengjiayi 已提交
3823

3824 3825 3826
    Raises:
        TypeError, if out data type is different with the input data type.
    
G
guosheng 已提交
3827 3828 3829
    Examples:
        .. code-block:: python

3830
            import paddle.fluid as fluid
G
guosheng 已提交
3831 3832 3833
            # x is a Tensor variable with following elements:
            #    [[0.2, 0.3, 0.5, 0.9]
            #     [0.1, 0.2, 0.6, 0.7]]
Q
qiaolongfei 已提交
3834
            # Each example is followed by the corresponding output tensor.
3835
            x = fluid.data(name='x', shape=[2, 4], dtype='float32')
G
guosheng 已提交
3836 3837 3838 3839
            fluid.layers.reduce_sum(x)  # [3.5]
            fluid.layers.reduce_sum(x, dim=0)  # [0.3, 0.5, 1.1, 1.6]
            fluid.layers.reduce_sum(x, dim=-1)  # [1.9, 1.6]
            fluid.layers.reduce_sum(x, dim=1, keep_dim=True)  # [[1.9], [1.6]]
W
whs 已提交
3840

3841
            # y is a Tensor variable with shape [2, 2, 2] and elements as below:
W
whs 已提交
3842 3843
            #      [[[1, 2], [3, 4]],
            #      [[5, 6], [7, 8]]]
Q
qiaolongfei 已提交
3844
            # Each example is followed by the corresponding output tensor.
3845
            y = fluid.data(name='y', shape=[2, 2, 2], dtype='float32')
3846 3847
            fluid.layers.reduce_sum(y, dim=[1, 2]) # [10, 26]
            fluid.layers.reduce_sum(y, dim=[0, 1]) # [16, 20]
W
whs 已提交
3848

G
guosheng 已提交
3849 3850
    """
    helper = LayerHelper('reduce_sum', **locals())
3851 3852
    check_type_and_dtype(input, 'input', Variable,
                         ['float32', 'float64', 'int32', 'int64'], 'reduce_sum')
X
Xin Pan 已提交
3853
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
3854 3855
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
G
guosheng 已提交
3856 3857 3858 3859 3860
    helper.append_op(
        type='reduce_sum',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3861
            'dim': dim if dim != None else [0],
G
guosheng 已提交
3862 3863 3864 3865
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
3866 3867


C
caoying03 已提交
3868
def reduce_mean(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
3869
    """
Y
Yibing Liu 已提交
3870
    Computes the mean of the input tensor's elements along the given dimension.
G
guosheng 已提交
3871 3872

    Args:
3873 3874 3875
        input (Variable): The input variable which is a Tensor, the data type is float32,
            float64, int32, int64.
        dim (list|int, optional): The dimension along which the mean is computed. If
Y
Yibing Liu 已提交
3876 3877
            `None`, compute the mean over all elements of :attr:`input`
            and return a variable with a single element, otherwise it
Y
yangyaming 已提交
3878
            must be in the range :math:`[-rank(input), rank(input))`. If
3879
            :math:`dim[i] < 0`, the dimension to reduce is
Y
Yibing Liu 已提交
3880
            :math:`rank(input) + dim[i]`.
3881
        keep_dim (bool, optional): Whether to reserve the reduced dimension in the
Y
yangyaming 已提交
3882
            output Tensor. The result tensor will have one fewer dimension
3883 3884 3885 3886 3887
            than the :attr:`input` unless :attr:`keep_dim` is true, default 
            value is False.
        name(str, optional): The default value is None.  Normally there is no need for
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`
    
G
guosheng 已提交
3888
    Returns:
3889 3890 3891 3892 3893 3894
        Variable: Tensor, results of average on the specified dim of input tensor,
        it's data type is the same as input's Tensor.
    
    Raises:
        TypeError, if out data type is different with the input data type.
    
G
guosheng 已提交
3895 3896 3897
    Examples:
        .. code-block:: python

3898
            import paddle.fluid as fluid
G
guosheng 已提交
3899 3900 3901 3902
            # x is a Tensor variable with following elements:
            #    [[0.2, 0.3, 0.5, 0.9]
            #     [0.1, 0.2, 0.6, 0.7]]
            # Each example is followed by the correspending output tensor.
3903
            x = fluid.data(name='x', shape=[2, 4], dtype='float32')
G
guosheng 已提交
3904 3905 3906
            fluid.layers.reduce_mean(x)  # [0.4375]
            fluid.layers.reduce_mean(x, dim=0)  # [0.15, 0.25, 0.55, 0.8]
            fluid.layers.reduce_mean(x, dim=-1)  # [0.475, 0.4]
3907
            fluid.layers.reduce_mean(x, dim=1, keep_dim=True)  # [[0.475], [0.4]]
W
whs 已提交
3908

3909
            # y is a Tensor variable with shape [2, 2, 2] and elements as below:
W
whs 已提交
3910 3911 3912
            #      [[[1.0, 2.0], [3.0, 4.0]],
            #      [[5.0, 6.0], [7.0, 8.0]]]
            # Each example is followed by the correspending output tensor.
3913
            y = fluid.data(name='y', shape=[2, 2, 2], dtype='float32')
3914 3915
            fluid.layers.reduce_mean(y, dim=[1, 2]) # [2.5, 6.5]
            fluid.layers.reduce_mean(y, dim=[0, 1]) # [4.0, 5.0]
G
guosheng 已提交
3916 3917
    """
    helper = LayerHelper('reduce_mean', **locals())
3918 3919 3920
    check_type_and_dtype(input, 'input', Variable,
                         ['float32', 'float64', 'int32', 'int64'],
                         'reduce_mean')
X
Xin Pan 已提交
3921
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
3922 3923
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
G
guosheng 已提交
3924 3925 3926 3927 3928
    helper.append_op(
        type='reduce_mean',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3929
            'dim': dim if dim != None else [0],
G
guosheng 已提交
3930 3931 3932 3933
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
3934 3935


C
caoying03 已提交
3936
def reduce_max(input, dim=None, keep_dim=False, name=None):
3937
    """
Y
yangyaming 已提交
3938
    Computes the maximum of tensor elements over the given dimension.
3939 3940

    Args:
3941 3942 3943
        input (Variable): The input variable which is a Tensor, the data type is float32,
            float64, int32, int64.
        dim (list|int, optional): The dimension along which the maximum is computed.
Y
yangyaming 已提交
3944 3945 3946
            If :attr:`None`, compute the maximum over all elements of
            :attr:`input` and return a Tensor variable with a single element,
            otherwise must be in the range :math:`[-rank(input), rank(input))`.
W
whs 已提交
3947
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
3948
        keep_dim (bool, optional): Whether to reserve the reduced dimension in the
Y
yangyaming 已提交
3949
            output Tensor. The result tensor will have one fewer dimension
3950 3951 3952 3953
            than the :attr:`input` unless :attr:`keep_dim` is true, default
            value is False.
        name(str, optional): The default value is None.  Normally there is no need for 
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`
3954 3955

    Returns:
3956 3957
        Variable: Tensor, results of maximum on the specified dim of input tensor,
        it's data type is the same as input's Tensor.
Y
yangyaming 已提交
3958

3959 3960 3961
    Examples:
        .. code-block:: python

3962
            import paddle.fluid as fluid
3963 3964 3965 3966
            # x is a Tensor variable with following elements:
            #    [[0.2, 0.3, 0.5, 0.9]
            #     [0.1, 0.2, 0.6, 0.7]]
            # Each example is followed by the correspending output tensor.
3967
            x = fluid.data(name='x', shape=[2, 4], dtype='float32')
3968 3969 3970 3971
            fluid.layers.reduce_max(x)  # [0.9]
            fluid.layers.reduce_max(x, dim=0)  # [0.2, 0.3, 0.6, 0.9]
            fluid.layers.reduce_max(x, dim=-1)  # [0.9, 0.7]
            fluid.layers.reduce_max(x, dim=1, keep_dim=True)  # [[0.9], [0.7]]
W
whs 已提交
3972

3973
            # y is a Tensor variable with shape [2, 2, 2] and elements as below:
W
whs 已提交
3974 3975 3976
            #      [[[1.0, 2.0], [3.0, 4.0]],
            #      [[5.0, 6.0], [7.0, 8.0]]]
            # Each example is followed by the correspending output tensor.
3977
            y = fluid.data(name='y', shape=[2, 2, 2], dtype='float32')
3978 3979
            fluid.layers.reduce_max(y, dim=[1, 2]) # [4.0, 8.0]
            fluid.layers.reduce_max(y, dim=[0, 1]) # [7.0, 8.0]
3980 3981
    """
    helper = LayerHelper('reduce_max', **locals())
X
Xin Pan 已提交
3982
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
3983 3984
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3985 3986 3987 3988 3989
    helper.append_op(
        type='reduce_max',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3990
            'dim': dim if dim != None else [0],
3991 3992 3993 3994 3995 3996
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


C
caoying03 已提交
3997
def reduce_min(input, dim=None, keep_dim=False, name=None):
3998
    """
Y
yangyaming 已提交
3999
    Computes the minimum of tensor elements over the given dimension.
4000 4001

    Args:
4002 4003 4004
        input (Variable): The input variable which is a Tensor, the data type is float32,
            float64, int32, int64.
        dim (list|int, optional): The dimensions along which the minimum is computed.
Y
yangyaming 已提交
4005 4006 4007
            If :attr:`None`, compute the minimum over all elements of
            :attr:`input` and return a Tensor variable with a single element,
            otherwise must be in the range :math:`[-rank(input), rank(input))`.
W
whs 已提交
4008
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
4009
        keep_dim (bool, optional): Whether to reserve the reduced dimension in the
Y
yangyaming 已提交
4010
            output Tensor. The result tensor will have one fewer dimension
4011 4012 4013 4014
            than the :attr:`input` unless :attr:`keep_dim` is true, default
            value is False.
        name(str, optional): The default value is None.  Normally there is no need for 
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`
4015 4016

    Returns:
4017 4018
        Variable: Tensor, result of minimum on the specified dim of input tensor,
        it's data type is the same as input's Tensor.
Y
yangyaming 已提交
4019

4020 4021 4022
    Examples:
        .. code-block:: python

4023
            import paddle.fluid as fluid
4024 4025 4026 4027
            # x is a Tensor variable with following elements:
            #    [[0.2, 0.3, 0.5, 0.9]
            #     [0.1, 0.2, 0.6, 0.7]]
            # Each example is followed by the correspending output tensor.
4028
            x = fluid.data(name='x', shape=[2, 4], dtype='float32')
4029 4030 4031 4032
            fluid.layers.reduce_min(x)  # [0.1]
            fluid.layers.reduce_min(x, dim=0)  # [0.1, 0.2, 0.5, 0.7]
            fluid.layers.reduce_min(x, dim=-1)  # [0.2, 0.1]
            fluid.layers.reduce_min(x, dim=1, keep_dim=True)  # [[0.2], [0.1]]
W
whs 已提交
4033

4034
            # y is a Tensor variable with shape [2, 2, 2] and elements as below:
W
whs 已提交
4035 4036 4037
            #      [[[1.0, 2.0], [3.0, 4.0]],
            #      [[5.0, 6.0], [7.0, 8.0]]]
            # Each example is followed by the correspending output tensor.
4038
            y = fluid.data(name='y', shape=[2, 2, 2], dtype='float32')
4039 4040
            fluid.layers.reduce_min(y, dim=[1, 2]) # [1.0, 5.0]
            fluid.layers.reduce_min(y, dim=[0, 1]) # [1.0, 2.0]
4041 4042
    """
    helper = LayerHelper('reduce_min', **locals())
X
Xin Pan 已提交
4043
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
4044 4045
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
4046 4047 4048 4049 4050
    helper.append_op(
        type='reduce_min',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
4051
            'dim': dim if dim != None else [0],
4052 4053 4054 4055
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
4056 4057


4058 4059 4060 4061 4062
def reduce_prod(input, dim=None, keep_dim=False, name=None):
    """
    Computes the product of tensor elements over the given dimension.

    Args:
4063 4064 4065
        input (Variable): The input variable which is a Tensor, the data type is float32,
            float64, int32, int64.
        dim (list|int, optional): The dimensions along which the product is performed. If
4066 4067
            :attr:`None`, multipy all elements of :attr:`input` and return a
            Tensor variable with a single element, otherwise must be in the
W
whs 已提交
4068 4069
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
4070
        keep_dim (bool, optional): Whether to reserve the reduced dimension in the
4071
            output Tensor. The result tensor will have one fewer dimension
4072 4073 4074 4075
            than the :attr:`input` unless :attr:`keep_dim` is true, default
            value is False.
        name(str, optional): The default value is None.  Normally there is no need for 
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`
4076 4077

    Returns:
4078 4079 4080
        Variable: Tensor, result of product on the specified dim of input tensor,
        it's data type is the same as input's Tensor.
    
4081 4082 4083
    Examples:
        .. code-block:: python

4084
            import paddle.fluid as fluid
4085 4086 4087 4088
            # x is a Tensor variable with following elements:
            #    [[0.2, 0.3, 0.5, 0.9]
            #     [0.1, 0.2, 0.6, 0.7]]
            # Each example is followed by the correspending output tensor.
4089
            x = fluid.data(name='x', shape=[2, 4], dtype='float32')
4090 4091 4092
            fluid.layers.reduce_prod(x)  # [0.0002268]
            fluid.layers.reduce_prod(x, dim=0)  # [0.02, 0.06, 0.3, 0.63]
            fluid.layers.reduce_prod(x, dim=-1)  # [0.027, 0.0084]
Y
yangyaming 已提交
4093
            fluid.layers.reduce_prod(x, dim=1,
Z
zhouhanqing 已提交
4094
                                     keep_dim=True)  # [[0.027], [0.0084]]
W
whs 已提交
4095

4096
            # y is a Tensor variable with shape [2, 2, 2] and elements as below:
W
whs 已提交
4097 4098 4099
            #      [[[1.0, 2.0], [3.0, 4.0]],
            #      [[5.0, 6.0], [7.0, 8.0]]]
            # Each example is followed by the correspending output tensor.
4100
            y = fluid.data(name='y', shape=[2, 2, 2], dtype='float32')
4101 4102
            fluid.layers.reduce_prod(y, dim=[1, 2]) # [24.0, 1680.0]
            fluid.layers.reduce_prod(y, dim=[0, 1]) # [105.0, 384.0]
4103 4104
    """
    helper = LayerHelper('reduce_prod', **locals())
X
Xin Pan 已提交
4105
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
4106 4107
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
4108 4109 4110 4111 4112
    helper.append_op(
        type='reduce_prod',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
4113
            'dim': dim if dim != None else [0],
4114 4115 4116 4117 4118 4119
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


Z
zhoukunsheng 已提交
4120 4121
def reduce_all(input, dim=None, keep_dim=False, name=None):
    """
4122
    This OP computes the ``logical and`` of tensor elements over the given dimension, and output the result.
Z
zhoukunsheng 已提交
4123 4124

    Args:
4125 4126
        input (Variable): The input variable which is a Tensor or LoDTensor, the input data type should be `bool`.
        dim (list|int|optional): The dimension along which the logical and is computed.
Z
zhoukunsheng 已提交
4127 4128 4129
            If :attr:`None`, compute the logical and over all elements of
            :attr:`input` and return a Tensor variable with a single element,
            otherwise must be in the range :math:`[-rank(input), rank(input))`.
4130
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`. The default value is None. 
Z
zhoukunsheng 已提交
4131 4132
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
4133
            than the :attr:`input` unless :attr:`keep_dim` is true. The default value is False.
Z
zhoukunsheng 已提交
4134
        name(str|None): A name for this layer(optional). If set None, the layer
4135
                       will be named automatically. The default value is None. 
Z
zhoukunsheng 已提交
4136

4137 4138
    Returns: 
        Variable, the output data type is bool. : The reduced tensor variable with ``logical and`` in given dims.
Z
zhoukunsheng 已提交
4139 4140 4141

    Examples:
        .. code-block:: python
Z
zhoukunsheng 已提交
4142
        
4143
            import paddle.fluid as fluid
4144 4145 4146
            import paddle.fluid.layers as layers
            import numpy as np

Z
zhoukunsheng 已提交
4147 4148 4149
            # x is a bool Tensor variable with following elements:
            #    [[True, False]
            #     [True, True]]
4150 4151 4152 4153 4154 4155
            x = layers.assign(np.array([[1, 0], [1, 1]], dtype='int32'))
            x = layers.cast(x, 'bool')

            out = layers.reduce_all(x)  # False 
            out = layers.reduce_all(x, dim=0)  # [True, False]
            out = layers.reduce_all(x, dim=-1)  # [False, True]
4156 4157
            # keep_dim=False, x.shape=(2,2), out.shape=(2,)

4158
            out = layers.reduce_all(x, dim=1, keep_dim=True)  # [[False], [True]]
4159
            # keep_dim=True, x.shape=(2,2), out.shape=(2,1)
Z
zhoukunsheng 已提交
4160 4161 4162 4163 4164 4165 4166 4167 4168 4169 4170 4171 4172 4173 4174 4175 4176 4177 4178 4179

    """
    helper = LayerHelper('reduce_all', **locals())
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
    helper.append_op(
        type='reduce_all',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
            'dim': dim if dim != None else [0],
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


def reduce_any(input, dim=None, keep_dim=False, name=None):
    """
4180
    This OP computes the ``logical or`` of tensor elements over the given dimension, and output the result.
Z
zhoukunsheng 已提交
4181 4182

    Args:
4183 4184 4185
        input (Variable): The input variable which is a Tensor or LoDTensor, the input data type should be `bool`.
        dim (list|int|optional): The dimension along which the logical and is computed.
            If :attr:`None`, compute the logical and over all elements of
Z
zhoukunsheng 已提交
4186 4187
            :attr:`input` and return a Tensor variable with a single element,
            otherwise must be in the range :math:`[-rank(input), rank(input))`.
4188
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`. The default value is None. 
Z
zhoukunsheng 已提交
4189 4190
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
4191
            than the :attr:`input` unless :attr:`keep_dim` is true. The default value is False.
Z
zhoukunsheng 已提交
4192 4193
        name(str|None): A name for this layer(optional). If set None, the layer

4194 4195
    Returns: 
        Variable, the output data type is bool. : The reduced tensor variable with ``logical or`` in given dims.
Z
zhoukunsheng 已提交
4196 4197 4198

    Examples:
        .. code-block:: python
Z
zhoukunsheng 已提交
4199

4200
            import paddle.fluid as fluid
4201 4202 4203
            import paddle.fluid.layers as layers
            import numpy as np

Z
zhoukunsheng 已提交
4204 4205 4206
            # x is a bool Tensor variable with following elements:
            #    [[True, False]
            #     [False, False]]
4207 4208 4209 4210 4211 4212
            x = layers.assign(np.array([[1, 0], [0, 0]], dtype='int32'))
            x = layers.cast(x, 'bool')

            out = layers.reduce_any(x)  # True
            out = layers.reduce_any(x, dim=0)  # [True, False]
            out = layers.reduce_any(x, dim=-1)  # [True, False]
4213 4214
            # keep_dim=False, x.shape=(2,2), out.shape=(2,)

4215
            out = layers.reduce_any(x, dim=1,
Z
zhoukunsheng 已提交
4216
                                     keep_dim=True)  # [[True], [False]]
4217
            # keep_dim=True, x.shape=(2,2), out.shape=(2,1)
Z
zhoukunsheng 已提交
4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229 4230

    """
    helper = LayerHelper('reduce_any', **locals())
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
    helper.append_op(
        type='reduce_any',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
            'dim': dim if dim != None else [0],
            'keep_dim': keep_dim,
4231 4232 4233 4234 4235
            'reduce_all': True if dim == None else False
        })
    return out


C
caoying03 已提交
4236
def split(input, num_or_sections, dim=-1, name=None):
G
guosheng 已提交
4237
    """
4238
    Split the input tensor into multiple sub-Tensors.
G
guosheng 已提交
4239 4240

    Args:
4241
        input (Variable): The input variable which is an N-D Tensor or LoDTensor, data type being float32, float64, int32 or int64.
4242
        num_or_sections (int|list|tuple): If :attr:`num_or_sections` is an integer,
4243 4244
            then the integer indicates the number of equal sized sub-Tensors
            that the Tensor will be divided into. If :attr:`num_or_sections`
4245 4246 4247 4248 4249
            is a list or tuple, the length of it indicates the number of
            sub-Tensors and the elements in it indicate the sizes of sub-Tensors'
            :attr:`dim` dimension orderly. The length of the list mustn't be larger than the Tensor's size of :attr:`dim` .
        dim (int32|Varible, optional): A scalar with type ``int32`` or a ``Tensor`` with shape [1] and type ``int32``. The dimension along which to split. If :math:`dim < 0`, the
            dimension to split along is :math:`rank(input) + dim`. Default is -1.
4250
        name(str, optional): The default value is None.  Normally there is no need for user to set this property.  For more information, please refer to :ref:`api_guide_Name` .
G
guosheng 已提交
4251 4252

    Returns:
4253
        list(Variable): The list of segmented Tensor variables.
G
guosheng 已提交
4254

4255 4256 4257 4258
    Raises:
        TypeError: num_or_sections is not int, list or tuple.
        TypeError: dim is not int or Variable.

4259
    Example:
G
guosheng 已提交
4260 4261
        .. code-block:: python

4262 4263
            import paddle.fluid as fluid

4264 4265
            # input is a variable which shape is [3, 9, 5]
            input = fluid.data(
4266 4267
                 name="input", shape=[3, 9, 5], dtype="float32")

4268 4269 4270 4271
            x0, x1, x2 = fluid.layers.split(input, num_or_sections=3, dim=1)
            # x0.shape [3, 3, 5]
            # x1.shape [3, 3, 5]
            # x2.shape [3, 3, 5]
4272

4273 4274 4275 4276 4277 4278 4279 4280 4281
            x0, x1, x2 = fluid.layers.split(input, num_or_sections=[2, 3, 4], dim=1)
            # x0.shape [3, 2, 5]
            # x1.shape [3, 3, 5]
            # x2.shape [3, 4, 5]

            x0, x1, x2 = fluid.layers.split(input, num_or_sections=[2, 3, -1], dim=1)
            # x0.shape [3, 2, 5]
            # x1.shape [3, 3, 5]
            # x2.shape [3, 4, 5]
G
guosheng 已提交
4282
    """
4283 4284 4285 4286 4287 4288 4289 4290 4291
    if not isinstance(num_or_sections, (int, list, tuple)):
        raise TypeError(
            "The type of 'num_or_sections' in split must be int, list or "
            "tuple, but received %s." % (type(num_or_sections)))
    if not isinstance(dim, (int, Variable)):
        raise TypeError(
            "The type of 'dim' in split must be int or Variable, but "
            "received %s." % (type(dim)))

G
guosheng 已提交
4292 4293
    helper = LayerHelper('split', **locals())
    input_shape = input.shape
4294 4295 4296 4297 4298 4299 4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312 4313 4314 4315 4316 4317 4318 4319 4320 4321 4322 4323 4324
    inputs = {'X': input}
    attrs = {'num': num_or_sections if isinstance(num_or_sections, int) else 0}

    def _get_SectionsTensorList(one_list):
        tensor_list = []
        unk_dim_idx = -1
        for idx, dim_size in enumerate(one_list):
            if isinstance(dim_size, Variable):
                dim_size.stop_gradient = True
                tensor_list.append(dim_size)
            else:
                assert (isinstance(dim_size, int))
                if dim_size == -1:
                    assert unk_dim_idx == -1, (
                        "Only one value of 'num_or_section' in split can "
                        "be -1. But received num_or_section[%d] is also -1." %
                        idx)
                    unk_dim_idx = idx
                temp_out = helper.create_variable_for_type_inference('int32')
                fill_constant(
                    [1], 'int32', dim_size, force_cpu=True, out=temp_out)
                tensor_list.append(temp_out)
        return tensor_list

    if isinstance(dim, Variable):
        dim.stop_gradient = True
        inputs['AxisTensor'] = dim
    else:
        dim = (len(input_shape) + dim) if dim < 0 else dim
        attrs['axis'] = dim

G
guosheng 已提交
4325 4326
    if isinstance(num_or_sections, int):
        assert num_or_sections > 1, 'num_or_sections must be more than 1.'
4327 4328 4329 4330 4331
        if isinstance(dim, int) and input_shape[dim] > 0:
            assert input_shape[dim] % num_or_sections ==0, \
                "The input's size along the split dimension " \
                "must be evenly divisible by Attr(num_or_sections). " \
                "But %d is not evenly divisible by %d. " % (num_or_sections,input_shape[dim])
G
guosheng 已提交
4332 4333
        num = num_or_sections
    else:
4334 4335 4336
        if isinstance(dim, int) and input_shape[dim] > 0:
            assert len(num_or_sections) <= input_shape[
                dim], 'len(num_or_sections) must not be more than input.shape[dim].'
G
guosheng 已提交
4337
        num = len(num_or_sections)
4338 4339 4340 4341 4342 4343 4344 4345 4346
        attrs['sections'] = list(
            map(lambda ele: -1 if isinstance(ele, Variable) else ele,
                num_or_sections))
        contain_var = not all(not isinstance(ele, Variable)
                              for ele in num_or_sections)
        if contain_var:
            inputs['SectionsTensorList'] = _get_SectionsTensorList(
                num_or_sections)

G
guosheng 已提交
4347
    outs = [
X
Xin Pan 已提交
4348
        helper.create_variable_for_type_inference(dtype=helper.input_dtype())
G
guosheng 已提交
4349 4350 4351
        for i in range(num)
    ]
    helper.append_op(
4352
        type='split', inputs=inputs, outputs={'Out': outs}, attrs=attrs)
G
guosheng 已提交
4353
    return outs
C
caoying03 已提交
4354 4355 4356 4357


def l2_normalize(x, axis, epsilon=1e-12, name=None):
    """
R
ruri 已提交
4358
    This op normalizes `x` along dimension `axis` using an L2
C
caoying03 已提交
4359 4360
    norm. For a 1-D tensor (`dim` is fixed to 0), this layer computes

4361
    .. math::
4362 4363

        y = \\frac{x}{ \sqrt{\sum {x^2} + epsion }}
C
caoying03 已提交
4364 4365 4366 4367 4368

    For `x` with more dimensions, this layer independently normalizes each 1-D
    slice along dimension `axis`.

    Args:
R
ruri 已提交
4369
        x(Variable|list): The input tensor could be N-D tensor, and the input data type could be float32 or float64.
4370
        axis(int): The axis on which to apply normalization. If `axis < 0`, \
4371 4372
            the dimension to normalization is rank(X) + axis. -1 is the
            last dimension.
4373
        epsilon(float): The epsilon value is used to avoid division by zero, \
翟飞跃 已提交
4374
            the default value is 1e-12.
R
ruri 已提交
4375 4376
	name(str, optional): The default value is None.  Normally there is no need for user to set this property.  For more information, please refer to :ref:`api_guide_Name`
    
C
caoying03 已提交
4377
    Returns:
R
ruri 已提交
4378
        Variable: The output has the same shape and data type with `x`.
C
caoying03 已提交
4379 4380

    Examples:
4381

C
caoying03 已提交
4382
        .. code-block:: python
R
ruri 已提交
4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394
	    
	    # declarative mode
	    import paddle.fluid as fluid
	    import numpy as np
	    input = fluid.data(name="input", shape=[2,3])
	    output = fluid.layers.l2_normalize(x=input,axis=0)
	    place = fluid.CPUPlace()
	    exe = fluid.Executor(place)
	    exe.run(fluid.default_startup_program())
 
	    input_data = np.random.rand(2,3).astype("float32")
	    print(input_data)
C
caoying03 已提交
4395

R
ruri 已提交
4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419
	    # [[0.5171216  0.12704141 0.56018186]
	    # [0.93251234 0.5382788  0.81709313]]
	
	    output_data = exe.run(fluid.default_main_program(),
                feed={"input":input_data},
                fetch_list=[output],
                return_numpy=True)
 
	    print(output_data)

	    # [array([[0.48496857, 0.22970329, 0.56545246],
	    # [0.8745316 , 0.9732607 , 0.82478094]], dtype=float32)]

	    # imperative mode
	    import paddle.fluid.dygraph as dg

	    with dg.guard(place) as g:
    		input = dg.to_variable(input_data)
    		output = fluid.layers.l2_normalize(x=input, axis=-1)
    		print(output.numpy())
	    	
		# [[0.66907585 0.16437206 0.7247892 ]
		# [0.6899054  0.3982376  0.6045142 ]]
		
C
caoying03 已提交
4420 4421
    """

F
fengjiayi 已提交
4422 4423
    if len(x.shape) == 1:
        axis = 0
C
caoying03 已提交
4424 4425
    helper = LayerHelper("l2_normalize", **locals())

X
Xin Pan 已提交
4426 4427
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    norm = helper.create_variable_for_type_inference(dtype=x.dtype)
C
caoying03 已提交
4428
    helper.append_op(
4429 4430 4431 4432
        type="norm",
        inputs={"X": x},
        outputs={"Out": out,
                 "Norm": norm},
C
caoying03 已提交
4433
        attrs={
4434 4435
            "axis": 1 if axis is None else axis,
            "epsilon": epsilon,
C
caoying03 已提交
4436 4437
        })
    return out
4438 4439


S
sneaxiy 已提交
4440
def matmul(x, y, transpose_x=False, transpose_y=False, alpha=1.0, name=None):
G
guosheng 已提交
4441
    """
Y
ying 已提交
4442 4443 4444 4445
    Applies matrix multiplication to two tensors.

    Currently, the input tensors' rank can be any, but when the rank of any
    inputs is bigger than 3, this two inputs' rank should be equal.
G
guosheng 已提交
4446

C
chengduoZH 已提交
4447
    The actual behavior depends on the shapes of :math:`x`, :math:`y` and the
4448
    flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically:
G
guosheng 已提交
4449

4450 4451 4452 4453 4454
    - If a transpose flag is specified, the last two dimensions of the tensor
      are transposed. If the tensor is rank-1 of shape :math:`[D]`, then for
      :math:`x` it is treated as :math:`[1, D]` in nontransposed form and as
      :math:`[D, 1]` in transposed form, whereas for :math:`y` it is the
      opposite: It is treated as :math:`[D, 1]` in nontransposed form and as
4455
      :math:`[1, D]` in transposed form.
G
guosheng 已提交
4456

C
chengduoZH 已提交
4457
    - After transpose, the two tensors are 2-D or n-D and matrix multiplication
4458
      performs in the following way.
G
guosheng 已提交
4459

4460
      - If both are 2-D, they are multiplied like conventional matrices.
C
chengduoZH 已提交
4461
      - If either is n-D, it is treated as a stack of matrices residing in the
Y
ying 已提交
4462
        last two dimensions and a batched matrix multiply supporting broadcast
4463
        applies on the two tensors.
G
guosheng 已提交
4464

Y
ying 已提交
4465 4466
    Also note that if the raw tensor :math:`x` or :math:`y` is rank-1 and
    nontransposed, the prepended or appended dimension :math:`1` will be
C
chengduoZH 已提交
4467
    removed after matrix multiplication.
G
guosheng 已提交
4468 4469 4470

    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
4471 4472 4473
        y (Variable): The input variable which is a Tensor or LoDTensor.
        transpose_x (bool): Whether to transpose :math:`x` before multiplication.
        transpose_y (bool): Whether to transpose :math:`y` before multiplication.
S
sneaxiy 已提交
4474
        alpha (float): The scale of output. Default 1.0.
4475
        name(str|None): A name for this layer(optional). If set None, the layer
4476
            will be named automatically.
G
guosheng 已提交
4477 4478

    Returns:
石晓伟 已提交
4479
        Variable: The product Tensor (or LoDTensor) variable.
G
guosheng 已提交
4480

G
guosheng 已提交
4481 4482 4483
    Examples:
        .. code-block:: python

4484
            # Examples to clarify shapes of the inputs and output
C
chengduoZH 已提交
4485
            # x: [B, ..., M, K], y: [B, ..., K, N]
4486
            # fluid.layers.matmul(x, y)  # out: [B, ..., M, N]
Y
ying 已提交
4487

4488
            # x: [B, M, K], y: [B, K, N]
4489
            # fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
4490

4491
            # x: [B, M, K], y: [K, N]
4492
            # fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
4493

4494
            # x: [M, K], y: [K, N]
4495
            # fluid.layers.matmul(x, y)  # out: [M, N]
Y
ying 已提交
4496 4497

            # x: [B, M, K], y: [K]
4498
            # fluid.layers.matmul(x, y)  # out: [B, M]
Y
ying 已提交
4499

4500
            # x: [K], y: [K]
4501
            # fluid.layers.matmul(x, y)  # out: [1]
4502

Y
ying 已提交
4503
            # x: [M], y: [N]
4504 4505
            # fluid.layers.matmul(x, y, True, True)  # out: [M, N]

4506
            import paddle.fluid as fluid
4507 4508 4509
            x = fluid.layers.data(name='x', shape=[2, 3], dtype='float32')
            y = fluid.layers.data(name='y', shape=[3, 2], dtype='float32')
            out = fluid.layers.matmul(x, y, True, True)
G
guosheng 已提交
4510
    """
Y
ying 已提交
4511 4512

    def __check_input(x, y):
4513 4514
        var_names = {'x': x, 'y': y}
        for name, val in var_names.items():
4515 4516
            check_type_and_dtype(val, name, Variable,
                                 ['float16', 'float32', 'float64'], 'matmul')
Y
ying 已提交
4517 4518 4519 4520 4521
        x_shape = list(x.shape)
        y_shape = list(y.shape)
        if len(x_shape) == 1:
            x_shape = [1] + x_shape
        if len(y_shape) == 1:
Y
ying 已提交
4522
            y_shape = y_shape + [1]
Y
ying 已提交
4523 4524 4525 4526 4527 4528 4529

        # check the inner 2 dimensions
        if transpose_x:
            x_shape[-2], x_shape[-1] = x_shape[-1], x_shape[-2]
        if transpose_y:
            y_shape[-2], y_shape[-1] = y_shape[-1], y_shape[-2]
        if x_shape[-1] != y_shape[-2]:
4530 4531 4532 4533 4534
            assert (x_shape[-1] == -1) or (y_shape[-2] == -1),                         \
                "After performing an optional transpose, Input X's width should be "   \
                "equal to Y's width for multiplication "                               \
                "prerequisites. But received X's shape: %s, Y's shape: %s\n" %         \
                (x_shape, y_shape)
Y
ying 已提交
4535

C
chengduo 已提交
4536
        if len(y_shape) > 2 and len(x_shape) > 2:
Y
ying 已提交
4537
            for i, dim_x in enumerate(x_shape[:-2]):
P
phlrain 已提交
4538 4539 4540
                # don't check neg shape
                if dim_x < 0 or y_shape[i] < 0:
                    continue
Y
ying 已提交
4541
                if dim_x != y_shape[i]:
4542 4543 4544 4545 4546
                    raise ValueError(
                        "When the matrix is larger than 2 dimensions, the higher "
                        "dimensional values of the two matrices need to be equal. "
                        "But received x_shape[%d] != y_shape[%d]. X's shape: %s, "
                        "Y's shape: %s.\n" % (i, i, x_shape, y_shape))
Y
ying 已提交
4547 4548 4549

    __check_input(x, y)

4550
    helper = LayerHelper('matmul', **locals())
X
Xin Pan 已提交
4551
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
G
guosheng 已提交
4552
    helper.append_op(
4553 4554 4555 4556
        type='matmul',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
S
sneaxiy 已提交
4557 4558 4559
        attrs={
            'transpose_X': transpose_x,
            'transpose_Y': transpose_y,
S
sneaxiy 已提交
4560
            'alpha': float(alpha),
S
sneaxiy 已提交
4561
        })
4562
    return out
4563 4564


4565
def topk(input, k, name=None):
Q
qingqing01 已提交
4566
    """
4567
    This OP is used to find values and indices of the k largest entries
Q
qingqing01 已提交
4568 4569
    for the last dimension.

4570 4571
    If the input is a 1-D Tensor, finds the k largest entries and outputs
    their values and indices.
Q
qingqing01 已提交
4572 4573 4574 4575

    If the input is a Tensor with higher rank, this operator computes the top k
    entries along the last dimension.

F
fengjiayi 已提交
4576 4577
    .. code-block:: text

4578 4579 4580 4581 4582
        Case 1:

          Input:
            input.shape = [3, 4]
            input.data = [[5, 4, 2, 3],
F
fengjiayi 已提交
4583 4584 4585 4586
                     [9, 7, 10, 25],
                     [6, 2, 10, 1]]
            k = 2

4587
          Output:
F
fengjiayi 已提交
4588
            The first output:
4589 4590
            values.shape = [3, 2]
            values.data = [[5, 4],
F
fengjiayi 已提交
4591 4592 4593 4594
                      [10, 25],
                      [6, 10]]

            The second output:
4595 4596
            indices.shape = [3, 2]
            indices.data = [[0, 1],
F
fengjiayi 已提交
4597 4598 4599
                       [2, 3],
                       [0, 2]]

Q
qingqing01 已提交
4600
    Args:
4601 4602 4603 4604
        input(Variable): The input tensor. Support data types: float32, float64.
        k(int | Variable): The number of top elements to look for along the last dimension
                           of input tensor.
        name (str, optional): Please refer to :ref:`api_guide_Name`, Default None.
Q
qingqing01 已提交
4605 4606

    Returns:
4607 4608
        Values (Variable): Input tensor's k largest elements along each last dimensional slice. The dimension is: :math:`input.shape[:-1]+[k]`.
        Indices (Variable): Indices of k largest elements alone the last dimension of input. The dimension is same as values.
Q
qingqing01 已提交
4609

F
fengjiayi 已提交
4610
    Raises:
4611
        ValueError: If :math:`k < 1` or :math:`k > last dimension of input`.
Q
qingqing01 已提交
4612 4613 4614 4615

    Examples:
        .. code-block:: python

4616
            import paddle.fluid as fluid
4617
            import paddle.fluid.layers as layers
4618 4619 4620 4621 4622 4623 4624 4625 4626 4627 4628 4629 4630
            # set batch size=None
            input = fluid.data(name="input", shape=[None, 13, 11], dtype='float32')
            top5_values, top5_indices = layers.topk(input, k=5) # top5_values.shape[None, 13, 5], top5_indices.shape=[None, 13, 5]

            # 1D Tensor
            input1 = fluid.data(name="input1", shape=[None, 13], dtype='float32')
            top5_values, top5_indices = layers.topk(input1, k=5) #top5_values.shape=[None, 5], top5_indices.shape=[None, 5]

            # k=Variable
            input2 = fluid.data(name="input2", shape=[None, 13, 11], dtype='float32')
            vk = fluid.data(name="vk", shape=[None, 1], dtype='int32') # save k in vk.data[0]
            vk_values, vk_indices = layers.topk(input2, k=vk) #vk_values.shape=[None, 13, k], vk_indices.shape=[None, 13, k]

Q
qingqing01 已提交
4631 4632
    """
    helper = LayerHelper("top_k", **locals())
X
Xin Pan 已提交
4633 4634
    values = helper.create_variable_for_type_inference(dtype=input.dtype)
    indices = helper.create_variable_for_type_inference(dtype="int64")
W
whs 已提交
4635 4636 4637 4638 4639 4640
    inputs = {"X": [input]}
    attrs = None
    if isinstance(k, Variable):
        inputs['K'] = k
    else:
        attrs = {'k': k}
Q
qingqing01 已提交
4641 4642
    helper.append_op(
        type="top_k",
W
whs 已提交
4643
        inputs=inputs,
Q
qingqing01 已提交
4644 4645
        outputs={"Out": [values],
                 "Indices": [indices]},
W
whs 已提交
4646
        attrs=attrs)
Q
qingqing01 已提交
4647 4648 4649 4650 4651
    values.stop_gradient = True
    indices.stop_gradient = True
    return values, indices


4652 4653 4654 4655 4656
def ctc_greedy_decoder(input,
                       blank,
                       input_length=None,
                       padding_value=0,
                       name=None):
4657
    """
S
SunGaofeng 已提交
4658
    This op is used to decode sequences by greedy policy by the following steps:
Y
yi.wu 已提交
4659

S
SunGaofeng 已提交
4660
    1. Get the indexes of maximum value for each row in input. a.k.a.
Y
ying 已提交
4661 4662 4663
       numpy.argmax(input, axis=0).
    2. For each sequence in result of step1, merge repeated tokens between two
       blanks and delete all blanks.
4664

S
SunGaofeng 已提交
4665 4666 4667 4668
    This op is implemented in two modes: lod and padding, either of them can be used.
    The input can be either LoDTensor or Tensor, corresponding to lod and padding 
    mode respectively.

4669 4670 4671 4672 4673
    A simple example as below:

    .. code-block:: text

        Given:
S
SunGaofeng 已提交
4674
        (1) for lod mode:
4675 4676 4677 4678 4679 4680 4681 4682 4683 4684 4685

        input.data = [[0.6, 0.1, 0.3, 0.1],
                      [0.3, 0.2, 0.4, 0.1],
                      [0.1, 0.5, 0.1, 0.3],
                      [0.5, 0.1, 0.3, 0.1],

                      [0.5, 0.1, 0.3, 0.1],
                      [0.2, 0.2, 0.2, 0.4],
                      [0.2, 0.2, 0.1, 0.5],
                      [0.5, 0.1, 0.3, 0.1]]

4686
        input.lod = [[4, 4]]
M
minqiyang 已提交
4687

W
whs 已提交
4688
        Computation:
4689

W
whs 已提交
4690 4691 4692 4693 4694 4695
        step1: Apply argmax to first input sequence which is input.data[0:4]. Then we get:
               [[0], [2], [1], [0]]
        step2: merge repeated tokens and remove blank which is 0. Then we get first output sequence:
               [[2], [1]]

        Finally:
4696 4697 4698 4699 4700

        output.data = [[2],
                       [1],
                       [3]]

4701
        output.lod = [[2, 1]]
4702

S
SunGaofeng 已提交
4703
        (2) for padding mode:
4704 4705 4706 4707 4708 4709 4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722 4723 4724 4725 4726 4727 4728 4729

         input.data = [[[0.6, 0.1, 0.3, 0.1],
                        [0.3, 0.2, 0.4, 0.1],
                        [0.1, 0.5, 0.1, 0.3],
                        [0.5, 0.1, 0.3, 0.1]],

                       [[0.5, 0.1, 0.3, 0.1],
                        [0.2, 0.2, 0.2, 0.4],
                        [0.2, 0.2, 0.1, 0.5],
                        [0.5, 0.1, 0.3, 0.1]]]

        input_length.data = [[4], [4]]
        input.shape = [2, 4, 4]

        step1: Apply argmax to first input sequence which is input.data[0:4]. Then we get:
               [[0], [2], [1], [0]], for input.data[4:8] is [[0], [3], [3], [0]], shape is [2,4,1]
        step2: Change the argmax result to use padding mode, then argmax result is 
                [[0, 2, 1, 0], [0, 3, 3, 0]], shape is [2, 4], lod is [], input_length is [[4], [4]]
        step3: Apply ctc_align to padding argmax result, padding_value is 0

        Finally:
        output.data = [[2, 1, 0, 0],
                       [3, 0, 0, 0]]
        output_length.data = [[2], [1]]


S
SunGaofeng 已提交
4730
    Parameters:
4731

S
SunGaofeng 已提交
4732 4733
        input(Variable): the probabilities of variable-length sequences. When in lod mode, 
                         it is a 2-D LoDTensor with LoD information. It's shape is [Lp, num_classes + 1] 
Y
ying 已提交
4734
                         where Lp is the sum of all input sequences' length and
4735 4736
                         num_classes is the true number of classes. When in padding mode,
                         it is a 3-D Tensor with padding, It's shape is [batch_size, N, num_classes + 1].
S
SunGaofeng 已提交
4737
                         (not including the blank label). The data type can be float32 or float64.
Y
ying 已提交
4738
        blank(int): the blank label index of Connectionist Temporal
S
SunGaofeng 已提交
4739
                    Classification (CTC) loss, which is in the half-opened
Y
ying 已提交
4740
                    interval [0, num_classes + 1).
S
SunGaofeng 已提交
4741 4742
        input_length(Variable, optional): 2-D LoDTensor, shape is [batch_size, 1], data type is int64.
                                 It is used for padding mode. In lod mode, input_length is None.
4743
        padding_value(int): padding value.
S
SunGaofeng 已提交
4744 4745 4746
        name(str, optional): The default value is None.  
                             Normally there is no need for user to set this property.  
                             For more information, please refer to :ref:`api_guide_Name` 
4747 4748

    Returns:
S
SunGaofeng 已提交
4749 4750 4751 4752 4753 4754 4755 4756 4757 4758 4759 4760 4761 4762 4763 4764 4765
        For lod mode, returns the result of CTC greedy decoder, 2-D LoDTensor, shape is [Lp, 1], \
        data type is int64. 'Lp' is the sum of all output sequences' length. If all the sequences \
        in result were empty, the result LoDTensor will be [-1] with  empty \
        LoD [[]].

        For padding mode, returns a tuple of (output, output_length), which was describled as below: 

        output, 2-D Tensor, shape is [batch_size, N], data type is int64.

        output_length, 2-D Tensor, shape is [batch_size, 1], data type is int64. It is the length of \
                           each sequence of output for padding mode.

    Return type:
        For lod mode: Variable

        For padding mode: tuple of two Variables (output, output_length).

4766 4767 4768 4769

    Examples:
        .. code-block:: python

4770
            # for lod mode
S
SunGaofeng 已提交
4771
            import paddle.fluid as fluid
S
SunGaofeng 已提交
4772
            x = fluid.data(name='x', shape=[None, 8], dtype='float32', lod_level=1)
4773
            cost = fluid.layers.ctc_greedy_decoder(input=x, blank=0)
4774 4775

            # for padding mode
S
SunGaofeng 已提交
4776 4777
            x_pad = fluid.data(name='x_pad', shape=[10, 4, 8], dtype='float32')
            x_pad_len = fluid.data(name='x_pad_len', shape=[10, 1], dtype='int64')
4778 4779 4780
            out, out_len = fluid.layers.ctc_greedy_decoder(input=x_pad, blank=0,
                            input_length=x_pad_len)

W
wanghaoshuang 已提交
4781
    """
4782
    helper = LayerHelper("ctc_greedy_decoder", **locals())
Q
qingqing01 已提交
4783
    _, topk_indices = topk(input, k=1)
4784 4785

    # ctc align op
X
Xin Pan 已提交
4786
    ctc_out = helper.create_variable_for_type_inference(dtype="int64")
4787 4788 4789 4790 4791 4792 4793 4794 4795 4796 4797 4798 4799 4800 4801 4802 4803 4804 4805 4806 4807 4808 4809 4810 4811

    if input_length is None:
        helper.append_op(
            type="ctc_align",
            inputs={"Input": [topk_indices]},
            outputs={"Output": [ctc_out]},
            attrs={"merge_repeated": True,
                   "blank": blank})
        return ctc_out
    else:
        ctc_out_len = helper.create_variable_for_type_inference(dtype="int64")
        ctc_input = squeeze(topk_indices, [2])

        helper.append_op(
            type="ctc_align",
            inputs={"Input": [ctc_input],
                    "InputLength": [input_length]},
            outputs={"Output": [ctc_out],
                     "OutputLength": [ctc_out_len]},
            attrs={
                "merge_repeated": True,
                "blank": blank,
                "padding_value": padding_value
            })
        return ctc_out, ctc_out_len
4812 4813


Y
fix ci.  
ying 已提交
4814
def transpose(x, perm, name=None):
Y
ying 已提交
4815
    """
4816
    Permute the data dimensions of `input` according to `perm`.
Y
ying 已提交
4817 4818 4819 4820 4821

    The `i`-th dimension  of the returned tensor will correspond to the
    perm[i]-th dimension of `input`.

    Args:
4822 4823
        x (Variable): The input Tensor. It is a N-D Tensor of data types float32, float64, int32.
        perm (list): Permute the input accoring to the data of perm.
4824
        name (str): The name of this layer. It is optional.
Y
ying 已提交
4825 4826

    Returns:
4827 4828 4829 4830 4831 4832 4833 4834 4835 4836 4837 4838 4839 4840 4841 4842 4843 4844 4845 4846 4847 4848 4849 4850
        Variable: A transposed n-D Tensor, with data type being float32, float64, int32, int64.

    For Example:

        .. code-block:: text

         x = [[[ 1  2  3  4] [ 5  6  7  8] [ 9 10 11 12]]
             [[13 14 15 16] [17 18 19 20] [21 22 23 24]]]
         shape(x) =  [2,3,4]

         # Example 1
         perm0 = [1,0,2]
         y_perm0 = [[[ 1  2  3  4] [13 14 15 16]]
                   [[ 5  6  7  8]  [17 18 19 20]]
                   [[ 9 10 11 12]  [21 22 23 24]]]
         shape(y_perm0) = [3,2,4]

         # Example 2
         perm1 = [2,1,0]
         y_perm1 = [[[ 1 13] [ 5 17] [ 9 21]]
                   [[ 2 14] [ 6 18] [10 22]]
                   [[ 3 15]  [ 7 19]  [11 23]]
                   [[ 4 16]  [ 8 20]  [12 24]]]
         shape(y_perm1) = [4,3,2]
Y
ying 已提交
4851 4852

    Examples:
4853

Y
ying 已提交
4854 4855
        .. code-block:: python

4856
            # use append_batch_size=False to avoid prepending extra
4857
            # batch size in shape
4858
            import paddle.fluid as fluid
4859
            x = fluid.layers.data(name='x', shape=[2, 3, 4],
4860
                            dtype='float32', append_batch_size=False)
4861
            x_transposed = fluid.layers.transpose(x, perm=[1, 0, 2])
4862 4863
            print x_transposed.shape
            #(3L, 2L, 4L)
Y
ying 已提交
4864

4865
    """
4866 4867 4868 4869
    check_type_and_dtype(x, 'x', Variable,
                         ['float16', 'float32', 'float64', 'int32', 'int64'],
                         'transpose')
    check_type(perm, 'perm', list, 'transpose')
Y
fix ci.  
ying 已提交
4870
    if len(perm) != len(x.shape):
Y
ying 已提交
4871
        raise ValueError(
4872 4873 4874 4875
            "Input(perm) is the permutation of dimensions of Input(x), "
            "its length should be equal to dimensions of Input(x), "
            "but received dimension of Input(x) is %s, "
            "the length of Input(perm) is %s." % (len(x.shape), len(perm)))
Y
ying 已提交
4876 4877 4878
    for idx, dim in enumerate(perm):
        if dim >= len(x.shape):
            raise ValueError(
4879 4880 4881
                "Each element in Input(perm) should be less than Input(x)'s dimension, "
                "but %d-th element in Input(perm) is %d which exceeds Input(x)'s "
                "dimension %d." % (idx, perm[idx], len(x.shape)))
Y
ying 已提交
4882 4883

    helper = LayerHelper('transpose', **locals())
X
Xin Pan 已提交
4884 4885
    out = helper.create_variable_for_type_inference(x.dtype)
    x_shape = helper.create_variable_for_type_inference(x.dtype)
Y
ying 已提交
4886
    helper.append_op(
4887
        type='transpose2',
Y
fix ci.  
ying 已提交
4888
        inputs={'X': [x]},
4889 4890
        outputs={'Out': [out],
                 'XShape': [x_shape]},
Y
ying 已提交
4891 4892
        attrs={'axis': perm})
    return out
4893 4894


4895 4896 4897 4898 4899 4900 4901
def im2sequence(input,
                filter_size=1,
                stride=1,
                padding=0,
                input_image_size=None,
                out_stride=1,
                name=None):
4902
    """
4903
    Extracts image patches from the input tensor to form a tensor of shape
L
Liufang Sang 已提交
4904 4905 4906
    {input.batch_size * output_height * output_width, filter_size_height *
    filter_size_width * input.channels}. This op use filter to scan images
    and convert these images to sequences. After expanding, the number of time step are
4907 4908
    output_height * output_width for an image, in which output_height and
    output_width are calculated by below equation:
4909 4910 4911

    .. math::

L
Liufang Sang 已提交
4912 4913 4914 4915
        output\_height  = 1 + \
            (padding\_up + padding\_down + input\_height  - filter\_size\_height  + stride\_height - 1) / stride\_height \\\\
        output\_width  = 1 + \
            (padding\_left + padding\_right + input\_width  - filter\_size\_width  + stride\_width - 1) / stride\_width
4916

L
Liufang Sang 已提交
4917
    And the dimension of each time step is filter_size_height * filter_size_width * input.channels.
4918

L
Liufang Sang 已提交
4919 4920
    Parameters:
        input (Variable): The input should be a 4-D Tensor in :math:`NCHW` format. The data type is float32.
W
wanghaoshuang 已提交
4921

L
Liufang Sang 已提交
4922 4923 4924
        filter_size(int32 | List[int32]): The filter size. If filter_size is a List,
            it must contain two integers, :math:`[filter\_size\_height, filter\_size\_width]` .
            Otherwise, the filter size will be a square :math:`[filter\_size, filter\_size]` . Default is 1.
4925

L
Liufang Sang 已提交
4926 4927
        stride(int32 | List[int32]): The stride size. If stride is a List, it must
            contain two integers, :math:`[stride\_height, stride\_width]` . Otherwise, the stride size will be a square :math:`[stride\_size, stride\_size]` . Default is 1.
4928

L
Liufang Sang 已提交
4929 4930 4931 4932 4933 4934 4935
        padding(int32 | List[int32]): The padding size. If padding is a List, it can
            contain four integers like :math:`[padding\_up, padding\_left, padding\_down, padding\_right]` to indicate
            paddings of four direction.  Or it can contain two integers :math:`[padding\_height, padding\_width]` which means
            padding_up = padding_down = padding_height and
            padding_left = padding_right = padding_width. Otherwise, a scalar padding means
            padding_up = padding_down = padding_left = padding_right = padding. 
            Default is 0.
4936

L
Liufang Sang 已提交
4937 4938 4939 4940 4941 4942 4943 4944 4945 4946 4947 4948 4949 4950 4951 4952
        input_image_size(Variable, optional): the input contains image real size.It's dim
            is :math:`[batchsize, 2]` . It is just for batch inference when not None. Default is None.

        out_stride(int32 | List[int32]): The scaling of image through CNN. It is valid only when input_image_size is not None.
            If out_stride is List,  it must contain two intergers,
            :math:`[out\_stride\_height, out\_stride\_W]` . Otherwise,
            the out_stride_height = out_stride_width = out_stride. Default is 1.

        name (str, optional): The default value is None.  Normally there is no need for
                    user to set this property.  For more information, please refer to :ref:`api_guide_Name` .
    
    Returns: 
            The output is a 2-D LoDTensor with shape {input.batch\_size * output\_height * output\_width, \ 
            filter\_size\_height * filter\_size\_width * input.channels}. The data type is float32.

    Return Type: Variable
4953 4954 4955 4956 4957 4958 4959 4960 4961 4962 4963 4964 4965 4966 4967 4968 4969 4970 4971 4972 4973 4974 4975 4976 4977 4978 4979

    Examples:

        .. code-block:: text

            Given:

            x = [[[[ 6.  2.  1.]
                   [ 8.  3.  5.]
                   [ 0.  2.  6.]]

                  [[ 2.  4.  4.]
                   [ 6.  3.  0.]
                   [ 6.  4.  7.]]]

                 [[[ 6.  7.  1.]
                   [ 5.  7.  9.]
                   [ 2.  4.  8.]]

                  [[ 1.  2.  1.]
                   [ 1.  3.  5.]
                   [ 9.  0.  8.]]]]

            x.dims = {2, 2, 3, 3}

            And:

W
wanghaoshuang 已提交
4980 4981 4982
            filter = [2, 2]
            stride = [1, 1]
            padding = [0, 0]
4983 4984 4985 4986 4987 4988 4989 4990 4991 4992 4993 4994

            Then:

            output.data = [[ 6.  2.  8.  3.  2.  4.  6.  3.]
                           [ 2.  1.  3.  5.  4.  4.  3.  0.]
                           [ 8.  3.  0.  2.  6.  3.  6.  4.]
                           [ 3.  5.  2.  6.  3.  0.  4.  7.]
                           [ 6.  7.  5.  7.  1.  2.  1.  3.]
                           [ 7.  1.  7.  9.  2.  1.  3.  5.]
                           [ 5.  7.  2.  4.  1.  3.  9.  0.]
                           [ 7.  9.  4.  8.  3.  5.  0.  8.]]

4995
            output.dims = {8, 8}
4996

4997
            output.lod = [[4, 4]]
4998

T
Tink_Y 已提交
4999
    Examples:
5000 5001 5002

        .. code-block:: python

B
Bai Yifan 已提交
5003
            import paddle.fluid as fluid
L
Liufang Sang 已提交
5004
            data = fluid.data(name='data', shape=[None, 3, 32, 32],
B
Bai Yifan 已提交
5005
                                     dtype='float32')
5006
            output = fluid.layers.im2sequence(
B
Bai Yifan 已提交
5007 5008
                input=data, stride=[1, 1], filter_size=[2, 2])

5009 5010

    """
L
lujun 已提交
5011
    assert not in_dygraph_mode(), (
5012
        "sequence layer is not supported in dygraph mode yet.")
W
wanghaoshuang 已提交
5013 5014 5015 5016 5017 5018 5019 5020 5021 5022

    if isinstance(filter_size, int):
        filter_size = [filter_size, filter_size]
    if isinstance(stride, int):
        stride = [stride, stride]
    if isinstance(padding, int):
        padding = [padding, padding]
    if len(padding) == 2:
        padding.append(padding[0])
        padding.append(padding[1])
5023
    inputs = {"X": input}
5024
    attrs = {"kernels": filter_size, "strides": stride, "paddings": padding}
5025 5026 5027 5028 5029
    if input_image_size:
        if isinstance(out_stride, int):
            out_stride = [out_stride, out_stride]
        inputs["Y"] = input_image_size
        attrs["out_stride"] = out_stride
5030
    helper = LayerHelper('im2sequence', **locals())
X
Xin Pan 已提交
5031
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
5032
    helper.append_op(
5033
        type='im2sequence', inputs=inputs, outputs={'Out': out}, attrs=attrs)
5034
    return out
5035 5036


Y
yuyang18 已提交
5037
@templatedoc()
5038
def row_conv(input, future_context_size, param_attr=None, act=None):
Y
yuyang18 已提交
5039 5040
    """
    ${comment}
5041 5042

    Args:
Y
yuyang18 已提交
5043
        input (${x_type}): ${x_comment}.
Y
yangyaming 已提交
5044 5045
        future_context_size (int): Future context size. Please note, the shape
            of convolution kernel is [future_context_size + 1, D].
5046 5047 5048 5049 5050
        param_attr (ParamAttr): Attributes of parameters, including
            name, initializer etc.
        act (str): Non-linear activation to be applied to output variable.

    Returns:
Y
yuyang18 已提交
5051
        ${out_comment}.
5052 5053

    Examples:
D
Double_V 已提交
5054
        >>>  # for LodTensor inputs
Y
yuyang18 已提交
5055
        >>> import paddle.fluid as fluid
D
Double_V 已提交
5056
        >>> x = fluid.data(name='x', shape=[9, 16],
Y
yuyang18 已提交
5057 5058
        >>>                        dtype='float32', lod_level=1)
        >>> out = fluid.layers.row_conv(input=x, future_context_size=2)
D
Double_V 已提交
5059 5060 5061
        >>> # for Tensor inputs
        >>> x = fluid.data(name='x', shape=[9, 4, 16], dtype='float32')
        >>> out = fluid.layers.row_conv(input=x, future_context_size=2)
5062 5063 5064 5065 5066 5067
    """
    helper = LayerHelper('row_conv', **locals())
    dtype = helper.input_dtype()
    filter_shape = [future_context_size + 1, input.shape[1]]
    filter_param = helper.create_parameter(
        attr=helper.param_attr, shape=filter_shape, dtype=dtype)
X
Xin Pan 已提交
5068
    out = helper.create_variable_for_type_inference(dtype)
5069 5070 5071 5072 5073
    helper.append_op(
        type='row_conv',
        inputs={'X': [input],
                'Filter': [filter_param]},
        outputs={'Out': [out]})
Y
yangyaming 已提交
5074
    return helper.append_activation(out)
5075 5076


Y
yuyang18 已提交
5077
@templatedoc()
5078 5079
def multiplex(inputs, index):
    """
Y
yuyang18 已提交
5080

5081
    Based on the given index parameter, the OP selects a specific row from each input Tensor to construct the output Tensor.
L
lujun 已提交
5082

5083
    If the input of this OP contains :math:`m` Tensors, where :math:`I_{i}` means the i-th input Tensor, :math:`i` between :math:`[0,m)` .
L
lujun 已提交
5084

5085
    And :math:`O` means the output, where :math:`O[i]` means the i-th row of the output, then the output satisfies that :math:`O[i] = I_{index[i]}[i]` .
L
lujun 已提交
5086

5087
    For Example:
L
lujun 已提交
5088

5089
            .. code-block:: text
L
lujun 已提交
5090

5091
                Given:
L
lujun 已提交
5092

5093 5094 5095 5096
                inputs = [[[0,0,3,4], [0,1,3,4], [0,2,4,4], [0,3,3,4]],
                          [[1,0,3,4], [1,1,7,8], [1,2,4,2], [1,3,3,4]],
                          [[2,0,3,4], [2,1,7,8], [2,2,4,2], [2,3,3,4]],
                          [[3,0,3,4], [3,1,7,8], [3,2,4,2], [3,3,3,4]]]
L
lujun 已提交
5097

5098
                index = [[3],[0],[1],[2]]
L
lujun 已提交
5099

5100 5101 5102 5103
                out = [[3,0,3,4],    # out[0] = inputs[index[0]][0] = inputs[3][0] = [3,0,3,4]
                       [0,1,3,4],    # out[1] = inputs[index[1]][1] = inputs[0][1] = [0,1,3,4]
                       [1,2,4,2],    # out[2] = inputs[index[2]][2] = inputs[1][2] = [1,2,4,2]
                       [2,3,3,4]]    # out[3] = inputs[index[3]][3] = inputs[2][3] = [2,3,3,4]
L
lujun 已提交
5104 5105


5106 5107 5108
    Args:
       inputs (list): The input Tensor list. The list elements are N-D Tensors of data types float32, float64, int32, int64. All input Tensor shapes should be the same and rank must be at least 2.
       index (Variable): Used to select some rows in the input Tensor to construct an index of the output Tensor. It is a 2-D Tensor with data type int32 or int64 and shape [M, 1], where M is the number of input Tensors.
L
lujun 已提交
5109

5110
    Returns:
5111
        Variable(Tensor): Output of multiplex OP, with data type being float32, float64, int32, int64.
X
xuezhong 已提交
5112 5113

    Examples:
5114

X
xuezhong 已提交
5115 5116
        .. code-block:: python

5117
            import paddle.fluid as fluid
5118
            import numpy as np
5119

5120 5121 5122 5123
            x1 = fluid.data(name='x1', shape=[None, 2], dtype='float32')
            x2 = fluid.data(name='x2', shape=[None, 2], dtype='float32')
            index = fluid.data(name='index', shape=[None, 1], dtype='int32')
            out = fluid.layers.multiplex(inputs=[x1, x2], index=index)
X
xuezhong 已提交
5124

5125 5126 5127 5128 5129 5130 5131 5132 5133
            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())

            img1 = np.array([[1, 2], [3, 4]]).astype(np.float32)
            img2 = np.array([[5, 6], [7, 8]]).astype(np.float32)
            index = np.array([[1], [0]]).astype(np.int32)

            res = exe.run(fluid.default_main_program(), feed={'x1':img1, 'x2':img2, 'index':index}, fetch_list=[out])
            print(res) # [array([[5., 6.], [3., 4.]], dtype=float32)]
X
xuezhong 已提交
5134

5135 5136 5137 5138 5139 5140 5141 5142
    """
    helper = LayerHelper('multiplex', **locals())

    if not isinstance(inputs, list) and len(inputs) < 2:
        raise ValueError("inputs should be a list object and contains at least "
                         "2 elements.")

    out = helper.create_variable_for_type_inference(inputs[0].dtype)
5143
    helper.append_op(
5144 5145 5146 5147 5148
        type='multiplex',
        inputs={'X': inputs,
                'Ids': index},
        outputs={'Out': [out]})
    return out
X
xuezhong 已提交
5149 5150


5151 5152
def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None):
    """
Y
Yibing Liu 已提交
5153 5154
    This layer computes the smooth L1 loss for Variable :attr:`x` and :attr:`y`.
    It takes the first dimension of :attr:`x` and :attr:`y` as batch size.
Q
qingqing01 已提交
5155
    For each instance, it computes the smooth L1 loss element by element first
5156
    and then sums all the losses. So the shape of ouput Variable is
5157
    [batch_size, 1].
5158

5159 5160
    Args:
        x (Variable): A tensor with rank at least 2. The input value of smooth
Q
qingqing01 已提交
5161
            L1 loss op with shape [batch_size, dim1, ..., dimN].
5162
            A LoDTensor or Tensor with type float32.
5163
        y (Variable): A tensor with rank at least 2. The target value of smooth
Y
Yibing Liu 已提交
5164
            L1 loss op with same shape as :attr:`x`.
5165
            A LoDTensor or Tensor with type float32.
5166
        inside_weight (Variable|None):  A tensor with rank at least 2. This
5167 5168
            input is optional and should have same shape with :attr:`x`. If
            provided, the result of (:attr:`x` - :attr:`y`) will be multiplied
Y
Yibing Liu 已提交
5169
            by this tensor element by element.
5170
            A Tensor with type float32.
5171
        outside_weight (Variable|None): A tensor with rank at least 2. This
5172 5173
            input is optional and should have same shape with :attr:`x`. If
            provided, the out smooth L1 loss will be multiplied by this tensor
Y
Yibing Liu 已提交
5174
            element by element.
5175
            A Tensor with type float32.
5176
        sigma (float|None): Hyper parameter of smooth L1 loss layer. A float
5177 5178
           scalar with default value 1.0.

5179
    Returns:
5180
        Variable: The output smooth L1 loss with shape [batch_size, 1].  A Tensor with type float32.
5181 5182 5183 5184

    Examples:
        .. code-block:: python

5185
            import paddle.fluid as fluid
5186 5187 5188 5189 5190 5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201 5202
            import numpy as np
            data = fluid.data(name="x", shape=[-1, 3], dtype="float32")
            label = fluid.data(name="y", shape=[-1, 3], dtype="float32")
            result = fluid.layers.smooth_l1(data,label)
            place = fluid.CPUPlace()
            exe = fluid.Executor(place)
            exe.run(fluid.default_startup_program())
            x = np.random.rand(3,3).astype("float32")
            y = np.random.rand(3,3).astype("float32")
            output= exe.run(feed={"x":x, "y":y},
                             fetch_list=[result])
            print(output)
        
            #[array([[0.08220536],
            #       [0.36652038],
            #      [0.20541131]], dtype=float32)]

5203
    """
5204

5205
    helper = LayerHelper('smooth_l1_loss', **locals())
X
Xin Pan 已提交
5206 5207
    diff = helper.create_variable_for_type_inference(dtype=x.dtype)
    loss = helper.create_variable_for_type_inference(dtype=x.dtype)
5208 5209 5210 5211 5212 5213 5214 5215 5216 5217
    helper.append_op(
        type='smooth_l1_loss',
        inputs={
            'X': x,
            'Y': y,
            'InsideWeight': inside_weight,
            'OutsideWeight': outside_weight
        },
        outputs={'Diff': diff,
                 'Out': loss},
5218
        attrs={'sigma': sigma if sigma is not None else 1.0})
5219
    return loss
5220 5221


5222
def one_hot(input, depth, allow_out_of_range=False):
5223
    """
5224 5225 5226 5227 5228 5229 5230 5231 5232 5233 5234 5235 5236 5237 5238 5239 5240 5241 5242 5243 5244 5245 5246 5247 5248 5249 5250 5251 5252 5253 5254 5255 5256 5257 5258 5259 5260 5261 5262 5263 5264 5265 5266 5267 5268 5269 5270 5271 5272 5273 5274 5275 5276 5277

    **WARING:** This OP requires the last dimension of Tensor shape must be equal to 1.
    This OP will be deprecated in a future release. It is recommended to use fluid. :ref:`api_fluid_one_hot` .

    The operator converts each id in the input to an one-hot vector with a
    :attr:`depth` length. The value in the vector dimension corresponding to the id
    is 1, and the value in the remaining dimension is 0.

    The shape of output Tensor or LoDTensor is generated by adding :attr:`depth` dimension
    behind the last dimension of the input shape.

    .. code-block:: text

        Example 1 (allow_out_of_range=False):

        input:
            X.shape = [4, 1]
            X.data = [[1], [1], [3], [0]]
            depth = 4

        output:
            Out.shape = [4, 4]
            Out.data = [[0., 1., 0., 0.],
                        [0., 1., 0., 0.],
                        [0., 0., 0., 1.],
                        [1., 0., 0., 0.]]

        Example 2 (allow_out_of_range=True):

        input:
            X.shape = [4, 1]
            X.data = [[1], [1], [5], [0]]
            depth = 4
            allow_out_of_range = True

        output:
            Out.shape = [4, 4]
            Out.data = [[0., 1., 0., 0.],
                        [0., 1., 0., 0.], 
                        [0., 0., 0., 0.], # This id is 5, which goes beyond depth, so set it all-zeros data.
                        [1., 0., 0., 0.]]

        Example 3 (allow_out_of_range=False):

        input:
            X.shape = [4, 1]
            X.data = [[1], [1], [5], [0]]
            depth = 4
            allow_out_of_range = False

        output: Throw an exception for Illegal value
            The second dimension in X is 5, which is greater than depth.  
            Allow_out_of_range =False means that does not allow the word id to exceed depth,
            so it throws an exception.
5278 5279

    Args:
5280 5281 5282 5283 5284
        input(Variable): Tensor or LoDTensor with shape :math:`[N_1, N_2, ..., N_k, 1]` ,
            which contains at least one dimension and the last dimension must be 1.
            The data type is int32 or int64.
        depth(scalar): An integer defining the :attr:`depth` of the one hot dimension. If input 
            is word id, depth is generally the dictionary size.
5285
        allow_out_of_range(bool): A bool value indicating whether the input
5286 5287 5288 5289
            indices could be out of range :math:`[0, depth)` . When input indices are
            out of range, exceptions :code:`Illegal value` is raised if :attr:`allow_out_of_range`
            is False, or zero-filling representations is created if it is set True.
            Default: False.
5290 5291

    Returns:
5292
        Variable: The one-hot representations of input. A Tensor or LoDTensor with type float32.
5293 5294

    Examples:
C
caoying03 已提交
5295
        .. code-block:: python
5296

5297
            import paddle.fluid as fluid
5298 5299 5300
            # Correspond to the first example above, where label.shape is [4, 1] and one_hot_label.shape is [4, 4].
            label = fluid.data(name="label", shape=[4, 1], dtype="int64")
            one_hot_label = fluid.layers.one_hot(input=label, depth=4)
5301 5302
    """
    helper = LayerHelper("one_hot", **locals())
5303

X
Xin Pan 已提交
5304
    one_hot_out = helper.create_variable_for_type_inference(dtype='float32')
5305 5306 5307

    if in_dygraph_mode():
        inputs = {'X': input}
Y
Yi Liu 已提交
5308
        attrs = {'depth': depth, 'allow_out_of_range': allow_out_of_range}
5309 5310
    else:
        if not isinstance(depth, Variable):
G
Guo Sheng 已提交
5311
            # user attribute
5312
            inputs = {'X': input}
Y
Yi Liu 已提交
5313
            attrs = {'depth': depth, 'allow_out_of_range': allow_out_of_range}
5314
        else:
H
Hongyu Liu 已提交
5315
            depth.stop_gradient = True
5316
            inputs = {'X': input, 'depth_tensor': depth}
Y
Yi Liu 已提交
5317
            attrs = {'allow_out_of_range': allow_out_of_range}
5318 5319
    helper.append_op(
        type="one_hot",
5320 5321
        inputs=inputs,
        attrs=attrs,
5322 5323
        outputs={'Out': one_hot_out})
    one_hot_out.stop_gradient = True
5324
    return one_hot_out
Y
Yu Yang 已提交
5325 5326


Y
Yu Yang 已提交
5327
def autoincreased_step_counter(counter_name=None, begin=1, step=1):
Y
Yu Yang 已提交
5328
    """
Y
Yibing Liu 已提交
5329 5330 5331
    Create an auto-increase variable. which will be automatically increased 
    by 1 in every iteration. By default, the first return of this counter is 1, 
    and the step size is 1.
Y
Yu Yang 已提交
5332 5333

    Args:
Y
Yibing Liu 已提交
5334 5335 5336
        counter_name(str, optional): The counter name. Default '@STEP_COUNTER@'.
        begin(int, optional): The first return value of this counter. Default 1.
        step(int, optional): The step size. Default 1.
Y
Yu Yang 已提交
5337

5338
    Returns:
Y
Yibing Liu 已提交
5339
        Variable: The auto-increased Variable with data type int64.
Y
yi.wu 已提交
5340 5341 5342 5343

    Examples:
        .. code-block:: python

5344
           import paddle.fluid as fluid
Y
yi.wu 已提交
5345
           global_step = fluid.layers.autoincreased_step_counter(
Y
Yibing Liu 已提交
5346
               counter_name='@LR_DECAY_COUNTER@', begin=0, step=1)
Y
Yu Yang 已提交
5347 5348
    """
    helper = LayerHelper('global_step_counter')
Y
Yu Yang 已提交
5349 5350
    if counter_name is None:
        counter_name = '@STEP_COUNTER@'
Y
Yu Yang 已提交
5351
    counter, is_new_var = helper.create_or_get_global_variable(
H
hong 已提交
5352 5353 5354 5355 5356
        name=counter_name,
        dtype='int64',
        shape=[1],
        persistable=True,
        belong_to_optimizer=True)
Y
Yu Yang 已提交
5357 5358 5359
    if is_new_var:
        helper.set_variable_initializer(
            counter, initializer=Constant(
Y
Yu Yang 已提交
5360
                value=begin - 1, force_cpu=True))
W
Wu Yi 已提交
5361
        helper.main_program.global_block()._prepend_op(
Y
Yu Yang 已提交
5362 5363
            type='increment',
            inputs={'X': [counter]},
Y
Yu Yang 已提交
5364
            outputs={'Out': [counter]},
5365
            attrs={'step': float(step)})
Y
Yu Yang 已提交
5366 5367 5368
        counter.stop_gradient = True

    return counter
Y
yangyaming 已提交
5369 5370


5371
def reshape(x, shape, actual_shape=None, act=None, inplace=False, name=None):
C
caoying03 已提交
5372
    """
5373
    This operator changes the shape of ``x`` without changing its data.
C
caoying03 已提交
5374

5375 5376 5377 5378
    The target shape can be given by ``shape`` or ``actual_shape``.
    When ``shape`` and ``actual_shape`` are set at the same time,
    ``actual_shape`` has a higher priority than ``shape``
    but at this time ``shape`` can only be an integer list or tuple, and ``shape`` still should be set correctly to
5379
    gurantee shape inference in compile-time.
C
caoying03 已提交
5380

5381
    Some tricks exist when specifying the target shape.
C
caoying03 已提交
5382

5383 5384 5385 5386
    1. -1 means the value of this dimension is inferred from the total element
    number of x and remaining dimensions. Thus one and only one dimension can
    be set -1.

5387
    2. 0 means the actual dimension value is going to be copied from the
5388
    corresponding dimension of x. The indice of 0s in shape can not exceed
5389
    the dimension of x.
5390 5391

    Here are some examples to explain it.
C
caoying03 已提交
5392 5393

    1. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
W
wanghaoshuang 已提交
5394
    is [6, 8], the reshape operator will transform x into a 2-D tensor with
5395
    shape [6, 8] and leaving x's data unchanged.
C
caoying03 已提交
5396

5397
    2. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
5398 5399
    specified is [2, 3, -1, 2], the reshape operator will transform x into a
    4-D tensor with shape [2, 3, 4, 2] and leaving x's data unchanged. In this
W
wanghaoshuang 已提交
5400 5401
    case, one dimension of the target shape is set to -1, the value of this
    dimension is inferred from the total element number of x and remaining
5402
    dimensions.
C
caoying03 已提交
5403

5404
    3. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
5405 5406 5407 5408
    is [-1, 0, 3, 2], the reshape operator will transform x into a 4-D tensor
    with shape [2, 4, 3, 2] and leaving x's data unchanged. In this case,
    besides -1, 0 means the actual dimension value is going to be copied from
    the corresponding dimension of x.
C
caoying03 已提交
5409

5410 5411
    **Note**:
        The parameter ``actual_shape`` will be deprecated in the future and only use ``shape`` instead to represent the target shape.
5412

C
caoying03 已提交
5413
    Args:
5414 5415 5416 5417 5418 5419 5420 5421 5422 5423 5424 5425 5426 5427 5428 5429 5430
        x(Variable): A ``Tensor`` or ``LoDTensor`` . The data type is ``float32``, ``float64``, ``int32`` or ``int64``.
        shape(list|tuple|Variable): Define the target shape. At most one dimension of the target shape can be -1.
                        The data type is ``int32`` . If ``shape`` is a list or tuple, the elements of it should be integers or Tensors with shape [1].
                        If ``shape`` is an Variable, it should be an 1-D Tensor .
        actual_shape(variable, optional): An 1-D ``Tensor`` or ``LoDTensor`` . The data type is ``int32`` . If provided, reshape
                                according to this given shape rather than ``shape`` specifying shape.
                                That is to say ``actual_shape`` has a higher priority
                                than ``shape(list|tuple)`` but not ``shape(Variable)``. \
                                This argument ``actual_shape`` will be removed in a future version. \
                                Instructions for updating: ``actual_shape`` will be removed in future versions and replaced by ``shape``.
        act (str, optional): The non-linear activation to be applied to the reshaped input. Default None.
        inplace(bool, optional): If ``inplace`` is True, the input and output of ``layers.reshape``
                       are the same variable. Otherwise, the input and output of
                       ``layers.reshape`` are different variable. Default False. Note that if ``x``
                       is more than one OPs' input, ``inplace`` must be False.
        name(str, optional): The default value is None. Normally there is no need for user to set this property.
                            For more information, please refer to :ref:`api_guide_Name` .
C
caoying03 已提交
5431

5432
    Returns:
5433
        Variable: A ``Tensor`` or ``LoDTensor``. The data type is same as ``x``. It is a new tensor variable if ``inplace`` is ``False``, otherwise it is ``x``. If ``act`` is None, return the reshaped tensor variable, otherwise return the activated tensor variable.
C
caoying03 已提交
5434

X
Xin Pan 已提交
5435
    Raises:
5436 5437 5438 5439
        TypeError: If actual_shape is neither Variable nor None.
        ValueError: If more than one elements of ``shape`` is -1.
        ValueError: If the element of ``shape`` is 0, the corresponding dimension should be less than or equal to the dimension of ``x``.
        ValueError: If the elements in ``shape`` is negative except -1.
X
Xin Pan 已提交
5440

C
caoying03 已提交
5441 5442
    Examples:
        .. code-block:: python
G
guosheng 已提交
5443

5444
            import paddle.fluid as fluid
5445 5446 5447

            # example 1:
            # attr shape is a list which doesn't contain tensor Variable.
5448 5449
            data_1 = fluid.data(
              name='data_1', shape=[2, 4, 6], dtype='float32')
5450
            reshaped_1 = fluid.layers.reshape(
5451 5452
              x=data_1, shape=[-1, 0, 3, 2], inplace=True)
            # the shape of reshaped_1 is [2,4,3,2].
5453 5454 5455 5456 5457 5458

            # example 2:
            # attr shape is a list which contains tensor Variable.
            data_2 = fluid.layers.fill_constant([2,25], "int32", 3)
            dim = fluid.layers.fill_constant([1], "int32", 5)
            reshaped_2 = fluid.layers.reshape(data_2, shape=[dim, 10])
5459
            # the shape of reshaped_2 is [5,10].
C
caoying03 已提交
5460
    """
5461 5462 5463 5464 5465
    check_type_and_dtype(x, 'x', Variable,
                         ['float16', 'float32', 'float64', 'int32', 'int64'],
                         'reshape')
    check_type(shape, 'shape', (list, tuple, Variable), 'reshape')
    check_type(actual_shape, 'actual_shape', (Variable, type(None)), 'reshape')
5466
    helper = LayerHelper("reshape2", **locals())
5467 5468 5469 5470 5471 5472 5473 5474 5475 5476 5477 5478 5479 5480 5481 5482 5483 5484 5485 5486 5487 5488 5489 5490 5491 5492 5493 5494 5495 5496 5497 5498
    inputs = {"X": x}
    attrs = {}

    def contain_var(one_list):
        for ele in one_list:
            if isinstance(ele, Variable):
                return True
        return False

    def get_new_shape_tensor(list_shape):
        new_shape_tensor = []
        for dim in list_shape:
            if isinstance(dim, Variable):
                dim.stop_gradient = True
                new_shape_tensor.append(dim)
            else:
                assert (isinstance(dim, int))
                temp_out = helper.create_variable_for_type_inference('int32')
                fill_constant([1], 'int32', dim, force_cpu=True, out=temp_out)
                new_shape_tensor.append(temp_out)
        return new_shape_tensor

    def get_attr_shape(list_shape):
        unk_dim_idx = -1
        attrs_shape = []
        for dim_idx, dim_size in enumerate(list_shape):
            if isinstance(dim_size, Variable):
                attrs_shape.append(-1)
            else:
                attrs_shape.append(dim_size)
                if dim_size == -1:
                    assert unk_dim_idx == -1, (
5499 5500
                        "Only one dimension value of 'shape' in reshape can "
                        "be -1. But received shape[%d] is also -1." % dim_idx)
5501 5502 5503
                    unk_dim_idx = dim_idx
                elif dim_size == 0:
                    assert dim_idx < len(x.shape), (
5504 5505 5506 5507
                        "The index of 0 in `shape` must be less than "
                        "the input tensor X's dimensions. "
                        "But received shape[%d] = 0, X's dimensions = %d." %
                        (dim_idx, len(x.shape)))
5508 5509
                else:
                    assert dim_size > 0, (
5510 5511 5512 5513
                        "Each dimension value of 'shape' in reshape must not "
                        "be negtive except one unknown dimension. "
                        "But received shape[%d] = %s." %
                        (dim_idx, str(dim_size)))
5514 5515
        return attrs_shape

5516 5517 5518 5519
    if in_dygraph_mode():
        inputs = {'X': x}
        attrs = {'shape': shape}
    else:
5520 5521 5522 5523 5524
        if isinstance(shape, Variable):
            shape.stop_gradient = True
            inputs["Shape"] = shape
        elif isinstance(shape, (list, tuple)):
            assert len(shape) > 0, (
5525 5526
                "The size of 'shape' in reshape can't be zero, "
                "but received %s." % len(shape))
5527 5528 5529 5530 5531 5532
            attrs["shape"] = get_attr_shape(shape)
            if contain_var(shape):
                inputs['ShapeTensor'] = get_new_shape_tensor(shape)
            elif isinstance(actual_shape, Variable):
                actual_shape.stop_gradient = True
                inputs["Shape"] = actual_shape
5533

5534 5535
    out = x if inplace and not in_dygraph_mode(
    ) else helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
5536
    x_shape = helper.create_variable_for_type_inference(dtype=x.dtype)
C
caoying03 已提交
5537
    helper.append_op(
5538
        type="reshape2",
X
Xin Pan 已提交
5539
        inputs=inputs,
5540
        attrs=attrs,
5541 5542
        outputs={"Out": out,
                 "XShape": x_shape})
C
caoying03 已提交
5543

D
dzhwinter 已提交
5544
    return helper.append_activation(out)
5545

5546

5547
def squeeze(input, axes, name=None):
Y
Yibing Liu 已提交
5548
    """
5549 5550 5551
    This OP will squeeze single-dimensional entries of input tensor's shape. If axes is provided, will
    remove the dims by axes, the dims selected by axes should be one. If not provide axes, all dims equal
    to one will be deleted.
M
minqiyang 已提交
5552

H
haowang101779990 已提交
5553

5554
    .. code-block:: text 
H
haowang101779990 已提交
5555

5556
        Case1:
H
haowang101779990 已提交
5557

5558
          Input:
H
haowang101779990 已提交
5559 5560
            X.shape = (1, 3, 1, 5)
            axes = [0]
5561
          Output:
H
haowang101779990 已提交
5562 5563
            Out.shape = (3, 1, 5)

5564
        Case2:
H
haowang101779990 已提交
5565

5566
          Input:
H
haowang101779990 已提交
5567 5568
            X.shape = (1, 3, 1, 5)
            axes = []
5569
          Output:
H
haowang101779990 已提交
5570
            Out.shape = (3, 5)
M
minqiyang 已提交
5571

5572 5573 5574 5575 5576 5577 5578 5579
        Case3:

          Input:
            X.shape = [1,3,1,5]
            axes = [-2]
          Output:
            Out.shape = [1,3,5]

Y
Yibing Liu 已提交
5580
    Args:
5581 5582 5583 5584 5585
        input (Variable): The input Tensor. Support data type: float32, float64, int8, int32, int64.
                          axes (list): One integer or List of integers, indicating the dimensions to be squeezed.
                          Axes range is :math:`[-rank(input), rank(input))`.
                          If axes is negative, :math:`axes=axes+rank(input)`.
        name (str, optional): Please refer to :ref:`api_guide_Name`, Default None.
Y
Yibing Liu 已提交
5586 5587

    Returns:
5588
        Variable: Output squeezed Tensor. Data type is same as input Tensor.
Y
Yibing Liu 已提交
5589 5590 5591 5592

    Examples:
        .. code-block:: python

5593
            import paddle.fluid as fluid
5594
            import paddle.fluid.layers as layers
5595 5596 5597 5598
            # set batch size=None
            x = fluid.data(name='x', shape=[None, 5, 1, 10])
            y = layers.squeeze(input=x, axes=[2]) # y.shape=[None, 5, 10]

Y
Yibing Liu 已提交
5599 5600
    """
    helper = LayerHelper("squeeze", **locals())
5601 5602 5603 5604
    check_type_and_dtype(input, 'input', Variable,
                         ['float32', 'float64', 'int8', 'int32', 'int64'],
                         'squeeze')
    check_type(axes, 'axes', list, 'squeeze')
X
Xin Pan 已提交
5605 5606
    out = helper.create_variable_for_type_inference(dtype=input.dtype)
    x_shape = helper.create_variable_for_type_inference(dtype=input.dtype)
Y
Yibing Liu 已提交
5607
    helper.append_op(
5608
        type="squeeze2",
5609
        inputs={"X": input},
Y
Yibing Liu 已提交
5610
        attrs={"axes": axes},
5611 5612
        outputs={"Out": out,
                 "XShape": x_shape})
Y
Yibing Liu 已提交
5613

5614 5615 5616
    return out


5617
def unsqueeze(input, axes, name=None):
Y
Yibing Liu 已提交
5618
    """
5619
    Insert single-dimensional entries to the shape of a Tensor. Takes one
M
minqiyang 已提交
5620 5621
    required argument axes, a list of dimensions that will be inserted.
    Dimension indices in axes are as seen in the output tensor.
Y
Yibing Liu 已提交
5622

M
minqiyang 已提交
5623
    For example:
H
haowang101779990 已提交
5624 5625 5626

    .. code-block:: text

M
minqiyang 已提交
5627
      Given a tensor such that tensor with shape [3, 4, 5],
Y
Yibing Liu 已提交
5628
      then Unsqueezed tensor with axes=[0, 4] has shape [1, 3, 4, 5, 1].
M
minqiyang 已提交
5629

Y
Yibing Liu 已提交
5630
    Args:
5631
        input (Variable): The input Tensor to be unsqueezed. It is a N-D Tensor of data types float32, float64, int32.
5632
        axes (int|list|tuple|Variable): Indicates the dimensions to be inserted. The data type is ``int32`` . If ``axes`` is a list or tuple, the elements of it should be integers or Tensors with shape [1]. If ``axes`` is an Variable, it should be an 1-D Tensor .
5633
        name (str|None): Name for this layer.
Y
Yibing Liu 已提交
5634 5635

    Returns:
5636
        Variable: Output unsqueezed Tensor, with data type being float32, float64, int32, int64.
Y
Yibing Liu 已提交
5637 5638 5639 5640

    Examples:
        .. code-block:: python

5641 5642 5643
            import paddle.fluid as fluid
            x = fluid.layers.data(name='x', shape=[5, 10])
            y = fluid.layers.unsqueeze(input=x, axes=[1])
5644

Y
Yibing Liu 已提交
5645
    """
5646 5647 5648 5649 5650 5651 5652 5653 5654 5655 5656 5657 5658 5659 5660 5661 5662 5663 5664 5665 5666 5667 5668 5669 5670 5671 5672 5673 5674 5675 5676 5677 5678
    if not isinstance(axes, (int, list, tuple, Variable)):
        raise TypeError(
            "The type of 'axes' in unsqueeze must be int, list, tuple or Variable, but "
            "received %s." % (type(axes)))
    helper = LayerHelper("unsqueeze2", **locals())
    inputs = {"X": input}
    attrs = {}

    def _to_Variable_list(one_list):
        Variable_list = []
        for ele in one_list:
            if isinstance(ele, Variable):
                ele.stop_gradient = True
                Variable_list.append(ele)
            else:
                assert (isinstance(ele, int))
                temp_out = helper.create_variable_for_type_inference('int32')
                fill_constant([1], 'int32', ele, force_cpu=True, out=temp_out)
                Variable_list.append(temp_out)
        return Variable_list

    if isinstance(axes, int):
        axes = [axes]
    if isinstance(axes, Variable):
        axes.stop_gradient = True
        inputs["AxesTensor"] = axes
    elif isinstance(axes, (list, tuple)):
        contain_var = not all(not isinstance(ele, Variable) for ele in axes)
        if contain_var:
            inputs["AxesTensorList"] = _to_Variable_list(axes)
        else:
            attrs["axes"] = axes

X
Xin Pan 已提交
5679 5680
    out = helper.create_variable_for_type_inference(dtype=input.dtype)
    x_shape = helper.create_variable_for_type_inference(dtype=input.dtype)
Y
Yibing Liu 已提交
5681
    helper.append_op(
5682
        type="unsqueeze2",
5683 5684
        inputs=inputs,
        attrs=attrs,
5685 5686
        outputs={"Out": out,
                 "XShape": x_shape})
Y
Yibing Liu 已提交
5687

5688 5689
    return out

5690

Y
yangyaming 已提交
5691
def lod_reset(x, y=None, target_lod=None):
Y
yangyaming 已提交
5692
    """
Y
Yibing Liu 已提交
5693
    Set LoD of :attr:`x` to a new one specified by :attr:`y` or
5694 5695 5696 5697
    :attr:`target_lod`. When :attr:`y` provided, :attr:`y.lod` would be
    considered as target LoD first, otherwise :attr:`y.data` would be
    considered as target LoD. If :attr:`y` is not provided, target LoD should
    be specified by :attr:`target_lod`. If target LoD is specified by
5698
    :attr:`y.data` or :attr:`target_lod`, only one level LoD is supported.
Y
yangyaming 已提交
5699 5700 5701 5702 5703 5704

    .. code-block:: text

        * Example 1:

            Given a 1-level LoDTensor x:
5705
                x.lod =  [[ 2,           3,                   1 ]]
Y
yangyaming 已提交
5706 5707 5708
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

5709
            target_lod: [4, 2]
Y
yangyaming 已提交
5710 5711

            then we get a 1-level LoDTensor:
5712
                out.lod =  [[4,                          2]]
Y
yangyaming 已提交
5713 5714 5715 5716 5717 5718
                out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                out.dims = [6, 1]

        * Example 2:

            Given a 1-level LoDTensor x:
5719
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
5720 5721 5722 5723
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a Tensor:
5724
                y.data = [[2, 4]]
Y
yangyaming 已提交
5725 5726 5727
                y.dims = [1, 3]

            then we get a 1-level LoDTensor:
5728
                out.lod =  [[2,            4]]
Y
yangyaming 已提交
5729 5730 5731 5732 5733 5734
                out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                out.dims = [6, 1]

        * Example 3:

            Given a 1-level LoDTensor x:
5735
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
5736 5737 5738 5739
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a 2-level LoDTensor:
5740
                y.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
5741 5742 5743 5744
                y.data = [[1.1], [2.1], [3.1], [4.1], [5.1], [6.1]]
                y.dims = [6, 1]

            then we get a 2-level LoDTensor:
5745
                out.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
5746 5747 5748 5749
                out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                out.dims = [6, 1]

    Args:
5750
        x (Variable): Input variable which could be a Tensor or LoDTensor.
5751
        y (Variable|None): If provided, output's LoD would be derived
Y
Yibing Liu 已提交
5752
                           from :attr:`y`.
Y
yangyaming 已提交
5753
        target_lod (list|tuple|None): One level LoD which should be considered
Y
Yibing Liu 已提交
5754
                                      as target LoD when :attr:`y` not provided.
Y
yangyaming 已提交
5755 5756

    Returns:
Y
Yibing Liu 已提交
5757
        Variable: Output variable with LoD specified by this layer.
Y
yangyaming 已提交
5758 5759

    Raises:
Y
Yibing Liu 已提交
5760
        ValueError: If :attr:`y` and :attr:`target_lod` are both None.
Y
yangyaming 已提交
5761 5762 5763 5764

    Examples:
        .. code-block:: python

5765
            import paddle.fluid as fluid
5766 5767 5768
            x = fluid.layers.data(name='x', shape=[10])
            y = fluid.layers.data(name='y', shape=[10, 20], lod_level=2)
            out = fluid.layers.lod_reset(x=x, y=y)
Y
yangyaming 已提交
5769 5770
    """
    helper = LayerHelper("lod_reset", **locals())
X
Xin Pan 已提交
5771
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
Y
yangyaming 已提交
5772 5773 5774 5775 5776 5777 5778 5779 5780 5781 5782
    if y is not None:
        helper.append_op(
            type="lod_reset", inputs={'X': x,
                                      'Y': y}, outputs={'Out': out})
    elif target_lod is not None:
        helper.append_op(
            type="lod_reset",
            inputs={'X': x},
            attrs={'target_lod': target_lod},
            outputs={'Out': out})
    else:
5783 5784 5785 5786 5787 5788 5789 5790 5791 5792 5793 5794 5795 5796 5797 5798 5799 5800 5801 5802 5803 5804 5805 5806 5807 5808
        raise ValueError("y and target_lod should not be both none.")
    return out


def lod_append(x, level):
    """
    Append level to LoD of :attr:`x`.

    .. code-block:: text

        * Example 1:

            given a 1-level LoDTensor x:
                x.lod =  [[ 2,           3,                   1 ]]
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            level: [1, 1, 1, 1, 1, 1, 1]

            then we get a 2-level LoDTensor:
                x.lod =  [[ 2, 3, 1 ], [1, 1, 1, 1, 1, 1]]
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

    Args:
        x (Variable): Input variable which could be a tensor or LoDTensor.
5809
        level (list|tuple|Variable): The LoD level to be appended into LoD of x.
5810 5811 5812 5813 5814 5815

    Returns:
        Variable: Output variable with new LoD level.

    Raises:
        ValueError: If :attr:`y` is None or and :attr:`level` is not Iterator.
Y
yangyaming 已提交
5816

5817 5818 5819 5820 5821 5822 5823 5824 5825 5826
    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
            x = fluid.layers.data(name='x', shape=[6, 10], lod_level=1)
            out = fluid.layers.lod_append(x, [1,1,1,1,1,1])
    """
    from collections import Iterable
    if x is None:
        raise ValueError("Input(x) can't be None.")
5827 5828 5829
    if (not isinstance(level, Iterable)) and (not isinstance(level, Variable)):
        raise ValueError("Input(level) must be list, tuple or Variable.")

5830 5831
    helper = LayerHelper("lod_append", **locals())
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
5832 5833 5834 5835 5836 5837 5838 5839

    inputs = {'X': x}
    attrs = {'append': True}

    if isinstance(level, Variable):
        inputs['Y'] = level
    else:
        attrs['target_lod'] = level
5840
    helper.append_op(
5841
        type="lod_reset", inputs=inputs, attrs=attrs, outputs={'Out': out})
Y
yangyaming 已提交
5842
    return out
D
dragonwarrior 已提交
5843 5844


5845 5846
def lrn(input, n=5, k=1.0, alpha=1e-4, beta=0.75, name=None,
        data_format='NCHW'):
D
dragonwarrior 已提交
5847
    """
5848 5849 5850
    This operator implements the Local Response Normalization Layer.
    This layer performs a type of "lateral inhibition" by normalizing over local input regions.
    For more information, please refer to `ImageNet Classification with Deep Convolutional Neural Networks <https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf>`_
D
dragonwarrior 已提交
5851 5852 5853 5854 5855

    The formula is as follows:

    .. math::

5856
        Output(i, x, y) = Input(i, x, y) / \\left(k + \\alpha \\sum\\limits^{\\min(C-1, i + n/2)}_{j = \\max(0, i - n/2)}(Input(j, x, y))^2\\right)^{\\beta}
D
dragonwarrior 已提交
5857 5858 5859

    In the above equation:

5860 5861 5862 5863
    - :math:`n` : The number of channels to sum over.
    - :math:`k` : The offset (avoid being divided by 0).
    - :math:`\\alpha` : The scaling parameter.
    - :math:`\\beta` : The exponent parameter.
D
dragonwarrior 已提交
5864 5865 5866


    Args:
5867 5868 5869
        input (Variable): Input feature, 4D-Tensor with the shape of [N,C,H,W] or [N, H, W, C], 
            where N is the batch size, C is the input channel, H is Height, W is weight. The data 
            type is float32. The rank of this tensor must be 4, otherwise it will raise ValueError.
5870 5871 5872 5873
        n (int, optional): The number of channels to sum over. Default: 5
        k (float, optional): An offset, positive. Default: 1.0
        alpha (float, optional): The scaling parameter, positive. Default:1e-4
        beta (float, optional): The exponent, positive. Default:0.75
5874 5875
        name (str, optional): The default value is None. Normally there is no need for user to set 
            this property. For more information, please refer to :ref:`api_guide_Name` 
5876 5877 5878 5879 5880
        data_format (str, optional): Specify the data format of the input, and the data format of the output 
            will be consistent with that of the input. An optional string from: `"NCHW"`, `"NHWC"`.
            The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`.
        
D
dragonwarrior 已提交
5881
    Returns:
5882 5883
        Variable: A tensor variable storing the transformation result with the same shape and data type as input.

D
dragonwarrior 已提交
5884 5885 5886

    Examples:

5887 5888 5889 5890 5891 5892 5893 5894
    .. code-block:: python

        import paddle.fluid as fluid
        data = fluid.data(
            name="data", shape=[None, 3, 112, 112], dtype="float32")
        lrn = fluid.layers.lrn(input=data)
        print(lrn.shape)  # [-1, 3, 112, 112]
        print(lrn.dtype)  # float32
D
dragonwarrior 已提交
5895 5896 5897 5898 5899 5900 5901 5902
    """
    helper = LayerHelper('lrn', **locals())
    dtype = helper.input_dtype()
    input_shape = input.shape
    dims = len(input_shape)

    if dims != 4:
        raise ValueError(
5903
            "Input's dimension size of Op(lrn) must be 4, but received %d." %
D
dragonwarrior 已提交
5904
            (dims))
5905 5906 5907 5908
    if data_format not in ['NCHW', 'NHWC']:
        raise ValueError(
            "Attr(data_format) of Op(lrn) got wrong value: received " +
            data_format + " but only NCHW or NHWC supported.")
D
dragonwarrior 已提交
5909

X
Xin Pan 已提交
5910 5911 5912
    mid_out = helper.create_variable_for_type_inference(
        dtype=dtype, stop_gradient=True)
    lrn_out = helper.create_variable_for_type_inference(dtype)
D
dragonwarrior 已提交
5913 5914 5915 5916 5917 5918 5919
    helper.append_op(
        type="lrn",
        inputs={"X": input},
        outputs={
            "Out": lrn_out,
            "MidOut": mid_out,
        },
5920 5921 5922 5923 5924 5925 5926
        attrs={
            "n": n,
            "k": k,
            "alpha": alpha,
            "beta": beta,
            "data_format": data_format
        })
D
dragonwarrior 已提交
5927 5928

    return lrn_out
G
guosheng 已提交
5929 5930 5931 5932


def pad(x, paddings, pad_value=0., name=None):
    """
S
SunGaofeng 已提交
5933 5934
    This op will pad a tensor with a constant value given by :attr:`pad_value`, and the
    padded shape is specified by :attr:`paddings`.
G
guosheng 已提交
5935

S
SunGaofeng 已提交
5936 5937 5938 5939
    Specifically, the number of values padded before the elements of :attr:`x`
    in dimension :attr:`i` is indicated by :attr:`paddings[2*i]`, and the number
    of values padded after the elements of :attr:`x` in dimension :attr:`i` is
    indicated by :attr:`paddings[2*i+1]`.
G
guosheng 已提交
5940 5941 5942 5943 5944 5945 5946 5947 5948 5949 5950 5951 5952 5953 5954 5955 5956 5957 5958

    See below for an example.

    .. code-block:: text

        Given:
            x = [[1, 2], [3, 4]]

            paddings = [0, 1, 1, 2]

            pad_value = 0

        Return:

            out = [[0, 1, 2, 0, 0]
                   [0, 3, 4, 0, 0]
                   [0, 0, 0, 0, 0]]

    Args:
S
SunGaofeng 已提交
5959
        x (Variable): Tensor, data type is float32.
G
guosheng 已提交
5960
        paddings (list): A list of integers. Its elements specify the padded
S
SunGaofeng 已提交
5961 5962
                         width before and after each dimension in turn.
                         The length of :attr:`paddings` must be equal to 
G
guosheng 已提交
5963 5964
                         :math:`rank(x) \\times 2`.
        pad_value (float): The constant value used to pad.
S
SunGaofeng 已提交
5965 5966 5967
        name(str, optional): The default value is None.  
                             Normally there is no need for user to set this property.  
                             For more information, please refer to :ref:`api_guide_Name`
G
guosheng 已提交
5968 5969

    Returns:
S
SunGaofeng 已提交
5970 5971 5972 5973
        The padded tensor, with the same data type and rank as :attr:`x`

    Return Type:
        Variable
G
guosheng 已提交
5974 5975 5976

    Examples:
        .. code-block:: python
G
guosheng 已提交
5977

S
SunGaofeng 已提交
5978 5979
            # x is a rank 2 tensor variable with shape [100, 224].
            # out will be a tensor of shape [101, 227] 
S
SunGaofeng 已提交
5980
            import paddle.fluid as fluid
S
SunGaofeng 已提交
5981
            x = fluid.data(name='data', shape=[100, 224], dtype='float32')
G
guosheng 已提交
5982 5983 5984 5985 5986
            out = fluid.layers.pad(
                x=x, paddings=[0, 1, 1, 2], pad_value=0.)
    """
    helper = LayerHelper('pad', input=x, **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
5987
    out = helper.create_variable_for_type_inference(dtype)
G
guosheng 已提交
5988 5989 5990 5991 5992 5993 5994
    helper.append_op(
        type='pad',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'paddings': paddings,
               'pad_value': float(pad_value)})
    return out
5995 5996


C
chengduo 已提交
5997 5998
def pad_constant_like(x, y, pad_value=0., name=None):
    """
S
SunGaofeng 已提交
5999
    Pad :attr:`y` with :attr:`pad_value`, the number of values padded to
C
chengduo 已提交
6000
    the edges of each axis is specified by the difference of the shape
S
SunGaofeng 已提交
6001 6002
    of :attr:`x` and :attr:`y` . ((0, shape_x_0 - shape_y_0), ... (0, shape_x_n - shape_y_n))
    specify padding widths for each axis. The input should be a k-D tensor(k > 0 and k < 7).
C
chengduo 已提交
6003 6004 6005 6006 6007 6008 6009 6010 6011 6012 6013 6014 6015 6016 6017 6018 6019 6020 6021 6022 6023 6024 6025 6026

    See below for an example.

    .. code-block:: text

        Given:
            X = [[[[ 0,  1,  2],
                   [ 3,  4,  5]],
                  [[ 6,  7,  8],
                   [ 9, 10, 11]],
                  [[12, 13, 14],
                   [15, 16, 17]]],
                 [[[18, 19, 20],
                   [21, 22, 23]],
                  [[24, 25, 26],
                   [27, 28, 29]],
                  [[30, 31, 32],
                   [33, 34, 35]]]]
            X.shape = (2, 3, 2, 3)

            Y = [[[[35, 36, 37]],
                  [[38, 39, 40]],
                  [[41, 42, 43]]]]
            Y.shape = (1, 3, 1, 3)
T
Tink_Y 已提交
6027 6028
		And
            pad_value = -1,
C
chengduo 已提交
6029

T
Tink_Y 已提交
6030 6031 6032 6033 6034 6035 6036 6037 6038 6039 6040 6041 6042 6043
        Return:
            Out = [[[[35, 36, 37],
                     [-1, -1, -1]],
                    [[38, 39, 40],
                     [-1, -1, -1]],
                    [[41, 42, 43],
                     [-1, -1, -1]]],
                  [[[-1, -1, -1],
                    [-1, -1, -1]],
                   [[-1, -1, -1],
                    [-1, -1, -1]],
                   [[-1, -1, -1],
                    [-1, -1, -1]]]]
            Out.shape = (2, 3, 2, 3)
C
chengduo 已提交
6044 6045

    Args:
S
SunGaofeng 已提交
6046 6047 6048
        x (Variable): Tensor, its shape spicifies the shape of output.
        y (Variable): Tensor, its rank is the same with :attr:`x`, and for each dimension :math:`i` , 
                      :math:`y\_shape[i] <= x\_shape[i]` . The data type can be float32 or float64.
C
chengduo 已提交
6049
        pad_value (float): The constant value used to pad.
S
SunGaofeng 已提交
6050 6051 6052
        name(str, optional): The default value is None.  
                             Normally there is no need for user to set this property.  
                             For more information, please refer to :ref:`api_guide_Name`
C
chengduo 已提交
6053 6054

    Returns:
S
SunGaofeng 已提交
6055 6056 6057 6058
        The padded tensor, with the same shape as :attr:`x` and the same data type as :attr:`y`

    Return Type:
        Variable
C
chengduo 已提交
6059 6060 6061 6062 6063 6064

    Examples:
        .. code-block:: python

            # x is a rank 4 tensor variable, x.shape = (2, 3, 2, 3)
            # y is a rank 4 tensor variable, y.shape = (1, 3, 1, 3)
S
SunGaofeng 已提交
6065
            import paddle.fluid as fluid
S
SunGaofeng 已提交
6066 6067
            x = fluid.data(name='x', shape=[2,3,2,3], dtype='float32')
            y = fluid.data(name='y', shape=[1,3,1,3], dtype='float32')
C
chengduo 已提交
6068 6069 6070 6071 6072
            out = fluid.layers.pad_constant_like(x=x, y=y, pad_value=0.)
            # out is a rank 4 tensor variable, and out.shape = [2, 3 ,2 , 3]
    """
    helper = LayerHelper('pad_constant_like', input=x, **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
6073
    out = helper.create_variable_for_type_inference(dtype)
C
chengduo 已提交
6074 6075 6076 6077 6078 6079 6080 6081 6082
    helper.append_op(
        type='pad_constant_like',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
        attrs={'pad_value': float(pad_value)})
    return out


6083 6084 6085 6086 6087 6088
def label_smooth(label,
                 prior_dist=None,
                 epsilon=0.1,
                 dtype="float32",
                 name=None):
    """
D
DuYao 已提交
6089 6090
    Label smoothing is a mechanism to regularize the classifier layer and is called 
    label-smoothing regularization (LSR). 
6091

6092 6093 6094 6095 6096 6097 6098 6099 6100 6101 6102 6103 6104 6105 6106 6107 6108
    Label smoothing is proposed to encourage the model to be less confident,
    since optimizing the log-likelihood of the correct label directly may
    cause overfitting and reduce the ability of the model to adapt. Label
    smoothing replaces the ground-truth label :math:`y` with the weighted sum
    of itself and some fixed distribution :math:`\mu`. For class :math:`k`,
    i.e.

    .. math::

        \\tilde{y_k} = (1 - \epsilon) * y_k + \epsilon * \mu_k,

    where :math:`1 - \epsilon` and :math:`\epsilon` are the weights
    respectively, and :math:`\\tilde{y}_k` is the smoothed label. Usually
    uniform distribution is used for :math:`\mu`.

    See more details about label smoothing in https://arxiv.org/abs/1512.00567.

D
DuYao 已提交
6109
    Parameters:
6110
        label(Variable): The input variable containing the label data. The
D
DuYao 已提交
6111 6112 6113 6114 6115 6116 6117 6118 6119 6120 6121 6122 6123 6124 6125
                        label data should use one-hot representation. It's 
                        a multidimensional tensor with a shape of 
                        :math:`[N_1, ..., Depth]`, where Depth is class number.
        prior_dist(Variable, optional): The prior distribution to be used to smooth
                        labels. If not provided, an uniform distribution
                        is used. It's a multidimensional tensor with a shape of
                        :math:`[1, class\_num]` . The default value is None.
        epsilon(float, optional): The weight used to mix up the original ground-truth
                        distribution and the fixed distribution. The default value is 
                        0.1.
        dtype(np.dtype|core.VarDesc.VarType|str, optional): The data type can be set
                        as 'float32', 'float64'. The default value is 'float32'.
        name(str, optional): The default value is None. Normally there is no need for user 
                        to set this property. For more information, please refer to 
                        :ref:`api_guide_Name`.
6126 6127 6128 6129 6130 6131

    Returns:
        Variable: The tensor variable containing the smoothed labels.

    Examples:
        .. code-block:: python
6132
            
6133
            import paddle.fluid as fluid
6134
            import paddle.fluid.layers as layers
6135 6136 6137 6138 6139 6140 6141 6142 6143 6144

            label = layers.data(name="label", shape=[1], dtype="float32")
            one_hot_label = layers.one_hot(input=label, depth=10)
            smooth_label = layers.label_smooth(
                label=one_hot_label, epsilon=0.1, dtype="float32")
    """
    if epsilon > 1. or epsilon < 0.:
        raise ValueError("The value of epsilon must be between 0 and 1.")
    helper = LayerHelper("label_smooth", **locals())
    label.stop_gradient = True
X
Xin Pan 已提交
6145
    smooth_label = helper.create_variable_for_type_inference(dtype)
6146 6147 6148 6149 6150 6151 6152
    helper.append_op(
        type="label_smooth",
        inputs={"X": label,
                "PriorDist": prior_dist} if prior_dist else {"X": label},
        outputs={"Out": smooth_label},
        attrs={"epsilon": float(epsilon)})
    return smooth_label
6153 6154


W
wopeizl 已提交
6155 6156 6157
@templatedoc()
def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0):
    """
6158 6159 6160 6161 6162 6163 6164 6165 6166 6167 6168
    This operator implements the roi_pooling layer. 
    Region of interest pooling (also known as RoI pooling) is to perform max pooling on inputs of nonuniform sizes to obtain fixed-size feature maps (e.g. 7*7).
    
    The operator has three steps:
    
        1. Dividing each region proposal into equal-sized sections with the pooled_width and pooled_height;
        2. Finding the largest value in each section;
        3. Copying these max values to the output buffer.
    
    For more information, please refer to https://stackoverflow.com/questions/43430056/what-is-roi-layer-in-fast-rcnn
    
W
wopeizl 已提交
6169
    Args:
6170 6171 6172 6173 6174 6175
        input (Variable): Input feature, 4D-Tensor with the shape of [N,C,H,W], where N is the batch size, C is the input channel, H is Height, W is weight. The data type is float32 or float64.
        rois (Variable): ROIs (Regions of Interest) to pool over. 2D-LoDTensor with the shape of [num_rois,4], the lod level is 1. Given as [[x1, y1, x2, y2], ...], (x1, y1) is the top left coordinates, and (x2, y2) is the bottom right coordinates.
        pooled_height (int, optional): The pooled output height, data type is int32. Default: 1
        pooled_width (int, optional): The pooled output height, data type is int32. Default: 1
        spatial_scale (float, optional): Multiplicative spatial scale factor to translate ROI coords from their input scale to the scale used when pooling. Default: 1.0
    
W
wopeizl 已提交
6176
    Returns:
6177 6178 6179
        Variable: The pooled feature, 4D-Tensor with the shape of [num_rois, C, pooled_height, pooled_width].
    
    
W
wopeizl 已提交
6180
    Examples:
6181 6182 6183 6184 6185 6186 6187 6188 6189 6190 6191 6192 6193 6194 6195 6196 6197 6198
    
    ..  code-block:: python
    
        import paddle.fluid as fluid
        import numpy as np
    
        DATATYPE='float32'
    
        place = fluid.CPUPlace()
        #place = fluid.CUDAPlace(0)
    
        input_data = np.array([i for i in range(1,17)]).reshape(1,1,4,4).astype(DATATYPE)
        roi_data =fluid.create_lod_tensor(np.array([[1., 1., 2., 2.], [1.5, 1.5, 3., 3.]]).astype(DATATYPE),[[2]], place)
    
        x = fluid.data(name='input', shape=[None,1,4,4], dtype=DATATYPE)
        rois = fluid.data(name='roi', shape=[None,4], dtype=DATATYPE)
    
        pool_out = fluid.layers.roi_pool(
6199 6200
                input=x,
                rois=rois,
6201 6202
                pooled_height=1,
                pooled_width=1,
6203
                spatial_scale=1.0)
6204 6205 6206 6207 6208
    
        exe = fluid.Executor(place)
        out, = exe.run(feed={'input':input_data ,'roi':roi_data}, fetch_list=[pool_out.name])
        print(out)   #array([[[[11.]]], [[[16.]]]], dtype=float32)
        print(np.array(out).shape)  # (2, 1, 1, 1)
W
wopeizl 已提交
6209 6210 6211 6212 6213 6214 6215 6216 6217 6218 6219 6220 6221 6222 6223 6224 6225
    """
    helper = LayerHelper('roi_pool', **locals())
    dtype = helper.input_dtype()
    pool_out = helper.create_variable_for_type_inference(dtype)
    argmaxes = helper.create_variable_for_type_inference(dtype='int32')
    helper.append_op(
        type="roi_pool",
        inputs={"X": input,
                "ROIs": rois},
        outputs={"Out": pool_out,
                 "Argmax": argmaxes},
        attrs={
            "pooled_height": pooled_height,
            "pooled_width": pooled_width,
            "spatial_scale": spatial_scale
        })
    return pool_out
W
whs 已提交
6226 6227


J
jerrywgz 已提交
6228 6229 6230 6231 6232 6233
@templatedoc()
def roi_align(input,
              rois,
              pooled_height=1,
              pooled_width=1,
              spatial_scale=1.0,
J
jerrywgz 已提交
6234 6235
              sampling_ratio=-1,
              name=None):
J
jerrywgz 已提交
6236 6237 6238 6239 6240
    """
    ${comment}

    Args:
        input (Variable): ${x_comment}
6241
        rois (Variable): ROIs (Regions of Interest) to pool over.It should be
W
wangguanzhong 已提交
6242 6243 6244 6245 6246 6247 6248 6249 6250 6251 6252
            a 2-D LoDTensor of shape (num_rois, 4), the lod level is 1. The 
            data type is float32 or float64. Given as [[x1, y1, x2, y2], ...], 
            (x1, y1) is the top left coordinates, and (x2, y2) is the bottom
            right coordinates. 
        pooled_height (int32, optional): ${pooled_height_comment} Default: 1
        pooled_width (int32, optional): ${pooled_width_comment} Default: 1
        spatial_scale (float32, optional): ${spatial_scale_comment} Default: 1.0
        sampling_ratio(int32, optional): ${sampling_ratio_comment} Default: -1
        name(str, optional): For detailed information, please refer 
            to :ref:`api_guide_Name`. Usually name is no need to set and 
            None by default. 
J
jerrywgz 已提交
6253 6254

    Returns:
W
wangguanzhong 已提交
6255 6256 6257 6258 6259
        Variable:

        Output: ${out_comment}.


J
jerrywgz 已提交
6260 6261 6262
    Examples:
        .. code-block:: python

6263
            import paddle.fluid as fluid
6264 6265 6266 6267
            x = fluid.data(
                name='data', shape=[None, 256, 32, 32], dtype='float32')
            rois = fluid.data(
                name='rois', shape=[None, 4], dtype='float32')
6268 6269 6270
            align_out = fluid.layers.roi_align(input=x,
                                               rois=rois,
                                               pooled_height=7,
J
jerrywgz 已提交
6271 6272 6273 6274 6275 6276
                                               pooled_width=7,
                                               spatial_scale=0.5,
                                               sampling_ratio=-1)
    """
    helper = LayerHelper('roi_align', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
6277
    align_out = helper.create_variable_for_type_inference(dtype)
J
jerrywgz 已提交
6278 6279 6280 6281 6282 6283 6284 6285 6286 6287 6288 6289 6290 6291
    helper.append_op(
        type="roi_align",
        inputs={"X": input,
                "ROIs": rois},
        outputs={"Out": align_out},
        attrs={
            "pooled_height": pooled_height,
            "pooled_width": pooled_width,
            "spatial_scale": spatial_scale,
            "sampling_ratio": sampling_ratio
        })
    return align_out


S
SunGaofeng 已提交
6292
def dice_loss(input, label, epsilon=0.00001, name=None):
W
whs 已提交
6293
    """
S
SunGaofeng 已提交
6294 6295 6296 6297
    Dice loss for comparing the similarity between the input predictions and the label.
    This implementation is for binary classification, where the input is sigmoid
    predictions of each pixel, usually used for segmentation task. The dice loss can
    be defined as the following equation:
W
whs 已提交
6298 6299 6300 6301 6302 6303 6304 6305

    .. math::

        dice\_loss &= 1 - \\frac{2 * intersection\_area}{total\_area} \\\\
                  &= \\frac{(total\_area - intersection\_area) - intersection\_area}{total\_area} \\\\
                  &= \\frac{(union\_area - intersection\_area)}{total\_area}


S
SunGaofeng 已提交
6306 6307 6308 6309 6310 6311
    Parameters:
        input (Variable): Tensor, rank>=2, shape is :math:`[N_1, N_2, ..., N_D]`, where :math:`N_1` is
                          the batch_size, :math:`N_D` is 1. It is usually the output predictions of sigmoid activation.
                          The data type can be float32 or float64.
        label (Variable): Tensor, the groud truth with the same rank as input, shape is :math:`[N_1, N_2, ..., N_D]`. 
                          where :math:`N_1` is the batch_size, :math:`N_D` is 1. The data type can be float32 or float64.
W
whs 已提交
6312 6313 6314
        epsilon (float): The epsilon will be added to the numerator and denominator.
                         If both input and label are empty, it makes sure dice is 1.
                         Default: 0.00001
S
SunGaofeng 已提交
6315 6316 6317
        name(str, optional): The default value is None.  
                             Normally there is no need for user to set this property.  
                             For more information, please refer to :ref:`api_guide_Name`
W
whs 已提交
6318 6319

    Returns:
S
SunGaofeng 已提交
6320 6321 6322
        The dice loss with shape [1], data type is the same as `input` .
    Return Type:
        Varaible
W
whs 已提交
6323

S
SunGaofeng 已提交
6324
    Example:
6325 6326
        .. code-block:: python

S
SunGaofeng 已提交
6327
            import paddle.fluid as fluid
S
SunGaofeng 已提交
6328 6329 6330
            x = fluid.data(name='data', shape = [3, 224, 224, 1], dtype='float32')
            label = fluid.data(name='label', shape=[3, 224, 224, 1], dtype='float32')
            predictions = fluid.layers.sigmoid(x)
S
SunGaofeng 已提交
6331
            loss = fluid.layers.dice_loss(input=predictions, label=label)
W
whs 已提交
6332 6333
    """
    label = one_hot(label, depth=input.shape[-1])
6334
    reduce_dim = list(range(1, len(input.shape)))
W
whs 已提交
6335 6336 6337 6338 6339 6340
    inse = reduce_sum(input * label, dim=reduce_dim)
    dice_denominator = reduce_sum(
        input, dim=reduce_dim) + reduce_sum(
            label, dim=reduce_dim)
    dice_score = 1 - inse * 2 / (dice_denominator + epsilon)
    return reduce_mean(dice_score)
6341 6342


6343 6344 6345 6346
def image_resize(input,
                 out_shape=None,
                 scale=None,
                 name=None,
6347
                 resample='BILINEAR',
6348 6349
                 actual_shape=None,
                 align_corners=True,
6350 6351
                 align_mode=1,
                 data_format='NCHW'):
6352
    """
R
ruri 已提交
6353
    This op resizes a batch of images.
F
stash  
fengjiayi 已提交
6354

6355 6356 6357 6358
    The input must be a 4-D Tensor of the shape (num_batches, channels, in_h, in_w) 
    or (num_batches, in_h, in_w, channels), or a 5-D Tensor of the shape 
    (num_batches, channels, in_d, in_h, in_w) or (num_batches, in_d, in_h, in_w, channels), 
    and the resizing only applies on the three dimensions(depth, hight and width).
6359

6360
    **Warning:** the parameter :attr:`actual_shape` will be deprecated in the
6361 6362
    future and only use :attr:`out_shape` instead.

6363
    Supporting resample methods:
Q
update  
qiaolongfei 已提交
6364

6365
        'BILINEAR' : Bilinear interpolation
T
Tink_Y 已提交
6366

K
Kaipeng Deng 已提交
6367 6368
        'TRILINEAR' : Trilinear interpolation

6369
        'NEAREST' : Nearest neighbor interpolation
F
stash  
fengjiayi 已提交
6370

6371 6372 6373 6374 6375 6376 6377 6378 6379 6380
    Nearest neighbor interpolation is to perform nearest neighbor interpolation
    in both the 3rd dimention(in height direction) and the 4th dimention(in width 
    direction) on input tensor.
            
    Bilinear interpolation is an extension of linear interpolation for 
    interpolating functions of two variables (e.g. H-direction and 
    W-direction in this op) on a rectilinear 2D grid. The key idea is 
    to perform linear interpolation first in one direction, and then 
    again in the other direction.

K
Kaipeng Deng 已提交
6381 6382 6383 6384 6385
    Trilinear interpolation is an extension of linear interpolation for 
    interpolating functions of three variables (e.g. D-direction, 
    H-direction and W-direction in this op) on a rectilinear 3D grid. 
    The linear interpolation is performed on three directions.

T
tink2123 已提交
6386
    Align_corners and align_mode are optinal parameters,the calculation method 
6387 6388 6389 6390
    of interpolation can be selected by them.

    Example:

T
Tink_Y 已提交
6391
    .. code-block:: text
6392

T
Tink_Y 已提交
6393
        For scale:
6394
          
T
Tink_Y 已提交
6395
            if align_corners = True && out_size > 1 :
6396

T
Tink_Y 已提交
6397 6398 6399 6400 6401 6402 6403 6404 6405 6406 6407
              scale_factor = (in_size-1.0)/(out_size-1.0)
            
            else:
              
              scale_factor = float(in_size/out_size)
            
          
        Nearest neighbor interpolation:
          
          if:
              align_corners = False
6408

T
Tink_Y 已提交
6409 6410
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
6411

T
Tink_Y 已提交
6412 6413
              H_out = floor (H_{in} * scale_{factor})
              W_out = floor (W_{in} * scale_{factor})
6414

T
Tink_Y 已提交
6415 6416
          else:
              align_corners = True
6417

T
Tink_Y 已提交
6418 6419
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
6420

T
Tink_Y 已提交
6421 6422
              H_out = round(H_{in} * scale_{factor})
              W_out = round(W_{in} * scale_{factor})
6423

T
Tink_Y 已提交
6424 6425 6426 6427 6428 6429 6430 6431 6432 6433
        Bilinear interpolation:

          if:
              align_corners = False , align_mode = 0
              
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
              
              H_out = (H_{in}+0.5) * scale_{factor} - 0.5
              W_out = (W_{in}+0.5) * scale_{factor} - 0.5
6434

T
Tink_Y 已提交
6435 6436 6437 6438
          else:
           
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
6439

T
Tink_Y 已提交
6440 6441
              H_out = H_{in} * scale_{factor}
              W_out = W_{in} * scale_{factor}
6442

K
Kaipeng Deng 已提交
6443 6444 6445 6446 6447 6448 6449 6450 6451 6452 6453 6454 6455 6456 6457 6458 6459 6460 6461 6462 6463 6464
        Trilinear interpolation:

          if:
              align_corners = False , align_mode = 0
              
              input : (N,C,D_in,H_in,W_in)
              output: (N,C,D_out,H_out,W_out) where:
              
              D_out = (D_{in}+0.5) * scale_{factor} - 0.5
              H_out = (H_{in}+0.5) * scale_{factor} - 0.5
              W_out = (W_{in}+0.5) * scale_{factor} - 0.5


          else:
           
              input : (N,C,D_in,H_in,W_in)
              output: (N,C,D_out,H_out,W_out) where:

              D_out = D_{in} * scale_{factor}
              H_out = H_{in} * scale_{factor}
              W_out = W_{in} * scale_{factor}
          
6465 6466 6467 6468 6469 6470
    For details of nearest neighbor interpolation, please refer to Wikipedia: 
    https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation.

    For details of bilinear interpolation, please refer to Wikipedia: 
    https://en.wikipedia.org/wiki/Bilinear_interpolation.

K
Kaipeng Deng 已提交
6471 6472 6473
    For details of trilinear interpolation, please refer to Wikipedia: 
    https://en.wikipedia.org/wiki/Trilinear_interpolation.

6474 6475


R
ruri 已提交
6476
    Parameters:
6477 6478
        input (Variable): 4-D or 5-D Tensor, its data type is float32, float64, or uint8,
                          its data format is specified by :attr:`data_format`.
6479
        out_shape(list|tuple|Variable|None): Output shape of image resize
6480 6481 6482 6483
             layer, the shape is (out_h, out_w) when input is a 4-D Tensor and is
             (out_d, out_h, out_w) when input is a 5-D Tensor. Default: None. If 
             a list, each element can be an integer or a Tensor Variable of shape: [1].
             If a Tensor Variable, its dimensions size should be a 1.
6484 6485 6486
        scale(float|Variable|None): The multiplier for the input height or width. At
             least one of :attr:`out_shape` or :attr:`scale` must be set.
             And :attr:`out_shape` has a higher priority than :attr:`scale`.
D
dengkaipeng 已提交
6487
             Default: None.
6488 6489
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
K
Kaipeng Deng 已提交
6490 6491
        resample(str): The resample method. It supports 'BILINEAR', 'TRILINEAR'
                       and 'NEAREST' currently. Default: 'BILINEAR'
6492 6493 6494
        actual_shape(Variable): An optional input to specify output shape
                                dynamically. If provided, image resize
                                according to this given shape rather than
6495
                                :attr:`out_shape` and :attr:`scale` specifying
6496 6497
                                shape. That is to say actual_shape has the
                                highest priority. It is recommended to use
6498 6499 6500 6501 6502 6503
                                :attr:`out_shape` if you want to specify output 
                                shape dynamically, because :attr:`actual_shape` 
                                will be deprecated. When using actual_shape to 
                                specify output shape, one of :attr:`out_shape` 
                                and :attr:`scale` should also be set, otherwise 
                                errors would be occured in graph constructing stage.
6504
                                Default: None
6505 6506 6507 6508
        align_corners(bool) :  An optional bool, If True, the centers of the 4 corner pixels of the 
                               input and output tensors are aligned, preserving the values at the 
                               corner pixels.
                               Default: True
T
tink2123 已提交
6509
        align_mode(int)  :  An optional for bilinear interpolation. can be \'0\' 
T
tink2123 已提交
6510
                            for src_idx = scale*(dst_indx+0.5)-0.5 , can be \'1\' for 
6511
                            src_idx = scale*dst_index.
6512 6513 6514 6515 6516
        data_format (str, optional): Specify the data format of the input, and the data format of the output 
            will be consistent with that of the input. An optional string from: `"NCHW"`, `"NHWC"`, `"NCDHW"`,
            `"NDHWC"`. The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`. When it is `"NCHW"`, the data is stored 
            in the order of: `[batch_size, input_channels, input_depth, input_height, input_width]`.
6517 6518

    Returns:
6519 6520
        A 4-D Tensor of the shape (num_batches, channels, out_h, out_w) or (num_batches, out_h, out_w, channels),
        or 5-D Tensor of the shape (num_batches, channels, out_d, out_h, out_w) or (num_batches, out_d, out_h, out_w, channels).
F
stash  
fengjiayi 已提交
6521

6522 6523 6524
    Raises:
        TypeError: out_shape should be a list or tuple or Variable.
        TypeError: actual_shape should either be Variable or None.
K
Kaipeng Deng 已提交
6525 6526 6527 6528
        ValueError: The 'resample' of image_resize can only be 'BILINEAR',
                    'TRILINEAR' or 'NEAREST' currently.
        ValueError: 'BILINEAR' and 'NEAREST' only support 4-D tensor.
        ValueError: 'TRILINEAR' only support 5-D tensor.
6529
        ValueError: One of out_shape and scale must not be None.
K
Kaipeng Deng 已提交
6530 6531
        ValueError: out_shape length should be 2 for input 4-D tensor.
        ValueError: out_shape length should be 3 for input 5-D tensor.
D
dengkaipeng 已提交
6532
        ValueError: scale should be greater than zero.
6533 6534
        TypeError: align_corners shoule be a bool value
        ValueError: align_mode can only be '0' or '1'
6535
        ValueError: data_format can only be 'NCHW', 'NHWC', 'NCDHW' or 'NDHWC'.
6536

6537 6538
    Examples:
        .. code-block:: python
R
ruri 已提交
6539 6540 6541 6542 6543 6544 6545 6546 6547 6548 6549 6550 6551 6552 6553 6554 6555 6556 6557 6558 6559 6560 6561 6562 6563 6564 6565 6566 6567 6568 6569 6570
	
	    #declarative mode
	    import paddle.fluid as fluid
	    import numpy as np
	    input = fluid.data(name="input", shape=[None,3,6,10])

	    #1
	    output = fluid.layers.image_resize(input=input,out_shape=[12,12])

	    #2
	    #x = np.array([2]).astype("int32")
	    #dim1 = fluid.data(name="dim1", shape=[1], dtype="int32")
	    #fluid.layers.assign(input=x, output=dim1)
	    #output = fluid.layers.image_resize(input=input,out_shape=[12,dim1])

	    #3
	    #x = np.array([3,12]).astype("int32")
	    #shape_tensor = fluid.data(name="shape_tensor", shape=[2], dtype="int32")
	    #fluid.layers.assign(input=x, output=shape_tensor)
	    #output = fluid.layers.image_resize(input=input,out_shape=shape_tensor)

	    #4
	    #x = np.array([0.5]).astype("float32")
	    #scale_tensor = fluid.data(name="scale", shape=[1], dtype="float32")
	    #fluid.layers.assign(x,scale_tensor)
	    #output = fluid.layers.image_resize(input=input,scale=scale_tensor)

	    place = fluid.CPUPlace()
	    exe = fluid.Executor(place)
	    exe.run(fluid.default_startup_program())
 
	    input_data = np.random.rand(2,3,6,10).astype("float32")
6571

R
ruri 已提交
6572 6573 6574 6575 6576 6577
	    output_data = exe.run(fluid.default_main_program(),
                feed={"input":input_data},
                fetch_list=[output],
                return_numpy=True)
 
	    print(output_data[0].shape)
6578

R
ruri 已提交
6579 6580 6581 6582 6583 6584 6585 6586
	    #1
	    # (2, 3, 12, 12)
	    #2
	    # (2, 3, 12, 2)
	    #3
	    # (2, 3, 3, 12)
	    #4
	    # (2, 3, 3, 5)
6587

R
ruri 已提交
6588 6589
	    #imperative mode
	    import paddle.fluid.dygraph as dg
6590

R
ruri 已提交
6591 6592 6593 6594
	    with dg.guard(place) as g:
    		input = dg.to_variable(input_data)
    		output = fluid.layers.image_resize(input=input, out_shape=[12,12])
    		print(output.shape)
6595

R
ruri 已提交
6596
		# [2L, 3L, 12L, 12L]
6597

6598
    """
6599 6600
    resample_methods = {
        'BILINEAR': 'bilinear',
K
Kaipeng Deng 已提交
6601
        'TRILINEAR': 'trilinear',
6602 6603
        'NEAREST': 'nearest',
    }
6604 6605
    if resample not in resample_methods:
        raise ValueError(
K
Kaipeng Deng 已提交
6606 6607
            "The 'resample' of image_resize can only be 'BILINEAR', 'TRILINEAR' "
            "or 'NEAREST' currently.")
6608
    resample_type = resample_methods[resample]
6609

K
Kaipeng Deng 已提交
6610 6611 6612 6613 6614
    if resample in ['BILINEAR', 'NEAREST'] and len(input.shape) != 4:
        raise ValueError("'BILINEAR' and 'NEAREST' only support 4-D tensor.")
    if resample == 'TRILINEAR' and len(input.shape) != 5:
        raise ValueError("'TRILINEAR'only support 5-D tensor.")

6615 6616 6617 6618 6619
    if not isinstance(align_corners, bool):
        raise TypeError("Attr align_corners should be a bool value")
    if align_mode != 0 and align_mode != 1:
        raise ValueError("align_mode can only be 0 or 1")

6620
    if out_shape is None and scale is None:
6621
        raise ValueError("One of out_shape and scale must not be None.")
6622
    helper = LayerHelper('{}_interp'.format(resample_type), **locals())
6623
    dtype = helper.input_dtype()
6624

6625 6626 6627 6628 6629 6630 6631 6632 6633
    if len(input.shape) == 4 and data_format not in ['NCHW', 'NHWC']:
        raise ValueError(
            "Got wrong value for param `data_format`: " + data_format +
            " received but only `NCHW` or `NHWC` supported for 4-D input.")
    elif len(input.shape) == 5 and data_format not in ['NCDHW', 'NDHWC']:
        raise ValueError(
            "Got wrong value for param `data_format`: " + data_format +
            " received but only `NCDHW` or `NDHWC` supported for 5-D input.")

6634 6635 6636
    def _is_list_or_turple_(data):
        return (isinstance(data, list) or isinstance(data, tuple))

6637 6638 6639 6640 6641
    if data_format == 'NCHW' or data_format == 'NCDHW':
        data_layout = 'NCHW'
    if data_format == 'NHWC' or data_format == 'NDHWC':
        data_layout = 'NHWC'

6642
    inputs = {"X": input}
D
dengkaipeng 已提交
6643
    attrs = {
6644 6645 6646
        "out_d": -1,
        "out_h": -1,
        "out_w": -1,
D
dengkaipeng 已提交
6647 6648
        "interp_method": resample_type,
        "align_corners": align_corners,
6649 6650
        "align_mode": align_mode,
        "data_layout": data_layout
D
dengkaipeng 已提交
6651 6652
    }

6653
    if out_shape is not None:
6654
        if isinstance(out_shape, Variable):
6655
            out_shape.stop_gradient = True
6656
            inputs['OutSize'] = out_shape
6657 6658
        else:
            if not (_is_list_or_turple_(out_shape)):
D
dengkaipeng 已提交
6659 6660
                raise TypeError(
                    "out_shape should be a list or tuple or Variable.")
6661 6662 6663 6664 6665 6666 6667 6668 6669 6670 6671 6672 6673 6674 6675 6676 6677 6678 6679 6680 6681 6682 6683 6684 6685 6686 6687 6688
            # Validate the shape
            contain_var = False
            for dim_idx, dim_size in enumerate(out_shape):
                if isinstance(dim_size, Variable):
                    contain_var = True
                    continue
                assert dim_size > 0, (
                    "Each dimension size given in out_shape must be greater than 0."
                )

            if contain_var:
                new_size_tensor = []
                size_list = []
                for dim in out_shape:
                    if isinstance(dim, Variable):
                        dim.stop_gradient = True
                        new_size_tensor.append(dim)
                        size_list.append(-1)
                    else:
                        assert (isinstance(dim, int))
                        temp_out = helper.create_variable_for_type_inference(
                            'int32')
                        fill_constant(
                            [1], 'int32', dim, force_cpu=True, out=temp_out)
                        new_size_tensor.append(temp_out)
                        size_list.append(dim)
                inputs['SizeTensor'] = new_size_tensor

K
Kaipeng Deng 已提交
6689 6690 6691 6692
            if len(input.shape) == 4:
                if len(out_shape) != 2:
                    raise ValueError("out_shape length should be 2 for "
                                     "input 4-D tensor.")
6693 6694 6695 6696 6697 6698 6699
                if contain_var:
                    attrs['out_h'] = size_list[0]
                    attrs['out_w'] = size_list[1]
                else:
                    out_shape = list(map(int, out_shape))
                    attrs['out_h'] = out_shape[0]
                    attrs['out_w'] = out_shape[1]
K
Kaipeng Deng 已提交
6700 6701 6702 6703
            if len(input.shape) == 5:
                if len(out_shape) != 3:
                    raise ValueError("out_shape length should be 3 for "
                                     "input 5-D tensor.")
6704 6705 6706 6707 6708 6709 6710 6711 6712
                if contain_var:
                    attrs['out_d'] = size_list[0]
                    attrs['out_h'] = size_list[1]
                    attrs['out_w'] = size_list[2]
                else:
                    out_shape = list(map(int, out_shape))
                    attrs['out_d'] = out_shape[0]
                    attrs['out_h'] = out_shape[1]
                    attrs['out_w'] = out_shape[2]
6713

6714
    else:
6715 6716 6717
        if isinstance(scale, Variable):
            scale.stop_gradient = True
            inputs["Scale"] = scale
6718
        elif isinstance(scale, float) or isinstance(scale, int):
6719
            if scale <= 0:
6720
                raise ValueError("Attr(scale) should be greater than zero.")
6721
            attrs['scale'] = float(scale)
6722 6723 6724
        else:
            raise TypeError(
                "Attr(scale)'s type should be float, int or Variable.")
6725

6726
    if isinstance(actual_shape, Variable):
6727 6728 6729 6730 6731
        warnings.warn(
            "actual_shape will be deprecated, it is recommended to use "
            "out_shape instead of actual_shape to specify output shape dynamically."
        )
        actual_shape.stop_gradient = True
6732 6733 6734 6735
        inputs["OutSize"] = actual_shape
    elif actual_shape is not None:
        raise TypeError("actual_shape should either be Variable or None.")

X
Xin Pan 已提交
6736
    out = helper.create_variable_for_type_inference(dtype)
6737
    helper.append_op(
6738
        type='{}_interp'.format(resample_type),
6739
        inputs=inputs,
6740
        outputs={"Out": out},
D
dengkaipeng 已提交
6741
        attrs=attrs)
6742
    return out
F
stash  
fengjiayi 已提交
6743 6744


6745
@templatedoc(op_type="bilinear_interp")
6746 6747 6748 6749
def resize_bilinear(input,
                    out_shape=None,
                    scale=None,
                    name=None,
6750 6751
                    actual_shape=None,
                    align_corners=True,
6752 6753
                    align_mode=1,
                    data_format='NCHW'):
6754
    """
R
ruri 已提交
6755
    This op resizes the input by performing bilinear interpolation based on given
6756
    output shape which specified by actual_shape, out_shape and scale
6757 6758
    in priority order.

6759 6760 6761
    **Warning:** the parameter :attr:`actual_shape` will be deprecated in 
    the future and only use :attr:`out_shape` instead.

6762 6763 6764 6765
    Bilinear interpolation is an extension of linear interpolation for
    interpolating functions of two variables (e.g. H-direction and
    W-direction in this op) on a rectilinear 2D grid. The key idea is
    to perform linear interpolation first in one direction, and then
6766 6767
    again in the other direction.

6768
    For details of bilinear interpolation, please refer to Wikipedia:
6769
    https://en.wikipedia.org/wiki/Bilinear_interpolation
Y
yuyang18 已提交
6770

T
tink2123 已提交
6771
    Align_corners and align_mode are optinal parameters,the calculation 
6772 6773 6774 6775
    method of interpolation can be selected by them.

    Example:

T
Tink_Y 已提交
6776
    .. code-block:: text
6777

T
Tink_Y 已提交
6778
        For scale:
6779
          
T
Tink_Y 已提交
6780
            if align_corners = True && out_size > 1 :
6781

T
Tink_Y 已提交
6782 6783 6784 6785
              scale_factor = (in_size-1.0)/(out_size-1.0)
            
            else:
              
6786
              scale_factor = float(in_size/out_size)
6787

T
Tink_Y 已提交
6788 6789 6790 6791 6792 6793 6794 6795 6796 6797
        Bilinear interpolation:

          if:
              align_corners = False , align_mode = 0
              
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
              
              H_out = (H_{in}+0.5) * scale_{factor} - 0.5
              W_out = (W_{in}+0.5) * scale_{factor} - 0.5
6798

T
Tink_Y 已提交
6799
          else:
T
tink2123 已提交
6800

T
Tink_Y 已提交
6801 6802 6803 6804
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
              H_out = H_{in} * scale_{factor}
              W_out = W_{in} * scale_{factor}
6805

R
ruri 已提交
6806 6807
    Parameters:
        input(Variable): 4-D Tensor(NCHW), its data type is float32, float64, or uint8,
6808
                          its data format is specified by :attr:`data_format`.
D
dengkaipeng 已提交
6809
        out_shape(list|tuple|Variable|None): Output shape of resize bilinear
6810
            layer, the shape is (out_h, out_w).Default: None. If a list, each 
6811 6812
            element can be an integer or a Tensor Variable with shape: [1]. If a 
            Tensor Variable, its dimension size should be 1.
6813
        scale(float|Variable|None): The multiplier for the input height or width. At
D
dengkaipeng 已提交
6814
             least one of :attr:`out_shape` or :attr:`scale` must be set. 
D
dengkaipeng 已提交
6815
             And :attr:`out_shape` has a higher priority than :attr:`scale`. 
D
dengkaipeng 已提交
6816
             Default: None.
6817 6818 6819
        actual_shape(Variable): An optional input to specify output shape
                                dynamically. If provided, image resize
                                according to this given shape rather than
6820
                                :attr:`out_shape` and :attr:`scale` specifying
6821 6822
                                shape. That is to say actual_shape has the
                                highest priority. It is recommended to use
6823 6824 6825 6826 6827 6828
                                :attr:`out_shape` if you want to specify output 
                                shape dynamically, because :attr:`actual_shape` 
                                will be deprecated. When using actual_shape to 
                                specify output shape, one of :attr:`out_shape` 
                                and :attr:`scale` should also be set, otherwise 
                                errors would be occured in graph constructing stage.
6829
                                Default: None
6830 6831
        align_corners(bool): ${align_corners_comment}
        align_mode(bool): ${align_mode_comment}
6832 6833 6834 6835
        data_format (str, optional): Specify the data format of the input, and the data format of the output 
            will be consistent with that of the input. An optional string from: `"NCHW"`, `"NHWC"`.
            The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`.
R
ruri 已提交
6836
        name(str, optional): The default value is None.  Normally there is no need for user to set this property.  For more information, please refer to :ref:`api_guide_Name`
Y
yuyang18 已提交
6837 6838

    Returns:
R
ruri 已提交
6839 6840
	Variable: 4-D tensor(NCHW or NHWC).
    
6841 6842
    Examples:
        .. code-block:: python
R
ruri 已提交
6843 6844 6845 6846 6847 6848 6849 6850 6851 6852 6853 6854 6855 6856 6857 6858 6859 6860 6861 6862 6863 6864 6865 6866 6867 6868 6869 6870 6871 6872 6873 6874
	
	    #declarative mode
	    import paddle.fluid as fluid
	    import numpy as np
	    input = fluid.data(name="input", shape=[None,3,6,10])

	    #1
	    output = fluid.layers.resize_bilinear(input=input,out_shape=[12,12])

	    #2
	    #x = np.array([2]).astype("int32")
	    #dim1 = fluid.data(name="dim1", shape=[1], dtype="int32")
	    #fluid.layers.assign(input=x, output=dim1)
	    #output = fluid.layers.resize_bilinear(input=input,out_shape=[12,dim1])

	    #3
	    #x = np.array([3,12]).astype("int32")
	    #shape_tensor = fluid.data(name="shape_tensor", shape=[2], dtype="int32")
	    #fluid.layers.assign(input=x, output=shape_tensor)
	    #output = fluid.layers.resize_bilinear(input=input,out_shape=shape_tensor)

	    #4
	    #x = np.array([0.5]).astype("float32")
	    #scale_tensor = fluid.data(name="scale", shape=[1], dtype="float32")
	    #fluid.layers.assign(x,scale_tensor)
	    #output = fluid.layers.resize_bilinear(input=input,scale=scale_tensor)

	    place = fluid.CPUPlace()
	    exe = fluid.Executor(place)
	    exe.run(fluid.default_startup_program())
 
	    input_data = np.random.rand(2,3,6,10).astype("float32")
6875

R
ruri 已提交
6876 6877 6878 6879 6880 6881
	    output_data = exe.run(fluid.default_main_program(),
                feed={"input":input_data},
                fetch_list=[output],
                return_numpy=True)
 
	    print(output_data[0].shape)
6882

R
ruri 已提交
6883 6884 6885 6886 6887 6888 6889 6890
	    #1
	    # (2, 3, 12, 12)
	    #2
	    # (2, 3, 12, 2)
	    #3
	    # (2, 3, 3, 12)
	    #4
	    # (2, 3, 3, 5)
6891

R
ruri 已提交
6892 6893
	    #imperative mode
	    import paddle.fluid.dygraph as dg
6894

R
ruri 已提交
6895 6896 6897 6898
	    with dg.guard(place) as g:
    		input = dg.to_variable(input_data)
    		output = fluid.layers.resize_bilinear(input=input, out_shape=[12,12])
    		print(output.shape)
6899

R
ruri 已提交
6900
		# [2L, 3L, 12L, 12L]
6901

6902 6903
    """

6904
    return image_resize(input, out_shape, scale, name, 'BILINEAR', actual_shape,
6905
                        align_corners, align_mode, data_format)
6906 6907


K
Kaipeng Deng 已提交
6908 6909 6910 6911 6912 6913 6914
@templatedoc(op_type="trilinear_interp")
def resize_trilinear(input,
                     out_shape=None,
                     scale=None,
                     name=None,
                     actual_shape=None,
                     align_corners=True,
6915 6916
                     align_mode=1,
                     data_format='NCDHW'):
K
Kaipeng Deng 已提交
6917
    """
R
ruri 已提交
6918
    This op resizes the input by performing trilinear interpolation based on given
K
Kaipeng Deng 已提交
6919 6920 6921
    output shape which specified by actual_shape, out_shape and scale
    in priority order.

6922 6923 6924
    **Warning:** the parameter :attr:`actual_shape` will be deprecated 
    in the future and only use :attr:`out_shape` instead.

K
Kaipeng Deng 已提交
6925 6926 6927 6928 6929 6930 6931 6932 6933 6934 6935 6936 6937 6938 6939 6940 6941 6942 6943 6944 6945 6946 6947 6948 6949 6950 6951 6952
    Trilinear interpolation is an extension of linear interpolation for 
    interpolating functions of three variables (e.g. D-direction, 
    H-direction and W-direction in this op) on a rectilinear 3D grid. 
    The linear interpolation is performed on three directions.

    For details of trilinear interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Trilinear_interpolation

    Align_corners and align_mode are optinal parameters,the calculation 
    method of interpolation can be selected by them.

    Example:

    .. code-block:: text

        For scale:
          
            if align_corners = True && out_size > 1 :

              scale_factor = (in_size-1.0)/(out_size-1.0)
            
            else:
              
              scale_factor = float(in_size/out_size)     

        Bilinear interpolation:

          if:
6953

K
Kaipeng Deng 已提交
6954 6955 6956 6957 6958 6959 6960 6961 6962 6963 6964 6965 6966 6967 6968 6969 6970 6971
              align_corners = False , align_mode = 0
              
              input : (N,C,D_in,H_in,W_in)
              output: (N,C,D_out,H_out,W_out) where:
              
              D_out = (D_{in}+0.5) * scale_{factor} - 0.5
              H_out = (H_{in}+0.5) * scale_{factor} - 0.5
              W_out = (W_{in}+0.5) * scale_{factor} - 0.5

          else:

              input : (N,C,D_in,H_in,W_in)
              output: (N,C,D_out,H_out,W_out) where:

              D_out = D_{in} * scale_{factor}
              H_out = H_{in} * scale_{factor}
              W_out = W_{in} * scale_{factor}

R
ruri 已提交
6972
    Parameters:
6973 6974
        input(${x_type}): 5-D Tensor, its data type is float32, float64, or uint8,
                          its data format is specified by :attr:`data_format`.
R
ruri 已提交
6975
        out_shape(list|tuple|Variable|None): The output shape of resized tensor, the shape is (out_d, out_h, out_w). Default: None. Every element should be an integer or a Tensor Variable with shape: [1] if it is a list. If it is a Tensor Variable, its dimension size should be 1.
6976
        scale(float|Variable|None): The multiplier for the input depth, height or width.
K
Kaipeng Deng 已提交
6977 6978 6979
             At least one of :attr:`out_shape` or :attr:`scale` must be set. 
             And :attr:`out_shape` has a higher priority than :attr:`scale`. 
             Default: None.
R
ruri 已提交
6980
        name(str, optional): The default value is None.  Normally there is no need for user to set this property.  For more information, please refer to :ref:`api_guide_Name`
K
Kaipeng Deng 已提交
6981 6982 6983 6984 6985 6986
        actual_shape(Variable): An optional input to specify output shape
                                dynamically. If provided, image resize
                                according to this given shape rather than
                                :attr:`out_shape` and :attr:`scale` specifying
                                shape. That is to say actual_shape has the
                                highest priority. It is recommended to use
6987 6988 6989 6990 6991 6992
                                :attr:`out_shape` if you want to specify output 
                                shape dynamically, because :attr:`actual_shape` 
                                will be deprecated. When using actual_shape to 
                                specify output shape, one of :attr:`out_shape` 
                                and :attr:`scale` should also be set, otherwise 
                                errors would be occured in graph constructing stage.
K
Kaipeng Deng 已提交
6993 6994 6995
                                Default: None
        align_corners(bool): ${align_corners_comment}
        align_mode(bool): ${align_mode_comment}
6996 6997 6998 6999
        data_format (str, optional): Specify the data format of the input, and the data format of the output 
            will be consistent with that of the input. An optional string from: `"NCDHW"`, `"NDHWC"`.
            The default is `"NCDHW"`. When it is `"NCDHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_depth, input_height, input_width]`.
K
Kaipeng Deng 已提交
7000 7001

    Returns:
R
ruri 已提交
7002
        Variable: A 5-D Tensor(NCDHW or NDHWC) 
K
Kaipeng Deng 已提交
7003 7004 7005

    Examples:
        .. code-block:: python
R
ruri 已提交
7006 7007 7008 7009 7010 7011 7012 7013 7014 7015 7016 7017 7018 7019 7020 7021 7022 7023 7024 7025 7026 7027 7028 7029 7030 7031 7032 7033 7034 7035 7036 7037
	
	    #declarative mode
	    import paddle.fluid as fluid
	    import numpy as np
	    input = fluid.data(name="input", shape=[None,3,6,8,10])

	    #1
	    output = fluid.layers.resize_trilinear(input=input,out_shape=[12,12,12])

	    #2
	    #x = np.array([2]).astype("int32")
	    #dim1 = fluid.data(name="dim1", shape=[1], dtype="int32")
	    #fluid.layers.assign(input=x, output=dim1)
	    #output = fluid.layers.resize_trilinear(input=input,out_shape=[12,dim1,4])

	    #3
	    #x = np.array([3,12,12]).astype("int32")
	    #shape_tensor = fluid.data(name="shape_tensor", shape=[3], dtype="int32")
	    #fluid.layers.assign(input=x, output=shape_tensor)
	    #output = fluid.layers.resize_trilinear(input=input,out_shape=shape_tensor)

	    #4
	    #x = np.array([0.5]).astype("float32")
	    #scale_tensor = fluid.data(name="scale", shape=[1], dtype="float32")
	    #fluid.layers.assign(x,scale_tensor)
	    #output = fluid.layers.resize_trilinear(input=input,scale=scale_tensor)

	    place = fluid.CPUPlace()
	    exe = fluid.Executor(place)
	    exe.run(fluid.default_startup_program())
 
	    input_data = np.random.rand(2,3,6,8,10).astype("float32")
K
Kaipeng Deng 已提交
7038

R
ruri 已提交
7039 7040 7041 7042 7043 7044 7045 7046 7047 7048 7049 7050 7051 7052 7053 7054 7055 7056
	    output_data = exe.run(fluid.default_main_program(),
                feed={"input":input_data},
                fetch_list=[output],
                return_numpy=True)
 
	    print(output_data[0].shape)

	    #1
	    # (2, 3, 12, 12, 12)
	    #2
	    # (2, 3, 12, 2, 4)
	    #3
	    # (2, 3, 3, 12, 12)
	    #4
	    # (2, 3, 3, 4, 5)

	    #imperative mode
	    import paddle.fluid.dygraph as dg
7057

R
ruri 已提交
7058 7059 7060 7061
	    with dg.guard(place) as g:
    		input = dg.to_variable(input_data)
    		output = fluid.layers.resize_trilinear(input=input, out_shape=[12,12,12])
    		print(output.shape)
7062

R
ruri 已提交
7063
		# [2L, 3L, 12L, 12L, 12L]
7064 7065 7066



K
Kaipeng Deng 已提交
7067 7068 7069
    """

    return image_resize(input, out_shape, scale, name, 'TRILINEAR',
7070
                        actual_shape, align_corners, align_mode, data_format)
K
Kaipeng Deng 已提交
7071 7072


7073
@templatedoc(op_type="nearest_interp")
7074 7075 7076 7077
def resize_nearest(input,
                   out_shape=None,
                   scale=None,
                   name=None,
7078
                   actual_shape=None,
7079 7080
                   align_corners=True,
                   data_format='NCHW'):
7081
    """
R
ruri 已提交
7082
    This op resizes the input by performing nearest neighbor interpolation in both the
7083 7084
    height direction and the width direction based on given output shape 
    which is specified by actual_shape, out_shape and scale in priority order.
7085

7086 7087 7088
    **Warning:** the parameter :attr:`actual_shape` will be deprecated in the 
    future and only use :attr:`out_shape` instead.

7089 7090
    Example:

T
Tink_Y 已提交
7091 7092 7093 7094 7095 7096 7097 7098 7099 7100 7101 7102
    .. code-block:: text

        For scale:
          
            if align_corners = True && out_size > 1 :
              scale_factor = (in_size-1.0)/(out_size-1.0)
            
            else:
              
              scale_factor = float(in_size/out_size)
          
        Nearest neighbor interpolation:
7103
          
T
Tink_Y 已提交
7104 7105
          if:
              align_corners = False
7106

T
Tink_Y 已提交
7107 7108
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
7109

T
Tink_Y 已提交
7110 7111
              H_out = floor(H_{in} * scale_{factor})
              W_out = floor(W_{in} * scale_{factor})
7112

T
Tink_Y 已提交
7113 7114
          else:
              align_corners = True
7115

T
Tink_Y 已提交
7116 7117
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
7118

T
Tink_Y 已提交
7119 7120
              H_out = round(H_{in} * scale_{factor})
              W_out = round(W_{in} * scale_{factor})
7121 7122


7123
    For details of nearest neighbor interpolation, please refer to Wikipedia:
7124
    https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation
Y
yuyang18 已提交
7125

R
ruri 已提交
7126
    Parameters:
7127 7128
        input(${x_type}): 4-D Tensor, its data type is float32, float64, or uint8,
                          its data format is specified by :attr:`data_format`.
R
ruri 已提交
7129
        out_shape(list|tuple|Variable|None): The output shape of resized tensor, the shape is (out_h, out_w). Default: None. Every element should be an integer or a tensor Variable with shape: [1] if it is a list. If it is a tensor Variable, its dimension size should be 1.
7130
        scale(float|Variable|None): The multiplier for the input height or width. At
D
dengkaipeng 已提交
7131
             least one of :attr:`out_shape` or :attr:`scale` must be set. 
D
dengkaipeng 已提交
7132
             And :attr:`out_shape` has a higher priority than :attr:`scale`. 
R
ruri 已提交
7133 7134 7135
             Default: None. 
        name(str, optional): The default value is None.  Normally there is no need for user to set this property.  For more information, please refer to :ref:`api_guide_Name`
	actual_shape(Variable): An optional input to specify output shape
7136 7137
                                dynamically. If provided, image resize
                                according to this given shape rather than
7138
                                :attr:`out_shape` and :attr:`scale` specifying
7139 7140
                                shape. That is to say actual_shape has the
                                highest priority. It is recommended to use
7141 7142 7143 7144 7145 7146
                                :attr:`out_shape` if you want to specify output 
                                shape dynamically, because :attr:`actual_shape` 
                                will be deprecated. When using actual_shape to 
                                specify output shape, one of :attr:`out_shape` 
                                and :attr:`scale` should also be set, otherwise 
                                errors would be occured in graph constructing stage.
7147
                                Default: None
7148
        align_corners(bool): ${align_corners_comment}
7149 7150 7151 7152
        data_format (str, optional): Specify the data format of the input, and the data format of the output 
            will be consistent with that of the input. An optional string from: `"NCHW"`, `"NHWC"`.
            The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`.
Y
yuyang18 已提交
7153 7154

    Returns:
R
ruri 已提交
7155
	Variable: 4-D tensor(NCHW or NHWC).
7156 7157 7158

    Examples:
        .. code-block:: python
R
ruri 已提交
7159 7160 7161 7162 7163 7164 7165 7166 7167 7168 7169 7170 7171 7172 7173 7174 7175 7176 7177 7178 7179 7180 7181 7182 7183 7184 7185 7186 7187 7188 7189 7190
	
	    #declarative mode
	    import paddle.fluid as fluid
	    import numpy as np
	    input = fluid.data(name="input", shape=[None,3,6,10])

	    #1
	    output = fluid.layers.resize_nearest(input=input,out_shape=[12,12])

	    #2
	    #x = np.array([2]).astype("int32")
	    #dim1 = fluid.data(name="dim1", shape=[1], dtype="int32")
	    #fluid.layers.assign(input=x, output=dim1)
	    #output = fluid.layers.resize_nearest(input=input,out_shape=[12,dim1])

	    #3
	    #x = np.array([3,12]).astype("int32")
	    #shape_tensor = fluid.data(name="shape_tensor", shape=[2], dtype="int32")
	    #fluid.layers.assign(input=x, output=shape_tensor)
	    #output = fluid.layers.resize_nearest(input=input,out_shape=shape_tensor)

	    #4
	    #x = np.array([0.5]).astype("float32")
	    #scale_tensor = fluid.data(name="scale", shape=[1], dtype="float32")
	    #fluid.layers.assign(x,scale_tensor)
	    #output = fluid.layers.resize_nearest(input=input,scale=scale_tensor)

	    place = fluid.CPUPlace()
	    exe = fluid.Executor(place)
	    exe.run(fluid.default_startup_program())
 
	    input_data = np.random.rand(2,3,6,10).astype("float32")
7191

R
ruri 已提交
7192 7193 7194 7195 7196 7197 7198 7199 7200 7201 7202 7203 7204 7205 7206
	    output_data = exe.run(fluid.default_main_program(),
                feed={"input":input_data},
                fetch_list=[output],
                return_numpy=True)
 
	    print(output_data[0].shape)

	    #1
	    # (2, 3, 12, 12)
	    #2
	    # (2, 3, 12, 2)
	    #3
	    # (2, 3, 3, 12)
	    #4
	    # (2, 3, 3, 5)
7207

R
ruri 已提交
7208 7209
	    #imperative mode
	    import paddle.fluid.dygraph as dg
7210

R
ruri 已提交
7211 7212 7213 7214 7215 7216
	    with dg.guard(place) as g:
    		input = dg.to_variable(input_data)
    		output = fluid.layers.resize_nearest(input=input, out_shape=[12,12])
    		print(output.shape)

		# [2L, 3L, 12L, 12L]
7217 7218 7219



7220 7221
    """

7222 7223 7224 7225 7226 7227 7228 7229 7230 7231
    return image_resize(
        input,
        out_shape,
        scale,
        name,
        'NEAREST',
        actual_shape,
        align_corners,
        align_mode=1,
        data_format=data_format)
7232 7233 7234 7235


def image_resize_short(input, out_short_len, resample='BILINEAR'):
    """
R
ruri 已提交
7236
    This op resizes a batch of images. The short edge of input images will be
7237 7238
    resized to the given 'out_short_len'. The long edge of input images
    will be resized proportionately to make images' length-width ratio
7239 7240
    constant.

R
ruri 已提交
7241 7242
    Parameters:
        input (Variable): 4-D tensor(NCHW), The input tensor of image resize layer.
7243
        out_short_len(int): The length of output images' short edge.
7244
        resample (str): resample method, default: BILINEAR.
F
fengjiayi 已提交
7245

7246
    Returns:
R
ruri 已提交
7247
        Variable: 4-D tensor(NCHW).
R
ruri 已提交
7248 7249 7250 7251

    Examples:
        .. code-block:: python

7252
            import paddle.fluid as fluid
R
ruri 已提交
7253
            input = fluid.data(name="input", shape=[None,3,6,9], dtype="float32")
R
ruri 已提交
7254
            out = fluid.layers.image_resize_short(input, out_short_len=3)
7255 7256 7257 7258 7259 7260 7261 7262 7263 7264
    """
    in_shape = input.shape
    if len(in_shape) != 4:
        raise ValueError(
            "The rank of input must be 4 (num_batches, channels, in_h, in_w).")
    hw = in_shape[2:4]
    short_idx = hw.index(min(hw))
    long_idx = 1 - short_idx
    out_shape = list(hw)
    out_shape[short_idx] = out_short_len
F
fengjiayi 已提交
7265 7266 7267
    out_shape[long_idx] = int(
        float(out_shape[long_idx]) * (float(out_short_len) / float(hw[
            short_idx])) + 0.5)
7268 7269 7270
    return image_resize(input=input, out_shape=out_shape, resample=resample)


7271
def gather(input, index, overwrite=True):
W
whs 已提交
7272
    """
Q
qiaolongfei 已提交
7273 7274
    **Gather Layer**

7275
    Output is obtained by gathering entries of the outer-most dimension
W
whs 已提交
7276 7277 7278 7279
    of X indexed by `index` and concatenate them together.

    .. math::

7280
        Out = X[Index]
W
whs 已提交
7281 7282 7283 7284 7285 7286 7287


    .. code-block:: text


                Given:

7288 7289
                X = [[1, 2],
                     [3, 4],
W
whs 已提交
7290 7291 7292 7293 7294 7295 7296 7297 7298 7299
                     [5, 6]]

                Index = [1, 2]

                Then:

                Out = [[3, 4],
                       [5, 6]]

    Args:
Y
Yibing Liu 已提交
7300 7301 7302 7303 7304
        input (Variable): The source input tensor with rank>=1. Supported data type is 
            int32, int64, float32, float64 and uint8 (only for CPU), 
            float16 (only for GPU).
        index (Variable): The index input tensor with rank=1. Data type is int32 or int64.
        overwrite (bool, optional): The mode that updating the grad when has same index.
7305 7306 7307 7308 7309
            If True, use the overwrite mode to update the grad of the same index,
	    if False, use the accumulate mode to update the grad of the same index. 
	    Default value is True.
	    

W
whs 已提交
7310 7311 7312 7313 7314

    Returns:
        output (Variable): The output is a tensor with the same rank as input.

    Examples:
W
whs 已提交
7315

W
whs 已提交
7316 7317
        .. code-block:: python

7318
            import paddle.fluid as fluid
Y
Yibing Liu 已提交
7319 7320
            x = fluid.data(name='x', shape=[-1, 5], dtype='float32')
            index = fluid.data(name='index', shape=[-1, 1], dtype='int32')
W
whs 已提交
7321 7322 7323 7324
            output = fluid.layers.gather(x, index)
    """
    helper = LayerHelper('gather', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
7325
    out = helper.create_variable_for_type_inference(dtype)
W
whs 已提交
7326 7327 7328 7329
    helper.append_op(
        type="gather",
        inputs={"X": input,
                "Index": index},
7330 7331
        outputs={"Out": out},
        attrs={'overwrite': overwrite})
W
whs 已提交
7332 7333 7334
    return out


7335 7336 7337 7338 7339 7340 7341 7342 7343 7344 7345 7346 7347 7348 7349 7350 7351 7352 7353 7354 7355 7356 7357 7358 7359 7360 7361 7362 7363 7364 7365 7366 7367 7368 7369 7370 7371 7372 7373 7374 7375 7376 7377 7378 7379 7380 7381 7382 7383 7384 7385 7386
def gather_nd(input, index, name=None):
    """
    **Gather Nd Layer**

    This function is actually a high-dimensional extension of :code:`gather` 
    and supports for simultaneous indexing by multiple axes. :attr:`index` is a 
    K-dimensional integer tensor, which is regarded as a (K-1)-dimensional 
    tensor of :attr:`index` into :attr:`input`, where each element defines 
    a slice of params:

    .. math::

        output[(i_0, ..., i_{K-2})] = input[index[(i_0, ..., i_{K-2})]]

    Obviously, :code:`index.shape[-1] <= input.rank` . And, the output tensor has
    shape :code:`index.shape[:-1] + input.shape[index.shape[-1]:]` .

    .. code-block:: text

            Given:
                input = [[[ 0,  1,  2,  3],
                          [ 4,  5,  6,  7],
                          [ 8,  9, 10, 11]],
                         [[12, 13, 14, 15],
                          [16, 17, 18, 19],
                          [20, 21, 22, 23]]]
                input.shape = (2, 3, 4)

            * Case 1:
                index = [[1]]
                
                gather_nd(input, index)  
                         = [input[1, :, :]] 
                         = [[12, 13, 14, 15],
                            [16, 17, 18, 19],
                            [20, 21, 22, 23]]

            * Case 2:
                index = [[0,2]]

                gather_nd(input, index)
                         = [input[0, 2, :]]
                         = [8, 9, 10, 11]

            * Case 3:
                index = [[1, 2, 3]]

                gather_nd(input, index)
                         = [input[1, 2, 3]]
                         = [23]

    Args:
7387 7388 7389
        input (Variable): The source input. Its dtype should be int32, int64, float32, float64.
        index (Variable): The index input with rank > 1, index.shape[-1] <= input.rank.
                          Its dtype should be int32, int64.
7390
        name (str|None): A name for this layer(optional). If set None, the
7391
                         layer will be named automatically.
7392 7393 7394 7395 7396 7397 7398 7399 7400

    Returns:
        output (Variable): A tensor with the shape index.shape[:-1] + input.shape[index.shape[-1]:]

    Examples:

        .. code-block:: python

            import paddle.fluid as fluid
7401 7402
            x = fluid.data(name='x', shape=[3, 4, 5], dtype='float32')
            index = fluid.data(name='index', shape=[2, 2], dtype='int32')
7403 7404 7405 7406 7407 7408 7409 7410 7411 7412 7413 7414 7415 7416 7417 7418 7419 7420
            output = fluid.layers.gather_nd(x, index)

    """
    helper = LayerHelper('gather_nd', **locals())
    dtype = helper.input_dtype()
    if name is None:
        output = helper.create_variable_for_type_inference(dtype)
    else:
        output = helper.create_variable(
            name=name, dtype=dtype, persistable=False)
    helper.append_op(
        type="gather_nd",
        inputs={"X": input,
                "Index": index},
        outputs={"Out": output})
    return output


7421
def scatter(input, index, updates, name=None, overwrite=True):
7422 7423 7424
    """
    **Scatter Layer**

7425
    Output is obtained by updating the input on selected indices based on updates.
7426

7427 7428 7429 7430 7431 7432 7433 7434 7435 7436 7437 7438 7439 7440 7441 7442 7443 7444 7445 7446 7447 7448 7449 7450
    .. code-block:: python
        import numpy as np
                
        #input:
        input = np.array([[1, 1], [2, 2], [3, 3]])
        index = np.array([2, 1, 0, 1])
        # shape of updates should be the same as input
        # shape of updates with dim > 1 should be the same as input
        updates = np.array([[1, 1], [2, 2], [3, 3], [4, 4]])
        overwrite = False

        # calculation:
        if not overwrite:
            for i in range(len(index)):
                input[index[i]] = np.zeros((2))

        for i in range(len(index)):
            if (overwrite):
                input[index[i]] = updates[i]
            else:
                input[index[i]] += updates[i]
        # output:
        out = np.array([[3, 3], [6, 6], [1, 1]])
        out.shape # [3, 2]
7451 7452

    Args:
7453 7454 7455 7456 7457
        input (Variable): The input N-D Tensor with rank>=1. Data type can be float32.
        index (Variable): The index 1-D Tensor. Data type can be int32, int64. The length of index cannot exceed updates's length, and the value in index cannot exceed input's length.
        updates (Variable): update input with updates parameter based on index. shape should be the same as input, and dim value with dim > 1 shoule be the same as input.
        name(str, optional): The default value is None.  Normally there is no need for user to set this property.  For more information, please refer to :ref:`api_guide_Name` .
        overwrite (bool): The mode that updating the output when there are same indices.
7458 7459
            If True, use the overwrite mode to update the output of the same index,
	    if False, use the accumulate mode to update the output of the same index. 
7460
	    Default value is True.
7461 7462

    Returns:
7463
        Variable(Tensor|LoDTensor): The output is a Tensor with the same shape as input.
7464 7465 7466 7467 7468

    Examples:

        .. code-block:: python

7469
            import numpy as np
7470 7471
            import paddle.fluid as fluid

7472 7473 7474
            input = fluid.layers.data(name='data', shape=[3, 2], dtype='float32', append_batch_size=False)
            index = fluid.layers.data(name='index', shape=[4], dtype='int64', append_batch_size=False)
            updates = fluid.layers.data(name='update', shape=[4, 2], dtype='float32', append_batch_size=False)
7475

7476 7477 7478 7479 7480 7481 7482 7483 7484 7485 7486 7487 7488 7489
            output = fluid.layers.scatter(input, index, updates, overwrite=False)

            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())

            in_data = np.array([[1, 1], [2, 2], [3, 3]]).astype(np.float32)
            index_data = np.array([2, 1, 0, 1]).astype(np.int64)
            update_data = np.array([[1, 1], [2, 2], [3, 3], [4, 4]]).astype(np.float32)

            res = exe.run(fluid.default_main_program(), feed={'data':in_data, "index":index_data, "update":update_data}, fetch_list=[output])
            print(res)
            # [array([[3., 3.],
            #   [6., 6.],
            #   [1., 1.]], dtype=float32)]
7490 7491 7492
    """
    helper = LayerHelper('scatter', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
7493
    out = helper.create_variable_for_type_inference(dtype)
7494 7495 7496 7497 7498
    helper.append_op(
        type="scatter",
        inputs={"X": input,
                "Ids": index,
                "Updates": updates},
7499
        attrs={'overwrite': overwrite},
7500 7501 7502 7503
        outputs={"Out": out})
    return out


7504 7505 7506 7507 7508
def scatter_nd_add(ref, index, updates, name=None):
    """
    **Scatter_nd_add Layer**

    Output is obtained by applying sparse addition to a single value
7509 7510 7511
    or slice in a Variable. 

    :attr:`ref` is a Tensor with rank :math:`R` 
7512 7513 7514 7515
    and :attr:`index` is a Tensor with rank :math:`K` . Thus, :attr:`index` 
    has shape :math:`[i_0, i_1, ..., i_{K-2}, Q]` where :math:`Q \leq R` . :attr:`updates` 
    is a Tensor with rank :math:`K - 1 + R - Q` and its
    shape is :math:`index.shape[:-1] + ref.shape[index.shape[-1]:]` .
7516

7517 7518 7519 7520 7521 7522 7523 7524 7525 7526 7527 7528 7529 7530 7531 7532 7533 7534 7535 7536 7537 7538 7539 7540 7541 7542 7543 7544 7545 7546 7547
    According to the :math:`[i_0, i_1, ..., i_{K-2}]` of :attr:`index` ,
    add the corresponding :attr:`updates` slice to the :attr:`ref` slice
    which is obtained by the last one dimension of :attr:`index` .

    .. code-block:: text
        
        Given:

        * Case 1:
            ref = [0, 1, 2, 3, 4, 5]
            index = [[1], [2], [3], [1]]
            updates = [9, 10, 11, 12]

          we get:
             
            output = [0, 22, 12, 14, 4, 5]

        * Case 2:
            ref = [[65, 17], [-14, -25]]
            index = [[], []]
            updates = [[[-1, -2], [1, 2]],
                       [[3, 4], [-3, -4]]]
            ref.shape = (2, 2)
            index.shape = (2, 0)
            updates.shape = (2, 2, 2)

          we get:
             
            output = [[67, 19], [-16, -27]]

    Args:
7548
        ref (Variable): The ref input. Its dtype should be int32, int64, float32, float64.
7549 7550
        index (Variable): The index input with rank > 1 and index.shape[-1] <= ref.rank.
                          Its dtype should be int32 or int64 as it is used as indexes.
7551 7552 7553
        updates (Variable): The updated value of scatter_nd_add op, and it must have the same dtype
                            as ref. It must have the shape index.shape[:-1] + ref.shape[index.shape[-1]:].
        name (str|None): The output variable name. If set None, the layer will be named automatically.
7554 7555

    Returns:
7556
        output (Variable): The output is a tensor with the same shape and dtype as ref.
7557 7558 7559 7560 7561 7562 7563

    Examples:

        .. code-block:: python

            import paddle.fluid as fluid

7564 7565 7566
            ref = fluid.data(name='ref', shape=[3, 5, 9, 10], dtype='float32')
            index = fluid.data(name='index', shape=[3, 2], dtype='int32')
            updates = fluid.data(name='update', shape=[3, 9, 10], dtype='float32')
7567 7568 7569 7570 7571 7572 7573

            output = fluid.layers.scatter_nd_add(ref, index, updates)
    """
    if ref.dtype != updates.dtype:
        raise ValueError("ref and updates must have same data type.")

    helper = LayerHelper('scatter_nd_add', **locals())
7574
    dtype = helper.input_dtype(input_param_name='ref')
7575 7576 7577 7578 7579 7580 7581 7582 7583 7584 7585 7586 7587 7588 7589 7590 7591 7592 7593 7594 7595 7596 7597 7598 7599 7600 7601 7602 7603 7604
    if name is None:
        output = helper.create_variable_for_type_inference(dtype)
    else:
        output = helper.create_variable(
            name=name, dtype=dtype, persistable=False)
    helper.append_op(
        type="scatter_nd_add",
        inputs={"X": ref,
                "Index": index,
                "Updates": updates},
        outputs={"Out": output})
    return output


def scatter_nd(index, updates, shape, name=None):
    """
    **Scatter_nd Layer**

    Output is obtained by scattering the :attr:`updates` in a new tensor according 
    to :attr:`index` . This op is similar to :code:`scatter_nd_add`, except the 
    tensor of :attr:`shape` is zero-initialized. Correspondingly, :code:`scatter_nd(index, updates, shape)` 
    is equal to :code:`scatter_nd_add(fluid.layers.zeros(shape, updates.dtype), index, updates)` . 
    If :attr:`index` has repeated elements, then the corresponding updates are accumulated. 
    Because of the numerical approximation issues, the different order of repeated elements 
    in :attr:`index` may cause different results. The specific calculation method can be 
    seen :code:`scatter_nd_add` . This op is the inverse of the :code:`gather_nd` op.

    Args:
        index (Variable): The index input with rank > 1 and index.shape[-1] <= len(shape).
                          Its dtype should be int32 or int64 as it is used as indexes.
7605
        updates (Variable): The updated value of scatter_nd op. Its dtype should be int32, int64, float32, float64.
7606 7607
                            It must have the shape index.shape[:-1] + shape[index.shape[-1]:]
        shape(tuple|list): Shape of output tensor.
7608
        name (str|None): The output variable name. If set None, the layer will be named automatically.
7609 7610 7611 7612 7613 7614 7615 7616 7617 7618

    Returns:
        output (Variable): The output is a tensor with the same type as :attr:`updates` .

    Examples:

        .. code-block:: python

            import paddle.fluid as fluid

7619 7620
            index = fluid.data(name='index', shape=[3, 2], dtype='int64')
            updates = fluid.data(name='update', shape=[3, 9, 10], dtype='float32')
7621 7622 7623 7624 7625 7626 7627
            shape = [3, 5, 9, 10]

            output = fluid.layers.scatter_nd(index, updates, shape)
    """
    return scatter_nd_add(zeros(shape, updates.dtype), index, updates, name)


Y
yuyang18 已提交
7628 7629 7630 7631 7632 7633 7634 7635 7636 7637 7638 7639 7640
@templatedoc()
def random_crop(x, shape, seed=None):
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        shape(${shape_type}): ${shape_comment}
        seed(int|${seed_type}|None): ${seed_comment} By default, the seed will
            get from `random.randint(-65536, 65535)`.

    Returns:
        ${out_comment}
7641

7642
    Examples:
Q
qingqing01 已提交
7643 7644 7645 7646 7647 7648 7649 7650 7651 7652 7653 7654 7655
        .. code-block:: python

            import paddle.fluid as fluid
            img = fluid.data("img", [None, 3, 256, 256])
            # cropped_img is [-1, 3, 224, 224]
            cropped_img = fluid.layers.random_crop(img, shape=[3, 224, 224])

            # cropped_img2 shape: [-1, 2, 224, 224]
            # cropped_img2 = fluid.layers.random_crop(img, shape=[2, 224, 224])

            # cropped_img3 shape: [-1, 3, 128, 224]
            # cropped_img3 = fluid.layers.random_crop(img, shape=[128, 224])

Y
yuyang18 已提交
7656
    """
F
stash  
fengjiayi 已提交
7657
    helper = LayerHelper("random_crop", **locals())
F
fengjiayi 已提交
7658
    dtype = x.dtype
X
Xin Pan 已提交
7659
    out = helper.create_variable_for_type_inference(dtype)
Y
yuyang18 已提交
7660
    if seed is None:
7661
        seed = np.random.randint(-65536, 65536)
F
fengjiayi 已提交
7662
    op_attrs = {"shape": shape}
F
stash  
fengjiayi 已提交
7663
    if isinstance(seed, int):
F
fengjiayi 已提交
7664 7665 7666 7667 7668
        op_attrs["startup_seed"] = seed
        seed = helper.create_variable(
            name=unique_name.generate("random_crop_seed"),
            dtype="int64",
            persistable=True)
F
stash  
fengjiayi 已提交
7669 7670 7671 7672
    elif not isinstance(seed, Variable):
        raise ValueError("'seed' must be a Variable or an int.")
    helper.append_op(
        type="random_crop",
F
fix  
fengjiayi 已提交
7673
        inputs={"X": x,
F
stash  
fengjiayi 已提交
7674 7675
                "Seed": seed},
        outputs={"Out": out,
F
fengjiayi 已提交
7676 7677
                 "SeedOut": seed},
        attrs=op_attrs)
F
stash  
fengjiayi 已提交
7678
    return out
W
whs 已提交
7679 7680


7681
def log(x, name=None):
W
wanghaoshuang 已提交
7682 7683 7684 7685 7686
    """
    Calculates the natural log of the given input tensor, element-wise.

    .. math::

7687
        Out = \\ln(x)
W
wanghaoshuang 已提交
7688 7689

    Args:
W
Wilber 已提交
7690 7691 7692
        x (Variable): Input LoDTensor or Tensor. Must be one of the following types: float32, float64.
        name (str|None): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`
    
W
wanghaoshuang 已提交
7693 7694

    Returns:
W
Wilber 已提交
7695
        Variable: The natural log of the input LoDTensor or Tensor computed element-wise.
W
wanghaoshuang 已提交
7696 7697 7698 7699 7700

    Examples:

        .. code-block:: python

7701
            import paddle.fluid as fluid
W
Wilber 已提交
7702 7703 7704 7705 7706 7707 7708 7709 7710 7711 7712 7713 7714
            import numpy as np

            # Graph Organizing
            x = fluid.layers.data(name="x", shape=[1], dtype="float32")
            res = fluid.layers.log(x)

            # Create an executor using CPU as an example
            exe = fluid.Executor(fluid.CPUPlace())

            # Execute
            x_i = np.array([[1], [2]]).astype(np.float32)
            res_val, = exe.run(fluid.default_main_program(), feed={'x':x_i}, fetch_list=[res])
            print(res_val) # [[0.], [0.6931472]]
W
wanghaoshuang 已提交
7715 7716
    """
    helper = LayerHelper('log', **locals())
W
wanghaoshuang 已提交
7717
    dtype = helper.input_dtype(input_param_name='x')
X
Xin Pan 已提交
7718
    out = helper.create_variable_for_type_inference(dtype)
W
wanghaoshuang 已提交
7719
    helper.append_op(type="log", inputs={"X": x}, outputs={"Out": out})
W
wanghaoshuang 已提交
7720 7721 7722
    return out


Z
zhupengyang 已提交
7723
@templatedoc()
7724
def relu(x, name=None):
W
wanghaoshuang 已提交
7725
    """
Z
zhupengyang 已提交
7726
    ${comment}
W
wanghaoshuang 已提交
7727 7728

    Args:
Z
zhupengyang 已提交
7729 7730 7731 7732
        x(Variable): ${x_comment}
        name(str, optional): The default value is None. Normally there is no
            need for user to set this property. For more information, please
            refer to :ref:`api_guide_Name`.
W
wanghaoshuang 已提交
7733 7734

    Returns:
Z
zhupengyang 已提交
7735
        Variable: ${out_comment}
W
wanghaoshuang 已提交
7736 7737 7738 7739 7740

    Examples:

        .. code-block:: python

7741
            import paddle.fluid as fluid
Z
zhupengyang 已提交
7742 7743 7744 7745 7746 7747 7748 7749 7750
            import numpy as np
            in1 = np.array([[-1,0],[1,2.6]])
            with fluid.dygraph.guard():
                x1 = fluid.dygraph.to_variable(in1)
                out1 = fluid.layers.relu(x1)
                print(out1.numpy())
                # [[0.  0. ]
                #  [1.  2.6]]
"""
W
wanghaoshuang 已提交
7751
    helper = LayerHelper('relu', **locals())
W
wanghaoshuang 已提交
7752
    dtype = helper.input_dtype(input_param_name='x')
X
Xin Pan 已提交
7753
    out = helper.create_variable_for_type_inference(dtype)
X
Xin Pan 已提交
7754 7755
    helper.append_op(
        type="relu", inputs={"X": helper.input('x')}, outputs={"Out": out})
W
wanghaoshuang 已提交
7756
    return out
7757 7758


C
chengduo 已提交
7759 7760
def selu(x, scale=None, alpha=None, name=None):
    """
7761 7762 7763 7764 7765 7766 7767 7768 7769 7770 7771 7772 7773 7774
    Selu Operator.

    The equation is:
    
    .. math::
        selu= \\lambda*
        \\begin{cases}
            x                      &\\quad \\text{ if } x>0 \n
            \\alpha * e^x - \\alpha  &\\quad \\text{ if } x<=0
        \\end{cases}
    

    The input `X` can carry the LoD (Level of Details) information,
    or not. And the output shares the LoD information with input `X`.
C
chengduo 已提交
7775 7776

    Args:
7777 7778
        x (Variable): The input N-D Tensor.
        scale(float, optional): lambda in selu activation function,
C
chengduo 已提交
7779 7780 7781
            the default value is 1.0507009873554804934193349852946.
            For more information about this value, please refer
            to: https://arxiv.org/abs/1706.02515.
7782
        alpha(float, optional): alpha in selu activation function,
C
chengduo 已提交
7783 7784 7785
            the default value is 1.6732632423543772848170429916717.
            For more information about this value, please refer
            to: https://arxiv.org/abs/1706.02515.
7786 7787
        name(str, optional): The default value is None.  Normally there is no need for user to set this property.  For more information, please refer to :ref:`api_guide_Name` .

C
chengduo 已提交
7788 7789

    Returns:
7790
        Variable(Tensor|LoDTensor): The output Tensor or LoDTensor with the same shape and LoD information as input.
C
chengduo 已提交
7791 7792 7793 7794

    Examples:

        .. code-block:: python
7795 7796
             
            import paddle.fluid as fluid
7797 7798 7799 7800 7801 7802 7803 7804 7805 7806 7807 7808
            import numpy as np

            inputs = fluid.layers.data(name="x", shape=[2, 2], dtype="float32")
            output = fluid.layers.selu(inputs)

            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())

            img = np.array([[0, 1],[2, 3]]).astype(np.float32)

            res = exe.run(fluid.default_main_program(), feed={'x':img}, fetch_list=[output])
            print(res) # [array([[0.      , 1.050701],[2.101402, 3.152103]], dtype=float32)]
C
chengduo 已提交
7809 7810 7811 7812 7813 7814 7815 7816 7817 7818 7819 7820 7821 7822 7823
    """
    helper = LayerHelper('selu', **locals())
    dtype = helper.input_dtype(input_param_name='x')
    out = helper.create_variable_for_type_inference(dtype)
    attrs = {}
    if scale is not None:
        attrs["scale"] = scale
    if alpha is not None:
        attrs["alpha"] = alpha

    helper.append_op(
        type="selu", inputs={"X": x}, outputs={"Out": out}, attrs=attrs)
    return out


W
whs 已提交
7824 7825 7826
def mean_iou(input, label, num_classes):
    """
    Mean Intersection-Over-Union is a common evaluation metric for
7827 7828 7829 7830
    semantic image segmentation, which first computes the IOU for each
    semantic class and then computes the average over classes.
    IOU is defined as follows:

W
whs 已提交
7831
    .. math::
7832

H
haowang101779990 已提交
7833
        IOU = \\frac{true\_positive}{(true\_positive + false\_positive + false\_negative)}.
W
whs 已提交
7834

7835
    The predictions are accumulated in a confusion matrix and mean-IOU
W
whs 已提交
7836 7837 7838
    is then calculated from it.


L
Liufang Sang 已提交
7839 7840
    Parameters:
        input (Variable): A n-D Tensor of prediction results for semantic labels with type int32 or int64.
7841
        label (Variable): A Tensor of ground truth labels with type int32 or int64.
W
whs 已提交
7842
                           Its shape should be the same as input.
L
Liufang Sang 已提交
7843 7844 7845 7846 7847 7848 7849 7850 7851 7852 7853 7854
        num_classes (int32): The possible number of labels.

    Returns: 
	Three Variables.

        - mean_iou(Variable) : A 1-D Tensor representing the mean intersection-over-union with shape [1]. \
			    Data type is float32.
        - out_wrong(Variable) : A 1-D Tensor with shape [num_classes]. Data type is int32. \
			     The wrong numbers of each class.
        - out_correct(Variable): A 1-D  Tensor with shape [num_classes]. Data type is int32. The correct numbers of each class.
 
   
W
whs 已提交
7855 7856 7857
    Examples:

        .. code-block:: python
7858

B
Bai Yifan 已提交
7859
            import paddle.fluid as fluid
L
Liufang Sang 已提交
7860
            iou_shape = [None, 32, 32]
7861
            num_classes = 5
L
Liufang Sang 已提交
7862 7863 7864
            predict = fluid.data(name='predict', shape=iou_shape, dtype='int64')
            label = fluid.data(name='label', shape=iou_shape, dtype='int64')
            mean_iou, out_wrong, out_correct = fluid.layers.mean_iou(predict, label,
7865
                                                          num_classes)
W
whs 已提交
7866 7867 7868
    """
    helper = LayerHelper('mean_iou', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
7869 7870 7871
    out_mean_iou = helper.create_variable_for_type_inference(dtype='float32')
    out_wrong = helper.create_variable_for_type_inference(dtype='int32')
    out_correct = helper.create_variable_for_type_inference(dtype='int32')
W
whs 已提交
7872 7873
    helper.append_op(
        type="mean_iou",
W
whs 已提交
7874 7875
        inputs={"Predictions": input,
                "Labels": label},
W
whs 已提交
7876
        outputs={
W
whs 已提交
7877 7878 7879
            "OutMeanIou": out_mean_iou,
            "OutWrong": out_wrong,
            "OutCorrect": out_correct
W
whs 已提交
7880 7881 7882
        },
        attrs={"num_classes": num_classes})
    return out_mean_iou, out_wrong, out_correct
7883 7884 7885 7886 7887 7888


def crop(x, shape=None, offsets=None, name=None):
    """
    Crop input into output, as specified by offsets and shape.

S
SunGaofeng 已提交
7889 7890
    **Warning:** THIS OP IS DEPRECATED. It will be removed in the future version.
    Instructions for updating: Use :ref:`api_fluid_layers_crop_tensor` instead.
7891

7892 7893 7894 7895 7896 7897 7898 7899 7900 7901 7902 7903 7904 7905 7906 7907 7908 7909 7910 7911 7912 7913 7914 7915 7916 7917 7918 7919
    .. code-block:: text

        * Case 1:
            Given
                X = [[0, 1, 2, 0, 0]
                     [0, 3, 4, 0, 0]
                     [0, 0, 0, 0, 0]],
            and
                shape = [2, 2],
                offsets = [0, 1],
            output is:
                Out = [[1, 2],
                       [3, 4]].
        * Case 2:
            Given
                X = [[0, 1, 2, 5, 0]
                     [0, 3, 4, 6, 0]
                     [0, 0, 0, 0, 0]],
            and shape is tensor
                shape = [[0, 0, 0]
                         [0, 0, 0]]
            and
                offsets = [0, 1],

            output is:
                Out = [[1, 2, 5],
                       [3, 4, 6]].

S
SunGaofeng 已提交
7920 7921 7922 7923 7924 7925
    Parameters:
        x (Variable): Tensor, data type can be float32 or float64.
        shape (Variable|list/tuple of integers): The output shape is specified
            by `shape`, which can be a Tensor or a list/tuple of integers.
            If it is a Tensor, it's rank must be the same as `x` , only 
            it's shape will be used, and the value of it will be ignored. This way
7926
            is suitable for the case that the output shape may be changed each
S
SunGaofeng 已提交
7927
            iteration. If it is a list/tuple of integers, it's length must be the same
7928
            as the rank of `x`
S
SunGaofeng 已提交
7929 7930 7931
        offsets (Variable|list/tuple of integers|None): Specifies the cropping
            offsets at each dimension. It can be a Tensor or a list/tuple
            of integers. If it is a Tensor, it's rank must be the same as `x`.
7932
            This way is suitable for the case that the offsets may be changed
S
SunGaofeng 已提交
7933 7934 7935 7936 7937
            each iteration. If it is a list/tuple of integers, it's length must be the
            same as the rank of `x`. If None, the offsets are 0 at each dimension.
        name(str, optional): For detailed information, please refer 
            to :ref:`api_guide_Name` . Usually name is no need to set and 
            None by default. 
7938 7939

    Returns:
S
SunGaofeng 已提交
7940 7941 7942 7943
        The cropped Tensor, which has the same rank and data type with `x`

    Return Type:
        Variable
7944 7945 7946 7947 7948 7949 7950 7951

    Raises:
        ValueError: If shape is not a list, tuple or Variable.

    Examples:

        .. code-block:: python

S
SunGaofeng 已提交
7952
            import paddle.fluid as fluid
S
SunGaofeng 已提交
7953 7954
            x = fluid.data(name="x", shape=[3, 3, 5], dtype="float32")
            y = fluid.data(name="y", shape=[2, 2, 3], dtype="float32")
7955 7956 7957
            crop = fluid.layers.crop(x, shape=y)

            # or
S
SunGaofeng 已提交
7958 7959
            z = fluid.data(name="z", shape=[3, 3, 5], dtype="float32")
            crop = fluid.layers.crop(z, shape=[2, 2, 3])
7960 7961 7962 7963 7964

    """
    helper = LayerHelper('crop', **locals())

    if not (isinstance(shape, list) or isinstance(shape, tuple) or \
7965
            isinstance(shape, Variable)):
7966 7967 7968 7969 7970
        raise ValueError("The shape should be a list, tuple or Variable.")

    if offsets is None:
        offsets = [0] * len(x.shape)

X
Xin Pan 已提交
7971
    out = helper.create_variable_for_type_inference(x.dtype)
7972 7973 7974 7975 7976 7977 7978 7979 7980 7981 7982 7983 7984 7985 7986 7987 7988
    ipts = {'X': x}
    attrs = {}
    if isinstance(shape, Variable):
        ipts['Y'] = shape
    else:
        attrs['shape'] = shape
    if isinstance(offsets, Variable):
        ipts['Offsets'] = offsets
    else:
        attrs['offsets'] = offsets

    helper.append_op(
        type='crop',
        inputs=ipts,
        outputs={'Out': out},
        attrs=None if len(attrs) == 0 else attrs)
    return out
7989 7990


7991 7992 7993 7994 7995 7996
def crop_tensor(x, shape=None, offsets=None, name=None):
    """
    Crop input into output, as specified by offsets and shape.

    .. code-block:: text

7997 7998
        * Case 1 (input is a 2-D Tensor):
            Input:
7999
                X.shape = [3, 5]
8000 8001 8002 8003 8004 8005 8006
                X.data = [[0, 1, 2, 0, 0],
                          [0, 3, 4, 0, 0],
                          [0, 0, 0, 0, 0]]
            Parameters:
                shape = [2, 2]
                offsets = [0, 1]
            Output:
8007 8008 8009
                Out.shape = [2, 2]
                Out.data = [[1, 2],
                            [3, 4]]
8010 8011 8012 8013 8014 8015 8016 8017 8018 8019
        * Case 2 (input is a 3-D Tensor):
            Input:
                X.shape = [2, 3, 4]
                X.data =  [[[0, 1, 2, 3],
                            [0, 5, 6, 7],
                            [0, 0, 0, 0]],
                           [[0, 3, 4, 5],
                            [0, 6, 7, 8],
                            [0, 0, 0, 0]]]
            Parameters:
8020
                shape = [2, 2, -1]
8021 8022
                offsets = [0, 0, 1]
            Output:
8023 8024 8025 8026 8027
                Out.shape = [2, 2, 3]
                Out.data  = [[[1, 2, 3],
                              [5, 6, 7]],
                             [[3, 4, 5],
                              [6, 7, 8]]]
8028 8029

    Parameters:
8030
        x (Variable): 1-D to 6-D Tensor, the data type is float32, float64, int32 or int64.
8031 8032 8033 8034
        shape (list|tuple|Variable): The output shape is specified
            by `shape`. Its data type is int32. If a list/tuple, it's length must be
            the same as the dimension size of `x`. If a Variable, it shoule be a 1-D Tensor.
            When it is a list, each element can be an integer or a Tensor of shape: [1].
8035 8036
            If Variable contained, it is suitable for the case that the shape may
            be changed each iteration.
8037 8038 8039 8040 8041 8042 8043 8044
        offsets (list|tuple|Variable, optional): Specifies the cropping
            offsets at each dimension. Its data type is int32. If a list/tuple, it's length
            must be the same as the dimension size of `x`. If a Variable, it shoule be a 1-D
            Tensor. When it is a list, each element can be an integer or a Tensor of shape: [1].
            If Variable contained, it is suitable for the case that the offsets may be changed
            each iteration. Default: None, the offsets are 0 at each dimension.
        name(str, optional): The default value is None. Normally there is no need for user to set
            this property. For more information, please refer to :ref:`api_guide_Name` .
8045 8046

    Returns:
8047
        Variable: The cropped Tensor has same data type with `x`.
8048 8049

    Raises:
8050 8051 8052 8053 8054 8055
        TypeError: If the data type of `x` is not in: float32, float64, int32, int64.
        TypeError: If `shape` is not a list, tuple or Variable.
        TypeError: If the data type of `shape` is not int32.
        TypeError: If `offsets` is not None and not a list, tuple or Variable.
        TypeError: If the data type of `offsets` is not int32.
        ValueError: If the element in `offsets` is less than zero.
8056 8057 8058 8059 8060 8061

    Examples:

        .. code-block:: python

            import paddle.fluid as fluid
8062
            x = fluid.data(name="x", shape=[None, 3, 5], dtype="float32")
8063 8064
            # x.shape = [-1, 3, 5], where -1 indicates batch size, and it will get the exact value in runtime.

8065 8066
            # shape is a 1-D Tensor
            crop_shape = fluid.data(name="crop_shape", shape=[3], dtype="int32")
8067 8068 8069 8070
            crop0 = fluid.layers.crop_tensor(x, shape=crop_shape)
            # crop0.shape = [-1, -1, -1], it means crop0.shape[0] = x.shape[0] in runtime.

            # or shape is a list in which each element is a constant
8071
            crop1 = fluid.layers.crop_tensor(x, shape=[-1, -1, 3], offsets=[0, 1, 0])
8072 8073
            # crop1.shape = [-1, 2, 3]

8074 8075 8076 8077 8078
            # or shape is a list in which each element is a constant or Variable
            y = fluid.data(name="y", shape=[3, 8, 8], dtype="float32")
            dim1 = fluid.data(name="dim1", shape=[1], dtype="int32")
            crop2 = fluid.layers.crop_tensor(y, shape=[3, dim1, 4])
            # crop2.shape = [3, -1, 4]
8079

8080 8081
            # offsets is a 1-D Tensor
            crop_offsets = fluid.data(name="crop_offsets", shape=[3], dtype="int32")
8082 8083 8084
            crop3 = fluid.layers.crop_tensor(x, shape=[-1, 2, 3], offsets=crop_offsets)
            # crop3.shape = [-1, 2, 3]

8085 8086
            # offsets is a list in which each element is a constant or Variable
            offsets_var =  fluid.data(name="dim1", shape=[1], dtype="int32")
8087 8088 8089 8090 8091
            crop4 = fluid.layers.crop_tensor(x, shape=[-1, 2, 3], offsets=[0, 1, offsets_var])
            # crop4.shape = [-1, 2, 3]

    """
    helper = LayerHelper('crop_tensor', **locals())
8092 8093 8094 8095 8096 8097
    check_type_and_dtype(x, 'x', Variable,
                         ['float32', 'float64', 'int32', 'int64'],
                         'crop_tensor')
    check_type(shape, 'shape', (list, tuple, Variable), 'crop_tensor')
    check_type(offsets, 'offsets', (list, tuple, Variable, type(None)),
               'crop_tensor')
8098 8099 8100 8101 8102 8103 8104 8105

    if offsets is None:
        offsets = [0] * len(x.shape)

    out = helper.create_variable_for_type_inference(x.dtype)
    ipts = {'X': x}
    attrs = {}

8106
    def _contain_var(input_list):
8107 8108 8109 8110 8111
        for ele in input_list:
            if isinstance(ele, Variable):
                return True
        return False

8112 8113 8114 8115 8116 8117 8118 8119 8120 8121 8122 8123 8124 8125 8126 8127 8128 8129 8130 8131 8132 8133 8134 8135
    def _attr_shape_check(shape_val):
        if not isinstance(shape_val, int):
            raise TypeError(
                "Attr(shape)'s dtype of Op(crop_tensor) should be int32, but received: %s."
                % type(shape_val))
        if shape_val == 0:
            raise ValueError(
                "Attr(shape) of Op(crop_tensor) should not be zero, but received: %s."
                % str(shape_val))
        if shape_val < -1:
            raise ValueError(
                "When the element in Attr(shape) of Op(crop_tensor) is negative, only -1 is supported, but received: %s."
                % str(shape_val))

    def _attr_offsets_check(offset_val):
        if not isinstance(offset_val, int):
            raise TypeError(
                "Attr(offsets)'s dtype of Op(crop_tensor) should be int32, but received: %s."
                % type(offset_val))
        if offset_val < 0:
            raise ValueError(
                "Attr(offsets) of Op(crop_tensor) should be greater or equal to zero, but received: %s."
                % str(offset_val))

8136 8137 8138
    if isinstance(offsets, Variable):
        offsets.stop_gradient = True
        ipts['Offsets'] = offsets
8139 8140
        attrs['offsets'] = [-1] * len(x.shape)
    elif _contain_var(offsets):
8141
        new_offsets_tensor = []
8142
        offsets_attr = []
8143 8144 8145 8146
        for dim in offsets:
            if isinstance(dim, Variable):
                dim.stop_gradient = True
                new_offsets_tensor.append(dim)
8147
                offsets_attr.append(-1)
8148
            else:
8149
                _attr_offsets_check(dim)
8150 8151 8152
                temp_out = helper.create_variable_for_type_inference('int32')
                fill_constant([1], 'int32', dim, force_cpu=True, out=temp_out)
                new_offsets_tensor.append(temp_out)
8153
                offsets_attr.append(dim)
8154
        ipts['OffsetsTensor'] = new_offsets_tensor
8155
        attrs['offsets'] = offsets_attr
8156
    else:
8157 8158
        for offset in offsets:
            _attr_offsets_check(offset)
8159 8160 8161 8162 8163
        attrs['offsets'] = offsets

    if isinstance(shape, Variable):
        shape.stop_gradient = True
        ipts['Shape'] = shape
8164
    elif _contain_var(shape):
8165 8166
        new_shape_tensor = []
        shape_attr = []
8167
        for dim_size in shape:
8168 8169 8170
            if isinstance(dim_size, Variable):
                dim_size.stop_gradient = True
                new_shape_tensor.append(dim_size)
8171
                shape_attr.append(0)
8172
            else:
8173
                _attr_shape_check(dim_size)
8174 8175 8176 8177 8178 8179 8180 8181
                temp_out = helper.create_variable_for_type_inference('int32')
                fill_constant(
                    [1], 'int32', dim_size, force_cpu=True, out=temp_out)
                new_shape_tensor.append(temp_out)
                shape_attr.append(dim_size)
        ipts['ShapeTensor'] = new_shape_tensor
        attrs['shape'] = shape_attr
    else:
8182 8183
        for dim_size in shape:
            _attr_shape_check(dim_size)
8184 8185 8186 8187 8188 8189 8190 8191 8192 8193
        attrs['shape'] = shape

    helper.append_op(
        type='crop_tensor',
        inputs=ipts,
        outputs={'Out': out},
        attrs=None if len(attrs) == 0 else attrs)
    return out


W
whs 已提交
8194 8195 8196 8197 8198 8199 8200 8201
def affine_grid(theta, out_shape, name=None):
    """
    It generates a grid of (x,y) coordinates using the parameters of
    the affine transformation that correspond to a set of points where
    the input feature map should be sampled to produce the transformed
    output feature map.

    Args:
8202 8203 8204 8205 8206 8207
        theta (Variable) - A Tensor with shape [N, 2, 3]. It contains a batch of affine transform parameters.
                           The data type can be float32 or float64.
        out_shape (Variable | list | tuple): The shape of target output with format [batch_size, channel, height, width].
                                             ``out_shape`` can be a Tensor or a list or tuple. The data
                                             type must be int32.
        name(str|None): The default value is None.  Normally there is no need for user to set this property.  For more information, please refer to :ref:`api_guide_Name`.
W
whs 已提交
8208 8209

    Returns:
8210
        Variable: A Tensor with shape [batch_size, H, W, 2] while 'H' and 'W' are the height and width of feature map in affine transformation. The data type is the same as `theta`. 
W
whs 已提交
8211 8212 8213 8214 8215 8216 8217

    Raises:
        ValueError: If the type of arguments is not supported.

    Examples:

        .. code-block:: python
H
haowang101779990 已提交
8218

S
SunGaofeng 已提交
8219
            import paddle.fluid as fluid
8220 8221 8222 8223 8224 8225 8226 8227 8228 8229 8230 8231 8232 8233
            import numpy as np
            place = fluid.CPUPlace()
            theta = fluid.data(name="x", shape=[None, 2, 3], dtype="float32")
            out_shape = fluid.data(name="y", shape=[4], dtype="int32")
            grid_0 = fluid.layers.affine_grid(theta, out_shape)
            grid_1 = fluid.layers.affine_grid(theta, [5, 3, 28, 28])
            batch_size=2
            exe = fluid.Executor(place)
            exe.run(fluid.default_startup_program())
            output= exe.run(feed={"x": np.random.rand(batch_size,2,3).astype("float32"),
                                  "y": np.array([5, 3, 28, 28]).astype("int32")},
                                  fetch_list=[grid_0.name, grid_1.name])
            print(output[0])
            print(output[1])
W
whs 已提交
8234 8235 8236 8237
    """
    helper = LayerHelper('affine_grid')

    if not (isinstance(out_shape, list) or isinstance(out_shape, tuple) or \
8238
            isinstance(out_shape, Variable)):
W
whs 已提交
8239 8240 8241 8242 8243 8244 8245 8246 8247 8248 8249 8250 8251 8252 8253 8254 8255 8256 8257 8258 8259
        raise ValueError("The out_shape should be a list, tuple or Variable.")

    if not isinstance(theta, Variable):
        raise ValueError("The theta should be a Variable.")

    out = helper.create_variable_for_type_inference(theta.dtype)
    ipts = {'Theta': theta}
    attrs = {}
    if isinstance(out_shape, Variable):
        ipts['OutputShape'] = out_shape
    else:
        attrs['output_shape'] = out_shape

    helper.append_op(
        type='affine_grid',
        inputs=ipts,
        outputs={'Output': out},
        attrs=None if len(attrs) == 0 else attrs)
    return out


W
whs 已提交
8260 8261 8262 8263 8264 8265 8266 8267 8268 8269 8270
def pad2d(input,
          paddings=[0, 0, 0, 0],
          mode='constant',
          pad_value=0.0,
          data_format="NCHW",
          name=None):
    """
    Pad 2-d images accordding to 'paddings' and 'mode'.
    If mode is 'reflect', paddings[0] and paddings[1] must be no greater
    than height-1. And the width dimension has the same condition.

L
Liufang Sang 已提交
8271 8272 8273 8274 8275 8276 8277 8278 8279 8280 8281 8282 8283 8284 8285 8286 8287 8288 8289 8290 8291 8292 8293 8294
    Parameters:
        input (Variable): The input image with [N, C, H, W] format or [N, H, W, C] format, which is a 4-D Tensor with data type float32.
        paddings (Variable | List[int32]): The padding size. If padding is a List, it must
            contain four integers, (padding_top, padding_bottom, padding_left, padding_right).
            Otherwise, it is a 1-D Tensor with shape [4]. Data type is int32.
            Default is [0, 0, 0, 0].
        mode (str): Three modes: 'constant' (default), 'reflect', 'edge' .
        	When in 'constant' mode, this op uses a constant value to pad the input tensor.
        	When in 'reflect' mode, uses reflection of the input boundaries to pad the input tensor.
        	When in 'edge' mode, uses input boundaries to pad the input tensor.
        	Default is 'constant'
        pad_value (float32): The value to fill the padded areas in 'constant' mode . Default is 0.0
        data_format (str): An string from: "NHWC", "NCHW". Specify the data format of
                           the input data.
                           Default is  "NCHW"
        name (str, optional) : The default value is None.  Normally there is no need for
                    user to set this property.  For more information, please refer to :ref:`api_guide_Name` .

    Returns: a 4-D Tensor padded accordding to paddings and mode and data type is same as input.

    Return Type: Variable


    Examples:
T
Tink_Y 已提交
8295
        .. code-block:: text
W
whs 已提交
8296

T
Tink_Y 已提交
8297
	      Given that X is a channel of image from input:
M
minqiyang 已提交
8298

T
Tink_Y 已提交
8299 8300
	      X = [[1, 2, 3],
		   [4, 5, 6]]
M
minqiyang 已提交
8301

T
Tink_Y 已提交
8302
	      Case 0:
M
minqiyang 已提交
8303

T
Tink_Y 已提交
8304 8305 8306
		paddings = [0, 1, 2, 3],
		mode = 'constant'
		pad_value = 0
M
minqiyang 已提交
8307

T
Tink_Y 已提交
8308 8309 8310
		Out = [[0, 0, 1, 2, 3, 0, 0, 0]
		       [0, 0, 4, 5, 6, 0, 0, 0]
		       [0, 0, 0, 0, 0, 0, 0, 0]]
M
minqiyang 已提交
8311

T
Tink_Y 已提交
8312
	      Case 1:
M
minqiyang 已提交
8313

T
Tink_Y 已提交
8314 8315
		paddings = [0, 1, 2, 1],
		mode = 'reflect'
M
minqiyang 已提交
8316

T
Tink_Y 已提交
8317 8318 8319
		Out = [[3, 2, 1, 2, 3, 2]
		       [6, 5, 4, 5, 6, 5]
		       [3, 2, 1, 2, 3, 2]]
M
minqiyang 已提交
8320

T
Tink_Y 已提交
8321
	      Case 2:
M
minqiyang 已提交
8322

T
Tink_Y 已提交
8323 8324
		paddings = [0, 1, 2, 1],
		mode = 'edge'
M
minqiyang 已提交
8325

T
Tink_Y 已提交
8326 8327 8328
		Out = [[1, 1, 1, 2, 3, 3]
		       [4, 4, 4, 5, 6, 6]
		       [4, 4, 4, 5, 6, 6]]
M
minqiyang 已提交
8329

L
Liufang Sang 已提交
8330
    Code Examples:
W
whs 已提交
8331 8332
        .. code-block:: python

B
Bai Yifan 已提交
8333
          import paddle.fluid as fluid
L
Liufang Sang 已提交
8334
          data = fluid.data(name='data', shape=[None, 3, 32, 32],
B
Bai Yifan 已提交
8335 8336 8337
                                   dtype='float32')
          result = fluid.layers.pad2d(input=data, paddings=[1, 2, 3, 4],
                                      mode='reflect')
W
whs 已提交
8338 8339 8340
    """

    helper = LayerHelper('pad2d', **locals())
8341 8342 8343 8344

    assert mode in ['reflect', 'edge', 'constant'
                    ], "mode should be one of constant, reflect, edge."

W
whs 已提交
8345
    dtype = helper.input_dtype(input_param_name='input')
X
Xin Pan 已提交
8346
    out = helper.create_variable_for_type_inference(dtype)
8347 8348 8349 8350 8351 8352 8353 8354 8355
    inputs = {'X': input}
    attrs = {'mode': mode, 'pad_value': pad_value, 'data_format': data_format}

    if isinstance(paddings, Variable):
        inputs['Paddings'] = paddings
        attrs['paddings'] = []
    else:
        attrs['paddings'] = paddings

W
whs 已提交
8356
    helper.append_op(
8357
        type='pad2d', inputs=inputs, outputs={"Out": out}, attrs=attrs)
W
whs 已提交
8358 8359 8360 8361

    return out


8362 8363 8364 8365 8366 8367 8368
@templatedoc()
def elu(x, alpha=1.0, name=None):
    """
    ${comment}
    Args:
        x(${x_type}): ${x_comment}
        alpha(${alpha_type}|1.0): ${alpha_comment}
8369 8370
        name(str|None): The default value is None. Normally there is no need for user to set this property. 
                        For more information, please refer to :ref:`api_guide_Name`.
8371
    Returns:
8372
        ${out_type}: ${out_comment}
Z
ZhenWang 已提交
8373 8374 8375 8376 8377

    Examples:

        .. code-block:: python

8378
            import paddle.fluid as fluid
8379 8380 8381 8382 8383 8384 8385 8386 8387
            import numpy as np
         
            input_elu = np.array([[-1,6],[1,15.6]])
            with fluid.dygraph.guard():
                x = fluid.dygraph.to_variable(input_elu)
                y = fluid.layers.elu(x, alpha=0.2)
                print(y.numpy())
                # [[-0.12642411  6.        ]
                # [ 1.          15.6       ]]
8388 8389
    """
    helper = LayerHelper('elu', **locals())
8390 8391
    check_type_and_dtype(x, 'x', Variable, ['float16', 'float32', 'float64'],
                         'elu')
X
Xin Pan 已提交
8392
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
8393 8394 8395 8396 8397 8398 8399 8400 8401 8402 8403 8404
    helper.append_op(
        type='elu',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'alpha': alpha})
    return out


@templatedoc()
def relu6(x, threshold=6.0, name=None):
    """
    ${comment}
Z
zhupengyang 已提交
8405

8406 8407
    Args:
        x(${x_type}): ${x_comment}
Z
zhupengyang 已提交
8408 8409 8410 8411
        threshold(float, optional): ${threshold_comment}
        name(str, optional): The default value is None. Normally there is no
            need for user to set this property. For more information, please
            refer to :ref:`api_guide_Name`.
8412 8413 8414

    Returns:
        output(${out_type}): ${out_comment}
Z
ZhenWang 已提交
8415 8416 8417 8418 8419

    Examples:

        .. code-block:: python

8420
            import paddle.fluid as fluid
Z
zhupengyang 已提交
8421 8422 8423 8424 8425 8426 8427 8428
            import numpy as np
            in1 = np.array([[-1,0],[2.5,7.8]])
            with fluid.dygraph.guard():
                x1 = fluid.dygraph.to_variable(in1)
                out1 = fluid.layers.relu6(x=x1, threshold=6.0)
                print(out1.numpy())
                # [[0.  0. ]
                #  [2.5 6. ]]
8429 8430
    """
    helper = LayerHelper('relu6', **locals())
X
Xin Pan 已提交
8431
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
8432 8433 8434 8435 8436 8437 8438 8439 8440 8441 8442
    helper.append_op(
        type='relu6',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'threshold': threshold})
    return out


@templatedoc()
def pow(x, factor=1.0, name=None):
    """
8443 8444 8445 8446
    This is Pow Activation Operator.

    :math:`out = x^{factor}`

8447
    Args:
8448 8449 8450
        x(Variable): A ``Tensor`` or ``LoDTensor`` . The data type is ``float32`` or ``float64``.
        factor(float32|Variable, optional): A scalar with type ``float32`` or a ``Tensor`` with shape [1] and type ``float32``.  The exponential factor of Pow. Default 1.0.
        name(str, optional): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name` .
8451 8452

    Returns:
8453
        Variable: A ``Tensor`` or ``LoDTensor``. The data type is same as ``x``.
Z
ZhenWang 已提交
8454 8455 8456 8457 8458

    Examples:

        .. code-block:: python

8459
            import paddle.fluid as fluid
8460

8461
            x = fluid.data(name="x", shape=[32,32], dtype="float32")
8462 8463 8464

            # example 1: argument factor is float
            y_1 = fluid.layers.pow(x, factor=2.0)
8465
            # y_1 is x^{2.0}
8466 8467 8468 8469

            # example 2: argument factor is Variable
            factor_tensor = fluid.layers.fill_constant([1], "float32", 3.0)
            y_2 = fluid.layers.pow(x, factor=factor_tensor)
8470
            # y_2 is x^{3.0}
8471 8472
    """
    helper = LayerHelper('pow', **locals())
8473 8474 8475 8476 8477 8478 8479 8480
    inputs = {'X': x}
    attrs = {}
    if isinstance(factor, Variable):
        factor.stop_gradient = True
        inputs['FactorTensor'] = factor
    else:
        attrs['factor'] = factor

X
Xin Pan 已提交
8481
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
8482
    helper.append_op(
8483
        type='pow', inputs=inputs, outputs={'Out': out}, attrs=attrs)
8484 8485 8486 8487
    return out


@templatedoc()
8488
def stanh(x, scale_a=0.67, scale_b=1.7159, name=None):
8489 8490 8491 8492 8493 8494 8495 8496 8497 8498
    """
    ${comment}
    Args:
        x(${x_type}): ${x_comment}
        scale_a(${scale_a_type}|2.0 / 3.0): ${scale_a_comment}
        scale_b(${scale_b_type}|1.7159): ${scale_b_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
8499
        output(${out_type}): ${out_comment}. 
Z
ZhenWang 已提交
8500 8501 8502 8503 8504

    Examples:

        .. code-block:: python

8505
            import paddle.fluid as fluid
8506 8507 8508 8509 8510 8511 8512 8513 8514 8515 8516 8517 8518 8519 8520
            import numpy as np
            data = fluid.data(name="input", shape=[-1, 3])
            result = fluid.layers.stanh(data,scale_a=0.67, scale_b=1.72)
            place = fluid.CPUPlace()
            exe = fluid.Executor(place)
            exe.run(fluid.default_startup_program())
            x = np.random.random(size=(3, 3)).astype('float32')
            output= exe.run(feed={"input": x},
                         fetch_list=[result])
            print(output)

            #[array([[0.626466  , 0.89842904, 0.7501062 ],
            #       [0.25147712, 0.7484996 , 0.22902708],
            #       [0.62705994, 0.23110689, 0.56902856]], dtype=float32)]

8521 8522
    """
    helper = LayerHelper('stanh', **locals())
X
Xin Pan 已提交
8523
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
8524 8525 8526 8527 8528 8529 8530 8531 8532 8533 8534 8535 8536
    helper.append_op(
        type='stanh',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'scale_a': scale_a,
               'scale_b': scale_b})
    return out


@templatedoc()
def hard_sigmoid(x, slope=0.2, offset=0.5, name=None):
    """
    ${comment}
8537 8538 8539 8540 8541 8542 8543
    Parameters:
        x (${x_type}): ${x_comment}
        slope (float, optional): ${slope_comment}
        offset (float, optional): ${offset_comment}
        name (str, optional): The default value is None. Normally there is no
            need for user to set this property. For more information, please
            refer to :ref:`api_guide_Name`
8544 8545

    Returns:
8546
        ${out_type}: ${out_comment}
Z
ZhenWang 已提交
8547 8548 8549 8550 8551

    Examples:

        .. code-block:: python

8552
            import paddle.fluid as fluid
8553 8554
            data = fluid.layers.fill_constant(shape=[3, 2], value=0.5, dtype='float32') # [[0.5, 0.5], [0.5, 0.5], [0.5, 0.5]]
            result = fluid.layers.hard_sigmoid(data) # [[0.6, 0.6], [0.6, 0.6], [0.6, 0.6]]
8555 8556
    """
    helper = LayerHelper('hard_sigmoid', **locals())
X
Xin Pan 已提交
8557
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
8558 8559 8560 8561 8562 8563 8564 8565 8566 8567 8568 8569
    helper.append_op(
        type='hard_sigmoid',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'slope': slope,
               'offset': offset})
    return out


@templatedoc()
def swish(x, beta=1.0, name=None):
    """
8570 8571 8572 8573 8574 8575 8576
    Elementwise swish activation function. See `Searching for Activation Functions <https://arxiv.org/abs/1710.05941>`_ for more details.
    
    Equation:

    .. math::
        out = \\frac{x}{1 + e^{- beta * x}}
    
8577
    Args:
8578 8579 8580 8581 8582
        x(Variable): Tensor or LoDTensor, dtype: float32 or float64, the input of swish activation.
        
        beta(float): Constant beta of swish operator, default 1.0.
        
        name(str, optional): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`.
8583 8584

    Returns:
8585 8586

        Variable: Output of the swish activation, Tensor or LoDTensor, with the same dtype and shape with the input x.
Z
ZhenWang 已提交
8587 8588 8589 8590

    Examples:

        .. code-block:: python
8591 8592 8593 8594 8595 8596
            
            # declarative mode
            import numpy as np
            from paddle import fluid
            
            x = fluid.data(name="x", shape=(-1, 3), dtype="float32")
Z
ZhenWang 已提交
8597
            y = fluid.layers.swish(x, beta=2.0)
8598 8599 8600 8601 8602 8603 8604 8605 8606 8607 8608 8609 8610 8611 8612 8613 8614 8615 8616 8617 8618 8619 8620 8621 8622 8623 8624 8625 8626 8627 8628 8629 8630 8631 8632 8633 8634
            
            place = fluid.CPUPlace()
            exe = fluid.Executor(place)
            start = fluid.default_startup_program()
            main = fluid.default_main_program()
            
            data = np.random.randn(2, 3).astype("float32")
            exe.run(start)
            y_np, = exe.run(main, feed={"x": data}, fetch_list=[y])
            
            data
            # array([[-1.1239197 ,  1.3391294 ,  0.03921051],
            #        [ 1.1970421 ,  0.02440812,  1.2055548 ]], dtype=float32)
            y_np
            # array([[-0.2756806 ,  1.0610548 ,  0.01998957],
            #        [ 0.9193261 ,  0.01235299,  0.9276883 ]], dtype=float32)


        .. code-block:: python

            # imperative mode
            import numpy as np
            from paddle import fluid
            import paddle.fluid.dygraph as dg
            
            data = np.random.randn(2, 3).astype("float32")
            place = fluid.CPUPlace()
            with dg.guard(place) as g:
                x = dg.to_variable(data)
                y = fluid.layers.swish(x)
                y_np = y.numpy()
            data
            # array([[-0.0816701 ,  1.1603649 , -0.88325626],
            #        [ 0.7522361 ,  1.0978601 ,  0.12987892]], dtype=float32)
            y_np
            # array([[-0.03916847,  0.8835007 , -0.25835553],
            #        [ 0.51126915,  0.82324016,  0.06915068]], dtype=float32)
8635 8636
    """
    helper = LayerHelper('swish', **locals())
X
Xin Pan 已提交
8637
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
8638 8639 8640 8641 8642 8643 8644 8645
    helper.append_op(
        type='swish',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'slope': beta})
    return out


J
jerrywgz 已提交
8646 8647 8648 8649
def prelu(x, mode, param_attr=None, name=None):
    """
    Equation:

H
haowang101779990 已提交
8650 8651
    .. math::
        y = \max(0, x) + \\alpha * \min(0, x)
J
jerrywgz 已提交
8652

J
jerrywgz 已提交
8653 8654 8655 8656 8657 8658 8659 8660
    There are three modes for the activation:

    .. code-block:: text

        all: All elements share same alpha.
        channel: Elements in same channel share same alpha.
        element: All elements do not share alpha. Each element has its own alpha.

J
jerrywgz 已提交
8661
    Args:
W
wangguanzhong 已提交
8662 8663
        x (Variable): The input Tensor or LoDTensor with data type float32.
        mode (str): The mode for weight sharing. 
J
jerrywgz 已提交
8664
        param_attr(ParamAttr|None): The parameter attribute for the learnable
W
wangguanzhong 已提交
8665 8666 8667 8668 8669
          weight (alpha), it can be create by ParamAttr. None by default.
          For detailed information, please refer to :ref:`api_fluid_ParamAttr`.
        name(str|None): For detailed information, please refer 
          to :ref:`api_guide_Name`. Usually name is no need to set and 
          None by default. 
J
jerrywgz 已提交
8670 8671

    Returns:
W
wangguanzhong 已提交
8672 8673 8674 8675
        Variable:

        output(Variable): The tensor or LoDTensor with the same shape as input.
        The data type is float32.
J
jerrywgz 已提交
8676 8677 8678 8679 8680

    Examples:

        .. code-block:: python

J
jerrywgz 已提交
8681 8682
            import paddle.fluid as fluid
            from paddle.fluid.param_attr import ParamAttr
8683
            x = fluid.data(name="x", shape=[None,5,10,10], dtype="float32")
J
jerrywgz 已提交
8684
            mode = 'channel'
J
jerrywgz 已提交
8685 8686 8687
            output = fluid.layers.prelu(
                     x,mode,param_attr=ParamAttr(name='alpha'))

J
jerrywgz 已提交
8688 8689 8690 8691 8692 8693 8694 8695
    """
    helper = LayerHelper('prelu', **locals())
    if mode not in ['all', 'channel', 'element']:
        raise ValueError('mode should be one of all, channel, element.')
    alpha_shape = [1]
    if mode == 'channel':
        alpha_shape = [1, x.shape[1], 1, 1]
    elif mode == 'element':
8696
        alpha_shape = [1, x.shape[1], x.shape[2], x.shape[3]]
J
jerrywgz 已提交
8697 8698
    dtype = helper.input_dtype(input_param_name='x')
    alpha = helper.create_parameter(
Q
Qiao Longfei 已提交
8699
        attr=helper.param_attr,
J
jerrywgz 已提交
8700 8701 8702
        shape=alpha_shape,
        dtype='float32',
        is_bias=False,
8703
        default_initializer=Constant(0.25))
X
Xin Pan 已提交
8704
    out = helper.create_variable_for_type_inference(dtype)
J
jerrywgz 已提交
8705 8706 8707 8708 8709 8710 8711 8712 8713
    helper.append_op(
        type="prelu",
        inputs={"X": x,
                'Alpha': alpha},
        attrs={"mode": mode},
        outputs={"Out": out})
    return out


8714 8715 8716 8717 8718 8719 8720 8721
@templatedoc()
def brelu(x, t_min=0.0, t_max=24.0, name=None):
    """
    ${comment}
    Args:
        x(${x_type}): ${x_comment}
        t_min(${t_min_type}|0.0): ${t_min_comment}
        t_max(${t_max_type}|24.0): ${t_max_comment}
8722 8723
        name(str|None): The default value is None. Normally there is no need for user to set this property. 
                        For more information, please refer to :ref:`api_guide_Name`.
8724
    Returns:
8725
        ${out_type}: ${out_comment}
8726 8727 8728

    Examples:

8729
    .. code-block:: python
8730

8731
            import paddle.fluid as fluid
8732 8733 8734 8735 8736 8737 8738 8739 8740
            import numpy as np
            
            input_brelu = np.array([[-1,6],[1,15.6]])
            with fluid.dygraph.guard():
                x = fluid.dygraph.to_variable(input_brelu)
                y = fluid.layers.brelu(x, t_min=1.0, t_max=10.0)
                print(y.numpy())
                #[[ 1.  6.]
                #[ 1. 10.]] 
8741 8742
    """
    helper = LayerHelper('brelu', **locals())
X
Xin Pan 已提交
8743
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
8744 8745 8746 8747 8748 8749 8750 8751 8752 8753 8754 8755 8756 8757 8758 8759
    helper.append_op(
        type='brelu',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'t_min': t_min,
               't_max': t_max})
    return out


@templatedoc()
def leaky_relu(x, alpha=0.02, name=None):
    """
    ${comment}
    Args:
        x(${x_type}): ${x_comment}
        alpha(${alpha_type}|0.02): ${alpha_comment}
W
Wilber 已提交
8760 8761
        name(str|None): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`

8762
    Returns:
8763
        output(${out_type}): ${out_comment}
8764 8765 8766 8767 8768

    Examples:

        .. code-block:: python

8769
            import paddle.fluid as fluid
W
Wilber 已提交
8770 8771 8772 8773 8774 8775 8776 8777 8778 8779 8780 8781 8782
            import numpy as np

            # Graph Organizing
            x = fluid.layers.data(name="x", shape=[2], dtype="float32")
            res = fluid.layers.leaky_relu(x, alpha=0.1)

            # Create an executor using CPU as an example
            exe = fluid.Executor(fluid.CPUPlace())

            # Execute
            x_i = np.array([[-1, 2], [3, -4]]).astype(np.float32)
            res_val, = exe.run(fluid.default_main_program(), feed={'x':x_i}, fetch_list=[res])
            print(res_val) # [[-0.1, 2], [3, -0.4]]
8783 8784
    """
    helper = LayerHelper('leaky_relu', **locals())
X
Xin Pan 已提交
8785
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
8786 8787 8788 8789 8790 8791 8792 8793 8794 8795
    helper.append_op(
        type='leaky_relu',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'alpha': alpha})
    return out


def soft_relu(x, threshold=40.0, name=None):
    """
8796 8797 8798 8799
    SoftRelu Activation Operator.

    $out = \ln(1 + \exp(\max(\min(x, threshold), -threshold)))$

8800
    Args:
8801 8802 8803 8804
        x(Variable): Input of soft_relu operator. Data type can be float32, float64.
        threshold(float, optional): The threshold value of soft_relu, default value being 40.0.
        name(str, optional): The default value is None.  Normally there is no need for user to set this property.  For more information, please refer to :ref:`api_guide_Name` .

8805
    Returns:
8806
        Variable(Tensor|LoDTensor)): Output of soft_relu operator, shape and LoD same as input.
8807 8808 8809

    Examples:

8810 8811 8812
        .. code-block:: python 
 
            import paddle.fluid as fluid
8813 8814 8815 8816 8817 8818 8819 8820 8821 8822 8823 8824
            import numpy as np

            inputs = fluid.layers.data(name="x", shape=[2, 2], dtype="float32")
            output = fluid.layers.soft_relu(inputs, threshold=20.0)

            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())

            img = np.array([[0, 1],[2, 3]]).astype(np.float32)

            res = exe.run(fluid.default_main_program(), feed={'x':img}, fetch_list=[output])
            print(res) # [array([[0.6931472, 1.3132616], [2.126928 , 3.0485873]], dtype=float32)]
8825 8826
    """
    helper = LayerHelper('soft_relu', **locals())
X
Xin Pan 已提交
8827
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
8828 8829 8830 8831 8832 8833 8834 8835
    helper.append_op(
        type='soft_relu',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'threshold': threshold})
    return out


8836 8837
def flatten(x, axis=1, name=None):
    """
8838 8839 8840
    **Flatten op**

    Flatten the input tensor into a 2D matrix.
M
minqiyang 已提交
8841

H
haowang101779990 已提交
8842
    For Example:
M
minqiyang 已提交
8843

H
haowang101779990 已提交
8844
    .. code-block:: text
8845

H
haowang101779990 已提交
8846 8847 8848 8849 8850 8851 8852 8853 8854 8855 8856 8857 8858 8859 8860 8861 8862 8863 8864 8865 8866
        Case 1:

          Given
            X.shape = (3, 100, 100, 4)

          and
            axis = 2

          We get:
            Out.shape = (3 * 100, 4 * 100)

        Case 2:

          Given
            X.shape = (3, 100, 100, 4)

          and
            axis = 0

          We get:
            Out.shape = (1, 3 * 100 * 100 * 4)
8867 8868

    Args:
8869 8870
        x (Variable): A tensor of rank >= axis. A tensor with type float32,
                      float64, int8, int32, int64.
8871 8872
        axis (int): Indicate up to which input dimensions (exclusive) should
                    be flattened to the outer dimension of the output.
8873
                    The value for axis must be in the range [0, R], where R
8874 8875 8876
                    is the rank of the input tensor. Default: 1.
        name(str, Optional): For details, please refer to :ref:`api_guide_Name`.
                        Generally, no setting is required. Default: None.
8877 8878

    Returns:
H
haowang101779990 已提交
8879 8880 8881
        Variable: A 2D tensor with the contents of the input tensor, with input \
                  dimensions up to axis flattened to the outer dimension of \
                  the output and remaining input dimensions flattened into the \
8882
                  inner dimension of the output. A Tensor with type same as input x.
8883 8884 8885

    Raises:
        ValueError: If x is not a variable.
8886
        ValueError: If axis is not in range [0, rank(x)].
8887 8888 8889 8890 8891

    Examples:

        .. code-block:: python

8892
            import paddle.fluid as fluid
B
Bai Yifan 已提交
8893
            x = fluid.data(name="x", shape=[4, 4, 3], dtype="float32")
8894
            # x shape is [4, 4, 3]
8895
            out = fluid.layers.flatten(x=x, axis=2)
8896
            # out shape is [16, 3]
8897 8898 8899 8900 8901 8902 8903 8904 8905
    """
    helper = LayerHelper('flatten', **locals())

    if not (isinstance(x, Variable)):
        raise ValueError("The input x should be a Variable")

    if not (isinstance(axis, int)) or axis > len(x.shape) or axis < 0:
        raise ValueError("The axis should be a int, and in range [0, rank(x)]")

X
Xin Pan 已提交
8906 8907
    out = helper.create_variable_for_type_inference(x.dtype)
    x_shape = helper.create_variable_for_type_inference(x.dtype)
8908
    helper.append_op(
8909
        type='flatten2',
8910
        inputs={"X": x},
8911 8912
        outputs={'Out': out,
                 'XShape': x_shape},
8913 8914
        attrs={"axis": axis})
    return out
X
Xin Pan 已提交
8915 8916 8917


def stack(x, axis=0):
S
sneaxiy 已提交
8918
    """
8919

8920
    This OP stacks all the inputs :code:`x` along axis.
C
chengduozh 已提交
8921

C
chengduozh 已提交
8922 8923 8924
    .. code-block:: text

        Case 1:
8925

C
chengduozh 已提交
8926
          Input:
8927
            x[0].shape = [1, 2]
C
chengduozh 已提交
8928
            x[0].data = [ [1.0 , 2.0 ] ]
8929
            x[1].shape = [1, 2]
C
chengduozh 已提交
8930
            x[1].data = [ [3.0 , 4.0 ] ]
8931
            x[2].shape = [1, 2]
C
chengduozh 已提交
8932 8933 8934 8935 8936 8937
            x[2].data = [ [5.0 , 6.0 ] ]

          Attrs:
            axis = 0

          Output:
8938
            Out.dims = [3, 1, 2]
C
chengduozh 已提交
8939 8940 8941
            Out.data =[ [ [1.0, 2.0] ],
                        [ [3.0, 4.0] ],
                        [ [5.0, 6.0] ] ]
8942

C
chengduozh 已提交
8943 8944

        Case 2:
8945 8946 8947 8948


          Input:
            x[0].shape = [1, 2]
C
chengduozh 已提交
8949
            x[0].data = [ [1.0 , 2.0 ] ]
8950
            x[1].shape = [1, 2]
C
chengduozh 已提交
8951
            x[1].data = [ [3.0 , 4.0 ] ]
8952
            x[2].shape = [1, 2]
C
chengduozh 已提交
8953
            x[2].data = [ [5.0 , 6.0 ] ]
8954

C
chengduozh 已提交
8955 8956 8957 8958 8959

          Attrs:
            axis = 1 or axis = -2

          Output:
8960
            Out.shape = [1, 3, 2]
C
chengduozh 已提交
8961 8962 8963
            Out.data =[ [ [1.0, 2.0]
                          [3.0, 4.0]
                          [5.0, 6.0] ] ]
8964

C
chengduozh 已提交
8965

S
sneaxiy 已提交
8966
    Args:
8967 8968 8969 8970 8971 8972 8973 8974 8975
        x (Variable|list(Variable)): Input :code:`x` can be a single Tensor, a :code:`list` of Tensors.
                                     If :code:`x` is a :code:`list`, the shapes of all these Tensors
                                     must be the same. Supposing input is N dims
                                     Tensors :math:`[d_0, d_1, ..., d_{n-1}]`, the output is N+1 dims
                                     Tensor :math:`[d_0, d_1, d_{axis-1}, len(x), d_{axis}, ..., d_{n-1}]`.
                                     Support data types: float32, float64, int32, int64.
        axis (int, optional): The axis along which all inputs are stacked. ``axis`` range is :math:`[-(R+1), R+1)`.
                              R is the first tensor of inputs. If ``axis`` < 0, :math:`axis=axis+rank(x[0])+1`.
                              The default value of axis is 0.
8976

S
sneaxiy 已提交
8977
    Returns:
8978
        Variable: The stacked Tensor, has same data type with input Tensors. Output dim is :math:`rank(x[0])+1`.
8979

8980 8981 8982
    Examples:
        .. code-block:: python

8983
            import paddle.fluid as fluid
8984
            import paddle.fluid.layers as layers
8985 8986 8987 8988 8989 8990 8991 8992 8993 8994
            # set batch size=None
            x1 = fluid.data(name='x1', shape=[None, 1, 2], dtype='int32')
            x2 = fluid.data(name='x2', shape=[None, 1, 2], dtype='int32')
            # stack Tensor list
            data = layers.stack([x1,x2]) # stack according to axis 0, data.shape=[2, None, 1, 2]

            data = layers.stack([x1,x2], axis=1) # stack according to axis 1, data.shape=[None, 2, 1, 2]

            # stack single Tensor
            data = layers.stack(x1)  # stack according to axis 0, data.shape=[1, None, 1, 2]
8995

S
sneaxiy 已提交
8996 8997
    """

X
Xin Pan 已提交
8998 8999 9000 9001 9002 9003
    helper = LayerHelper('stack', **locals())
    axis = 0 if axis is None else axis

    if not isinstance(x, list) and not isinstance(x, tuple):
        x = [x]

X
Xin Pan 已提交
9004
    out = helper.create_variable_for_type_inference(x[0].dtype)
X
Xin Pan 已提交
9005
    helper.append_op(
S
sneaxiy 已提交
9006 9007
        type='stack', inputs={'X': x}, outputs={'Y': out},
        attrs={'axis': axis})
9008

X
Xin Pan 已提交
9009
    return out
D
dzhwinter 已提交
9010 9011


J
Jiawei Wang 已提交
9012 9013 9014 9015 9016 9017 9018 9019 9020 9021 9022 9023 9024 9025 9026 9027 9028 9029 9030 9031 9032 9033 9034 9035 9036 9037 9038 9039 9040 9041 9042 9043 9044 9045 9046 9047 9048 9049 9050 9051 9052 9053 9054 9055 9056 9057 9058 9059 9060 9061 9062 9063 9064 9065 9066 9067 9068 9069 9070 9071 9072 9073 9074 9075 9076 9077 9078 9079 9080 9081
@templatedoc(op_type="filter_by_instag")
def filter_by_instag(ins, ins_tag, filter_tag, is_lod):
    """
    **Filter By Instag Layer**
   
    This function filter a batch of ins by instag, 
    There are multiple ins, and every ins belongs to some tags. 
    We can specify some tags we want. So the ins which belongs to that tags
    remains in the output, and others removed.
 
    For example, one batch has 4 ins. Every ins has its tag list. 
     
       | Ins   |   Ins_Tag |
       |:-----:|:------:|
       |  0    |   0, 1 |
       |  1    |   1, 3 |
       |  2    |   0, 3 |
       |  3    |   2, 6 |

    And Lod is [1,1,1,1]

    And the filter tags [1]

    From the definition above, ins which has tag 1 can pass the filter
    So Ins 0 and Ins 1 can pass and be seen in the output,
    Ins 2 and 3 cannot pass because they do not has tag 1.

    Actually, if is_lod is false, it is normal tensor that equals to 
    lod_tensor with all 1, similar to the example above.

    Args:
        ins (Variable): Input Variable (LoDTensor), usually it is 2D tensor
                        And first dimension can have lod info or not.
        ins_tag (Variable): Input Variable (LoDTensor), usually it is 1D list
                        And split them by lod info
        filter_tag (Variable): Input Variable (1D Tensor/List), usually it is 
                        list that holds the tags.
        is_lod (Bool): Boolean value to indicate ins is lod tensor or not.

    Returns:
        Variable: filtered ins (LoDTensor) and loss weight (Tensor)

    Examples:
        .. code-block:: python

          import paddle.fluid.layers as layers
          ins = layers.data(name='Ins', shape=[-1,32], lod_level=0, dtype='float64')
          ins_tag = layers.data(name='Ins_tag', shape=[-1,16], lod_level=0, dtype='int64')
          filter_tag = layers.data(name='Filter_tag', shape=[-1,16], dtype='int64')
          out, loss_weight = layers.filter_by_instag(ins,  ins_tag,  filter_tag, True)
        		
    """
    helper = LayerHelper('filter_by_instag', **locals())

    out = helper.create_variable_for_type_inference(dtype=ins.dtype)
    loss_weight = helper.create_variable_for_type_inference(dtype=np.float64)
    mmap = helper.create_variable_for_type_inference(dtype=ins_tag.dtype)
    helper.append_op(
        type='filter_by_instag',
        inputs={'Ins': ins,
                'Ins_tag': ins_tag,
                'Filter_tag': filter_tag},
        outputs={'Out': out,
                 'LossWeight': loss_weight,
                 'IndexMap': mmap},
        attrs={'is_lod': is_lod})

    return [out, loss_weight]


D
dzhwinter 已提交
9082 9083 9084 9085
def unstack(x, axis=0, num=None):
    """
    **UnStack Layer**

9086
    This layer unstacks input Tensor :code:`x` into several Tensors along :code:`axis`.
M
minqiyang 已提交
9087

D
dzhwinter 已提交
9088 9089 9090
    If :code:`axis` < 0, it would be replaced with :code:`axis+rank(x)`.
    If :code:`num` is None, it would be inferred from :code:`x.shape[axis]`,
    and if :code:`x.shape[axis]` <= 0 or is unknown, :code:`ValueError` is
M
minqiyang 已提交
9091
    raised.
D
dzhwinter 已提交
9092 9093

    Args:
9094
        x (Variable): Input Tensor. It is a N-D Tensors of data types float32, float64, int32, int64.
D
dzhwinter 已提交
9095 9096
        axis (int): The axis along which the input is unstacked.
        num (int|None): The number of output variables.
M
minqiyang 已提交
9097

D
dzhwinter 已提交
9098
    Returns:
9099 9100 9101 9102
        list(Variable): The unstacked Tensors list. The list elements are N-D Tensors of data types float32, float64, int32, int64.

    Raises:
        ValueError: If x.shape[axis] <= 0 or axis is not in range [-D, D).
M
minqiyang 已提交
9103

9104 9105 9106 9107
    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
9108 9109
            x = fluid.layers.data(name='x', shape=[2, 3, 5], dtype='float32')  # create a tensor with shape=[2, 3, 5]
            y = fluid.layers.unstack(x, axis=1)  # unstack with second axis, which results 3 tensors with shape=[2, 5]
D
dzhwinter 已提交
9110

9111
    """
D
dzhwinter 已提交
9112 9113 9114 9115 9116 9117 9118 9119
    helper = LayerHelper('unstack', **locals())
    if num is None:
        if axis is None or x.shape[axis] <= 0:
            raise ValueError('unknown unstack number')
        else:
            num = x.shape[axis]

    outs = []
Y
Yibing Liu 已提交
9120
    for _ in range(num):
X
Xin Pan 已提交
9121
        outs.append(helper.create_variable_for_type_inference(x.dtype))
D
dzhwinter 已提交
9122 9123 9124 9125 9126 9127 9128 9129

    helper.append_op(
        type='unstack',
        inputs={'X': [x]},
        outputs={'Y': outs},
        attrs={'axis': axis,
               'num': num})
    return outs
W
whs 已提交
9130 9131 9132


def expand(x, expand_times, name=None):
9133 9134 9135 9136
    """
    This operation tiles ``x`` multiple times according to the parameter ``expand_times``.
    The times number for each dimension of ``x`` is set by the parameter ``expand_times``.
    The rank of ``x`` should be less than or equal to 6. Please note that size of ``expand_times`` must be the same
W
whs 已提交
9137 9138 9139 9140 9141 9142
    with X's rank. Following is a using case:


    .. code-block:: text

        Input(X) is a 3-D tensor with shape [2, 3, 1]:
M
minqiyang 已提交
9143

W
whs 已提交
9144 9145 9146 9147
                [
                   [[1], [2], [3]],
                   [[4], [5], [6]]
                ]
M
minqiyang 已提交
9148

W
whs 已提交
9149
        Attr(expand_times):  [1, 2, 2]
M
minqiyang 已提交
9150

W
whs 已提交
9151
        Output(Out) is a 3-D tensor with shape [2, 6, 2]:
M
minqiyang 已提交
9152

W
whs 已提交
9153 9154 9155 9156
                [
                    [[1, 1], [2, 2], [3, 3], [1, 1], [2, 2], [3, 3]],
                    [[4, 4], [5, 5], [6, 6], [4, 4], [5, 5], [6, 6]]
                ]
M
minqiyang 已提交
9157

W
whs 已提交
9158
    Args:
9159 9160 9161 9162 9163
        x (Variable): A ``Tensor`` or ``LoDTensor`` with dimension in [1, 6]. The data type is ``bool``, ``float32``, ``float64`` or ``int32`` .
        expand_times (list|tuple|Variable): The data type is ``int32`` . If ``expand_times`` is a list or tuple, the elements of
                it should be integers or Tensors with shape [1]. If ``expand_times`` is an Variable, it should be an 1-D Tensor.
                Expand times number for each dimension of ``x`` .
        name (str, optional): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name` .
W
whs 已提交
9164 9165

    Returns:
9166
        Variable: A ``Tensor`` or ``LoDTensor``. The data type is same as ``x``. After expanding, size of each dimension of output is equal to the size of the corresponding dimension of ``x`` multiplying the corresponding value given by ``expand_times`` .
W
whs 已提交
9167

9168 9169 9170
    Raises:
        TypeError: The type of ``expand_times`` must be list, tuple or Variable.
        ValueError: The elements of ``expand_times`` cannot be negative.
W
whs 已提交
9171 9172 9173

    Examples:
        .. code-block:: python
L
liym27 已提交
9174

W
wangchaochaohu 已提交
9175
            import paddle.fluid as fluid
L
liym27 已提交
9176 9177 9178 9179

            # example 1:
            data_1 = fluid.layers.fill_constant(shape=[2, 3, 1], dtype='int32', value=0)
            expanded_1 = fluid.layers.expand(data_1, expand_times=[1, 2, 2])
9180
            # the shape of expanded_1 is [2, 6, 2].
L
liym27 已提交
9181 9182 9183 9184 9185

            # example 2:
            data_2 = fluid.layers.fill_constant(shape=[12, 14], dtype="int32", value=3)
            expand_times = fluid.layers.fill_constant(shape=[2], dtype="int32", value=4)
            expanded_2 = fluid.layers.expand(data_2, expand_times=expand_times)
9186
            # the shape of expanded_2 is [48, 56].
W
whs 已提交
9187
    """
9188 9189 9190 9191
    check_type_and_dtype(x, 'x', Variable,
                         ['bool', 'float32', 'float64', 'int32', 'int64'],
                         'expand')
    check_type(expand_times, 'expand_times', (list, tuple, Variable), 'expand')
W
wangchaochaohu 已提交
9192 9193 9194
    if convert_dtype(x.dtype) == 'bool' and x.stop_gradient == True:
        raise ValueError(
            "expand op bool date type must set the stop_gradient to be False")
L
liym27 已提交
9195

W
whs 已提交
9196
    helper = LayerHelper('expand', input=x, **locals())
L
liym27 已提交
9197 9198 9199 9200 9201 9202 9203 9204 9205 9206 9207 9208 9209 9210 9211 9212 9213 9214 9215 9216 9217 9218 9219 9220 9221 9222 9223 9224 9225 9226 9227 9228
    inputs = {"X": x}
    attrs = {}

    def contain_var(expand_times):
        for ele in expand_times:
            if isinstance(ele, Variable):
                return True
        return False

    def get_attr_expand_times(list_expand_times):
        attrs_expand_times = []
        for idx, times in enumerate(list_expand_times):
            if isinstance(times, Variable):
                attrs_expand_times.append(-1)
            else:
                attrs_expand_times.append(times)
                assert times > 0, (
                    "Each element given in expand_times must not be negtive.")
        return attrs_expand_times

    def get_new_expand_times_tensor(list_expand_times):
        new_expand_times_tensor = []
        for ele in list_expand_times:
            if isinstance(ele, Variable):
                ele.stop_gradient = True
                new_expand_times_tensor.append(ele)
            else:
                assert (isinstance(ele, int))
                temp_out = helper.create_variable_for_type_inference('int32')
                fill_constant([1], 'int32', ele, force_cpu=True, out=temp_out)
                new_expand_times_tensor.append(temp_out)
        return new_expand_times_tensor
9229 9230 9231 9232 9233

    if in_dygraph_mode():
        inputs = {'X': x}
        attrs = {'expand_times': expand_times}
    else:
L
liym27 已提交
9234 9235 9236 9237 9238 9239 9240 9241
        if isinstance(expand_times, Variable):
            expand_times.stop_gradient = True
            inputs['ExpandTimes'] = expand_times
        elif isinstance(expand_times, (list, tuple)):
            attrs['expand_times'] = get_attr_expand_times(expand_times)
            if contain_var(expand_times):
                inputs['expand_times_tensor'] = get_new_expand_times_tensor(
                    expand_times)
9242

L
liym27 已提交
9243 9244
    dtype = helper.input_dtype(input_param_name='x')
    out = helper.create_variable_for_type_inference(dtype)
W
whs 已提交
9245
    helper.append_op(
9246
        type='expand', inputs=inputs, outputs={'Out': out}, attrs=attrs)
W
whs 已提交
9247
    return out
S
sneaxiy 已提交
9248 9249


9250 9251 9252 9253 9254 9255 9256 9257 9258 9259 9260 9261 9262 9263 9264 9265 9266 9267 9268 9269 9270 9271 9272 9273 9274 9275 9276 9277 9278 9279 9280 9281 9282 9283 9284 9285 9286 9287 9288 9289 9290 9291 9292 9293 9294 9295 9296 9297 9298 9299 9300 9301 9302 9303 9304 9305 9306 9307 9308 9309 9310 9311 9312 9313 9314 9315 9316 9317 9318 9319
def expand_as(x, target_tensor, name=None):
    """
    expand_as operator tiles to the input by given expand tensor. You should set expand tensor
    for each dimension by providing tensor 'target_tensor'. The rank of X
    should be in [1, 6]. Please note that size of 'target_tensor' must be the same
    with X's rank. Following is a using case:


    .. code-block:: text

        Input(X) is a 3-D tensor with shape [2, 3, 1]:

                [
                   [[1], [2], [3]],
                   [[4], [5], [6]]
                ]

        target_tensor's shape:  [2, 6, 2] 

        Output(Out) is a 3-D tensor with shape [2, 6, 2]:

                [
                    [[1, 1], [2, 2], [3, 3], [1, 1], [2, 2], [3, 3]],
                    [[4, 4], [5, 5], [6, 6], [4, 4], [5, 5], [6, 6]]
                ]
                

    Args:
        x (Variable): A Tensor with dtype float64, float32, int32.
        A tensor with rank in [1, 6].
        target_tensor (Variable): A Tensor with dtype float64, float32, int32.
        target_tensor for expanding to Input(X). Only use target_tensor'shape.

    Returns:
        Variable: A Tensor with dtype float64, float32, int32. 
        After expanding, size of each dimension of Output(Out) is equal to the size 
        of the corresponding dimension of target_tensor multiplying the corresponding
        value given by target_tensor.


    Examples:
        .. code-block:: python
          
        import paddle.fluid as fluid
        import numpy as np

        data = fluid.layers.data(name="data", shape=[-1,10], dtype='float64')
        target_tensor = fluid.layers.data(
          name="target_tensor", shape=[-1,20], dtype='float64')
        result = fluid.layers.expand_as(x=data, target_tensor=target_tensor) 
        use_cuda = False
        place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
        exe = fluid.Executor(place)
        exe.run(fluid.default_startup_program())
        x = np.random.rand(3,10)
        y = np.random.rand(3,20)
        output= exe.run(feed={"data":x,"target_tensor":y},fetch_list=[result.name])
        print(output[0].shape)
        #(3,20)

    """

    helper = LayerHelper('expand_as', input=x, **locals())
    dtype = helper.input_dtype(input_param_name='x')
    out = helper.create_variable_for_type_inference(dtype)
    inputs = {'X': x, 'target_tensor': target_tensor}
    helper.append_op(type='expand_as', inputs=inputs, outputs={'Out': out})
    return out


G
fix  
gongweibao 已提交
9320 9321 9322
from paddle.fluid.framework import convert_np_dtype_to_dtype_


G
gongweibao 已提交
9323
@templatedoc()
G
fix  
gongweibao 已提交
9324 9325 9326 9327 9328 9329 9330 9331 9332
def uniform_random_batch_size_like(input,
                                   shape,
                                   dtype='float32',
                                   input_dim_idx=0,
                                   output_dim_idx=0,
                                   min=-1.0,
                                   max=1.0,
                                   seed=0):
    """
9333 9334 9335 9336 9337 9338
    This OP initializes a variable with random values sampled from a
    uniform distribution in the range [min, max). The input_dim_idx used to get the input dimension value which will be used to resize the output dimension.

    .. code-block:: text

        *Case 1:
G
fix  
gongweibao 已提交
9339

9340 9341 9342 9343 9344 9345 9346 9347 9348 9349 9350 9351 9352 9353 9354 9355 9356 9357 9358 9359 9360 9361 9362 9363 9364 9365
            Given:
                input =[[0.946741  , 0.1357001 , 0.38086128]]    # input.shape=[1,3]
                shape=[2,4]

            result.shape[output_dim_idx] = input.shape[input_dim_idx],
            output_dim_idx = 0, 
            input_dim_idx = 0,
            result.shape[0] = input.shape[0], 
            then:
                result=[[ 0.3443427 , -0.23056602,  0.3477049 ,  0.06139076]]    # result.shape=[1,4]
            
       *Case 2:
           
           Given:
               input =[[0.946741  , 0.1357001 , 0.38086128]]     # input.shape=[1,3]
               shape=[2,4]
               input_dim_idx=1
               output_dim_idx=1
         
           result.shape[output_dim_idx] = input.shape[input_dim_idx],
           output_dim_idx = 1, 
           input_dim_idx = 1,
           result.shape[1] = input.shape[1], 
           then:
               result=[[-0.23133647, -0.84195036,  0.21441269],
                       [-0.08774924,  0.25605237, -0.09403259]]    # result.shape=[2,3]
G
fix  
gongweibao 已提交
9366
    Args:
9367 9368 9369 9370 9371 9372 9373 9374
        input (Variable): A Tensor. Supported data types: float32, float64.
        shape (tuple|list): A python list or python tuple. The shape of the output Tensor, the data type is int.
        input_dim_idx (int, optional): An index used to get the input dimension value which will be used to resize the output dimension. Default  0. 
        output_dim_idx (int, optional): An index used to indicate the specific dimension that will be replaced by corresponding input dimension value. Default 0.
        min (float, optional): The lower bound on the range of random values to generate, the min is included in the range. Default -1.0.
        max (float, optional): The upper bound on the range of random values to generate, the max is excluded in the range. Default 1.0.
        seed (int, optional):  Random seed used for generating samples. 0 means use a seed generated by the system.Note that if seed is not 0, this operator will always generate the same random numbers every time.
        dtype(np.dtype|core.VarDesc.VarType|str, optional): The data type of output Tensor. Supported data types: float32, float64. Default float32.
G
fix  
gongweibao 已提交
9375
    Returns:
9376
        Variable: A Tensor of the specified shape filled with uniform_random values. The shape of the Tensor is determined by the shape parameter and the specified dimension of the input Tensor.
G
fix  
gongweibao 已提交
9377

9378 9379 9380
    Examples:
        .. code-block:: python

9381
            import paddle.fluid as fluid
9382 9383 9384 9385
            
            # example 1: 
            input = fluid.data(name="input", shape=[1, 3], dtype='float32')
            out_1 = fluid.layers.uniform_random_batch_size_like(input, [2, 4]) # out_1.shape=[1, 4]
9386

9387 9388 9389 9390
            # example 2: 
            out_2 = fluid.layers.uniform_random_batch_size_like(input, [2, 4], input_dim_idx=1, output_dim_idx=1) # out_2.shape=[2, 3]

            
G
fix  
gongweibao 已提交
9391 9392 9393
    """

    helper = LayerHelper('uniform_random_batch_size_like', **locals())
X
Xin Pan 已提交
9394
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
9395 9396 9397 9398 9399 9400 9401 9402 9403 9404 9405 9406 9407 9408 9409 9410
    c_dtype = convert_np_dtype_to_dtype_(dtype)
    helper.append_op(
        type='uniform_random_batch_size_like',
        inputs={'Input': input},
        outputs={'Out': out},
        attrs={
            'shape': shape,
            'input_dim_idx': input_dim_idx,
            'output_dim_idx': output_dim_idx,
            'min': min,
            'max': max,
            'seed': seed,
            'dtype': c_dtype
        })

    return out
G
fix  
gongweibao 已提交
9411 9412


G
gongweibao 已提交
9413
@templatedoc()
X
Xin Pan 已提交
9414
def gaussian_random(shape, mean=0.0, std=1.0, seed=0, dtype='float32'):
G
fix  
gongweibao 已提交
9415
    """
9416
    Generate a random tensor whose data is drawn from a Gaussian distribution.
G
fix  
gongweibao 已提交
9417 9418

    Args:
9419 9420 9421 9422 9423 9424 9425 9426 9427
        shape (Tuple[int] | List[int]): Shape of the generated random tensor.
        
        mean (float): Mean of the random tensor, defaults to 0.0.
            
        std (float): Standard deviation of the random tensor, defaults to 1.0.
        
        seed (int): ${seed_comment}
        
        dtype(np.dtype | core.VarDesc.VarType | str): Output data type, float32 or float64.
G
fix  
gongweibao 已提交
9428 9429

    Returns:
9430
        Variable: Random tensor whose data is drawn from a Gaussian distribution, dtype: flaot32 or float64 as specified.
G
fix  
gongweibao 已提交
9431

9432
    Examples:
9433 9434 9435 9436 9437 9438 9439 9440 9441 9442 9443 9444 9445 9446 9447
       .. code-block:: python
       
           # declarative mode 
           import numpy as np
           from paddle import fluid
   
           x = fluid.layers.gaussian_random((2, 3), std=2., seed=10)
   
           place = fluid.CPUPlace()
           exe = fluid.Executor(place)
           start = fluid.default_startup_program()
           main = fluid.default_main_program()
   
           exe.run(start)
           x_np, = exe.run(main, feed={}, fetch_list=[x])
9448

9449 9450 9451 9452 9453 9454 9455 9456 9457 9458 9459 9460 9461 9462 9463 9464 9465 9466
           x_np
           # array([[2.3060477, 2.676496 , 3.9911983],
           #        [0.9990833, 2.8675377, 2.2279181]], dtype=float32)

       .. code-block:: python

           # imperative mode
           import numpy as np
           from paddle import fluid
           import paddle.fluid.dygraph as dg
    
           place = fluid.CPUPlace()
           with dg.guard(place) as g:
               x = fluid.layers.gaussian_random((2, 4), mean=2., dtype="float32", seed=10)
               x_np = x.numpy()       
           x_np
           # array([[2.3060477 , 2.676496  , 3.9911983 , 0.9990833 ],
           #        [2.8675377 , 2.2279181 , 0.79029655, 2.8447366 ]], dtype=float32)
G
fix  
gongweibao 已提交
9467 9468 9469
    """

    helper = LayerHelper('gaussian_random', **locals())
X
Xin Pan 已提交
9470
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
9471 9472 9473 9474 9475 9476 9477 9478 9479 9480
    c_dtype = convert_np_dtype_to_dtype_(dtype)
    helper.append_op(
        type='gaussian_random',
        outputs={'Out': out},
        attrs={
            'shape': shape,
            'mean': mean,
            'std': std,
            'seed': seed,
            'dtype': c_dtype,
X
Xin Pan 已提交
9481
            'use_mkldnn': False
G
fix  
gongweibao 已提交
9482 9483 9484 9485 9486
        })

    return out


G
gongweibao 已提交
9487
@templatedoc()
G
fix  
gongweibao 已提交
9488
def sampling_id(x, min=0.0, max=1.0, seed=0, dtype='float32'):
G
fix  
gongweibao 已提交
9489
    """
R
ruri 已提交
9490
    This op is used for sampling id from multinomial distribution from the input, sampling one id for one sample.
G
fix  
gongweibao 已提交
9491

R
ruri 已提交
9492 9493 9494 9495 9496
    Parameters:
        x (Variable): 2-D tensor, [batch_size, input_feature_dimensions]
        min (Float): minimum , default 0.0.
        max (Float): maximum, default 1.0.
        seed (Float): Random seed, default 0. if seed is not 0, will generate same number every time. 
G
fix  
gongweibao 已提交
9497
        dtype(np.dtype|core.VarDesc.VarType|str): The type of output data : float32, float_16, int etc
G
fix  
gongweibao 已提交
9498 9499

    Returns:
R
ruri 已提交
9500
        Variable: sampling tensor.
G
fix  
gongweibao 已提交
9501

9502 9503 9504
    Examples:
        .. code-block:: python

9505
            import paddle.fluid as fluid
R
ruri 已提交
9506
            x = fluid.data(
9507 9508
                name="X",
                shape=[13, 11],
R
ruri 已提交
9509
                dtype='float32')
9510

Y
Yibing Liu 已提交
9511
            out = fluid.layers.sampling_id(x)
G
fix  
gongweibao 已提交
9512 9513 9514
    """

    helper = LayerHelper('sampling_id', **locals())
X
Xin Pan 已提交
9515
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
9516 9517 9518 9519 9520 9521 9522 9523 9524 9525 9526
    helper.append_op(
        type='sampling_id',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'min': min,
               'max': max,
               'seed': seed})

    return out


G
gongweibao 已提交
9527
@templatedoc()
G
fix  
gongweibao 已提交
9528 9529 9530 9531 9532 9533 9534 9535 9536
def gaussian_random_batch_size_like(input,
                                    shape,
                                    input_dim_idx=0,
                                    output_dim_idx=0,
                                    mean=0.0,
                                    std=1.0,
                                    seed=0,
                                    dtype='float32'):
    """
G
gongweibao 已提交
9537
    ${comment}
G
fix  
gongweibao 已提交
9538 9539

    Args:
G
gongweibao 已提交
9540 9541
        input (Variable): ${input_comment}
        shape (tuple|list): ${shape_comment}
Y
Yibing Liu 已提交
9542 9543 9544 9545 9546 9547
        input_dim_idx (int): ${input_dim_idx_comment}
        output_dim_idx (int): ${output_dim_idx_comment}
        mean (float): ${mean_comment}
        std (float): ${std_comment}
        seed (int): ${seed_comment}
        dtype(np.dtype|core.VarDesc.VarType|str): The type of output data, float32 or float_64.
G
fix  
gongweibao 已提交
9548 9549

    Returns:
G
gongweibao 已提交
9550
        out (Variable): ${out_comment}
9551 9552 9553 9554

    Examples:
        .. code-block:: python

9555
            import paddle.fluid as fluid
Y
Yibing Liu 已提交
9556
            input = fluid.data(name="input", shape=[13, 11], dtype='float32')
9557

Y
Yibing Liu 已提交
9558
            out = fluid.layers.gaussian_random_batch_size_like(
9559
                input, shape=[-1, 11], mean=1.0, std=2.0)
G
fix  
gongweibao 已提交
9560 9561 9562
    """

    helper = LayerHelper('gaussian_random_batch_size_like', **locals())
X
Xin Pan 已提交
9563
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
9564 9565 9566 9567 9568 9569 9570 9571 9572 9573 9574 9575 9576 9577 9578 9579 9580 9581
    c_dtype = convert_np_dtype_to_dtype_(dtype)
    helper.append_op(
        type='gaussian_random_batch_size_like',
        inputs={'Input': input},
        outputs={'Out': out},
        attrs={
            'shape': shape,
            'input_dim_idx': input_dim_idx,
            'output_dim_idx': output_dim_idx,
            'mean': mean,
            'std': std,
            'seed': seed,
            'dtype': c_dtype
        })

    return out


G
gongweibao 已提交
9582
@templatedoc()
X
Xin Pan 已提交
9583
def sum(x):
G
fix  
gongweibao 已提交
9584
    """
G
gongweibao 已提交
9585
    ${comment}
9586 9587 9588 9589 9590 9591 9592 9593 9594 9595 9596 9597 9598 9599 9600 9601 9602 9603 9604 9605 9606 9607 9608 9609 9610 9611 9612 9613 9614 9615
    
    Case 1:
    ::
        Input:
            Input. Shape = [2, 3]
            Input = [[1, 2, 3],
                     [4, 5, 6]]

        Output:
            The output. Shape = [2, 3]
            Output = [[1, 2, 3],
                      [4, 5, 6]]

    Case 2:
    ::
        Input:
            First input:
            Input1. Shape = [2, 3]
            Input1 = [[1, 2, 3],
                      [4, 5, 6]]

        The second input:
            Input2. Shape = [2, 3]
            Input2 = [[7, 8, 9],
                      [10, 11, 12]]

        Output:
            The output. Shape = [2, 3]
            Output = [[8, 10, 12],
                      [14, 16, 18]]
G
fix  
gongweibao 已提交
9616 9617

    Args:
9618
        x (Variable|list(Variable)): ${x_comment}
G
fix  
gongweibao 已提交
9619 9620

    Returns:
9621
        Variable: ${out_comment}
9622 9623 9624 9625

    Examples:
        .. code-block:: python

9626
            import paddle.fluid as fluid
9627 9628 9629 9630 9631 9632 9633 9634 9635 9636 9637 9638 9639 9640 9641 9642 9643 9644 9645 9646 9647 9648

            input0 = fluid.layers.fill_constant(shape=[2, 3], dtype='int64', value=5)
            input1 = fluid.layers.fill_constant(shape=[2, 3], dtype='int64', value=3)
            sum = fluid.layers.sum([input0, input1])

            # You can print out 'sum' via executor.
            out = fluid.layers.Print(sum, message="the sum of input0 and input1: ")
            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_main_program())

            # The printed result is:
            # 1570701754	the sum of input0 and input1: 	The place is:CPUPlace
            # Tensor[sum_0.tmp_0]
            #    shape: [2,3,]
            #    dtype: l
            #    data: 8,8,8,8,8,8,

            # the sum of input0 and input1 is 2-D Tensor with shape [2,3].
            # dtype is the corresponding C++ data type, which may vary in different environments.
            # Eg: if the data type of tensor is int64, then the corresponding C++ data type is int64_t, 
            #       so the dtype value is typeid(int64_t).Name(), which is 'x' on MacOS, 'l' on Linux, 
            #       and '__int64' on Windows. They both represent 64-bit integer variables.
G
fix  
gongweibao 已提交
9649 9650 9651
    """

    helper = LayerHelper('sum', **locals())
X
Xin Pan 已提交
9652 9653
    out = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype('x'))
G
fix  
gongweibao 已提交
9654 9655 9656 9657
    helper.append_op(
        type='sum',
        inputs={'X': x},
        outputs={'Out': out},
X
Xin Pan 已提交
9658
        attrs={'use_mkldnn': False})
G
fix  
gongweibao 已提交
9659 9660 9661 9662

    return out


G
gongweibao 已提交
9663
@templatedoc()
G
fix  
gongweibao 已提交
9664 9665
def slice(input, axes, starts, ends):
    """
9666
    This operator produces a slice of ``input`` along multiple axes. Similar to numpy:
9667
    https://docs.scipy.org/doc/numpy/reference/arrays.indexing.html
9668 9669 9670 9671 9672 9673 9674
    Slice uses ``axes``, ``starts`` and ``ends`` attributes to specify the start and
    end dimension for each axis in the list of axes and Slice uses this information
    to slice the input data tensor. If a negative value is passed to
    ``starts`` or ``ends`` such as :math:`-i`,  it represents the reverse position of the
    axis :math:`i-1` (here 0 is the initial position).
    If the value passed to ``starts`` or ``ends`` is greater than n
    (the number of elements in this dimension), it represents n.
9675
    For slicing to the end of a dimension with unknown size, it is recommended
9676
    to pass in INT_MAX. The size of ``axes`` must be equal to ``starts`` and ``ends``.
9677 9678 9679
    Following examples will explain how slice works:

    .. code-block:: text
G
fix  
gongweibao 已提交
9680

9681 9682 9683 9684 9685 9686 9687 9688
        Case1:
            Given:
                data = [ [1, 2, 3, 4], [5, 6, 7, 8], ]
                axes = [0, 1]
                starts = [1, 0]
                ends = [2, 3]
            Then:
                result = [ [5, 6, 7], ]
9689

9690 9691 9692 9693 9694
        Case2:
            Given:
                data = [ [1, 2, 3, 4], [5, 6, 7, 8], ]
                axes = [0, 1]
                starts = [0, 1]
9695
                ends = [-1, 1000]       # -1 denotes the reverse 0th position of dimension 0.
9696
            Then:
9697
                result = [ [2, 3, 4], ] # result = data[0:1, 1:4]
G
fix  
gongweibao 已提交
9698
    Args:
9699 9700 9701 9702 9703 9704 9705 9706 9707
        input (Variable): A ``Tensor`` or ``LoDTensor`` . The data type is ``float16``, ``float32``, ``float64``, ``int32`` or ``int64``.
        axes (list|tuple): The data type is ``int32`` . Axes that `starts` and `ends` apply to.
                            It's optional. If it is not provides, it will be treated as :math:`[0,1,...,len(starts)-1]`.
        starts (list|tuple|Variable): The data type is ``int32`` . If ``starts`` is a list or tuple, the elements of
                it should be integers or Tensors with shape [1]. If ``starts`` is an Variable, it should be an 1-D Tensor.
                It represents starting indices of corresponding axis in ``axes``.
        ends (list|tuple|Variable): The data type is ``int32`` . If ``ends`` is a list or tuple, the elements of
                it should be integers or Tensors with shape [1]. If ``ends`` is an Variable, it should be an 1-D Tensor .
                It represents ending indices of corresponding axis in ``axes``.
G
fix  
gongweibao 已提交
9708 9709

    Returns:
9710 9711 9712 9713 9714
        Variable:  A ``Tensor`` or ``LoDTensor``. The data type is same as ``input``.

    Raises:
        TypeError: The type of ``starts`` must be list, tuple or Variable.
        TypeError: The type of ``ends`` must be list, tuple or Variable.
G
fix  
gongweibao 已提交
9715

9716 9717 9718
    Examples:
        .. code-block:: python

9719
            import paddle.fluid as fluid
9720

9721 9722
            input = fluid.data(
                name="input", shape=[4, 5, 6], dtype='float32')
9723

9724 9725 9726 9727 9728 9729
            # example 1:
            # attr starts is a list which doesn't contain tensor Variable.
            axes = [0, 1, 2]
            starts = [-3, 0, 2]
            ends = [3, 2, 4]
            sliced_1 = fluid.layers.slice(input, axes=axes, starts=starts, ends=ends)
9730
            # sliced_1 is input[0:3, 0:2, 2:4].
9731 9732 9733 9734 9735

            # example 2:
            # attr starts is a list which contain tensor Variable.
            minus_3 = fluid.layers.fill_constant([1], "int32", -3)
            sliced_2 = fluid.layers.slice(input, axes=axes, starts=[minus_3, 0, 2], ends=ends)
9736
            # sliced_2 is input[0:3, 0:2, 2:4].
G
fix  
gongweibao 已提交
9737 9738
    """

9739 9740 9741 9742 9743 9744 9745
    if not isinstance(starts, (list, tuple, Variable)):
        raise ValueError(
            "Input starts must be an Variable, python list or tuple.")
    if not isinstance(ends, (list, tuple, Variable)):
        raise ValueError(
            "Input ends must be an Variable, python list or tuple.")

G
fix  
gongweibao 已提交
9746
    helper = LayerHelper('slice', **locals())
9747 9748 9749 9750 9751 9752 9753 9754 9755 9756 9757 9758 9759 9760 9761 9762 9763 9764 9765 9766 9767 9768 9769 9770 9771 9772 9773 9774 9775 9776 9777 9778 9779 9780 9781 9782 9783 9784 9785 9786 9787 9788 9789 9790 9791 9792 9793 9794 9795 9796 9797 9798 9799 9800 9801 9802 9803 9804 9805 9806 9807 9808 9809 9810 9811 9812 9813 9814 9815 9816

    def contain_var(one_list):
        for ele in one_list:
            if isinstance(ele, Variable):
                return True
        return False

    def get_new_list_tensor(old_list):
        new_list_tensor = []
        for dim in old_list:
            if isinstance(dim, Variable):
                dim.stop_gradient = True
                new_list_tensor.append(dim)
            else:
                assert (isinstance(dim, int))
                temp_out = helper.create_variable_for_type_inference('int32')
                fill_constant([1], 'int32', dim, force_cpu=True, out=temp_out)
                new_list_tensor.append(temp_out)
        return new_list_tensor

    inputs = {'Input': input}
    attrs = {'axes': axes}
    infer_flags = list(1 for i in range(len(axes)))

    if in_dygraph_mode():
        inputs = {'Input': input}
        attrs = {
            'axes': axes,
            'starts': starts,
            'ends': ends,
            'infer_flags': infer_flags
        }
    else:
        # starts
        if isinstance(starts, Variable):
            starts.stop_gradient = True
            inputs['StartsTensor'] = starts
            infer_flags = list(-1 for i in range(len(axes)))
        elif isinstance(starts, (list, tuple)):
            attrs['starts'] = []
            if not contain_var(starts):
                attrs['starts'] = starts
            else:
                inputs['StartsTensorList'] = get_new_list_tensor(starts)
                for i, dim in enumerate(starts):
                    if isinstance(dim, Variable):
                        attrs['starts'].append(-1)
                        infer_flags[i] = -1
                    else:
                        attrs['starts'].append(dim)

        # ends
        if isinstance(ends, Variable):
            ends.stop_gradient = True
            inputs['EndsTensor'] = ends
            infer_flags = list(-1 for i in range(len(axes)))
        elif isinstance(ends, (list, tuple)):
            attrs['ends'] = []
            if not contain_var(ends):
                attrs['ends'] = ends
            else:
                inputs['EndsTensorList'] = get_new_list_tensor(ends)
                for i, dim in enumerate(ends):
                    if isinstance(dim, Variable):
                        attrs['ends'].append(-1)
                        infer_flags[i] = -1
                    else:
                        attrs['ends'].append(dim)
        # infer_flags
        attrs['infer_flags'] = infer_flags
X
Xin Pan 已提交
9817 9818
    out = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype('input'))
G
fix  
gongweibao 已提交
9819
    helper.append_op(
9820
        type='slice', inputs=inputs, attrs=attrs, outputs={'Out': out})
G
fix  
gongweibao 已提交
9821 9822 9823 9824

    return out


W
wangchaochaohu 已提交
9825 9826 9827
@templatedoc()
def strided_slice(input, axes, starts, ends, strides):
    """
W
wangchaochaohu 已提交
9828 9829 9830 9831 9832 9833 9834 9835 9836 9837 9838 9839 9840
    This operator produces a slice of ``input`` along multiple axes. Similar to numpy:
    https://docs.scipy.org/doc/numpy/reference/arrays.indexing.html
    Slice uses ``axes``, ``starts`` and ``ends`` attributes to specify the start and
    end dimension for each axis in the list of axes and Slice uses this information
    to slice the input data tensor. If a negative value is passed to
    ``starts`` or ``ends`` such as :math:`-i`,  it represents the reverse position of the
    axis :math:`i-1` th(here 0 is the initial position). The ``strides`` represents steps of
    slicing and if the ``strides`` is negative, slice operation is in the opposite direction.
    If the value passed to ``starts`` or ``ends`` is greater than n
    (the number of elements in this dimension), it represents n.
    For slicing to the end of a dimension with unknown size, it is recommended
    to pass in INT_MAX. The size of ``axes`` must be equal to ``starts`` , ``ends`` and ``strides``.
    Following examples will explain how strided_slice works:
W
wangchaochaohu 已提交
9841 9842 9843 9844 9845 9846 9847 9848 9849

    .. code-block:: text

        Case1:
            Given:
                data = [ [1, 2, 3, 4], [5, 6, 7, 8], ]
                axes = [0, 1]
                starts = [1, 0]
                ends = [2, 3]
W
wangchaochaohu 已提交
9850
                strides = [1, 1]
W
wangchaochaohu 已提交
9851
            Then:
9852
                result = [ [5, 6, 7], ]
W
wangchaochaohu 已提交
9853 9854 9855 9856 9857
        
        Case2:
            Given:
                data = [ [1, 2, 3, 4], [5, 6, 7, 8], ]
                axes = [0, 1]
9858
                starts = [0, 1]
W
wangchaochaohu 已提交
9859 9860 9861 9862 9863 9864 9865 9866 9867
                ends = [2, 0]
                strides = [1, -1]
            Then:
                result = [ [8, 7, 6], ]
        
        Case3:
            Given:
                data = [ [1, 2, 3, 4], [5, 6, 7, 8], ]
                axes = [0, 1]
9868
                starts = [0, 1]
9869 9870
                ends = [-1, 1000]
                strides = [1, 3]
W
wangchaochaohu 已提交
9871
            Then:
9872 9873
                result = [ [2], ]
    Args:
W
wangchaochaohu 已提交
9874 9875 9876 9877 9878 9879 9880 9881 9882 9883 9884 9885
        input (Variable): An N-D ``Tensor`` or ``LoDTensor`` . The data type is ``float32``, ``float64``, ``int32`` or ``int64``.
        axes (list|tuple): The data type is ``int32`` . Axes that `starts` and `ends` apply to.
                            It's optional. If it is not provides, it will be treated as :math:`[0,1,...,len(starts)-1]`.
        starts (list|tuple|Variable): The data type is ``int32`` . If ``starts`` is a list or tuple, the elements of
                it should be integers or Tensors with shape [1]. If ``starts`` is an Variable, it should be an 1-D Tensor.
                It represents starting indices of corresponding axis in ``axes``.
        ends (list|tuple|Variable): The data type is ``int32`` . If ``ends`` is a list or tuple, the elements of
                it should be integers or Tensors with shape [1]. If ``ends`` is an Variable, it should be an 1-D Tensor .
                It represents ending indices of corresponding axis in ``axes``.
        strides (list|tuple|Variable): The data type is ``int32`` . If ``strides`` is a list or tuple, the elements of
                it should be integers or Tensors with shape [1]. If ``strides`` is an Variable, it should be an 1-D Tensor .
                It represents slice step of corresponding axis in ``axes``.
9886 9887

    Returns:
W
wangchaochaohu 已提交
9888 9889 9890 9891 9892 9893
        Variable:  A ``Tensor`` or ``LoDTensor`` with the same dimension as ``input``. The data type is same as ``input``.

    Raises:
        TypeError: The type of ``starts`` must be list, tuple or Variable.
        TypeError: The type of ``ends`` must be list, tuple or Variable.
        TypeError: The type of ``strides`` must be list, tuple or Variable.
9894

W
wangchaochaohu 已提交
9895 9896 9897 9898 9899
    Examples:
        .. code-block:: python

            import paddle.fluid as fluid

W
wangchaochaohu 已提交
9900
            input = fluid.data(
W
wangchaochaohu 已提交
9901 9902
                name="input", shape=[3, 4, 5, 6], dtype='float32')

9903 9904 9905 9906 9907
            # example 1:
            # attr starts is a list which doesn't contain tensor Variable.
            axes = [0, 1, 2]
            starts = [-3, 0, 2]
            ends = [3, 2, 4]
W
wangchaochaohu 已提交
9908 9909 9910 9911 9912
            strides_1 = [1, 1, 1]
            strides_2 = [1, 1, 2]
            sliced_1 = fluid.layers.strided_slice(input, axes=axes, starts=starts, ends=ends, strides=strides_1)
            # sliced_1 is input[:, 0:3:1, 0:2:1, 2:4:1].

9913 9914 9915 9916

            # example 2:
            # attr starts is a list which contain tensor Variable.
            minus_3 = fluid.layers.fill_constant([1], "int32", -3)
W
wangchaochaohu 已提交
9917 9918
            sliced_2 = fluid.layers.strided_slice(input, axes=axes, starts=[minus_3, 0, 2], ends=ends, strides=strides_2)
            # sliced_2 is input[:, 0:3:1, 0:2:1, 2:4:2].
W
wangchaochaohu 已提交
9919
    """
9920 9921 9922 9923 9924 9925 9926 9927 9928 9929
    if not isinstance(starts, (list, tuple, Variable)):
        raise ValueError(
            "Input starts must be an Variable, python list or tuple.")
    if not isinstance(ends, (list, tuple, Variable)):
        raise ValueError(
            "Input ends must be an Variable, python list or tuple.")
    if not isinstance(strides, (list, tuple, Variable)):
        raise ValueError(
            "Input strides must be an Variable, python list or tuple.")

W
wangchaochaohu 已提交
9930 9931
    helper = LayerHelper('strided_slice', **locals())

9932 9933 9934 9935 9936 9937 9938 9939 9940 9941 9942 9943 9944 9945 9946 9947 9948 9949 9950 9951 9952 9953 9954 9955 9956 9957
    def contain_var(one_list):
        for ele in one_list:
            if isinstance(ele, Variable):
                return True
        return False

    def get_new_list_tensor(old_list):
        new_list_tensor = []
        for dim in old_list:
            if isinstance(dim, Variable):
                dim.stop_gradient = True
                new_list_tensor.append(dim)
            else:
                assert (isinstance(dim, int))
                temp_out = helper.create_variable_for_type_inference('int32')
                fill_constant([1], 'int32', dim, force_cpu=True, out=temp_out)
                new_list_tensor.append(temp_out)
        return new_list_tensor

    inputs = {'Input': input}
    attrs = {'axes': axes}
    infer_flags = list(1 for i in range(len(axes)))

    if in_dygraph_mode():
        inputs = {'Input': input}
        attrs = {
W
wangchaochaohu 已提交
9958 9959 9960
            'axes': axes,
            'starts': starts,
            'ends': ends,
9961 9962 9963 9964 9965 9966 9967 9968 9969 9970 9971 9972 9973 9974 9975 9976 9977 9978 9979 9980 9981 9982 9983 9984 9985 9986 9987 9988 9989 9990 9991 9992 9993 9994 9995 9996 9997 9998 9999 10000 10001 10002 10003 10004 10005 10006 10007 10008 10009 10010 10011 10012 10013 10014 10015 10016 10017 10018
            'strides': strides,
            'infer_flags': infer_flags
        }
    else:
        # starts
        if isinstance(starts, Variable):
            starts.stop_gradient = True
            inputs['StartsTensor'] = starts
        elif isinstance(starts, (list, tuple)):
            attrs['starts'] = []
            if not contain_var(starts):
                attrs['starts'] = starts
            else:
                inputs['StartsTensorList'] = get_new_list_tensor(starts)
                for i, dim in enumerate(starts):
                    if isinstance(dim, Variable):
                        attrs['starts'].append(-1)
                        infer_flags[i] = -1
                    else:
                        attrs['starts'].append(dim)

        # ends
        if isinstance(ends, Variable):
            ends.stop_gradient = True
            inputs['EndsTensor'] = ends
        elif isinstance(ends, (list, tuple)):
            attrs['ends'] = []
            if not contain_var(ends):
                attrs['ends'] = ends
            else:
                inputs['EndsTensorList'] = get_new_list_tensor(ends)
                for i, dim in enumerate(ends):
                    if isinstance(dim, Variable):
                        attrs['ends'].append(-1)
                        infer_flags[i] = -1
                    else:
                        attrs['ends'].append(dim)
        # strides
        if isinstance(strides, Variable):
            strides.stop_gradient = True
            inputs['StridesTensor'] = strides
        elif isinstance(strides, (list, tuple)):
            attrs['strides'] = []
            if not contain_var(strides):
                attrs['strides'] = strides
            else:
                inputs['StridesTensorList'] = get_new_list_tensor(strides)
                for i, dim in enumerate(strides):
                    if isinstance(dim, Variable):
                        attrs['strides'].append(-1)
                        infer_flags[i] = -1
                    else:
                        attrs['strides'].append(dim)
        attrs['infer_flags'] = infer_flags
    out = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype('input'))
    helper.append_op(
        type='strided_slice', inputs=inputs, attrs=attrs, outputs={'Out': out})
W
wangchaochaohu 已提交
10019 10020 10021 10022

    return out


G
fix  
gongweibao 已提交
10023 10024
def shape(input):
    """
C
chengduozh 已提交
10025 10026
    **Shape Layer**

C
fix doc  
chengduozh 已提交
10027
    Get the shape of the input.
G
fix  
gongweibao 已提交
10028 10029

    Args:
10030
        input (Variable): The input N-D Tensor. Datatype can be float32, float64, int32, int64.
G
fix  
gongweibao 已提交
10031 10032

    Returns:
10033
        Variable (Tensor): The shape of the input variable.
G
fix  
gongweibao 已提交
10034

10035 10036 10037
    Examples:
        .. code-block:: python

10038
            import paddle.fluid as fluid
10039
            import numpy as np
10040

10041 10042 10043 10044 10045 10046 10047 10048 10049 10050
            inputs = fluid.layers.data(name="x", shape=[3, 100, 100], dtype="float32")
            output = fluid.layers.shape(inputs)

            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())

            img = np.ones((3, 100, 100)).astype(np.float32)

            res = exe.run(fluid.default_main_program(), feed={'x':img}, fetch_list=[output])
            print(res) # [array([  3, 100, 100], dtype=int32)]
G
fix  
gongweibao 已提交
10051 10052 10053
    """

    helper = LayerHelper('shape', **locals())
10054
    out = helper.create_variable_for_type_inference(dtype='int32')
G
fix  
gongweibao 已提交
10055
    helper.append_op(
G
fix  
gongweibao 已提交
10056
        type='shape', inputs={'Input': input}, outputs={'Out': out})
G
fix  
gongweibao 已提交
10057 10058

    return out
G
merge  
gongweibao 已提交
10059 10060


Z
zhoukunsheng 已提交
10061 10062
def rank(input):
    """
10063
    The OP returns the number of dimensions for a tensor, which is a 0-D int32 Tensor.
Z
zhoukunsheng 已提交
10064 10065

    Args:
10066
        input (Variable): The input N-D tensor with shape of :math:`[N_1, N_2, ..., N_k]`, the data type is arbitrary.
Z
zhoukunsheng 已提交
10067 10068

    Returns:
10069
        Variable, the output data type is int32.: The 0-D tensor with the dimensions of the input variable.
Z
zhoukunsheng 已提交
10070 10071 10072 10073

    Examples:
        .. code-block:: python

10074 10075
            import paddle.fluid as fluid

10076 10077
            input = fluid.data(name="input", shape=[3, 100, 100], dtype="float32")
            rank = fluid.layers.rank(input) # rank=(3,)
Z
zhoukunsheng 已提交
10078 10079 10080 10081 10082 10083 10084 10085
    """

    ndims = len(input.shape)
    out = assign(np.array(ndims, 'int32'))

    return out


Z
zhoukunsheng 已提交
10086 10087 10088 10089 10090 10091 10092 10093 10094 10095 10096 10097 10098 10099 10100 10101 10102 10103 10104 10105 10106 10107 10108 10109 10110 10111 10112 10113 10114
def size(input):
    """
    **Size Layer**

    Returns the number of elements for a tensor, which is a int64 Tensor with shape [1].

    Args:
        input (Variable): The input variable.

    Returns:
        Variable: The number of elements for the input variable.

    Examples:
        .. code-block:: python

            import paddle.fluid.layers as layers

            input = layers.data(
                name="input", shape=[3, 100], dtype="float32", append_batch_size=False)
            rank = layers.size(input) # 300
    """

    helper = LayerHelper('size', **locals())
    out = helper.create_variable_for_type_inference(dtype='int64')
    helper.append_op(type='size', inputs={'Input': input}, outputs={'Out': out})

    return out


S
sneaxiy 已提交
10115 10116 10117 10118
def _elementwise_op(helper):
    op_type = helper.layer_type
    x = helper.kwargs.get('x', None)
    y = helper.kwargs.get('y', None)
L
lujun 已提交
10119
    if in_dygraph_mode():
X
Xin Pan 已提交
10120 10121 10122
        x = base.to_variable(x)
        y = base.to_variable(y)

S
sneaxiy 已提交
10123 10124
    assert x is not None, 'x cannot be None in {}'.format(op_type)
    assert y is not None, 'y cannot be None in {}'.format(op_type)
10125 10126 10127 10128 10129 10130
    check_type_and_dtype(x, 'x', Variable,
                         ['float16', 'float32', 'float64', 'int32', 'int64'],
                         op_type)
    check_type_and_dtype(y, 'y', Variable,
                         ['float16', 'float32', 'float64', 'int32', 'int64'],
                         op_type)
10131

S
sneaxiy 已提交
10132 10133
    axis = helper.kwargs.get('axis', -1)
    use_mkldnn = helper.kwargs.get('use_mkldnn', False)
S
sneaxiy 已提交
10134 10135
    name = helper.kwargs.get('name', None)
    if name is None:
X
Xin Pan 已提交
10136
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
S
sneaxiy 已提交
10137 10138 10139
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)
S
sneaxiy 已提交
10140

S
sneaxiy 已提交
10141 10142 10143 10144 10145 10146 10147 10148 10149 10150
    helper.append_op(
        type=op_type,
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
        attrs={'axis': axis,
               'use_mkldnn': use_mkldnn})
    return helper.append_activation(out)


S
sneaxiy 已提交
10151
def scale(x, scale=1.0, bias=0.0, bias_after_scale=True, act=None, name=None):
S
sneaxiy 已提交
10152
    """
10153 10154 10155 10156 10157 10158 10159 10160 10161 10162 10163 10164 10165
    Scale operator.

    Putting scale and bias to the input Tensor as following:

    ``bias_after_scale`` is True:

    .. math::
                            Out=scale*X+bias

    ``bias_after_scale`` is False:

    .. math::
                            Out=scale*(X+bias)
S
sneaxiy 已提交
10166 10167

    Args:
10168
        x(Variable): Input N-D Tensor of scale operator. Data type can be float32, float64, int8, int16, int32, int64, uint8.
10169
        scale(float|Variable): The scale factor of the input, it should be a float number or a Variable with shape [1] and data type as float32.
10170 10171 10172 10173
        bias(float): The bias to be put on the input.
        bias_after_scale(bool): Apply bias addition after or before scaling. It is useful for numeric stability in some circumstances.
        act(str, optional): Activation applied to the output such as tanh, softmax, sigmoid, relu.
        name(str, optional): The default value is None. Normally there is no need for user to set this property.  For more information, please refer to :ref:`api_guide_Name` 
S
sneaxiy 已提交
10174 10175

    Returns:
10176
        Variable(Tensor|LoDTensor): Output tensor of scale operator, with shape and data type same as input.
10177 10178 10179 10180 10181

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
10182 10183 10184 10185 10186 10187 10188 10189 10190
            import numpy as np

            inputs = fluid.layers.data(name="x", shape=[2, 3], dtype='float32')
            output = fluid.layers.scale(inputs, scale = 2.0, bias = 1.0)

            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())

            img = np.array([[1, 2, 3], [4, 5, 6]]).astype(np.float32)
10191

10192 10193
            res = exe.run(fluid.default_main_program(), feed={'x':img}, fetch_list=[output])
            print(res) # [array([[ 3.,  5.,  7.], [ 9., 11., 13.]], dtype=float32)]
10194 10195 10196 10197 10198 10199 10200 10201

        .. code-block:: python

            # scale with parameter scale as Variable
            import paddle.fluid as fluid
            import numpy as np

            inputs = fluid.layers.data(name="x", shape=[2, 3], dtype='float32')
10202
            scale = fluid.layers.data(name="scale", shape=[1], dtype='float32',
10203 10204 10205 10206 10207 10208 10209 10210 10211 10212 10213 10214
                                      append_batch_size=False)
            output = fluid.layers.scale(inputs, scale = scale, bias = 1.0)

            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())

            img = np.array([[1, 2, 3], [4, 5, 6]]).astype(np.float32)
            scale_np = np.array([2.]).astype(np.float32)

            res = exe.run(fluid.default_main_program(), feed={'x':img, 'scale':scale_np}, fetch_list=[output])
            print(res) # [array([[ 3.,  5.,  7.], [ 9., 11., 13.]], dtype=float32)]

S
sneaxiy 已提交
10215 10216 10217
    """

    helper = LayerHelper('scale', **locals())
S
sneaxiy 已提交
10218
    if name is None:
X
Xin Pan 已提交
10219
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
S
sneaxiy 已提交
10220 10221 10222
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)
S
sneaxiy 已提交
10223

10224 10225 10226 10227 10228 10229 10230 10231 10232 10233
    inputs = {'X': x}
    attrs = {
        'bias': float(bias),
        'bias_after_scale': bias_after_scale,
    }
    if isinstance(scale, Variable):
        inputs['ScaleTensor'] = scale
    else:
        attrs['scale'] = float(scale)

S
sneaxiy 已提交
10234
    helper.append_op(
10235
        type='scale', inputs=inputs, outputs={'Out': out}, attrs=attrs)
S
sneaxiy 已提交
10236
    return helper.append_activation(out)
S
sneaxiy 已提交
10237 10238


X
Xin Pan 已提交
10239
def elementwise_add(x, y, axis=-1, act=None, name=None):
10240 10241 10242 10243 10244 10245 10246 10247 10248 10249
    """
Examples:

    .. code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
10250 10251
                "x": np.array([2, 3, 4]).astype('float32'),
                "y": np.array([1, 5, 2]).astype('float32')
10252 10253
            }

10254 10255
        x = fluid.data(name="x", shape=[3], dtype='float32')
        y = fluid.data(name="y", shape=[3], dtype='float32')
10256 10257 10258 10259 10260 10261 10262 10263 10264 10265 10266 10267 10268 10269 10270 10271 10272 10273 10274 10275 10276
        z = fluid.layers.elementwise_add(x, y)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])

        print(z_value) #[3., 8., 6.]


    .. code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
                "x": np.ones((2, 3, 4, 5)).astype('float32'),
                "y": np.zeros((3, 4)).astype('float32')
            }

10277 10278
        x = fluid.data(name="x", shape=[2,3,4,5], dtype='float32')
        y = fluid.data(name="y", shape=[3,4], dtype='float32')
10279 10280 10281 10282 10283 10284 10285 10286 10287 10288 10289 10290 10291 10292 10293 10294 10295 10296 10297 10298 10299 10300
        z = fluid.layers.elementwise_add(x, y, axis=1)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)

        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])

        print(z_value) # z.shape=[2,3,4,5]


    ..  code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
                "x": np.random.randint(1, 5, size=[2, 3, 4, 5]).astype('float32'),
                "y": np.random.randint(1, 5, size=[5]).astype('float32')
            }
        
10301 10302
        x = fluid.data(name="x", shape=[2,3,4,5], dtype='float32')
        y = fluid.data(name="y", shape=[5], dtype='float32')
10303 10304 10305 10306 10307 10308 10309 10310 10311 10312
        z = fluid.layers.elementwise_add(x, y, axis=3)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)

        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])
        print(z_value) # z.shape=[2,3,4,5]

    """
S
sneaxiy 已提交
10313 10314 10315
    return _elementwise_op(LayerHelper('elementwise_add', **locals()))


X
Xin Pan 已提交
10316
def elementwise_div(x, y, axis=-1, act=None, name=None):
10317 10318 10319 10320 10321 10322 10323 10324 10325 10326
    """
Examples:

    .. code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
10327 10328
                "x": np.array([2, 3, 4]).astype('float32'),
                "y": np.array([1, 5, 2]).astype('float32')
10329 10330
            }

10331 10332
        x = fluid.data(name="x", shape=[3], dtype='float32')
        y = fluid.data(name="y", shape=[3], dtype='float32')
10333 10334 10335 10336 10337 10338 10339 10340 10341 10342 10343 10344 10345 10346 10347 10348 10349 10350 10351 10352 10353
        z = fluid.layers.elementwise_div(x, y)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])

        print(z_value) #[2., 0.6, 2.]


    .. code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
                "x": np.ones((2, 3, 4, 5)).astype('float32'),
                "y": np.zeros((3, 4)).astype('float32')
            }

10354 10355
        x = fluid.data(name="x", shape=[2,3,4,5], dtype='float32')
        y = fluid.data(name="y", shape=[3,4], dtype='float32')
10356 10357 10358 10359 10360 10361 10362 10363 10364 10365 10366 10367 10368 10369 10370 10371 10372 10373 10374 10375 10376 10377
        z = fluid.layers.elementwise_div(x, y, axis=1)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)

        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])

        print(z_value) # z.shape=[2,3,4,5]


    ..  code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
                "x": np.random.randint(1, 5, size=[2, 3, 4, 5]).astype('float32'),
                "y": np.random.randint(1, 5, size=[5]).astype('float32')
            }
        
10378 10379
        x = fluid.data(name="x", shape=[2,3,4,5], dtype='float32')
        y = fluid.data(name="y", shape=[5], dtype='float32')
10380 10381 10382 10383 10384 10385 10386 10387 10388 10389
        z = fluid.layers.elementwise_div(x, y, axis=3)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        
        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])
        print(z_value) # z.shape=[2,3,4,5]

    """
S
sneaxiy 已提交
10390 10391 10392
    return _elementwise_op(LayerHelper('elementwise_div', **locals()))


X
Xin Pan 已提交
10393
def elementwise_sub(x, y, axis=-1, act=None, name=None):
10394 10395 10396 10397 10398 10399 10400 10401 10402 10403
    """
Examples:

    .. code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
10404 10405
                "x": np.array([2, 3, 4]).astype('float32'),
                "y": np.array([1, 5, 2]).astype('float32')
10406 10407
            }

10408 10409
        x = fluid.data(name="x", shape=[3], dtype='float32')
        y = fluid.data(name="y", shape=[3], dtype='float32')
10410 10411 10412 10413 10414 10415 10416 10417 10418 10419 10420 10421 10422 10423 10424 10425 10426 10427 10428 10429 10430
        z = fluid.layers.elementwise_sub(x, y)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])

        print(z_value) #[1., -2., 2.]


    .. code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
                "x": np.ones((2, 3, 4, 5)).astype('float32'),
                "y": np.zeros((3, 4)).astype('float32')
            }

10431 10432
        x = fluid.data(name="x", shape=[2,3,4,5], dtype='float32')
        y = fluid.data(name="y", shape=[3,4], dtype='float32')
10433 10434 10435 10436 10437 10438 10439 10440 10441 10442 10443 10444 10445 10446 10447 10448 10449 10450 10451 10452 10453 10454
        z = fluid.layers.elementwise_sub(x, y, axis=1)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)

        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])

        print(z_value) # z.shape=[2,3,4,5]


    ..  code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
                "x": np.random.randint(1, 5, size=[2, 3, 4, 5]).astype('float32'),
                "y": np.random.randint(1, 5, size=[5]).astype('float32')
            }
        
10455 10456
        x = fluid.data(name="x", shape=[2,3,4,5], dtype='float32')
        y = fluid.data(name="y", shape=[5], dtype='float32')
10457 10458 10459 10460 10461 10462 10463 10464 10465 10466
        z = fluid.layers.elementwise_sub(x, y, axis=3)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        
        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])
        print(z_value) # z.shape=[2,3,4,5]

    """
S
sneaxiy 已提交
10467 10468 10469
    return _elementwise_op(LayerHelper('elementwise_sub', **locals()))


X
Xin Pan 已提交
10470
def elementwise_mul(x, y, axis=-1, act=None, name=None):
10471 10472 10473 10474 10475 10476 10477 10478 10479 10480
    """
Examples:

    .. code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
10481 10482
                "x": np.array([2, 3, 4]).astype('float32'),
                "y": np.array([1, 5, 2]).astype('float32')
10483 10484
            }

10485 10486
        x = fluid.data(name="x", shape=[3], dtype='float32')
        y = fluid.data(name="y", shape=[3], dtype='float32')
10487 10488 10489 10490 10491 10492 10493 10494 10495 10496 10497 10498 10499 10500 10501 10502 10503 10504 10505 10506 10507
        z = fluid.layers.elementwise_mul(x, y)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])

        print(z_value) #[2., 15., 8.]


    .. code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
                "x": np.ones((2, 3, 4, 5)).astype('float32'),
                "y": np.zeros((3, 4)).astype('float32')
            }

10508 10509
        x = fluid.data(name="x", shape=[2,3,4,5], dtype='float32')
        y = fluid.data(name="y", shape=[3,4], dtype='float32')
10510 10511 10512 10513 10514 10515 10516 10517 10518 10519 10520 10521 10522 10523 10524 10525 10526 10527 10528 10529 10530 10531
        z = fluid.layers.elementwise_mul(x, y, axis=1)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)

        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])

        print(z_value) # z.shape=[2,3,4,5]


    ..  code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
                "x": np.random.randint(1, 5, size=[2, 3, 4, 5]).astype('float32'),
                "y": np.random.randint(1, 5, size=[5]).astype('float32')
            }
        
10532 10533
        x = fluid.data(name="x", shape=[2,3,4,5], dtype='float32')
        y = fluid.data(name="y", shape=[5], dtype='float32')
10534 10535 10536 10537 10538 10539 10540 10541 10542 10543
        z = fluid.layers.elementwise_mul(x, y, axis=3)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        
        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])
        print(z_value) # z.shape=[2,3,4,5]
 
    """
S
sneaxiy 已提交
10544 10545 10546
    return _elementwise_op(LayerHelper('elementwise_mul', **locals()))


X
Xin Pan 已提交
10547
def elementwise_max(x, y, axis=-1, act=None, name=None):
10548 10549 10550 10551 10552 10553 10554 10555 10556 10557
    """
Examples:

    .. code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
10558 10559
                "x": np.array([2, 3, 4]).astype('float32'),
                "y": np.array([1, 5, 2]).astype('float32')
10560 10561
            }

10562 10563
        x = fluid.data(name="x", shape=[3], dtype='float32')
        y = fluid.data(name="y", shape=[3], dtype='float32')
10564 10565 10566 10567 10568 10569 10570 10571 10572 10573 10574 10575 10576 10577 10578 10579 10580 10581 10582 10583 10584
        z = fluid.layers.elementwise_max(x, y)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])

        print(z_value) #[2, 5, 4]


    .. code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
                "x": np.ones((2, 3, 4, 5)).astype('float32'),
                "y": np.zeros((3, 4)).astype('float32')
            }

10585 10586
        x = fluid.data(name="x", shape=[2,3,4,5], dtype='float32')
        y = fluid.data(name="y", shape=[3,4], dtype='float32')
10587 10588 10589 10590 10591 10592 10593 10594 10595 10596 10597
        z = fluid.layers.elementwise_max(x, y, axis=1)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)

        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])

        print(z_value)#[[[[1., 1., 1., 1., 1.] .... [1., 1., 1., 1., 1.]]]]

    """
S
sneaxiy 已提交
10598 10599 10600
    return _elementwise_op(LayerHelper('elementwise_max', **locals()))


X
Xin Pan 已提交
10601
def elementwise_min(x, y, axis=-1, act=None, name=None):
10602 10603 10604 10605 10606 10607 10608 10609 10610 10611
    """
Examples:

    ..  code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
10612 10613
                "x": np.array([2, 3, 4]).astype('float32'),
                "y": np.array([1, 5, 2]).astype('float32')
10614 10615
            }

10616 10617
        x = fluid.data(name="x", shape=[3], dtype='float32')
        y = fluid.data(name="y", shape=[3], dtype='float32')
10618 10619 10620 10621 10622 10623 10624 10625 10626 10627 10628 10629 10630 10631 10632 10633 10634 10635 10636 10637
        z = fluid.layers.elementwise_max(x, y)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])

        print(z_value) #[1, 3, 2]

    ..  code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
                "x": np.ones((2, 3, 4, 5)).astype('float32'),
                "y": np.zeros((3, 4)).astype('float32')
            }

10638 10639
        x = fluid.data(name="x", shape=[2,3,4,5], dtype='float32')
        y = fluid.data(name="y", shape=[3,4], dtype='float32')
10640 10641 10642 10643 10644 10645 10646 10647 10648 10649 10650
        z = fluid.layers.elementwise_max(x, y, axis=1)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)

        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])

        print(z_value)#[[[[0., 0., 0., 0., 0.] .... [0., 0., 0., 0., 0.]]]]
    """

S
sneaxiy 已提交
10651 10652 10653
    return _elementwise_op(LayerHelper('elementwise_min', **locals()))


X
Xin Pan 已提交
10654
def elementwise_pow(x, y, axis=-1, act=None, name=None):
10655 10656 10657 10658 10659 10660 10661 10662 10663 10664
    """
Examples:

    ..  code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
10665 10666
                "x": np.array([2, 3, 4]).astype('float32'),
                "y": np.array([1, 5, 2]).astype('float32')
10667 10668
            }

10669 10670
        x = fluid.data(name="x", shape=[3], dtype='float32')
        y = fluid.data(name="y", shape=[3], dtype='float32')
10671 10672 10673 10674 10675 10676 10677 10678 10679 10680
        z = fluid.layers.elementwise_pow(x, y)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])

        print(z_value) #[2, 243, 16]
    """

S
sneaxiy 已提交
10681 10682 10683
    return _elementwise_op(LayerHelper('elementwise_pow', **locals()))


10684
def elementwise_mod(x, y, axis=-1, act=None, name=None):
10685 10686 10687 10688 10689 10690 10691 10692 10693 10694 10695 10696 10697 10698 10699 10700 10701 10702 10703 10704 10705 10706 10707 10708 10709
    """
Examples:

    ..  code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
                "x": np.array([10, 15, 8]).astype('int32'),
                "y": np.array([3, 6, 5]).astype('int32')
            }

        x = fluid.data(name="x", shape=[3], dtype='int32')
        y = fluid.data(name="y", shape=[3], dtype='int32')
        z = fluid.layers.elementwise_mod(x, y)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])

        print(z_value) #[1, 3, 3]
    """
10710 10711 10712 10713
    return _elementwise_op(LayerHelper('elementwise_mod', **locals()))


def elementwise_floordiv(x, y, axis=-1, act=None, name=None):
10714 10715 10716 10717 10718 10719 10720 10721 10722 10723 10724 10725 10726 10727 10728 10729 10730 10731 10732 10733 10734 10735 10736 10737 10738
    """
Examples:

    ..  code-block:: python

        import paddle.fluid as fluid
        import numpy as np

        def gen_data():
            return {
                "x": np.array([10, 15, 8]).astype('int32'),
                "y": np.array([3, 7, 5]).astype('int32')
            }

        x = fluid.data(name="x", shape=[3], dtype='int32')
        y = fluid.data(name="y", shape=[3], dtype='int32')
        z = fluid.layers.elementwise_floordiv(x, y)

        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        z_value = exe.run(feed=gen_data(),
                            fetch_list=[z.name])

        print(z_value) #[3, 2, 1]
    """
10739 10740 10741
    return _elementwise_op(LayerHelper('elementwise_floordiv', **locals()))


S
sneaxiy 已提交
10742
for func in [
10743 10744 10745 10746
        elementwise_add,
        elementwise_div,
        elementwise_sub,
        elementwise_mul,
10747 10748
        elementwise_max,
        elementwise_pow,
10749
        elementwise_min,
10750 10751
        elementwise_mod,
        elementwise_floordiv,
10752 10753 10754 10755 10756 10757 10758 10759 10760 10761 10762 10763 10764 10765 10766 10767 10768
]:
    op_proto = OpProtoHolder.instance().get_op_proto(func.__name__)
    func.__doc__ = _generate_doc_string_(
        op_proto,
        additional_args_lines=[
            "axis (int32, optional): If X.dimension != Y.dimension, \
            Y.dimension must be a subsequence of x.dimension. \
            And axis is the start dimension index for broadcasting Y onto X. ",
            "act (string, optional): Activation applied to the output. \
            Default is None. Details: :ref:`api_guide_activations_en` ",
            "name (string, optional): Name of the output. \
            Default is None. It's used to print debug info for developers. Details: \
            :ref:`api_guide_Name` "
        ],
        skip_attrs_set={"x_data_format", "y_data_format", "axis"
                        }) + """\n""" + str(func.__doc__)

10769
for func in []:
S
sneaxiy 已提交
10770 10771 10772 10773
    op_proto = OpProtoHolder.instance().get_op_proto(func.__name__)
    func.__doc__ = _generate_doc_string_(
        op_proto,
        additional_args_lines=[
S
sneaxiy 已提交
10774 10775
            "act (basestring|None): Activation applied to the output.",
            "name (basestring|None): Name of the output."
S
sneaxiy 已提交
10776
        ])
10777 10778 10779 10780 10781 10782 10783 10784 10785 10786 10787 10788 10789 10790 10791 10792 10793 10794 10795 10796 10797 10798 10799 10800 10801 10802 10803 10804 10805 10806 10807 10808 10809 10810 10811 10812 10813
    func.__doc__ = func.__doc__ + """

Examples:
  .. code-block:: python
    
    import paddle.fluid as fluid
    # example 1: shape(x) = (2, 3, 4, 5), shape(y) = (2, 3, 4, 5)
    x0 = fluid.layers.data(name="x0", shape=[2, 3, 4, 5], dtype='float32')
    y0 = fluid.layers.data(name="y0", shape=[2, 3, 4, 5], dtype='float32')
    z0 = fluid.layers.%s(x0, y0)

    # example 2: shape(X) = (2, 3, 4, 5), shape(Y) = (5)
    x1 = fluid.layers.data(name="x1", shape=[2, 3, 4, 5], dtype='float32')
    y1 = fluid.layers.data(name="y1", shape=[5], dtype='float32')
    z1 = fluid.layers.%s(x1, y1)

    # example 3: shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5), with axis=-1(default) or axis=2
    x2 = fluid.layers.data(name="x2", shape=[2, 3, 4, 5], dtype='float32')
    y2 = fluid.layers.data(name="y2", shape=[4, 5], dtype='float32')
    z2 = fluid.layers.%s(x2, y2, axis=2)

    # example 4: shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
    x3 = fluid.layers.data(name="x3", shape=[2, 3, 4, 5], dtype='float32')
    y3 = fluid.layers.data(name="y3", shape=[3, 4], dtype='float32')
    z3 = fluid.layers.%s(x3, y3, axis=1)

    # example 5: shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0
    x4 = fluid.layers.data(name="x4", shape=[2, 3, 4, 5], dtype='float32')
    y4 = fluid.layers.data(name="y4", shape=[2], dtype='float32')
    z4 = fluid.layers.%s(x4, y4, axis=0)

    # example 6: shape(X) = (2, 3, 4, 5), shape(Y) = (2, 1), with axis=0
    x5 = fluid.layers.data(name="x5", shape=[2, 3, 4, 5], dtype='float32')
    y5 = fluid.layers.data(name="y5", shape=[2], dtype='float32')
    z5 = fluid.layers.%s(x5, y5, axis=0)
    """ % (func.__name__, func.__name__, func.__name__, func.__name__,
           func.__name__, func.__name__)
M
minqiyang 已提交
10814 10815


10816
def _logical_op(op_name, x, y, out=None, name=None, binary_op=True):
M
minqiyang 已提交
10817 10818
    helper = LayerHelper(op_name, **locals())

M
minqiyang 已提交
10819 10820
    if binary_op:
        assert x.dtype == y.dtype
M
minqiyang 已提交
10821 10822 10823

    if out is None:
        if name is None:
X
Xin Pan 已提交
10824
            out = helper.create_variable_for_type_inference(dtype=x.dtype)
M
minqiyang 已提交
10825 10826 10827 10828 10829 10830 10831 10832 10833 10834 10835 10836 10837 10838 10839
        else:
            out = helper.create_variable(
                name=name, dtype=x.dtype, persistable=False)

    if binary_op:
        helper.append_op(
            type=op_name, inputs={"X": x,
                                  "Y": y}, outputs={"Out": out})
    else:
        helper.append_op(type=op_name, inputs={"X": x}, outputs={"Out": out})

    return out


@templatedoc()
10840
def logical_and(x, y, out=None, name=None):
M
minqiyang 已提交
10841
    """
W
Wilber 已提交
10842 10843 10844 10845 10846 10847 10848 10849
    logical_and Operator

    It operates element-wise on X and Y, and returns the Out. X, Y and Out are N-dim boolean LoDTensor or Tensor.
    Each element of Out is calculated by
    
    .. math::

        Out = X \land Y
M
minqiyang 已提交
10850 10851 10852 10853

    Args:
        x(${x_type}): ${x_comment}
        y(${y_type}): ${y_comment}
W
Wilber 已提交
10854 10855
        out(LoDTensor or Tensor): The LoDTensor or Tensor that specifies the output of the operator, which can be any Variable that has been created in the program. The default value is None, and a new Variable will be created to save the output.
        name(str|None): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`
M
minqiyang 已提交
10856 10857

    Returns:
W
Wilber 已提交
10858
        ${out_type}: ${out_comment}
10859 10860 10861 10862

    Examples:
        .. code-block:: python

10863
            import paddle.fluid as fluid
W
Wilber 已提交
10864 10865 10866 10867 10868 10869 10870 10871 10872 10873 10874 10875 10876 10877 10878 10879 10880 10881
            import numpy as np

            # Graph organizing
            x = fluid.layers.data(name='x', shape=[2], dtype='bool')
            y = fluid.layers.data(name='y', shape=[2], dtype='bool')
            res = fluid.layers.logical_and(x=x, y=y)
            # The comment lists another available method.
            # res = fluid.layers.fill_constant(shape=[2], dtype='bool', value=0)
            # fluid.layers.logical_and(x=x, y=y, out=res)

            # Create an executor using CPU as an example
            exe = fluid.Executor(fluid.CPUPlace())

            # Execute
            x_i = np.array([[1, 0], [0, 1]]).astype(np.bool)
            y_i = np.array([[1, 1], [0, 0]]).astype(np.bool)
            res_val, = exe.run(fluid.default_main_program(), feed={'x':x_i, 'y':y_i}, fetch_list=[res])
            print(res_val) # [[True, False], [False, False]]
M
minqiyang 已提交
10882 10883 10884 10885 10886 10887 10888
    """

    return _logical_op(
        op_name="logical_and", x=x, y=y, name=name, out=out, binary_op=True)


@templatedoc()
10889
def logical_or(x, y, out=None, name=None):
M
minqiyang 已提交
10890
    """
W
Wilber 已提交
10891 10892 10893 10894 10895 10896 10897 10898
    logical_or Operator

    It operates element-wise on X and Y, and returns the Out. X, Y and Out are N-dim boolean LoDTensor or Tensor.
    Each element of Out is calculated by
    
    .. math::

        Out = X \lor Y
M
minqiyang 已提交
10899 10900 10901 10902

    Args:
        x(${x_type}): ${x_comment}
        y(${y_type}): ${y_comment}
W
Wilber 已提交
10903 10904
        out(LoDTensor or Tensor): The LoDTensor or Tensor that specifies the output of the operator, which can be any Variable that has been created in the program. The default value is None, and a new Variable will be created to save the output.
        name(str|None): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`
M
minqiyang 已提交
10905 10906

    Returns:
W
Wilber 已提交
10907
        ${out_type}: ${out_comment}
10908 10909 10910 10911

    Examples:
        .. code-block:: python

10912
            import paddle.fluid as fluid
W
Wilber 已提交
10913 10914 10915 10916 10917 10918 10919 10920 10921 10922 10923 10924 10925 10926 10927 10928 10929 10930
            import numpy as np

            # Graph organizing
            x = fluid.layers.data(name='x', shape=[2], dtype='bool')
            y = fluid.layers.data(name='y', shape=[2], dtype='bool')
            res = fluid.layers.logical_or(x=x, y=y)
            # The comment lists another available method.
            # res = fluid.layers.fill_constant(shape=[2], dtype='bool', value=0)
            # fluid.layers.logical_or(x=x, y=y, out=res)

            # Create an executor using CPU as an example
            exe = fluid.Executor(fluid.CPUPlace())

            # Execute
            x_i = np.array([[1, 0], [0, 1]]).astype(np.bool)
            y_i = np.array([[1, 1], [0, 0]]).astype(np.bool)
            res_val, = exe.run(fluid.default_main_program(), feed={'x':x_i, 'y':y_i}, fetch_list=[res])
            print(res_val) # [[True, True], [False, True]]
M
minqiyang 已提交
10931 10932 10933 10934 10935 10936 10937
    """

    return _logical_op(
        op_name="logical_or", x=x, y=y, name=name, out=out, binary_op=True)


@templatedoc()
10938
def logical_xor(x, y, out=None, name=None):
M
minqiyang 已提交
10939
    """
W
Wilber 已提交
10940 10941 10942 10943 10944 10945 10946 10947
    logical_xor Operator

    It operates element-wise on X and Y, and returns the Out. X, Y and Out are N-dim boolean LoDTensor or Tensor.
    Each element of Out is calculated by
    
    .. math::

        Out = (X \lor Y) \land \lnot (X \land Y)
M
minqiyang 已提交
10948 10949 10950 10951

    Args:
        x(${x_type}): ${x_comment}
        y(${y_type}): ${y_comment}
W
Wilber 已提交
10952 10953
        out(LoDTensor or Tensor): The LoDTensor or Tensor that specifies the output of the operator, which can be any Variable that has been created in the program. The default value is None, and a new Variable will be created to save the output.
        name(str|None): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`
M
minqiyang 已提交
10954 10955

    Returns:
W
Wilber 已提交
10956
        ${out_type}: ${out_comment}
10957 10958 10959 10960

    Examples:
        .. code-block:: python

10961
            import paddle.fluid as fluid
W
Wilber 已提交
10962 10963 10964 10965 10966 10967 10968 10969 10970 10971 10972 10973 10974 10975 10976 10977 10978 10979
            import numpy as np

            # Graph organizing
            x = fluid.layers.data(name='x', shape=[2], dtype='bool')
            y = fluid.layers.data(name='y', shape=[2], dtype='bool')
            res = fluid.layers.logical_xor(x=x, y=y)
            # The comment lists another available method.
            # res = fluid.layers.fill_constant(shape=[2], dtype='bool', value=0)
            # fluid.layers.logical_xor(x=x, y=y, out=res)

            # Create an executor using CPU as an example
            exe = fluid.Executor(fluid.CPUPlace())

            # Execute
            x_i = np.array([[1, 0], [0, 1]]).astype(np.bool)
            y_i = np.array([[1, 1], [0, 0]]).astype(np.bool)
            res_val, = exe.run(fluid.default_main_program(), feed={'x':x_i, 'y':y_i}, fetch_list=[res])
            print(res_val) # [[False, True], [False, True]]
M
minqiyang 已提交
10980 10981 10982 10983 10984 10985 10986
    """

    return _logical_op(
        op_name="logical_xor", x=x, y=y, name=name, out=out, binary_op=True)


@templatedoc()
10987
def logical_not(x, out=None, name=None):
M
minqiyang 已提交
10988
    """
W
Wilber 已提交
10989 10990 10991 10992 10993 10994 10995 10996
    logical_not Operator

    It operates element-wise on X, and returns the Out. X and Out are N-dim boolean LoDTensor or Tensor.
    Each element of Out is calculated by
    
    .. math::

        Out = \lnot X
M
minqiyang 已提交
10997 10998 10999

    Args:
        x(${x_type}): ${x_comment}
W
Wilber 已提交
11000 11001
        out(LoDTensor/Tensor): The LoDTensor/Tensor that specifies the output of the operator, which can be any Variable that has been created in the program. The default value is None, and a new Variable will be created to save the output.
        name(str|None): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`
M
minqiyang 已提交
11002 11003

    Returns:
W
Wilber 已提交
11004
        ${out_type}: ${out_comment}
11005 11006 11007 11008

    Examples:
        .. code-block:: python

11009
            import paddle.fluid as fluid
W
Wilber 已提交
11010 11011 11012 11013 11014 11015 11016 11017 11018 11019 11020 11021 11022 11023 11024 11025
            import numpy as np

            # Graph organizing
            x = fluid.layers.data(name='x', shape=[2], dtype='bool')
            res = fluid.layers.logical_not(x)
            # The comment lists another availble method.
            # res = fluid.layers.fill_constant(shape=[2], dtype='bool', value=0)
            # fluid.layers.logical_not(x, out=res)

            # Create an executor using CPU as an example
            exe = fluid.Executor(fluid.CPUPlace())

            # Execute
            x_i = np.array([[1, 0]]).astype(np.bool)
            res_val, = exe.run(fluid.default_main_program(), feed={'x':x_i}, fetch_list=[res])
            print(res_val) # [[False, True]]
M
minqiyang 已提交
11026 11027 11028 11029
    """

    return _logical_op(
        op_name="logical_not", x=x, y=None, name=name, out=out, binary_op=False)
11030 11031 11032 11033 11034 11035 11036 11037 11038


@templatedoc()
def clip(x, min, max, name=None):
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
S
SunGaofeng 已提交
11039 11040 11041 11042 11043
        min(float): ${min_comment}
        max(float): ${max_comment}
        name(str, optional): The default value is None.  
                             Normally there is no need for user to set this property.  
                             For more information, please refer to :ref:`api_guide_Name`
11044 11045

    Returns:
S
SunGaofeng 已提交
11046 11047 11048 11049
        ${out_comment}

    Return Type:
        ${out_type}
11050 11051 11052 11053

    Examples:
        .. code-block:: python

S
SunGaofeng 已提交
11054
            import paddle.fluid as fluid
S
SunGaofeng 已提交
11055
            input = fluid.data(
11056 11057
                name='data', shape=[1], dtype='float32')
            reward = fluid.layers.clip(x=input, min=-1.0, max=1.0)
11058 11059 11060 11061 11062
    """

    helper = LayerHelper("clip", **locals())

    if name is None:
11063 11064
        name = unique_name.generate_with_ignorable_key(".".join(
            [helper.name, 'tmp']))
S
sneaxiy 已提交
11065 11066 11067

    out = helper.create_variable(
        type=x.type, name=name, dtype=x.dtype, persistable=False)
11068 11069 11070 11071 11072 11073 11074 11075 11076 11077 11078 11079 11080 11081 11082 11083 11084 11085 11086

    helper.append_op(
        type="clip",
        inputs={"X": x},
        attrs={"min": min,
               "max": max},
        outputs={"Out": out})

    return out


@templatedoc()
def clip_by_norm(x, max_norm, name=None):
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        max_norm(${max_norm_type}): ${max_norm_comment}
W
wangguanzhong 已提交
11087 11088 11089
        name(str, optional): For detailed information, please refer 
            to :ref:`api_guide_Name`. Usually name is no need to set and 
            None by default. 
11090 11091

    Returns:
W
wangguanzhong 已提交
11092 11093
        Variable:

11094
        out(${out_type}): ${out_comment}
11095

W
wangguanzhong 已提交
11096

11097 11098 11099
    Examples:
        .. code-block:: python

11100
            import paddle.fluid as fluid
11101 11102
            input = fluid.data(
                name='data', shape=[None, 1], dtype='float32')
11103
            reward = fluid.layers.clip_by_norm(x=input, max_norm=1.0)
11104 11105 11106 11107 11108
    """

    helper = LayerHelper("clip_by_norm", **locals())

    if name is None:
11109 11110
        name = unique_name.generate_with_ignorable_key(".".join(
            [helper.name, 'tmp']))
S
sneaxiy 已提交
11111 11112 11113

    out = helper.create_variable(
        type=x.type, name=name, dtype=x.dtype, persistable=False)
11114 11115 11116 11117 11118 11119 11120 11121

    helper.append_op(
        type="clip_by_norm",
        inputs={"X": x},
        attrs={"max_norm": max_norm},
        outputs={"Out": out})

    return out
X
Xin Pan 已提交
11122 11123 11124 11125 11126 11127 11128 11129 11130 11131 11132 11133 11134


@templatedoc()
def mean(x, name=None):
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        name(basestring|None): Name of the output.

    Returns:
        out(${out_type}): ${out_comment}
11135 11136 11137 11138

    Examples:
        .. code-block:: python

11139
            import paddle.fluid as fluid
11140 11141 11142
            input = fluid.layers.data(
                name='data', shape=[2, 3], dtype='float32')
            mean = fluid.layers.mean(input)
X
Xin Pan 已提交
11143 11144 11145
    """

    helper = LayerHelper("mean", **locals())
11146 11147
    check_type_and_dtype(x, 'x', Variable, ['float16', 'float32', 'float64'],
                         'mean')
X
Xin Pan 已提交
11148
    if name is None:
X
Xin Pan 已提交
11149
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
11150 11151 11152 11153 11154 11155 11156 11157 11158 11159
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    helper.append_op(
        type="mean", inputs={"X": x}, attrs={}, outputs={"Out": out})

    return out


C
chengduo 已提交
11160 11161 11162 11163 11164 11165 11166 11167 11168 11169 11170
@templatedoc()
def merge_selected_rows(x, name=None):
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        name(basestring|None): Name of the output.

    Returns:
        out(${out_type}): ${out_comment}
11171 11172 11173 11174

    Examples:
        .. code-block:: python

11175
            import paddle.fluid as fluid
11176 11177 11178 11179 11180
            b = fluid.default_main_program().global_block()
            var = b.create_var(
                name="X", dtype="float32", persistable=True,
                type=fluid.core.VarDesc.VarType.SELECTED_ROWS)
            y = fluid.layers.merge_selected_rows(var)
C
chengduo 已提交
11181 11182 11183 11184 11185 11186 11187 11188 11189 11190 11191 11192
    """

    helper = LayerHelper("merge_selected_rows", **locals())
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
        type="merge_selected_rows",
        inputs={"X": x},
        attrs={},
        outputs={"Out": out})
    return out


X
Xin Pan 已提交
11193 11194
def mul(x, y, x_num_col_dims=1, y_num_col_dims=1, name=None):
    """
L
liu zhengxi 已提交
11195 11196 11197 11198 11199 11200 11201 11202
    Mul Operator.
    This operator is used to perform matrix multiplication for input $x$ and $y$.
    The equation is:

    ..  math::
        Out = x * y

    Both the input $x$ and $y$ can carry the LoD (Level of Details) information, or not. But the output only shares the LoD information with input $x$.
X
Xin Pan 已提交
11203 11204

    Args:
L
liu zhengxi 已提交
11205 11206 11207 11208 11209
        x (Variable): The first input Tensor/LoDTensor of mul_op.
        y (Variable): The second input Tensor/LoDTensor of mul_op.
        x_num_col_dims (int, optional): The mul_op can take tensors with more than two dimensions as its inputs. If the input $x$ is a tensor with more than two dimensions, $x$ will be flattened into a two-dimensional matrix first. The flattening rule is: the first `num_col_dims` will be flattened to form the first dimension of the final matrix (the height of the matrix), and the rest `rank(x) - num_col_dims` dimensions are flattened to form the second dimension of the final matrix (the width of the matrix). As a result, height of the flattened matrix is equal to the product of $x$'s first `x_num_col_dims` dimensions' sizes, and width of the flattened matrix is equal to the product of $x$'s last `rank(x) - num_col_dims` dimensions' size. For example, suppose $x$ is a 6-dimensional tensor with the shape [2, 3, 4, 5, 6], and `x_num_col_dims` = 3. Thus, the flattened matrix will have a shape [2 x 3 x 4, 5 x 6] = [24, 30]. Default is 1. 
        y_num_col_dims (int, optional): The mul_op can take tensors with more than two dimensions as its inputs. If the input $y$ is a tensor with more than two dimensions, $y$ will be flattened into a two-dimensional matrix first. The attribute `y_num_col_dims` determines how $y$ is flattened. See comments of `x_num_col_dims` for more details. Default is 1. 
        name (str, optional): Name of the output. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`. Default is None. 
X
Xin Pan 已提交
11210 11211

    Returns:
L
liu zhengxi 已提交
11212
        Variable(Tensor/LoDTensor): The output Tensor/LoDTensor of mul op.
11213 11214

    Examples:
L
liu zhengxi 已提交
11215
        ..  code-block:: python
11216 11217 11218 11219 11220 11221 11222 11223 11224
            
            import paddle.fluid as fluid
            dataX = fluid.layers.data(name="dataX", append_batch_size = False, shape=[2, 5], dtype="float32")
            dataY = fluid.layers.data(name="dataY", append_batch_size = False, shape=[5, 3], dtype="float32")
            output = fluid.layers.mul(dataX, dataY,
                                      x_num_col_dims = 1,
                                      y_num_col_dims = 1)
            

X
Xin Pan 已提交
11225 11226 11227
    """

    helper = LayerHelper("mul", **locals())
11228 11229 11230 11231
    check_type_and_dtype(x, 'x', Variable, ['float16', 'float32', 'float64'],
                         'mul')
    check_type_and_dtype(y, 'y', Variable, ['float16', 'float32', 'float64'],
                         'mul')
X
Xin Pan 已提交
11232
    if name is None:
X
Xin Pan 已提交
11233
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
11234 11235 11236 11237 11238 11239 11240 11241 11242
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    helper.append_op(
        type="mul",
        inputs={"X": x,
                "Y": y},
        attrs={
X
fix  
Xin Pan 已提交
11243 11244
            "x_num_col_dims": x_num_col_dims,
            "y_num_col_dims": y_num_col_dims
X
Xin Pan 已提交
11245 11246 11247 11248 11249 11250
        },
        outputs={"Out": out})
    return out


@templatedoc()
11251
def maxout(x, groups, name=None, axis=1):
X
Xin Pan 已提交
11252 11253 11254 11255 11256
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
11257 11258
        groups(int): ${groups_comment}
        axis(int, optional): ${axis_comment}
W
wangguanzhong 已提交
11259 11260 11261
        name(str, optional): For detailed information, please refer 
            to :ref:`api_guide_Name`. Usually name is no need to set and 
            None by default.
X
Xin Pan 已提交
11262 11263

    Returns:
11264
        Variable: ${out_comment}
J
jerrywgz 已提交
11265

11266 11267
    Raises:
        ValueError: If `axis` is not 1, -1 or 3.
11268
        ValueError: If the number of input channels can not be divisible by `groups`.
W
wangguanzhong 已提交
11269

J
jerrywgz 已提交
11270 11271 11272
    Examples:
        .. code-block:: python

11273
            import paddle.fluid as fluid
11274
            input = fluid.data(
J
jerrywgz 已提交
11275
                name='data', 
11276
                shape=[None, 256, 32, 32], 
J
jerrywgz 已提交
11277 11278
                dtype='float32')
            out = fluid.layers.maxout(input, groups=2)
X
Xin Pan 已提交
11279 11280
    """
    helper = LayerHelper("maxout", **locals())
11281 11282 11283 11284 11285 11286
    if axis not in [1, -1, 3]:
        raise ValueError(
            "Attr(axis) should be 1 when data format is NCHW, -1 or 3 when data format is NHWC. Received "
            "Attr(axis): %s." % str(axis))
    if axis == -1:
        axis = 3
X
Xin Pan 已提交
11287 11288

    if name is None:
X
Xin Pan 已提交
11289
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
11290 11291 11292 11293 11294 11295 11296
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    helper.append_op(
        type="maxout",
        inputs={"X": x},
11297 11298
        attrs={"groups": groups,
               "axis": axis},
X
Xin Pan 已提交
11299 11300
        outputs={"Out": out})
    return out
11301 11302


J
JiabinYang 已提交
11303
def space_to_depth(x, blocksize, name=None):
J
JiabinYang 已提交
11304
    """
J
JiabinYang 已提交
11305
    Gives a blocksize to space_to_depth the input LoDtensor with Layout: [batch, channel, height, width]
11306

11307 11308 11309
    This op rearranges blocks of spatial data, into depth. More specifically, this op outputs a copy of \
        theinput LoDtensor where values from the height and width dimensions are moved to the channel \
        dimension.
J
JiabinYang 已提交
11310
    The attr blocksize indicates the input block size.
11311

11312 11313 11314
    space_to_depth will reorgnize the elements of input with shape[batch, channel, height, width] \
        according to blocksize to construct output with shape \
        [batch, channel * blocksize * blocksize, height/blocksize, width/blocksize]:
J
JiabinYang 已提交
11315

J
JiabinYang 已提交
11316 11317 11318 11319 11320
    - Non-overlapping blocks of size block_size x block size are rearranged into depth at each location.
    - The Y, X coordinates within each block of the input become the high order component of the output channel index
    - channel should be divisible by square of blocksize
    - height, width should be divsible by blocksize

11321 11322 11323 11324 11325 11326 11327 11328 11329 11330 11331 11332 11333 11334 11335 11336 11337
    This OP is useful for resizing the activations between convolutions \
        (but keeping all data)

    .. code-block:: text

        Given the input x with the shape [1, 1, 4, 4]:
        x.data = [[[[1,   2,  5,  6],
                    [3,   4,  7,  8],
                    [9,  10, 13, 14],
                    [11, 12, 15, 16]]]]
        blocksize = 2

        then get the output with the shape [1, 4, 2, 2]:
        out.data = [[[[1,   2],  [3,  4]],
                     [[5,   6],  [7,  8]],
                     [[9,  10], [11, 12]],
                     [[13, 14], [15, 16]]]]
J
JiabinYang 已提交
11338

J
JiabinYang 已提交
11339
    Args:
11340 11341 11342 11343 11344 11345
        x (Variable): The input, which should be 4 dims Tensor or LodTensor, with the shape \
            [batch, channel, height, width]
        blocksize (int): The blocksize to select the element on each feature map should be > 2
        name(str, optional): For detailed information, please refer \
            to :ref:`api_guide_Name`. Usually name is no need to set and \
            None by default.
J
JiabinYang 已提交
11346

11347 11348 11349 11350
    Returns: The output, which should be 4 dims Tensor or LodTensor, with the shape \
            [batch, channel * blocksize * blocksize, height/blocksize, width/blocksize]

    Return Type: Variable
J
JiabinYang 已提交
11351 11352

    Raises:
11353
        TypeError: blocksize type must be int64.
J
JiabinYang 已提交
11354 11355 11356

    Examples:
        .. code-block:: python
11357
    
11358 11359
            import paddle.fluid as fluid
            import numpy as np
J
JiabinYang 已提交
11360

11361 11362
            data = fluid.data(
                name='data', shape=[1, 4, 2, 2], dtype='float32')
J
JiabinYang 已提交
11363
            space_to_depthed = fluid.layers.space_to_depth(
J
JiabinYang 已提交
11364
                x=data, blocksize=2)
11365

11366
            exe = fluid.Executor(fluid.CPUPlace())
11367
            data_np = np.arange(0,16).reshape((1,4,2,2)).astype('float32')
11368 11369 11370 11371 11372 11373 11374

            print(data_np)
            #array([[[[ 0.,  1.], [ 2.,  3.]],
            #        [[ 4.,  5.], [ 6.,  7.]],
            #        [[ 8.,  9.], [10., 11.]],
            #        [[12., 13.], [14., 15.]]]], dtype=float32)

11375
            out_main = exe.run(fluid.default_main_program(),
11376 11377 11378 11379 11380 11381 11382 11383
                        feed={'data': data_np},
                        fetch_list=[space_to_depthed])

            print(out_main)
            #[array([[[[ 0.]], [[ 4.]], [[ 1.]], [[ 5.]],
            #         [[ 8.]], [[12.]], [[ 9.]], [[13.]],
            #         [[ 2.]], [[ 6.]], [[ 3.]], [[ 7.]],
            #         [[10.]], [[14.]], [[11.]], [[15.]]]], dtype=float32)]
11384

J
JiabinYang 已提交
11385 11386
    """

J
JiabinYang 已提交
11387
    helper = LayerHelper("space_to_depth", **locals())
J
JiabinYang 已提交
11388

J
JiabinYang 已提交
11389 11390
    if not (isinstance(blocksize, int)):
        raise ValueError("blocksize must be a python Int")
J
JiabinYang 已提交
11391 11392

    if name is None:
J
JiabinYang 已提交
11393 11394
        out = helper.create_variable_for_type_inference(
            dtype=x.dtype)  #fix create
J
JiabinYang 已提交
11395 11396 11397 11398 11399
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    helper.append_op(
J
JiabinYang 已提交
11400
        type="space_to_depth",
J
JiabinYang 已提交
11401
        inputs={"X": x},
J
JiabinYang 已提交
11402
        attrs={"blocksize": blocksize},
J
JiabinYang 已提交
11403
        outputs={"Out": out})
J
JiabinYang 已提交
11404 11405
    return out

J
JiabinYang 已提交
11406

11407 11408 11409 11410 11411 11412
def affine_channel(x,
                   scale=None,
                   bias=None,
                   data_layout='NCHW',
                   name=None,
                   act=None):
11413 11414 11415 11416 11417
    """
    Applies a separate affine transformation to each channel of the input.
    Useful for replacing spatial batch norm with its equivalent fixed
    transformation. The input also can be 2D tensor and applies a affine
    transformation in second dimension.
11418

11419 11420 11421
    Args:
        x (Variable): Feature map input can be a 4D tensor with order NCHW
            or NHWC. It also can be a 2D tensor and the affine transformation
L
LielinJiang 已提交
11422
            is applied in the second dimension.The data type is float32 or float64.
11423 11424
        scale (Variable): 1D input of shape (C), the c-th element is the scale
            factor of the affine transformation for the c-th channel of
L
LielinJiang 已提交
11425
            the input.The data type is float32 or float64.
11426 11427
        bias (Variable): 1D input of shape (C), the c-th element is the bias
            of the affine transformation for the c-th channel of the input.
L
LielinJiang 已提交
11428
            The data type is float32 or float64.
11429 11430 11431 11432 11433
        data_layout (str, optional): Specify the data format of the input, and the data format of the output 
            will be consistent with that of the input. An optional string from: `"NCHW"`, `"NHWC"`.
            The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`. If input is 2D Tensor, you can ignore 
            data_layout.
L
LielinJiang 已提交
11434 11435
        name (str, default None): The name of this layer. For more information,
            please refer to :ref:`api_guide_Name` .
11436
        act (str, default None): Activation to be applied to the output of this layer.
11437 11438

    Returns:
L
LielinJiang 已提交
11439
        Variable: A tensor which has the same shape, data layout and data type with x.
B
Bai Yifan 已提交
11440 11441 11442

    Examples:
        .. code-block:: python
L
LielinJiang 已提交
11443 11444

            import numpy as np
B
Bai Yifan 已提交
11445
            import paddle.fluid as fluid
L
LielinJiang 已提交
11446 11447 11448 11449 11450 11451 11452 11453 11454 11455

            use_gpu = False
            place = fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace()
            exe = fluid.Executor(place)

            data = fluid.data(name='data', shape=[None, 1, 2, 2], dtype='float32')
            input_scale = fluid.layers.create_parameter(shape=[1], dtype="float32",
                                    default_initializer=fluid.initializer.Constant(2.0))
            input_bias = fluid.layers.create_parameter(shape=[1],dtype="float32",
                                    default_initializer=fluid.initializer.Constant(0.5))
B
Bai Yifan 已提交
11456
            out = fluid.layers.affine_channel(data,scale=input_scale,
L
LielinJiang 已提交
11457 11458 11459 11460 11461 11462 11463 11464 11465 11466
                                    bias=input_bias)

            exe.run(fluid.default_startup_program())
            test_program = fluid.default_main_program().clone(for_test=True)

            [out_array] = exe.run(test_program,
                                  fetch_list=out,
                                  feed={'data': np.ones([1,1,2,2]).astype('float32')})
            # out_array is [[[[2.5, 2.5],
            #                [2.5, 2.5]]]] with shape: [1, 1, 2, 2]
B
Bai Yifan 已提交
11467

11468 11469 11470 11471
    """
    helper = LayerHelper("affine_channel", **locals())

    if name is None:
X
Xin Pan 已提交
11472
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
11473 11474 11475 11476 11477 11478 11479 11480 11481 11482 11483
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    helper.append_op(
        type="affine_channel",
        inputs={"X": x,
                'Scale': scale,
                'Bias': bias},
        attrs={"data_layout": data_layout},
        outputs={"Out": out})
11484
    return helper.append_activation(out)
11485 11486


B
barrierye 已提交
11487
def similarity_focus(input, axis, indexes, name=None):
11488
    """
B
barrierye 已提交
11489
    SimilarityFocus Operator
B
barrierye 已提交
11490 11491

    Generate a similarity focus mask with the same shape of input using the following method:
M
minqiyang 已提交
11492

11493 11494 11495
    1. Extract the 3-D tensor(here the first dimension is BatchSize) corresponding
       to the axis according to the indexes. For example, if axis=1 and indexes=[a],
       it will get the matrix T=X[:, a, :, :]. In this case, if the shape of input X
B
barrierye 已提交
11496
       is (BatchSize, A, B, C), the shape of tensor T is (BatchSize, B, C).
11497 11498 11499 11500 11501 11502 11503
    2. For each index, find the largest numbers in the tensor T, so that the same
       row and same column has at most one number(what it means is that if the
       largest number has been found in the i-th row and the j-th column, then
       the numbers in the i-th row or j-th column will be skipped. And then the
       next largest number will be selected from the remaining numbers. Obviously
       there will be min(B, C) numbers), and mark the corresponding position of the
       3-D similarity focus mask as 1, otherwise as 0. Do elementwise-or for
B
barrierye 已提交
11504
       each index.
B
barrierye 已提交
11505 11506 11507 11508
    3. Broadcast the 3-D similarity focus mask to the same shape of input X.

    Refer to `Similarity Focus Layer <http://www.aclweb.org/anthology/N16-1108>`_

B
barrierye 已提交
11509 11510 11511 11512 11513 11514 11515 11516 11517 11518 11519 11520 11521 11522 11523 11524 11525 11526 11527 11528 11529 11530 11531 11532 11533 11534 11535 11536 11537 11538 11539 11540 11541 11542 11543 11544 11545 11546 11547 11548 11549 11550 11551 11552 11553 11554 11555 11556 11557
    .. code-block:: text

        * Example :

            Given a 4-D tensor x with the shape (BatchSize, C, A, B), where C is
            the number of channels and the shape of feature map is (A, B):
                x.shape = (2, 3, 2, 2)
                x.data = [[[[0.8, 0.1],
                            [0.4, 0.5]],

                           [[0.9, 0.7],
                            [0.9, 0.9]],

                           [[0.8, 0.9],
                            [0.1, 0.2]]],


                          [[[0.2, 0.5],
                            [0.3, 0.4]],

                           [[0.9, 0.7],
                            [0.8, 0.4]],

                           [[0.0, 0.2],
                            [0.4, 0.7]]]]

            Given axis: 1 (the axis of the channel)
            Given indexes: [0]

            then we get a 4-D tensor out with the same shape of input x:
                out.shape = (2, 3, 2, 2)
                out.data = [[[[1.0, 0.0],
                              [0.0, 1.0]],

                             [[1.0, 0.0],
                              [0.0, 1.0]],

                             [[1.0, 0.0],
                              [0.0, 1.0]]],

                            [[[0.0, 1.0],
                              [1.0, 0.0]],

                             [[0.0, 1.0],
                              [1.0, 0.0]],

                             [[0.0, 1.0],
                              [1.0, 0.0]]]]

B
barrierye 已提交
11558
    Args:
11559
        input(Variable): The input tensor variable(default float). It should
Y
Yibing Liu 已提交
11560 11561
            be a 4-D tensor with shape [BatchSize, A, B, C]. Data type is 
            float32 or float64.
B
barrierye 已提交
11562
        axis(int): Indicating the dimension to be selected. It can only be
B
barrierye 已提交
11563
            1, 2 or 3.
B
barrierye 已提交
11564
        indexes(list): Indicating the indexes of the selected dimension.
B
barrierye 已提交
11565 11566

    Returns:
H
haowang101779990 已提交
11567 11568
        Variable: A tensor variable with the same shape and same type \
                  as the input.
11569

B
barrierye 已提交
11570 11571
    Examples:
        .. code-block:: python
H
haowang101779990 已提交
11572

11573
            import paddle.fluid as fluid
Y
Yibing Liu 已提交
11574
            data = fluid.data(
Y
Yibing Liu 已提交
11575 11576
                name='data', shape=[-1, 3, 2, 2], dtype='float32')
            fluid.layers.similarity_focus(input=data, axis=1, indexes=[0])
B
barrierye 已提交
11577 11578 11579 11580 11581 11582 11583 11584 11585 11586 11587 11588
    """
    helper = LayerHelper('similarity_focus', **locals())
    # check attrs
    if isinstance(axis, int) is False:
        raise TypeError("axis must be int type.")
    if isinstance(indexes, list) is False:
        raise TypeError("indexes must be list type.")
    if axis != 1 and axis != 2 and axis != 3:
        raise ValueError("axis must be 1, 2 or 3.")
    if len(indexes) == 0:
        raise ValueError("indexes can not be empty.")

B
barrierye 已提交
11589 11590 11591 11592 11593
    if name is None:
        out = helper.create_variable_for_type_inference(dtype=input.dtype)
    else:
        out = helper.create_variable(
            name=name, dtype=input.dtype, persistable=False)
B
barrierye 已提交
11594 11595 11596 11597 11598 11599 11600
    helper.append_op(
        type='similarity_focus',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={"axis": axis,
               "indexes": indexes})
    return out
B
barrierye 已提交
11601 11602


M
minqiyang 已提交
11603 11604
def hash(input, hash_size, num_hash=1, name=None):
    """
Z
zhupengyang 已提交
11605
    This OP hash the input to an integer less than the hash_size.
M
minqiyang 已提交
11606 11607
    The hash algorithm we used was xxHash - Extremely fast hash algorithm
    (https://github.com/Cyan4973/xxHash/tree/v0.6.5)
M
minqiyang 已提交
11608 11609

    Args:
Z
zhupengyang 已提交
11610 11611 11612 11613 11614 11615
        input(Variable): A **Two-Dimensional** LoDTensor with type int32, int64.
             **Only support LoDTensor**.
        num_hash(int, optional): The times of hash, default is 1.
        name(str, optional): The default value is None. Normally there is no
            need for user to set this property. For more information, please
            refer to :ref:`api_guide_Name`.
M
minqiyang 已提交
11616 11617

    Returns:
Z
zhupengyang 已提交
11618
       Variable: A LoDTensor with the same data type as input.
M
minqiyang 已提交
11619 11620

    Examples:
Z
zhupengyang 已提交
11621
        .. code-block:: python
H
haowang101779990 已提交
11622

11623
            import paddle.fluid as fluid
Z
zhupengyang 已提交
11624
            import numpy as np
11625

Z
zhupengyang 已提交
11626
            place = fluid.core.CPUPlace()
11627

Z
zhupengyang 已提交
11628 11629
            x = fluid.data(name="x", shape=[1], dtype="int32", lod_level=1)
            res = fluid.layers.hash(name="res",input=x, hash_size=1000, num_hash=4)
11630

Z
zhupengyang 已提交
11631 11632 11633 11634 11635 11636 11637 11638 11639 11640 11641 11642 11643 11644 11645 11646 11647
            exe = fluid.Executor(place)
            exe.run(fluid.default_startup_program())
            in1 = np.array([[1,2],[3,4]]).astype("int32")
            print(in1)
            x_i = fluid.core.LoDTensor()
            x_i.set(in1,place)
            x_i.set_recursive_sequence_lengths([[0,2]])
            res = exe.run(fluid.default_main_program(), feed={'x':x_i}, fetch_list=[res], return_numpy=False)
            print(np.array(res[0]))
            # [[[722]
            #   [407]
            #   [337]
            #   [395]]
            #  [[603]
            #   [590]
            #   [386]
            #   [901]]]
M
minqiyang 已提交
11648 11649
    """
    helper = LayerHelper('hash', **locals())
M
minqiyang 已提交
11650 11651
    out = helper.create_variable_for_type_inference(
        helper.input_dtype(), stop_gradient=True)
M
minqiyang 已提交
11652 11653 11654 11655 11656 11657 11658
    helper.append_op(
        type='hash',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={'num_hash': num_hash,
               'mod_by': hash_size})
    return out
G
gmcather 已提交
11659 11660


D
dengkaipeng 已提交
11661
@templatedoc()
11662 11663
def grid_sampler(x, grid, name=None):
    """
11664
    This operation samples input X by using bilinear interpolation based on
H
haowang101779990 已提交
11665
    flow field grid, which is usually gennerated by :code:`affine_grid` . The grid of
K
Kaipeng Deng 已提交
11666 11667 11668
    shape [N, H, W, 2] is the concatenation of (x, y) coordinates
    with shape [N, H, W] each, where x is indexing the 4th dimension
    (in width dimension) of input data x and y is indexng the 3rd
11669
    dimention (in height dimension), finally results is the bilinear
K
Kaipeng Deng 已提交
11670 11671
    interpolation value of 4 nearest corner points. The output tensor 
    shape will be [N, C, H, W].
11672

H
haowang101779990 已提交
11673
    .. code-block:: text
11674

H
haowang101779990 已提交
11675 11676
        Step 1:
        Get (x, y) grid coordinates and scale to [0, H-1/W-1].
11677

K
Kaipeng Deng 已提交
11678 11679 11680 11681
        .. code-block:: text

            grid_x = 0.5 * (grid[:, :, :, 0] + 1) * (W - 1)
            grid_y = 0.5 * (grid[:, :, :, 1] + 1) * (H - 1)
11682

H
haowang101779990 已提交
11683 11684 11685
        Step 2:
        Indices input data X with grid (x, y) in each [H, W] area, and bilinear
        interpolate point value by 4 nearest points.
11686

H
haowang101779990 已提交
11687 11688 11689 11690 11691 11692 11693 11694 11695
          wn ------- y_n ------- en
          |           |           |
          |          d_n          |
          |           |           |
         x_w --d_w-- grid--d_e-- x_e
          |           |           |
          |          d_s          |
          |           |           |
          ws ------- y_s ------- wn
11696

H
haowang101779990 已提交
11697 11698 11699 11700
        x_w = floor(x)              // west side x coord
        x_e = x_w + 1               // east side x coord
        y_n = floor(y)              // north side y coord
        y_s = y_s + 1               // south side y coord
11701

H
haowang101779990 已提交
11702 11703 11704 11705
        d_w = grid_x - x_w          // distance to west side
        d_e = x_e - grid_x          // distance to east side
        d_n = grid_y - y_n          // distance to north side
        d_s = y_s - grid_y          // distance to south side
11706

H
haowang101779990 已提交
11707 11708 11709 11710
        wn = X[:, :, y_n, x_w]      // north-west point value
        en = X[:, :, y_n, x_e]      // north-east point value
        ws = X[:, :, y_s, x_w]      // south-east point value
        es = X[:, :, y_s, x_w]      // north-east point value
11711

H
haowang101779990 已提交
11712 11713
        output = wn * d_e * d_s + en * d_w * d_s
               + ws * d_e * d_n + es * d_w * d_n
D
dengkaipeng 已提交
11714 11715

    Args:
K
Kaipeng Deng 已提交
11716 11717 11718 11719 11720 11721 11722 11723 11724
        x(Variable): The input tensor, which is a 4-D tensor with shape
                     [N, C, H, W], N is the batch size, C is the channel
                     number, H and W is the feature height and width.
                     The data type is float32 or float64.
        grid(Variable): Input grid tensor of shape [N, H, W, 2]. The
                        data type is float32 or float64.
        name(str, optional): For detailed information, please refer
                             to :ref:`api_guide_Name`. Usually name is no need to set and
                             None by default.
D
dengkaipeng 已提交
11725 11726

    Returns:
H
haowang101779990 已提交
11727
        Variable: Output of shape [N, C, H, W] data samples input X
K
Kaipeng Deng 已提交
11728 11729
                  using bilnear interpolation based on input grid.
                  The data type is same as input tensor.
11730

H
haowang101779990 已提交
11731 11732 11733 11734
    Examples:

        .. code-block:: python

K
Kaipeng Deng 已提交
11735 11736
            import paddle.fluid as fluid

K
Kaipeng Deng 已提交
11737 11738
            # use with affine_grid
            x = fluid.data(name='x', shape=[None, 10, 32, 32], dtype='float32')
K
Kaipeng Deng 已提交
11739 11740
            theta = fluid.layers.data(name='theta', shape=[2, 3], dtype='float32')
            grid = fluid.layers.affine_grid(theta=theta, out_shape=[3, 10, 32, 32])
H
haowang101779990 已提交
11741
            out = fluid.layers.grid_sampler(x=x, grid=grid)
11742

D
dengkaipeng 已提交
11743 11744 11745 11746 11747 11748 11749 11750 11751
    """
    helper = LayerHelper("grid_sampler", **locals())

    if not isinstance(x, Variable):
        return ValueError("The x should be a Variable")

    if not isinstance(grid, Variable):
        return ValueError("The grid should be a Variable")

11752
    out = helper.create_variable_for_type_inference(x.dtype)
D
dengkaipeng 已提交
11753 11754
    ipts = {'X': x, 'Grid': grid}

11755
    helper.append_op(type='grid_sampler', inputs=ipts, outputs={'Output': out})
11756 11757 11758
    return out


G
gmcather 已提交
11759 11760 11761 11762 11763 11764 11765 11766 11767 11768 11769 11770 11771
def log_loss(input, label, epsilon=1e-4, name=None):
    """
    **Negative Log Loss Layer**

    This layer accepts input predictions and target label and returns the
    negative log loss.

    .. math::

        Out = -label * \\log{(input + \\epsilon)}
              - (1 - label) * \\log{(1 - input + \\epsilon)}

    Args:
Y
Yibing Liu 已提交
11772
        input (Variable|list):  A 2-D tensor with shape [N x 1], where N is the
G
gmcather 已提交
11773
                                batch size. This input is a probability computed
Y
Yibing Liu 已提交
11774 11775 11776 11777 11778 11779 11780
                                by the previous operator. Data type float32.
        label (Variable|list):  The ground truth which is a 2-D tensor with
                                shape [N x 1], where N is the batch size. 
                                Data type float32.
        epsilon (float, optional): A small number for numerical stability. Default 1e-4.
        name(str|None): For detailed information, please refer to 
            :ref:`api_guide_Name` . Usually name is no need to set and None by default.
G
gmcather 已提交
11781 11782 11783 11784 11785 11786 11787

    Returns:
        Variable: A 2-D tensor with shape [N x 1], the negative log loss.

    Examples:
        .. code-block:: python

11788
          import paddle.fluid as fluid
Y
Yibing Liu 已提交
11789 11790
          label = fluid.data(name='label', shape=[-1, 1], dtype='int64')
          prob = fluid.data(name='prob', shape=[-1, 10], dtype='float32')
G
gmcather 已提交
11791 11792 11793 11794 11795 11796 11797 11798 11799 11800 11801 11802 11803 11804 11805 11806 11807 11808 11809 11810 11811
          cost = fluid.layers.log_loss(input=prob, label=label)
    """
    helper = LayerHelper('log_loss', **locals())

    if name is None:
        loss = helper.create_variable_for_type_inference(dtype=input.dtype)
    else:
        loss = helper.create_variable(
            name=name, dtype=input.dtype, persistable=False)

    helper.append_op(
        type='log_loss',
        inputs={'Predicted': [input],
                'Labels': [label]},
        outputs={'Loss': [loss]},
        attrs={'epsilon': epsilon})
    return loss


def add_position_encoding(input, alpha, beta, name=None):
    """
G
Guo Sheng 已提交
11812 11813
    This operator performs weighted sum of input feature at each position
    (position in the sequence) and the corresponding position encoding.
G
gmcather 已提交
11814

G
Guo Sheng 已提交
11815 11816
    For more details of position encoding, please refer to `Attention Is All You 
    Need <http://arxiv.org/pdf/1706.03762.pdf>`_ .
G
gmcather 已提交
11817

G
Guo Sheng 已提交
11818
    The formula is as follows:
G
gmcather 已提交
11819 11820

    .. math::
H
haowang101779990 已提交
11821 11822 11823
        PE(pos, 2i) &= \\sin{(pos / 10000^{2i / P})}   \\\\
        PE(pos, 2i + 1) &= \\cos{(pos / 10000^{2i / P})}  \\\\
        Out(:, pos, i) &= \\alpha * input(:, pos, i) + \\beta * PE(pos, i)
G
gmcather 已提交
11824 11825

    Where:
G
Guo Sheng 已提交
11826 11827 11828 11829 11830 11831 11832 11833 11834 11835 11836 11837 11838 11839 11840 11841 11842
      - :math:`PE(pos, 2i)` : the value at even index `2i` for encoding of position `pos`.
      - :math:`PE(pos, 2i + 1)` : the value at odd index `2i+1` for encoding of position `pos`

    Args:
        input(Variable): A Tensor or LoDTensor (lod level is 1). If it is a
            Tensor, the shape should be `[N, M, P]`, where `N` stands for
            batch size, `M` for sequence length, `P` for the size of feature
            dimension. If it is a LoDTensor, the shape should be `[N, P]`,
            where `N` stands for the total sequence lengths in this mini-batch,
            `P` for the size of feature. The data type should be float32 or float64.
        alpha(float): Indicate the weight coefficient for `input` when performing
            weighted sum.
        beta(float): Indicate the weight coefficient for position encoding when
            performing weighted sum.
        name(str, optional): For detailed information, please refer 
            to :ref:`api_guide_Name`. Usually name is no need to set and 
            None by default.
G
gmcather 已提交
11843 11844

    Returns:
G
Guo Sheng 已提交
11845
        Variable: A Tensor or LoDTensor. It has the same shape, data type and lod as `input`.
G
gmcather 已提交
11846 11847 11848 11849

    Examples:
        .. code-block:: python

11850 11851
          import paddle.fluid as fluid

G
Guo Sheng 已提交
11852
          tensor = fluid.data(
11853
              name='tensor',
G
Guo Sheng 已提交
11854 11855
              shape=[None, 64, 512],
              dtype='float32')
11856 11857
          position_tensor = fluid.layers.add_position_encoding(
              input=tensor, alpha=1.0, beta=1.0)
H
haowang101779990 已提交
11858

G
gmcather 已提交
11859 11860 11861 11862 11863 11864 11865 11866 11867 11868 11869 11870 11871 11872 11873 11874
    """
    helper = LayerHelper('add_position_encoding', **locals())
    dtype = helper.input_dtype()

    if name is None:
        out = helper.create_variable_for_type_inference(dtype=dtype)
    else:
        out = helper.create_variable(name=name, dtype=dtype, persistable=False)

    helper.append_op(
        type="add_position_encoding",
        inputs={"X": input},
        outputs={"Out": out},
        attrs={"alpha": alpha,
               "beta": beta})
    return out
Q
Qiao Longfei 已提交
11875 11876 11877 11878 11879 11880 11881 11882 11883 11884


def bilinear_tensor_product(x,
                            y,
                            size,
                            act=None,
                            name=None,
                            param_attr=None,
                            bias_attr=None):
    """
Y
Yibing Liu 已提交
11885
    **Bilinear Tensor Product Layer**
Q
Qiao Longfei 已提交
11886

Q
Qiao Longfei 已提交
11887
    This layer performs bilinear tensor product on two inputs.
Q
Qiao Longfei 已提交
11888 11889 11890
    For example:

    .. math::
H
haowang101779990 已提交
11891
       out_{i} = x * W_{i} * {y^\mathrm{T}}, i=0,1,...,size-1
Q
Qiao Longfei 已提交
11892

Q
Qiao Longfei 已提交
11893
    In this formula:
11894 11895
      - :math:`x`: the first input contains M elements, shape is [batch_size, M].
      - :math:`y`: the second input contains N elements, shape is [batch_size, N].
Y
Yibing Liu 已提交
11896
      - :math:`W_{i}`: the i-th learned weight, shape is [M, N].
H
haowang101779990 已提交
11897
      - :math:`out_{i}`: the i-th element of out, shape is [batch_size, size].
Q
Qiao Longfei 已提交
11898 11899 11900
      - :math:`y^\mathrm{T}`: the transpose of :math:`y_{2}`.

    Args:
Y
Yibing Liu 已提交
11901 11902 11903 11904
        x (Variable): 2-D input tensor with shape [batch_size, M]. Data type 
            is float32 or float64.
        y (Variable): 2-D input tensor with shape [batch_size, N]. Data type 
            should be same as **x**.
Q
Qiao Longfei 已提交
11905
        size (int): The dimension of this layer.
Y
Yibing Liu 已提交
11906 11907 11908 11909 11910 11911 11912 11913 11914
        act (str|None): Activation to be applied to the output of this layer. Default None.
        name(str|None): For detailed information, please refer to 
            :ref:`api_guide_Name` . Usually name is no need to set and None by default.
        param_attr (ParamAttr|None): To specify the weight parameter attribute. 
            Default: None, which means the default weight parameter property is 
            used. See usage for details in :ref:`api_fluid_ParamAttr` .
        bias_attr (ParamAttr|None): To specify the bias parameter attribute. 
            Default: None, which means the default bias parameter property is 
            used. See usage for details in :ref:`api_fluid_ParamAttr` .
Q
Qiao Longfei 已提交
11915
    Returns:
Y
Yibing Liu 已提交
11916
        Variable: A 2-D Tensor of shape [batch_size, size]. Data type is the same as input **x**.
Q
Qiao Longfei 已提交
11917 11918 11919 11920

    Examples:
        .. code-block:: python

11921
          import paddle.fluid as fluid
Y
Yibing Liu 已提交
11922 11923
          layer1 = fluid.data("t1", shape=[-1, 5], dtype="float32")
          layer2 = fluid.data("t2", shape=[-1, 4], dtype="float32")
Y
Yibing Liu 已提交
11924
          tensor = fluid.layers.bilinear_tensor_product(x=layer1, y=layer2, size=1000)
Q
Qiao Longfei 已提交
11925 11926
    """
    helper = LayerHelper('bilinear_tensor_product', **locals())
Q
Qiao Longfei 已提交
11927
    dtype = helper.input_dtype('x')
Q
Qiao Longfei 已提交
11928 11929 11930 11931

    param_shape = [size, x.shape[1], y.shape[1]]

    w = helper.create_parameter(
Q
Qiao Longfei 已提交
11932
        attr=helper.param_attr, shape=param_shape, dtype=dtype, is_bias=False)
Q
Qiao Longfei 已提交
11933 11934 11935 11936 11937 11938 11939 11940 11941 11942 11943 11944 11945 11946 11947 11948 11949

    if name is None:
        out = helper.create_variable_for_type_inference(dtype=dtype)
    else:
        out = helper.create_variable(name=name, dtype=dtype, persistable=False)

    inputs = {"X": x, "Y": y, "Weight": w}
    if helper.bias_attr:
        bias_size = [1, size]
        bias = helper.create_parameter(
            attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
        inputs["Bias"] = bias
    helper.append_op(
        type="bilinear_tensor_product", inputs=inputs, outputs={"Out": out})

    # add activation
    return helper.append_activation(out)
C
chengduo 已提交
11950 11951 11952 11953 11954


@templatedoc()
def get_tensor_from_selected_rows(x, name=None):
    """
11955 11956 11957 11958 11959 11960 11961 11962 11963 11964 11965 11966 11967 11968 11969 11970
    This operator gets tensor data from input with SelectedRows type, and outputs a LoDTensor.

    .. code-block:: text

        input x is SelectedRows:
           x.rows = [0, 5, 5, 4, 19]
           x.height = 20
           x.value = [[1, 1] [2, 2] [2, 2] [3, 3] [6, 6]]

        Ouput is LoDTensor:
           out.shape = [5, 2]
           out.data = [[1, 1],
                       [2, 2],
                       [2, 2],
                       [3, 3],
                       [6, 6]]
C
chengduo 已提交
11971 11972

    Args:
11973 11974 11975
        x(SelectedRows): Input with SelectedRows type. The data type is float32, float64, int32 or int64.
        name(str, optional): The default value is None.  Normally there is no need for user to set this property.
            For more information, please refer to :ref:`api_guide_Name` .
C
chengduo 已提交
11976 11977

    Returns:
11978
        Variable: LoDTensor transformed from SelectedRows. The data type is same with input.
B
bdzhuxiaoning 已提交
11979 11980 11981 11982 11983 11984 11985 11986

    Examples:
        .. code-block:: python
	    
            import paddle.fluid as fluid
            b = fluid.default_main_program().global_block()
            input = b.create_var(name="X", dtype="float32", persistable=True, type=fluid.core.VarDesc.VarType.SELECTED_ROWS)
            out = fluid.layers.get_tensor_from_selected_rows(input)
C
chengduo 已提交
11987 11988 11989 11990 11991 11992 11993 11994 11995 11996
    """

    helper = LayerHelper('get_tensor_from_selected_rows', **locals())
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
        type='get_tensor_from_selected_rows',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={})
    return out
11997 11998


S
shippingwang 已提交
11999
def shuffle_channel(x, group, name=None):
S
shippingwang 已提交
12000
    """
S
shippingwang 已提交
12001 12002 12003 12004 12005 12006
    This operator shuffles the channels of input x.
    It divide the input channels in each group into :attr:`group` subgroups,
    and obtain a new order by selecting element from every subgroup one by one.

    Please refer to the paper
    https://arxiv.org/pdf/1707.01083.pdf
S
shippingwang 已提交
12007
    
S
shippingwang 已提交
12008
    .. code-block:: text
12009

S
shippingwang 已提交
12010 12011 12012 12013 12014 12015 12016 12017 12018 12019 12020 12021 12022 12023 12024 12025 12026 12027 12028 12029 12030 12031 12032 12033 12034 12035 12036 12037
        Given a 4-D tensor input with the shape (N, C, H, W):
            input.shape = (1, 4, 2, 2)
            input.data =[[[[0.1, 0.2],
                           [0.2, 0.3]],

                          [[0.3, 0.4],
                           [0.4, 0.5]],

                          [[0.5, 0.6],
                           [0.6, 0.7]],

                          [[0.7, 0.8],
                           [0.8, 0.9]]]]
            Given group: 2
            then we get a 4-D tensor out whth the same shape of input:
            out.shape = (1, 4, 2, 2)
            out.data = [[[[0.1, 0.2],
                          [0.2, 0.3]],
                          
                         [[0.5, 0.6],
                          [0.6, 0.7]],
                          
                         [[0.3, 0.4],
                          [0.4, 0.5]],
                          
                         [[0.7, 0.8],
                          [0.8, 0.9]]]]
                        
S
shippingwang 已提交
12038
    Args: 
S
shippingwang 已提交
12039 12040
        x(Variable): The input tensor variable. It should be a 4-D tensor with shape [N, C, H, W]
        group(int): Indicating the conuts of subgroups, It should divide the number of channels.
S
shippingwang 已提交
12041 12042

    Returns:
S
shippingwang 已提交
12043 12044
        out(Variable): the channels shuffling result is a tensor variable with the 
        same shape and same type as the input.
S
shippingwang 已提交
12045 12046

    Raises:
S
shippingwang 已提交
12047
        ValueError: If group is not an int type variable.
S
shippingwang 已提交
12048 12049 12050

    Examples:
        .. code-block:: python
12051

12052
            import paddle.fluid as fluid
R
ruri 已提交
12053
            input = fluid.data(name='input', shape=[None,4,2,2], dtype='float32')
S
shippingwang 已提交
12054
            out = fluid.layers.shuffle_channel(x=input, group=2)
S
shippingwang 已提交
12055 12056 12057
    """
    helper = LayerHelper("shuffle_channel", **locals())

S
shippingwang 已提交
12058
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
S
shippingwang 已提交
12059 12060 12061 12062 12063 12064 12065 12066 12067

    if not isinstance(group, int):
        raise TypeError("group must be int type")

    helper.append_op(
        type="shuffle_channel",
        inputs={"X": x},
        outputs={"Out": out},
        attrs={"group": group})
S
shippingwang 已提交
12068
    return out
S
Add  
shippingwang 已提交
12069 12070


12071
@templatedoc()
D
dengkaipeng 已提交
12072
def temporal_shift(x, seg_num, shift_ratio=0.25, name=None):
12073 12074 12075 12076 12077 12078 12079 12080
    """
    **Temporal Shift Operator**
    
    ${comment}
                        
    Args: 
        x(Variable): ${x_comment}
        seg_num(int): ${seg_num_comment}
D
dengkaipeng 已提交
12081
        shift_ratio(float): ${shift_ratio_comment}
K
Kaipeng Deng 已提交
12082 12083 12084
        name(str, optional): For detailed information, please refer
                             to :ref:`api_guide_Name`. Usually name is no need to set and
                             None by default.
12085 12086 12087

    Returns:
        out(Variable): The temporal shifting result is a tensor variable with the 
K
Kaipeng Deng 已提交
12088
        same shape and same data type as the input.
12089 12090 12091 12092 12093 12094 12095

    Raises:
        TypeError: seg_num must be int type.

    Examples:
        .. code-block:: python

12096
            import paddle.fluid as fluid
K
Kaipeng Deng 已提交
12097
            input = fluid.data(name='input', shape=[None,4,2,2], dtype='float32')
D
dengkaipeng 已提交
12098
            out = fluid.layers.temporal_shift(x=input, seg_num=2, shift_ratio=0.2)
12099 12100 12101 12102 12103 12104 12105 12106 12107 12108 12109 12110
    """
    helper = LayerHelper("temporal_shift", **locals())

    out = helper.create_variable_for_type_inference(dtype=x.dtype)

    if not isinstance(seg_num, int):
        raise TypeError("seg_num must be int type.")

    helper.append_op(
        type="temporal_shift",
        inputs={"X": x},
        outputs={"Out": out},
D
dengkaipeng 已提交
12111 12112
        attrs={"seg_num": seg_num,
               "shift_ratio": shift_ratio})
12113 12114 12115
    return out


S
sneaxiy 已提交
12116
class PyFuncRegistry(object):
S
sneaxiy 已提交
12117 12118 12119
    _register_funcs = []

    def __init__(self, func):
S
sneaxiy 已提交
12120
        if func is None or not callable(func):
S
sneaxiy 已提交
12121 12122 12123
            raise TypeError('func must be a Python function')

        self._func = func
M
minqiyang 已提交
12124
        # find named args using reflection
S
sneaxiy 已提交
12125 12126 12127 12128 12129 12130 12131
        args = inspect.getargspec(self._func)
        if len(args[0]) == 0 and args[1] is None and args[2] is None:
            # Function with no inputs
            self._named_args = None
        else:
            self._named_args = args[0]
        self._id = core._append_python_callable_object_and_return_id(self)
S
sneaxiy 已提交
12132 12133 12134
        '''
        Why record self here?

M
minqiyang 已提交
12135 12136
        1. For debug usage. Users can call
           :code:`py_func.registered_func(idx)` method
S
sneaxiy 已提交
12137
           to find the registered function corresponding
M
minqiyang 已提交
12138
           to :code:`idx`.
S
sneaxiy 已提交
12139

M
minqiyang 已提交
12140 12141
        2. For increasing reference count of self.
           It seems that to release Python object
S
sneaxiy 已提交
12142
           whose reference count is 1 would cause
M
minqiyang 已提交
12143
           segmentation fault error in C++ side.
S
sneaxiy 已提交
12144 12145
           May be lack of Python GC in C++ side?
        '''
S
sneaxiy 已提交
12146
        PyFuncRegistry._register_funcs.append(self)
S
sneaxiy 已提交
12147 12148 12149 12150 12151 12152 12153 12154 12155 12156 12157 12158 12159 12160

    @classmethod
    def registered_func(cls, idx):
        return cls._register_funcs[idx]._func

    @classmethod
    def registered_func_num(cls):
        return len(cls._register_funcs)

    @property
    def id(self):
        return self._id

    def __call__(self, *args):
S
sneaxiy 已提交
12161 12162 12163 12164 12165 12166 12167 12168 12169
        if self._named_args is None:
            func_ret = self._func()
        else:
            kwargs = dict()
            idx = 0
            for arg in self._named_args:
                kwargs[arg] = args[idx]
                idx += 1
            func_ret = self._func(*args[idx:], **kwargs)
S
sneaxiy 已提交
12170

S
sneaxiy 已提交
12171 12172
        if not isinstance(func_ret, (list, tuple)):
            func_ret = (func_ret, )
S
sneaxiy 已提交
12173 12174

        ret = []
S
sneaxiy 已提交
12175 12176 12177
        for each_ret in func_ret:
            if each_ret is None or isinstance(each_ret, core.LoDTensor):
                ret.append(each_ret)
S
sneaxiy 已提交
12178 12179
                continue

S
sneaxiy 已提交
12180 12181
            if not isinstance(each_ret, np.ndarray):
                each_ret = np.array(each_ret)
S
sneaxiy 已提交
12182

S
sneaxiy 已提交
12183 12184 12185
            tensor = core.LoDTensor()
            tensor.set(each_ret, core.CPUPlace())
            ret.append(tensor)
S
sneaxiy 已提交
12186

S
sneaxiy 已提交
12187
        return tuple(ret)
S
sneaxiy 已提交
12188 12189


S
sneaxiy 已提交
12190 12191 12192
@templatedoc()
def py_func(func, x, out, backward_func=None, skip_vars_in_backward_input=None):
    """
12193 12194 12195 12196 12197 12198 12199 12200 12201 12202 12203 12204 12205 12206 12207 12208 12209 12210 12211 12212 12213 12214 12215 12216 12217 12218 12219 12220 12221 12222 12223 12224 12225 12226 12227 12228 12229 12230 12231 12232 12233 12234 12235
    This API is used to register customized OP to Fluid. The forward  function 
    of the registered OP is ``func`` and the backward function of that is 
    ``backward_func``. Paddle will call ``func`` at forward runtime  and call 
    ``backward_func`` at backward runtime(if ``backward_func`` is not  None). 
    ``x`` is the input of ``func``, whose type must be LoDTensor; ``out`` is 
    the output of ``func``, whose type can be either LoDTensor or NumPy array.

    The input of the backward function ``backward_func`` is ``x``, ``out`` and 
    the gradient of ``out``. If some variables of ``out`` have no gradient, the 
    relevant input variable of ``backward_func`` is None. If some variables of 
    ``x`` do not have a gradient, the user should return None in ``backward_func``.

    The data type and shape of ``out`` should also be set correctly before this 
    API is called, and the data type and shape of the gradient of ``out`` and 
    ``x`` will be inferred automatically.

    This API can also be used to debug the neural network by setting the ``func``
    as a function that only print variables.

    Args:
        func (callable): The forward function of the registered OP. When the network
            is running, the forward output ``out`` will be calculated according to this 
            function and the forward input ``x``.
        x (Variable): The input of the forward function ``func``, its type can be 
            Variable | tuple[Variable] | list[Variale], in which Variable is LoDTensor.
        out (Variable): The output of the forward function ``func``, its type can be
            Variable | tuple[Variable] | list[Variale], in which Variable can be either 
            LoDTensor or NumPy array. Since Paddle cannot automatically infer the shape
            and data type of ``out``, ``out`` must be created in advance.
        backward_func (callable, optional): The backward function of the registered OP. 
            Its default value is None, which means there is no reverse calculation. If 
            it is not None, ``backward_func`` is called to calculate the gradient of 
            ``x`` when the network is at backward runtime.
        skip_vars_in_backward_input (Variable, optional): It's used to limit the input 
            variable list of ``backward_func``, and it can be single Variable, tuple[Variable]
            or list[Variable]. It must belong to either ``x`` or ``out``. The default 
            value is None, which means that no variables need to be removed from ``x`` 
            and ``out``. If it is not None, these variables will not be the input of 
            ``backward_func``. This parameter is only useful when ``backward_func`` is 
            not None.
    
    Returns: 
        Variable: The output ``out`` of the forward function ``func``.
S
sneaxiy 已提交
12236 12237

    Examples:
12238
        .. code-block:: python
M
minqiyang 已提交
12239

12240 12241 12242 12243 12244 12245 12246 12247 12248 12249 12250 12251 12252 12253 12254 12255 12256 12257 12258 12259 12260 12261 12262 12263 12264 12265 12266 12267 12268 12269 12270 12271 12272 12273 12274 12275 12276
            import paddle.fluid as fluid
            import six

            def create_tmp_var(name, dtype, shape):
            return fluid.default_main_program().current_block().create_var(
            name=name, dtype=dtype, shape=shape)

            # Tanh activation function provided by Paddle C++ op
            # Here, tanh is used as an example to show how to use py_func
            def tanh(x):
                return np.tanh(x)

            # Skip forward input x
            def tanh_grad(y, dy):
                return np.array(dy) * (1 - np.square(np.array(y)))

            def debug_func(x):
                print(x)

            def simple_net(img, label):
                hidden = img
                for idx in six.moves.range(4):
                    hidden = fluid.layers.fc(hidden, size=200)
                    new_hidden = create_tmp_var(name='hidden_{}'.format(idx),
                        dtype=hidden.dtype, shape=hidden.shape)

                    # User-defined forward and backward 
                    hidden = fluid.layers.py_func(func=tanh, x=hidden,
                        out=new_hidden, backward_func=tanh_grad,
                        skip_vars_in_backward_input=hidden)

                    # User-defined debugging layer, which can print out variable details
                    fluid.layers.py_func(func=debug_func, x=hidden, out=None)

                prediction = fluid.layers.fc(hidden, size=10, act='softmax')
                loss = fluid.layers.cross_entropy(input=prediction, label=label)
                return fluid.layers.mean(loss)
S
sneaxiy 已提交
12277
    """
S
sneaxiy 已提交
12278
    helper = LayerHelper('py_func', **locals())
S
sneaxiy 已提交
12279 12280 12281
    if x is None:
        x = []
    elif isinstance(x, Variable):
S
sneaxiy 已提交
12282
        x = [x]
12283 12284 12285
    elif isinstance(x, tuple):
        x = list(x)
    elif not isinstance(x, (list, tuple, Variable)):
S
sneaxiy 已提交
12286
        raise TypeError('Input must be Variable/list(Variable)/tuple(Variable)')
S
sneaxiy 已提交
12287

S
sneaxiy 已提交
12288 12289 12290
    if out is None:
        out_list = []
    elif isinstance(out, Variable):
S
sneaxiy 已提交
12291
        out_list = [out]
12292 12293 12294
    elif isinstance(out, tuple):
        out_list = list(out)
    elif not isinstance(x, (list, tuple, Variable)):
S
sneaxiy 已提交
12295 12296
        raise TypeError(
            'Output must be Variable/list(Variable)/tuple(Variable)')
S
sneaxiy 已提交
12297

S
sneaxiy 已提交
12298 12299
    fwd_func_id = PyFuncRegistry(func).id
    bwd_func_id = PyFuncRegistry(
S
sneaxiy 已提交
12300
        backward_func).id if backward_func is not None else -1
S
sneaxiy 已提交
12301 12302

    for each_out in out_list:
S
sneaxiy 已提交
12303 12304
        if len(each_out.shape) == 0:
            raise ValueError(
S
sneaxiy 已提交
12305 12306
                'Output shapes of py_func op should be provided by users manually'
            )
S
sneaxiy 已提交
12307

S
sneaxiy 已提交
12308 12309 12310 12311 12312 12313 12314 12315 12316 12317 12318 12319 12320 12321 12322
    backward_skip_vars = set()
    if backward_func is not None and skip_vars_in_backward_input is not None:
        if isinstance(skip_vars_in_backward_input, Variable):
            skip_vars_in_backward_input = [skip_vars_in_backward_input]

        fwd_in_out = [v.name for v in x]
        fwd_in_out.extend([v.name for v in out_list])
        fwd_in_out = set(fwd_in_out)
        backward_skip_vars = set()
        for v in skip_vars_in_backward_input:
            if not v.name in fwd_in_out:
                raise ValueError(
                    'Variable {} is not found in forward inputs and outputs'
                    .format(v.name))
            backward_skip_vars.add(v.name)
S
sneaxiy 已提交
12323 12324 12325 12326

    helper.append_op(
        type='py_func',
        inputs={'X': x},
S
sneaxiy 已提交
12327 12328
        outputs={'Out': out_list},
        attrs={
S
sneaxiy 已提交
12329 12330 12331
            'forward_callable_id': fwd_func_id,
            'backward_callable_id': bwd_func_id,
            'backward_skip_vars': list(backward_skip_vars)
S
sneaxiy 已提交
12332
        })
S
sneaxiy 已提交
12333
    return out
S
sneaxiy 已提交
12334 12335 12336


# For debug usage
S
sneaxiy 已提交
12337 12338 12339 12340
py_func.registered_func = PyFuncRegistry.registered_func
py_func.registered_func_num = PyFuncRegistry.registered_func_num


12341 12342 12343 12344 12345 12346 12347 12348 12349 12350 12351
@templatedoc()
def psroi_pool(input,
               rois,
               output_channels,
               spatial_scale,
               pooled_height,
               pooled_width,
               name=None):
    """
    ${comment}

S
SunGaofeng 已提交
12352
    Parameters:
12353
        input (Variable): ${x_comment}
S
SunGaofeng 已提交
12354
        rois (Variable): LoDTensor, ROIs (Regions of Interest) to pool over.It should be
S
SunGaofeng 已提交
12355 12356 12357
                         a 2-D LoDTensor of shape (num_rois, 4), the lod level
                         is 1. Given as [[x1, y1, x2, y2], ...], (x1, y1) is
                         the top left coordinates, and (x2, y2) is the bottom
S
SunGaofeng 已提交
12358 12359
                         right coordinates. The data type is the same as `input`
        output_channels (int): ${output_channels_comment}
12360
        spatial_scale (float): ${spatial_scale_comment} Default: 1.0
S
SunGaofeng 已提交
12361 12362 12363 12364 12365
        pooled_height (int): ${pooled_height_comment} Default: 1
        pooled_width (int): ${pooled_width_comment} Default: 1
        name(str, optional): The default value is None.  
                             Normally there is no need for user to set this property.  
                             For more information, please refer to :ref:`api_guide_Name`
12366 12367

    Returns:
S
SunGaofeng 已提交
12368 12369 12370 12371
        ${out_comment}.

    Return Type:
        Variable
12372 12373 12374 12375

    Examples:
        .. code-block:: python

S
SunGaofeng 已提交
12376
            import paddle.fluid as fluid
S
SunGaofeng 已提交
12377 12378
            x = fluid.data(name='x', shape=[100, 490, 28, 28], dtype='float32')
            rois = fluid.data(name='rois', shape=[None, 4], lod_level=1, dtype='float32')
S
SunGaofeng 已提交
12379
            pool_out = fluid.layers.psroi_pool(x, rois, 10, 1.0, 7, 7)
12380 12381 12382 12383 12384 12385 12386 12387 12388 12389 12390 12391 12392 12393 12394 12395 12396 12397 12398 12399 12400 12401 12402 12403 12404
    """
    helper = LayerHelper('psroi_pool', **locals())
    # check attrs
    if not isinstance(output_channels, int):
        raise TypeError("output_channels must be int type")
    if not isinstance(spatial_scale, float):
        raise TypeError("spatial_scale must be float type")
    if not isinstance(pooled_height, int):
        raise TypeError("pooled_height must be int type")
    if not isinstance(pooled_width, int):
        raise TypeError("pooled_width must be int type")
    dtype = helper.input_dtype()
    out = helper.create_variable_for_type_inference(dtype)
    helper.append_op(
        type='psroi_pool',
        inputs={'X': input,
                'ROIs': rois},
        outputs={'Out': out},
        attrs={
            'output_channels': output_channels,
            'spatial_scale': spatial_scale,
            'pooled_height': pooled_height,
            'pooled_width': pooled_width
        })
    return out
12405 12406 12407 12408 12409 12410 12411 12412 12413 12414 12415 12416 12417 12418 12419 12420 12421 12422 12423 12424 12425 12426 12427 12428 12429 12430 12431 12432 12433 12434 12435 12436 12437 12438 12439 12440


@templatedoc()
def prroi_pool(input,
               rois,
               spatial_scale=1.0,
               pooled_height=1,
               pooled_width=1,
               name=None):
    """
    The precise roi pooling implementation for paddle?https://arxiv.org/pdf/1807.11590.pdf

    Args:
        input (Variable):The input of Deformable PSROIPooling.The shape of input tensor is
                        [N,C,H,W]. Where N is batch size,C is number of input channels,H
                        is height of the feature, and W is the width of the feature.
        rois (Variable): ROIs (Regions of Interest) to pool over.It should be
                        a 2-D LoDTensor of shape (num_rois, 4), the lod level
                        is 1. Given as [[x1, y1, x2, y2], ...], (x1, y1) is
                        the top left coordinates, and (x2, y2) is the bottom
                        right coordinates.
        spatial_scale (float): Ratio of input feature map height (or width) to raw image height (or width).
                             Equals the reciprocal of total stride in convolutional layers, Default: 1.0.
        pooled_height (integer): The pooled output height. Default: 1.
        pooled_width (integer): The pooled output width. Default: 1.
        name (str, default None): The name of this operation.

    Returns:
        Variable(Tensor): The shape of the returned Tensor is (num_rois, output_channels, pooled_h, pooled_w), with value type float32,float16..

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
            x = fluid.layers.data(name='x', shape=[490, 28, 28], dtype='float32')
            rois = fluid.layers.data(name='rois', shape=[4], lod_level=1, dtype='float32')
12441
            pool_out = fluid.layers.prroi_pool(x, rois, 1.0, 7, 7)
12442 12443 12444 12445 12446 12447 12448 12449 12450 12451 12452 12453 12454 12455 12456 12457 12458 12459 12460 12461 12462 12463
    """
    helper = LayerHelper('prroi_pool', **locals())
    # check attrs
    if not isinstance(spatial_scale, float):
        raise TypeError("spatial_scale must be float type")
    if not isinstance(pooled_height, int):
        raise TypeError("pooled_height must be int type")
    if not isinstance(pooled_width, int):
        raise TypeError("pooled_width must be int type")
    dtype = helper.input_dtype()
    out = helper.create_variable_for_type_inference(dtype)
    helper.append_op(
        type='prroi_pool',
        inputs={'X': input,
                'ROIs': rois},
        outputs={'Out': out},
        attrs={
            'spatial_scale': spatial_scale,
            'pooled_height': pooled_height,
            'pooled_width': pooled_width
        })
    return out
12464

M
minqiyang 已提交
12465

R
ruri 已提交
12466 12467 12468
def pixel_shuffle(x, upscale_factor):
    """

R
ruri 已提交
12469
    This op rearranges elements in a tensor of shape [N, C, H, W]
R
ruri 已提交
12470 12471 12472 12473 12474 12475 12476
    to a tensor of shape [N, C/r**2, H*r, W*r].
    This is useful for implementing efficient sub-pixel convolution
    with a stride of 1/r.
    Please refer to the paper: `Real-Time Single Image and Video Super-Resolution 
    Using an Efficient Sub-Pixel Convolutional Neural Network <https://arxiv.org/abs/1609.05158v2>`_ .
    by Shi et. al (2016) for more details.

R
ruri 已提交
12477
    Parameters:
R
ruri 已提交
12478

R
ruri 已提交
12479 12480
        x(Variable): 4-D tensor, the data type should be float32 or float64.
        upscale_factor(int): factor to increase spatial resolution.
R
ruri 已提交
12481 12482

    Returns:
12483
        Out(Variable): Reshaped tensor according to the new dimension.
R
ruri 已提交
12484 12485 12486 12487 12488 12489 12490

    Raises:
        ValueError: If the square of upscale_factor cannot divide the channels of input.

    Examples:
        .. code-block:: python

R
ruri 已提交
12491 12492 12493 12494 12495 12496 12497 12498 12499 12500 12501 12502 12503 12504 12505 12506 12507
	    # declarative mode
	    import paddle.fluid as fluid
	    import numpy as np
	    input = fluid.data(name="input", shape=[2,9,4,4])
	    output = fluid.layers.pixel_shuffle(x=input, upscale_factor=3)
	    place = fluid.CPUPlace()
	    exe = fluid.Executor(place)
	    exe.run(fluid.default_startup_program())
 
	    input_data = np.random.rand(2,9,4,4).astype("float32")
	    output_data = exe.run(fluid.default_main_program(),
                feed={"input":input_data},
                fetch_list=[output],
                return_numpy=True)
 
 	    # print(output.shape)
	    # (2L, 1L, 12L, 12L)
R
ruri 已提交
12508 12509 12510 12511 12512 12513 12514 12515 12516 12517 12518 12519 12520 12521 12522 12523 12524 12525

    """

    helper = LayerHelper("pixel_shuffle", **locals())

    out = helper.create_variable_for_type_inference(dtype=x.dtype)

    if not isinstance(upscale_factor, int):
        raise TypeError("upscale factor must be int type")

    helper.append_op(
        type="pixel_shuffle",
        inputs={"X": x},
        outputs={"Out": out},
        attrs={"upscale_factor": upscale_factor})
    return out


12526 12527 12528 12529 12530
def fsp_matrix(x, y):
    """

    **FSP matrix op**

12531
    This op is used to calculate the flow of solution procedure (FSP) matrix of two 4-D Tensor feature maps.
12532 12533 12534 12535 12536 12537 12538 12539 12540 12541 12542
    Given feature map x with shape [x_channel, h, w] and feature map y with shape
    [y_channel, h, w], we can get the fsp matrix of x and y in two steps:

    1. reshape x into matrix with shape [x_channel, h * w] and reshape and
       transpose y into matrix with shape [h * w, y_channel].
    2. multiply x and y to get fsp matrix with shape [x_channel, y_channel].

    The output is a batch of fsp matrices.

    Args:

12543 12544 12545
        x (Variable): A 4-D Tensor feature map with shape [batch_size, x_channel, height, width].
                      A Tensor with type float32, float64.
        y (Variable): A 4-D Tensor feature map with shape [batch_size, y_channel, height, width].
12546
                      The y_channel can be different with the x_channel of Input(X)
12547 12548
                      while the other dimensions must be the same with Input(X)'s. A Tensor with
                      type float32, float64.
12549 12550 12551 12552

    Returns:

        fsp matrix (Variable): The output of FSP op with shape [batch_size, x_channel, y_channel].
12553 12554
        The x_channel is the channel of x and the y_channel is the channel of y. A Tensor with
        type float32, float64.
12555 12556 12557 12558 12559

    Examples:

        .. code-block:: python

B
Bai Yifan 已提交
12560
            import paddle.fluid as fluid
B
Bai Yifan 已提交
12561
            data = fluid.data(name='data', shape=[None, 3, 32, 32])
B
Bai Yifan 已提交
12562 12563 12564 12565
            feature_map_0 = fluid.layers.conv2d(data, num_filters=2,
                                                filter_size=3)
            feature_map_1 = fluid.layers.conv2d(feature_map_0, num_filters=2,
                                                filter_size=1)
12566 12567 12568 12569 12570 12571 12572 12573
            loss = fluid.layers.fsp_matrix(feature_map_0, feature_map_1)

    """
    helper = LayerHelper('fsp_matrix', **locals())
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype(
        input_param_name='x'))
    helper.append_op(type='fsp', inputs={'X': x, 'Y': y}, outputs={'Out': out})
    return out
H
heqiaozhi 已提交
12574 12575 12576 12577


def continuous_value_model(input, cvm, use_cvm=True):
    """
H
fix doc  
heqiaozhi 已提交
12578

H
heqiaozhi 已提交
12579
    **continuous_value_model layers**
H
fix doc  
heqiaozhi 已提交
12580

Z
zhoushiyu 已提交
12581
    Now, this OP is used in CTR project to remove or dispose show and click value in :attr:`input`.
H
fix doc  
heqiaozhi 已提交
12582

Z
zhoushiyu 已提交
12583 12584 12585 12586 12587
    :attr:`input` is an embedding vector including show and click value, whose shape is :math:`[N, D]` (N is batch size. D is `2 + embedding dim` ).
    Show and click at first two dims of embedding vector D.
    If :attr:`use_cvm` is True, it will caculate :math:`log(show)` and :math:`log(click)` , and output shape is :math:`[N, D]` .
    If :attr:`use_cvm` is False, it will remove show and click from :attr:`input` , and output shape is :math:`[N, D - 2]` .
    :attr:`cvm` is show_click info, whose shape is :math:`[N, 2]` .
H
fix doc  
heqiaozhi 已提交
12588

Z
zhoushiyu 已提交
12589 12590 12591 12592 12593 12594 12595
    Args:
        input (Variable): The input variable. A 2-D LoDTensor with shape :math:`[N, D]` , where N is the batch size, D is `2 + the embedding dim` . `lod level = 1` .
        A Tensor with type float32, float64.
        cvm (Variable): Show and click variable. A 2-D Tensor with shape :math:`[N, 2]` , where N is the batch size, 2 is show and click.
        A Tensor with type float32, float64.
        use_cvm  (bool):  Use show_click or not. if use, the output dim is the same as input.
                          if not use, the output dim is `input dim - 2` (remove show and click)
H
fix doc  
heqiaozhi 已提交
12596

H
heqiaozhi 已提交
12597
    Returns:
H
fix doc  
heqiaozhi 已提交
12598

Z
zhoushiyu 已提交
12599 12600
        Variable: A 2-D LodTensor with shape :math:`[N, M]` . if :attr:`use_cvm` = True, M is equal to input dim D. if False, M is equal to `D - 2`. \
        A Tensor with same type as input.
H
fix doc  
heqiaozhi 已提交
12601

H
heqiaozhi 已提交
12602
    Examples:
H
fix doc  
heqiaozhi 已提交
12603

H
heqiaozhi 已提交
12604
        .. code-block:: python
H
fix doc  
heqiaozhi 已提交
12605

12606
          import paddle.fluid as fluid
Z
zhoushiyu 已提交
12607 12608
          input = fluid.data(name="input", shape=[64, 1], dtype="int64")
          label = fluid.data(name="label", shape=[64, 1], dtype="int64")
H
heqiaozhi 已提交
12609 12610 12611 12612 12613 12614 12615 12616
          embed = fluid.layers.embedding(
                            input=input,
                            size=[100, 11],
                            dtype='float32')
          ones = fluid.layers.fill_constant_batch_size_like(input=label, shape=[-1, 1], dtype="int64", value=1)
          show_clk = fluid.layers.cast(fluid.layers.concat([ones, label], axis=1), dtype='float32')
          show_clk.stop_gradient = True
          input_with_cvm = fluid.layers.continuous_value_model(embed, show_clk, True)
H
fix doc  
heqiaozhi 已提交
12617

H
heqiaozhi 已提交
12618 12619 12620 12621 12622 12623 12624 12625 12626
    """
    helper = LayerHelper('cvm', **locals())
    out = helper.create_variable(dtype=input.dtype)
    helper.append_op(
        type='cvm',
        inputs={'X': [input],
                'CVM': [cvm]},
        outputs={'Y': [out]},
        attrs={"use_cvm": use_cvm})
H
heqiaozhi 已提交
12627
    return out
Z
zhoukunsheng 已提交
12628 12629 12630 12631 12632 12633 12634


def where(condition):
    """
    Return an int64 tensor with rank 2, specifying the coordinate of true element in `condition`.

    Args:
12635
        condition(Variable): A bool tensor with rank at least 1, the data type is bool.
Z
zhoukunsheng 已提交
12636 12637

    Returns:
12638
        Variable, the output data type is int64. : The tensor variable storing a 2-D tensor, which involves all coordinate. 
Z
zhoukunsheng 已提交
12639 12640 12641 12642

    Examples:
        .. code-block:: python

12643
             import paddle.fluid as fluid
12644 12645 12646
             import paddle.fluid.layers as layers
             import numpy as np

Z
zhoukunsheng 已提交
12647
             # condition is a tensor [True, False, True]
12648 12649 12650
             condition = layers.assign(np.array([1, 0, 1], dtype='int32'))
             condition = layers.cast(condition, 'bool')
             out = layers.where(condition) # [[0], [2]]
Z
zhoukunsheng 已提交
12651 12652

             # condition is a tensor [[True, False], [False, True]]
12653 12654 12655
             condition = layers.assign(np.array([[1, 0], [0, 1]], dtype='int32'))
             condition = layers.cast(condition, 'bool')
             out = layers.where(condition) # [[0, 0], [1, 1]]
Z
zhoukunsheng 已提交
12656 12657

             # condition is a tensor [False, False, False]
12658 12659 12660 12661
             condition = layers.assign(np.array([0, 0, 0], dtype='int32'))
             condition = layers.cast(condition, 'bool')
             out = layers.where(condition) # [[]]

Z
zhoukunsheng 已提交
12662 12663 12664 12665 12666 12667 12668 12669 12670
    """
    helper = LayerHelper("where", **locals())

    out = helper.create_variable_for_type_inference(
        dtype=core.VarDesc.VarType.INT64)

    helper.append_op(
        type='where', inputs={'Condition': condition}, outputs={'Out': [out]})
    return out
Z
zhoukunsheng 已提交
12671 12672 12673 12674


def sign(x):
    """
12675
    This OP returns sign of every element in `x`: 1 for positive, -1 for negative and 0 for zero.
Z
zhoukunsheng 已提交
12676 12677

    Args:
12678 12679
        x(Variable|numpy.ndarray): The input variable could be N-D tensor or N-D numpy array, \
            the input data type is float32 or float64.
Z
zhoukunsheng 已提交
12680 12681

    Returns:
12682
        Variable, the output data type is the same as input data type. : The output sign tensor with identical shape to input :attr:`x`.
Z
zhoukunsheng 已提交
12683 12684 12685 12686

    Examples:
        .. code-block:: python

12687 12688 12689
          import paddle.fluid as fluid
          import numpy as np

12690 12691
          # [1.0, 0.0, -1.0]
          data = fluid.layers.sign(np.array([3.0, 0.0, -2.0], dtype='float32')) 
Z
zhoukunsheng 已提交
12692 12693 12694
    """

    helper = LayerHelper("sign", **locals())
12695 12696 12697 12698
    check_type(x, 'x', (Variable, np.ndarray), 'sign')
    if isinstance(x, np.ndarray):
        x = assign(x)
    check_dtype(x.dtype, 'x', ['float16', 'float32', 'float64'], 'sign')
Z
zhoukunsheng 已提交
12699 12700 12701 12702 12703
    out = helper.create_variable_for_type_inference(dtype=x.dtype)

    helper.append_op(type='sign', inputs={'X': [x]}, outputs={'Out': [out]})

    return out
12704 12705


Z
zhoukunsheng 已提交
12706 12707 12708 12709 12710 12711 12712 12713 12714 12715 12716 12717 12718 12719 12720 12721 12722 12723 12724 12725 12726 12727 12728 12729 12730 12731 12732 12733 12734 12735 12736 12737 12738 12739 12740 12741 12742 12743 12744
def unique(x, dtype='int32'):
    """
    **unique** 

    Return a unique tensor for `x` and an index tensor pointing to this unique tensor.

    Args:
        x(Variable): A 1-D input tensor.
        dtype(np.dtype|core.VarDesc.VarType|str): The type of index tensor: int32, int64.

    Returns:
        tuple: (out, index). `out` is the unique tensor for `x`, with identical dtype to `x`, and \
            `index` is an index tensor pointing to `out`, by which user can recover the original `x` tensor.

    Examples:
        .. code-block:: python

             import numpy as np
             import paddle.fluid as fluid
             x = fluid.assign(np.array([2, 3, 3, 1, 5, 3], dtype='int32'))
             out, index = fluid.layers.unique(x) # out is [2, 3, 1, 5]; index is [0, 1, 1, 2, 3, 1]
    """

    helper = LayerHelper("unique", **locals())

    out = helper.create_variable_for_type_inference(dtype=x.dtype)

    index = helper.create_variable_for_type_inference(dtype)

    helper.append_op(
        type='unique',
        inputs={'X': x},
        attrs={'dtype': convert_np_dtype_to_dtype_(dtype)},
        outputs={'Out': [out],
                 'Index': [index]})

    return out, index


12745 12746
def unique_with_counts(x, dtype='int32'):
    """
12747 12748
    This OP return a unique tensor for `x` , and count tensor that the count of unqiue result in raw input, \
    and an index tensor pointing to this unique tensor. 
12749

12750
    **NOTICE**: This op support the variable type of Tensor only.
12751 12752

    Args:
12753 12754
        x(Variable): A 1-D input tensor with input shape of :math:`[N]` , the input data type is float32, float64, int32, int64.
        dtype(np.dtype|core.VarDesc.VarType|str): The type of count and index tensor, it could be int32, int64. Defalut value is int32.
12755

12756 12757 12758 12759 12760 12761
    Returns: 
        tuple, the variable type in tuple is Tensor, the output :attr:`out` data type is the same as input :attr:`x`, \
        and data type of output :attr:`index` and :attr:`count` will be int32 or int64.: The :attr:`out` is unique tensor for input :attr:`x`,\
        the data shape is :math:`[K]`, the `K` may be different to the `N` in shape of :attr:`x`. :attr:`index` is an index tensor pointing\
        to :attr:`out`, the data shape is :math:`[N]` , the data shape is the same as input :attr:`x`. :attr:`count` is count of unqiue element in\
        the :attr:`x`, the data shape is :math:`[K]`, the data shape is the same as output :attr:`out`.
12762 12763 12764 12765 12766 12767 12768 12769 12770

    Examples:
        .. code-block:: python

             import numpy as np
             import paddle.fluid as fluid
             x = fluid.layers.assign(np.array([2, 3, 3, 1, 5, 3], dtype='int32'))
             out, index, count = fluid.layers.unique_with_counts(x) # out is [2, 3, 1, 5]; index is [0, 1, 1, 2, 3, 1]
                                                        # count is [1, 3, 1, 1]
12771
            # x.shape=(6,) out.shape=(4,), index.shape=(6,), count.shape=(4,)
12772 12773 12774 12775 12776 12777 12778 12779 12780 12781 12782 12783 12784 12785 12786 12787 12788 12789 12790 12791 12792 12793 12794 12795 12796 12797 12798 12799 12800
    """
    if not (dtype == 'int32' or dtype == 'int64'):
        raise TypeError(
            "Op unique_with_counts, index dtype must be int32 or int64")

    if x is None or len(x.shape) != 1:
        raise ValueError(
            "Op unique_with_counts, x must not be null and size of dim must be 1"
        )

    helper = LayerHelper("unique_with_counts", **locals())

    out = helper.create_variable_for_type_inference(dtype=x.dtype)

    index = helper.create_variable_for_type_inference(dtype)

    count = helper.create_variable_for_type_inference(dtype)

    helper.append_op(
        type='unique_with_counts',
        inputs={'X': x},
        attrs={'dtype': convert_np_dtype_to_dtype_(dtype)},
        outputs={'Out': [out],
                 'Index': [index],
                 'Count': [count]})

    return out, index, count


12801 12802 12803 12804 12805 12806 12807 12808 12809 12810 12811 12812 12813
def deformable_conv(input,
                    offset,
                    mask,
                    num_filters,
                    filter_size,
                    stride=1,
                    padding=0,
                    dilation=1,
                    groups=None,
                    deformable_groups=None,
                    im2col_step=None,
                    param_attr=None,
                    bias_attr=None,
12814
                    modulated=True,
12815 12816
                    name=None):
    """
12817
    **Deformable Convolution op**
12818 12819 12820

    Compute 2-D deformable convolution on 4-D input.
    Given input image x, output feature map y, the deformable convolution operation can be expressed as follow:
12821 12822 12823
   
    
    Deformable Convolution v2: 
12824 12825 12826 12827
    
    .. math::

        y(p) = \sum_{k=1}^{K}{w_k * x(p + p_k + \Delta p_k) * \Delta m_k}
12828 12829

    Deformable Convolution v1:
12830
    
12831 12832 12833 12834 12835
    .. math::

        y(p) = \sum_{k=1}^{K}{w_k * x(p + p_k + \Delta p_k)}
    
    Where :math:`\Delta p_k` and :math:`\Delta m_k` are the learnable offset and modulation scalar for the k-th location, 
12836
    Which :math:`\Delta m_k` is one in deformable convolution v1. Please refer to `Deformable ConvNets v2: More Deformable, Better Results
12837
    <https://arxiv.org/abs/1811.11168v2>`_ and `Deformable Convolutional Networks <https://arxiv.org/abs/1703.06211>`_.
12838 12839 12840 12841 12842 12843 12844 12845 12846 12847 12848 12849 12850 12851 12852 12853 12854 12855 12856 12857 12858 12859 12860 12861
    
    Example:
        - Input:

          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`

          Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)`

          Offset shape: :math:`(N, 2 * deformable\_groups * H_f * H_w, H_{in}, W_{in})`

          Mask shape: :math:`(N, deformable\_groups * H_f * H_w, H_{in}, W_{in})`

        - Output:

          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`

        Where

        .. math::

            H_{out}&= \\frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]} + 1 \\\\
            W_{out}&= \\frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]} + 1

    Args:
12862 12863
        input (Variable): The input image with [N, C, H, W] format. A Tensor with type
            float32, float64.
12864
        offset (Variable): The input coordinate offset of deformable convolution layer.
12865
            A Tensor with type float32, float64.
12866 12867 12868
        Mask (Variable, Optional): The input mask of deformable convolution layer.
            A Tensor with type float32, float64. It should be None when you use
            deformable convolution v1.
12869 12870
        num_filters(int): The number of filter. It is as same as the output
            image channel.
12871
        filter_size (int|tuple): The filter size. If filter_size is a tuple,
12872 12873 12874 12875 12876 12877 12878 12879 12880 12881 12882 12883 12884 12885 12886 12887 12888 12889 12890 12891 12892 12893 12894
            it must contain two integers, (filter_size_H, filter_size_W).
            Otherwise, the filter will be a square.
        stride (int|tuple): The stride size. If stride is a tuple, it must
            contain two integers, (stride_H, stride_W). Otherwise, the
            stride_H = stride_W = stride. Default: stride = 1.
        padding (int|tuple): The padding size. If padding is a tuple, it must
            contain two integers, (padding_H, padding_W). Otherwise, the
            padding_H = padding_W = padding. Default: padding = 0.
        dilation (int|tuple): The dilation size. If dilation is a tuple, it must
            contain two integers, (dilation_H, dilation_W). Otherwise, the
            dilation_H = dilation_W = dilation. Default: dilation = 1.
        groups (int): The groups number of the deformable conv layer. According to
            grouped convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
            the first half of the filters is only connected to the first half
            of the input channels, while the second half of the filters is only
            connected to the second half of the input channels. Default: groups=1.
        deformable_groups (int): The number of deformable group partitions.
            Default: deformable_groups = 1.
        im2col_step (int): Maximum number of images per im2col computation; 
            The total batch size should be divisable by this value or smaller
            than this value; if you face out of memory problem, you can try
            to use a smaller value here.
            Default: im2col_step = 64.
12895
        param_attr (ParamAttr, Optional): The parameter attribute for learnable parameters/weights
12896 12897 12898 12899 12900
            of deformable conv. If it is set to None or one attribute of ParamAttr,
            deformable conv will create ParamAttr as param_attr.
            If the Initializer of the param_attr is not set, the parameter is
            initialized with :math:`Normal(0.0, std)`, and the 
            :math:`std` is :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None.
12901
        bias_attr (ParamAttr|bool, Optional): The parameter attribute for the bias of
12902 12903 12904 12905
            deformable conv layer. If it is set to False, no bias will be added
            to the output units. If it is set to None or one attribute of ParamAttr, conv2d
            will create ParamAttr as bias_attr. If the Initializer of the bias_attr
            is not set, the bias is initialized zero. Default: None.
12906 12907
        modulated (bool): Make sure which version should be used between v1 and v2, where v2 is \
            used while True. Default: True.
12908 12909
        name(str, Optional): For details, please refer to :ref:`api_guide_Name`.
                        Generally, no setting is required. Default: None.
12910 12911
    Returns:
        Variable: The tensor variable storing the deformable convolution \
12912
                  result. A Tensor with type float32, float64.
12913 12914 12915 12916 12917 12918
    Raises:
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
    Examples:
        .. code-block:: python

12919 12920
          #deformable conv v2:
         
12921
          import paddle.fluid as fluid
12922 12923
          C_in, H_in, W_in = 3, 32, 32
          filter_size, deformable_groups = 3, 1
B
Bai Yifan 已提交
12924 12925 12926
          data = fluid.data(name='data', shape=[None, C_in, H_in, W_in], dtype='float32')
          offset = fluid.data(name='offset', shape=[None, 2*deformable_groups*filter_size**2, H_in, W_in], dtype='float32')
          mask = fluid.data(name='mask', shape=[None, deformable_groups*filter_size**2, H_in, W_in], dtype='float32')
12927
          out = fluid.layers.deformable_conv(input=data, offset=offset, mask=mask,
12928
                                             num_filters=2, filter_size=filter_size, padding=1, modulated=True)
12929 12930 12931 12932

          #deformable conv v1:

          import paddle.fluid as fluid
12933 12934
          C_in, H_in, W_in = 3, 32, 32
          filter_size, deformable_groups = 3, 1
B
Bai Yifan 已提交
12935 12936
          data = fluid.data(name='data', shape=[None, C_in, H_in, W_in], dtype='float32')
          offset = fluid.data(name='offset', shape=[None, 2*deformable_groups*filter_size**2, H_in, W_in], dtype='float32')
12937
          out = fluid.layers.deformable_conv(input=data, offset=offset, mask=None,
12938
                                             num_filters=2, filter_size=filter_size, padding=1, modulated=False)
12939 12940 12941 12942 12943 12944 12945 12946 12947 12948 12949 12950 12951 12952 12953 12954 12955 12956 12957 12958 12959 12960 12961 12962 12963 12964 12965 12966 12967 12968 12969 12970 12971 12972 12973 12974 12975 12976 12977 12978 12979
    """

    num_channels = input.shape[1]
    assert param_attr is not False, "param_attr should not be False here."

    helper = LayerHelper('deformable_conv', **locals())
    dtype = helper.input_dtype()

    if not isinstance(input, Variable):
        raise TypeError("Input of deformable_conv must be Variable")
    if not isinstance(offset, Variable):
        raise TypeError("Input Offset of deformable_conv must be Variable")

    if groups is None:
        num_filter_channels = num_channels
    else:
        if num_channels % groups != 0:
            raise ValueError("num_channels must be divisible by groups.")
        num_filter_channels = num_channels // groups

    filter_size = utils.convert_to_list(filter_size, 2, 'filter_size')
    stride = utils.convert_to_list(stride, 2, 'stride')
    padding = utils.convert_to_list(padding, 2, 'padding')
    dilation = utils.convert_to_list(dilation, 2, 'dilation')

    input_shape = input.shape
    filter_shape = [num_filters, int(num_filter_channels)] + filter_size

    def _get_default_param_initializer():
        filter_elem_num = filter_size[0] * filter_size[1] * num_channels
        std = (2.0 / filter_elem_num)**0.5
        return Normal(0.0, std, 0)

    filter_param = helper.create_parameter(
        attr=helper.param_attr,
        shape=filter_shape,
        dtype=dtype,
        default_initializer=_get_default_param_initializer())

    pre_bias = helper.create_variable_for_type_inference(dtype)

12980 12981 12982 12983 12984 12985 12986 12987 12988 12989 12990 12991 12992 12993 12994 12995 12996 12997 12998 12999 13000 13001 13002 13003 13004 13005 13006 13007 13008 13009 13010 13011 13012 13013 13014 13015
    if modulated:
        helper.append_op(
            type='deformable_conv',
            inputs={
                'Input': input,
                'Filter': filter_param,
                'Offset': offset,
                'Mask': mask,
            },
            outputs={"Output": pre_bias},
            attrs={
                'strides': stride,
                'paddings': padding,
                'dilations': dilation,
                'groups': groups,
                'deformable_groups': deformable_groups,
                'im2col_step': im2col_step,
            })

    else:
        helper.append_op(
            type='deformable_conv_v1',
            inputs={
                'Input': input,
                'Filter': filter_param,
                'Offset': offset,
            },
            outputs={"Output": pre_bias},
            attrs={
                'strides': stride,
                'paddings': padding,
                'dilations': dilation,
                'groups': groups,
                'deformable_groups': deformable_groups,
                'im2col_step': im2col_step,
            })
13016 13017 13018

    output = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
    return output
13019 13020 13021 13022 13023


def unfold(x, kernel_sizes, strides=1, paddings=0, dilations=1, name=None):
    """

S
SunGaofeng 已提交
13024
    This op returns a col buffer of sliding local blocks of input x, also known
13025 13026 13027 13028
    as im2col for batched 2D image tensors. For each block under the convolution filter,
    all element will be rearranged as a column. While the convolution filter silding over
    the input feature map, a series of such columns will be formed.

S
SunGaofeng 已提交
13029
    For each input :math:`x` with shape [N, C, H, W], the output shape [N, Cout, Lout]
13030 13031 13032 13033 13034 13035 13036 13037 13038 13039 13040 13041 13042 13043 13044 13045 13046
    can be calculated as following.

    .. math::

        dkernel[0] &= dilations[0] \\times (kernel\_sizes[0] - 1) + 1

        dkernel[1] &= dilations[1] \\times (kernel\_sizes[1] - 1) + 1

        hout &= \\frac{H + paddings[0] + paddings[2] - dkernel[0]}{strides[0]} + 1

        wout &= \\frac{W + paddings[1] + paddings[3] - dkernel[1]}{strides[1]} + 1

        Cout &= C \\times kernel\_sizes[0] \\times kernel\_sizes[1]

        Lout &= hout \\times wout


S
SunGaofeng 已提交
13047 13048 13049
    Parameters:
        x(Varaible):              4-D Tensor, input tensor of format [N, C, H, W], 
                                  data type can be float32 or float64
13050 13051 13052 13053 13054 13055 13056 13057 13058 13059 13060 13061 13062 13063 13064
        kernel_sizes(int|list):   The size of convolution kernel, should be [k_h, k_w]
                                  or an integer k treated as [k, k].
        strides(int|list):        The strides, should be [stride_h, stride_w]
                                  or an integer stride treated as [sride, stride].
                                  For default, strides will be [1, 1].
        paddings(int|list):       The paddings of each dimension, should be
                                  [padding_top, padding_left, padding_bottom, padding_right]
                                  or [padding_h, padding_w] or an integer padding.
                                  If [padding_h, padding_w] was given, it will expanded to
                                  [padding_h, padding_w, padding_h, padding_w]. If an integer
                                  padding was given, [padding, padding, padding, padding] will
                                  be used. For default, paddings will be [0, 0, 0, 0]
        dilations(int|list):      the dilations of convolution kernel, shold be
                                  [dilation_h, dilation_w], or an integer dialtion treated as
                                  [dilation, dilation]. For default, it will be [1, 1].
S
SunGaofeng 已提交
13065 13066 13067
        name(str, optional): The default value is None.  
                             Normally there is no need for user to set this property.  
                             For more information, please refer to :ref:`api_guide_Name`
13068 13069 13070

    
    Returns:
S
SunGaofeng 已提交
13071 13072 13073 13074 13075 13076 13077 13078
        The tensor variable corresponding to the sliding local blocks. 
        The output shape is [N, Cout, Lout] as decribled above. 
        Cout is the  total number of values within each block, 
        and Lout is the total number of such blocks. 
        The data type of output is the same as the input :math:`x`

    Return Type:
        Variable
13079 13080 13081 13082 13083 13084

    Examples:

        .. code-block:: python

            import paddle.fluid as fluid
S
SunGaofeng 已提交
13085
            x = fluid.data(name = 'data', shape = [100, 3, 224, 224], dtype = 'float32')
13086 13087 13088 13089 13090 13091 13092 13093 13094 13095 13096 13097 13098 13099 13100 13101 13102 13103 13104 13105 13106 13107 13108 13109 13110 13111 13112 13113 13114 13115 13116 13117 13118 13119 13120 13121 13122 13123 13124 13125 13126 13127 13128 13129 13130 13131 13132 13133 13134 13135 13136 13137 13138 13139
            y = fluid.layers.unfold(x, [3, 3], 1, 1, 1)
    """

    helper = LayerHelper("unfold", **locals())

    assert len(x.shape) == 4, \
            "input should be the format of [N, C, H, W]"

    if isinstance(kernel_sizes, int):
        kernel_sizes = [kernel_sizes, kernel_sizes]
    else:
        assert isinstance(kernel_sizes, list) and (len(kernel_sizes) == 2), \
            "kernel_sizes should either be an integer or a list of two integers"

    if isinstance(strides, int):
        strides = [strides, strides]
    else:
        assert isinstance(strides, list) and (len(strides) == 2), \
            "strides should either be an integer or a list of two integers"

    if isinstance(dilations, int):
        dilations = [dilations, dilations]
    else:
        assert isinstance(dilations, list) and (len(dilations) == 2), \
            "dilations should either be an integer or a list of two integers"

    if isinstance(paddings, int):
        paddings = [paddings] * 4
    elif isinstance(paddings, list):
        if len(paddings) == 2:
            paddings = paddings * 2
        elif len(paddings) == 4:
            pass
        else:
            raise ValueError(
                "paddings should either be an integer or a list of 2 or 4 integers"
            )
    else:
        raise ValueError(
            "Unexpected type of paddings, it should be either an integer or a list"
            "of 2 or 4 integers")

    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
        type="unfold",
        inputs={"X": x},
        outputs={"Y": out},
        attrs={
            "kernel_sizes": kernel_sizes,
            "strides": strides,
            "paddings": paddings,
            "dilations": dilations
        })
    return out
C
cjt222 已提交
13140 13141 13142 13143 13144 13145 13146 13147 13148 13149 13150 13151 13152 13153 13154 13155


def deformable_roi_pooling(input,
                           rois,
                           trans,
                           no_trans=False,
                           spatial_scale=1.0,
                           group_size=[1, 1],
                           pooled_height=1,
                           pooled_width=1,
                           part_size=None,
                           sample_per_part=1,
                           trans_std=0.1,
                           position_sensitive=False,
                           name=None):
    """
13156 13157 13158 13159 13160 13161 13162
    Deformable ROI Pooling Layer
  
    Performs deformable region-of-interest pooling on inputs. As described
    in `Deformable Convolutional Networks <https://arxiv.org/abs/1703.06211>`_, it will get offset for each bin after 
    roi pooling so that pooling at correct region. Batch_size will change to the number of region bounding boxes after deformable_roi_pooling.
  
    The operation has three steps:
C
cjt222 已提交
13163
    
13164 13165 13166 13167 13168 13169 13170 13171 13172 13173 13174 13175 13176 13177 13178 13179 13180 13181 13182 13183 13184 13185 13186 13187 13188 13189 13190 13191 13192 13193 13194 13195 13196 13197 13198 13199 13200 13201 13202
    1. Dividing each region proposal into equal-sized sections with the pooled_width and pooled_height.
  
    2. Add offset to pixel in ROI to get new location and the new value which are computed directly through
       bilinear interpolation with four nearest pixel.
     
    3. Sample several points in each bin to get average values as output.
  
  
    Args:
        input (Variable):The input of deformable roi pooling and it is tensor which value type is float32. The shape of input is
                         [N, C, H, W]. Where N is batch size, C is number of input channels,
                         H is height of the feature, and W is the width of the feature.
        rois (Variable): ROIs (Regions of Interest) with type float32 to pool over. It should be
                         a 2-D LoDTensor of shape (num_rois, 4), and the lod level
                         is 1. Given as [[x1, y1, x2, y2], ...], (x1, y1) is
                         the top left coordinates, and (x2, y2) is the bottom
                         right coordinates, which value type is float32.
        trans (Variable): Offset of features on ROIs while pooling which value type is float32. The format is [N, C, H, W], where 
                          N is number of ROIs, C is number of channels, which indicate the offset distance 
                          in the x and y directions, H is pooled height, and W is pooled width. 
        no_trans (bool): Whether to add offset to get new value or not while roi pooling, which value with type bool is True or False.
                         If value is True, no offset will be added in operation. Default: False.
        spatial_scale (float): Ratio of input feature map height (or width) to raw image height (or width), which value type is float32.
                         Equals the reciprocal of total stride in convolutional layers, Default: 1.0.
        group_size (list|tuple): The number of groups which input channels are divided and the input is list or tuple, which value type is int32. (eg.number of input channels 
                          is k1 * k2 * (C + 1), which k1 and k2 are group width and height and C+1 is number of output
                          chanels.) eg.(4, 6), which 4 is height of group and 6 is width of group. Default: [1, 1].
        pooled_height (int): The pooled output height which value type is int32. Default: 1.
        pooled_width (int): The pooled output width which value type is int32. Default: 1.
        part_size (list|tuple): The height and width of offset which values in list or tuple is int32, eg.(4, 6), which height is 4 and width is 6, and values always equal to pooled_height \
                         and pooled_width. Default: if None, default value is [pooled_height, pooled_width].
        sample_per_part (int): The number of samples in each bin which value type is int32. If value is bigger, it will consume more performance. Default: 1.
        trans_std (float): Coefficient of offset which value type is float32. It controls weight of offset. Default: 0.1.
        position_sensitive (bool): Whether to choose deformable psroi pooling mode or not, and value type is bool(True or False). If value is False, input dimension equals to output dimension. \
                                   If value is True, input dimension shoule be output dimension * pooled_height * pooled_width. Default: False.
        name (str|None): Name of layer. Default: None.
    Returns:
        Variable: Output of deformable roi pooling is that, if position sensitive is False, input dimension equals to output dimension. If position sensitive is True,\
                  input dimension should be the result of output dimension divided by pooled height and pooled width.
C
cjt222 已提交
13203 13204 13205 13206

    Examples:
      .. code-block:: python

13207 13208
        # position_sensitive=True
        import paddle.fluid as fluid
C
chengjuntao 已提交
13209 13210 13211 13212 13213 13214 13215 13216 13217 13218 13219 13220 13221 13222 13223 13224 13225 13226 13227 13228 13229 13230
        input = fluid.data(name="input",
                           shape=[2, 192, 64, 64], 
                           dtype='float32')                   
        rois = fluid.data(name="rois",
                          shape=[-1, 4],
                          dtype='float32', 
                          lod_level=1)
        trans = fluid.data(name="trans",
                           shape=[2, 384, 64, 64], 
                           dtype='float32') 
        x = fluid.layers.deformable_roi_pooling(input=input, 
                                                rois=rois, 
                                                trans=trans, 
                                                no_trans=False,
                                                spatial_scale=1.0, 
                                                group_size=(1, 1),
                                                pooled_height=8,
                                                pooled_width=8,
                                                part_size=(8, 8),
                                                sample_per_part=4, 
                                                trans_std=0.1,
                                                position_sensitive=True)
13231 13232
  
        # position_sensitive=False
13233
        import paddle.fluid as fluid
C
chengjuntao 已提交
13234 13235 13236 13237 13238 13239 13240 13241 13242 13243 13244 13245 13246 13247 13248 13249 13250 13251 13252 13253 13254 13255
        input = fluid.data(name="input",
                           shape=[2, 192, 64, 64], 
                           dtype='float32')                   
        rois = fluid.data(name="rois",
                          shape=[-1, 4],
                          dtype='float32', 
                          lod_level=1)
        trans = fluid.data(name="trans",
                           shape=[2, 384, 64, 64], 
                           dtype='float32') 
        x = fluid.layers.deformable_roi_pooling(input=input, 
                                                rois=rois, 
                                                trans=trans, 
                                                no_trans=False,
                                                spatial_scale=1.0, 
                                                group_size=(1, 1),
                                                pooled_height=8,
                                                pooled_width=8,
                                                part_size=(8, 8),
                                                sample_per_part=4, 
                                                trans_std=0.1,
                                                position_sensitive=False)
C
cjt222 已提交
13256 13257 13258 13259 13260 13261 13262 13263 13264 13265 13266 13267 13268 13269 13270 13271 13272 13273 13274 13275 13276 13277 13278 13279 13280 13281 13282 13283 13284 13285 13286 13287 13288 13289 13290 13291 13292
    """

    input_channels = input.shape[1]
    if position_sensitive == False:
        output_channels = input_channels
    else:
        output_channels = input_channels / pooled_height / pooled_width

    if part_size is None:
        part_height = pooled_height
        part_width = pooled_width
        part_size = [part_height, part_width]
    part_size = utils.convert_to_list(part_size, 2, 'part_size')
    group_size = utils.convert_to_list(group_size, 2, 'group_size')
    helper = LayerHelper('deformable_psroi_pooling', **locals())
    dtype = helper.input_dtype()
    output = helper.create_variable_for_type_inference(dtype)
    top_count = helper.create_variable_for_type_inference(dtype='int32')
    helper.append_op(
        type="deformable_psroi_pooling",
        inputs={"Input": input,
                "ROIs": rois,
                "Trans": trans},
        outputs={"Output": output,
                 "TopCount": top_count},
        attrs={
            "no_trans": no_trans,
            "spatial_scale": spatial_scale,
            "output_dim": output_channels,
            "group_size": group_size,
            "pooled_height": pooled_height,
            "pooled_width": pooled_width,
            "part_size": part_size,
            "sample_per_part": sample_per_part,
            "trans_std": trans_std
        })
    return output
13293 13294 13295 13296


def shard_index(input, index_num, nshards, shard_id, ignore_value=-1):
    """
13297
    This operator recomputes the `input` indices according to the offset of the
13298 13299 13300 13301 13302
    shard. The length of the indices is evenly divided into N shards, and if
    the `shard_id` matches the shard with the input index inside, the index is
    recomputed on the basis of the shard offset, elsewise it is set to
    `ignore_value`. The detail is as follows:
    :: 
13303
        
13304 13305
        shard_size = (index_num + nshards - 1) // nshards
        y = x % shard_size if x // shard_size == shard_id else ignore_value
13306

13307 13308
    NOTE: If the length of indices cannot be evely divided by the shard number,
    the size of the last shard will be less than the calculated `shard_size`
13309 13310

    Examples:
13311
    ::
13312
    
13313
        Input:
13314 13315
          X.shape = [4, 1]
          X.data = [[1], [6], [12], [19]]
13316 13317 13318
          index_num = 20
          nshards = 2
          ignore_value = -1
13319
        
13320
        if shard_id == 0, we get:
13321 13322 13323
          Out.shape = [4, 1]
          Out.data = [[1], [6], [-1], [-1]]
        
13324
        if shard_id == 1, we get:
13325 13326 13327 13328
          Out.shape = [4, 1]
          Out.data = [[-1], [-1], [2], [9]]
    
    Args:
13329 13330 13331 13332 13333
        - **input** (Variable): Input indices, last dimension must be 1.
        - **index_num** (scalar): An interger defining the range of the index.
        - **nshards** (scalar): The number of shards
        - **shard_id** (scalar): The index of the current shard
        - **ignore_value** (scalar): An ingeter value out of sharded index range
13334 13335

    Returns:
13336
        Variable: The sharded index of input.
13337 13338 13339 13340 13341

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
13342 13343
            batch_size = 32
            label = fluid.data(name="label", shape=[batch_size, 1], dtype="int64")
13344 13345 13346 13347 13348 13349 13350 13351 13352 13353 13354 13355 13356 13357 13358 13359 13360 13361 13362 13363 13364 13365 13366 13367
            shard_label = fluid.layers.shard_index(input=label,
                                                   index_num=20,
                                                   nshards=2,
                                                   shard_id=0)
    """
    op_type = 'shard_index'
    helper = LayerHelper(op_type, **locals())
    if shard_id < 0 or shard_id >= nshards:
        raise ValueError('The shard_id(%d) should be in [0, %d)' %
                         (shard_id, nshards))

    out = helper.create_variable_for_type_inference(dtype=input.dtype)
    helper.append_op(
        type=op_type,
        inputs={'X': [input]},
        outputs={'Out': out},
        attrs={
            'index_num': index_num,
            'nshards': nshards,
            'shard_id': shard_id,
            'ignore_value': ignore_value
        },
        stop_gradient=True)
    return out
H
huangjun12 已提交
13368 13369 13370 13371 13372


@templatedoc()
def hard_swish(x, threshold=6.0, scale=6.0, offset=3.0, name=None):
    """
13373 13374 13375
    This operator implements the hard_swish activation function.
    Hard_swish is proposed in MobileNetV3, and performs better in computational stability and efficiency compared to swish function.
    For more details please refer to: https://arxiv.org/pdf/1905.02244.pdf
H
huangjun12 已提交
13376

13377
    The formula is as follows:
H
huangjun12 已提交
13378

13379
    .. math::
H
huangjun12 已提交
13380

13381
        out = \\frac{x * (min(max(0, x+offset), threshold))}{scale}
H
huangjun12 已提交
13382

13383 13384 13385 13386 13387 13388 13389 13390 13391 13392 13393 13394 13395 13396 13397 13398 13399 13400 13401 13402 13403 13404 13405 13406 13407 13408 13409 13410 13411 13412 13413 13414 13415 13416
    In the above equation:

    ``threshold`` and ``scale`` should be positive, ``offset`` can be positive or negative. It is recommended to use default parameters.

    Args:
        x (Variable): Input feature, multi-dimensional Tensor. The data type should be float32 or float64.
        threshold (float, optional): The threshold in Relu function. Default: 6.0
        scale (float, optional): The scale factor. Default: 6.0
        offset (float, optional): The offset factor. Default: 3.0
        name (str, optional): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name` 
        
    Returns:
        Variable: The output tensor with the same shape and data type as input.
    
    
    Examples:
    
    .. code-block:: python
    
        import paddle.fluid as fluid
        import numpy as np
    
        DATATYPE='float32'
    
        x_data = np.array([i for i in range(1,5)]).reshape([1,1,4]).astype(DATATYPE)
    
        x = fluid.data(name="x", shape=[None,1,4], dtype=DATATYPE)
        y = fluid.layers.hard_swish(x)
    
        place = fluid.CPUPlace()
        #place = fluid.CUDAPlace(0)
        exe = fluid.Executor(place)
        out, = exe.run(feed={'x':x_data}, fetch_list=[y.name])
        print(out)  # [[0.66666667, 1.66666667,3., 4.]]
H
huangjun12 已提交
13417 13418 13419 13420 13421 13422 13423 13424 13425 13426 13427
    """
    helper = LayerHelper('hard_swish', **locals())
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
        type='hard_swish',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'threshold': threshold,
               'scale': scale,
               'offset': offset})
    return out
R
ruri 已提交
13428 13429


G
Guo Sheng 已提交
13430 13431 13432 13433 13434 13435 13436 13437 13438 13439 13440 13441 13442 13443 13444 13445 13446 13447 13448 13449 13450 13451 13452 13453 13454 13455 13456 13457 13458 13459 13460 13461 13462 13463 13464 13465 13466 13467 13468 13469 13470 13471 13472 13473 13474 13475 13476 13477 13478 13479 13480 13481 13482 13483 13484 13485 13486 13487 13488 13489 13490 13491 13492 13493 13494 13495 13496 13497 13498 13499 13500 13501 13502 13503 13504
def gather_tree(ids, parents):
    """
    To be used after beam search. After beam search, we get selected ids at
    each time step and the corresponding parents in the search tree. Both ids
    and parents have the layout :attr:`[max_time, batch_size, beam_size]`. Then
    :attr:`gather_tree` is used to backtrace from the last time step and
    generate the full sequences by collecting selected ids.

    Here is an example:

    .. code-block:: text

            Given:
                ids = [[[2 2]
                        [6 1]]
                       [[3 9]
                        [6 1]]
                       [[0 1]
                        [9 0]]]
                parents = [[[0 0]
                            [1 1]]
                           [[1 0]
                            [1 0]]
                           [[0 0]
                            [0 1]]]

            Then:                
                gather_tree(ids, parents)  
                         = [[[2 2]
                             [1 6]]
                            [[3 3]
                             [6 1]]
                            [[0 1]
                             [9 0]]]

    Args:
        ids(Variable): A Tensor with shape :attr:`[length, batch_size, beam_size]`
            and data type :attr:`int32` or :attr:`int64`. It contains the selected
            ids of all time steps.
        parents(Variable): A Tensor with the same shape and data type as :attr:`ids`,
            It contains the parents corresponding to selected ids when searching
            among beams.

    Returns:
        Variable: A Tensor with the same shape and data type as :attr:`ids`. \
            It contains the full sequences. The sequences are collected from \
            :attr:`ids` by backtracing according to :attr:`parents`.

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid

            ids = fluid.layers.data(name='ids',
                                    shape=[5, 2, 2],
                                    dtype='int64',
                                    append_batch_size=False)
            parents = fluid.layers.data(name='parents',
                                        shape=[5, 2, 2],
                                        dtype='int64',
                                        append_batch_size=False)
            final_sequences = fluid.layers.gather_tree(ids, parents)
    """
    helper = LayerHelper('gather_tree', **locals())
    out = helper.create_variable_for_type_inference(dtype=ids.dtype)

    helper.append_op(
        type="gather_tree",
        inputs={"Ids": ids,
                "Parents": parents},
        outputs={"Out": out})

    return out


13505 13506 13507
@templatedoc()
def uniform_random(shape, dtype='float32', min=-1.0, max=1.0, seed=0):
    """
13508 13509
    This OP initializes a variable with random values sampled from a
    uniform distribution in the range [min, max).
13510 13511 13512 13513 13514 13515 13516 13517 13518 13519 13520

    Examples:
    ::
    
        Input:
          shape = [1, 2]
        
        Output:
          result=[[0.8505902, 0.8397286]]

    Args:
13521 13522
        shape (list|tuple|Variable): The shape of the output Tensor,  if the shape is a list or tuple, 
                                     its elements can be an integer
13523 13524
                                     or a Tensor with the shape [1], and the type of the Tensor must be int32 or int64. 
                                     If the shape is a Variable, it is a 1-D Tensor, and the type of the Tensor must be int32 or int64.
13525
        dtype(np.dtype|core.VarDesc.VarType|str, optional): The type of the output Tensor. Supported data types: float32, float64.
13526
                                                  Default: float32.
13527 13528
        min (float, optional): The lower bound on the range of random values to generate, the min is included in the range. Default -1.0.
        max (float, optional): The upper bound on the range of random values to generate, the max is excluded in the range. Default 1.0.
13529 13530 13531 13532 13533
        seed (int, optional): Random seed used for generating samples. 0 means use a
            seed generated by the system. Note that if seed is not 0, this
            operator will always generate the same random numbers every time.
            Default 0.

13534 13535
    Returns: 
        Variable: A Tensor of the specified shape filled with uniform_random values.
13536

13537
    Raises:
13538 13539 13540 13541 13542 13543 13544 13545 13546 13547 13548 13549 13550 13551
        TypeError: The shape type should be list or tupple or variable.
    
    Examples:
        .. code-block:: python

            import paddle.fluid as fluid

            # example 1:
            # attr shape is a list which doesn't contain tensor Variable.
            result_1 = fluid.layers.uniform_random(shape=[3, 4])

            # example 2:
            # attr shape is a list which contains tensor Variable.
            dim_1 = fluid.layers.fill_constant([1],"int64",3)
13552 13553
            dim_2 = fluid.layers.fill_constant([1],"int32",5)
            result_2 = fluid.layers.uniform_random(shape=[dim_1, dim_2])
13554 13555

            # example 3:
13556
            # attr shape is a Variable, the data type must be int64 or int32.
13557
            var_shape = fluid.data(name='var_shape', shape=[2], dtype="int64")
13558
            result_3 = fluid.layers.uniform_random(var_shape)
13559 13560 13561 13562
            var_shape_int32 = fluid.data(name='var_shape_int32', shape=[2], dtype="int32")
            result_4 = fluid.layers.uniform_random(var_shape_int32)
             

13563 13564

    """
13565
    check_type(shape, 'shape', (list, tuple, Variable), 'uniform_random')
13566 13567
    if not isinstance(dtype, core.VarDesc.VarType):
        dtype = convert_np_dtype_to_dtype_(dtype)
13568
    check_dtype(dtype, 'dtype', ['float32', 'float64'], 'uniform_random')
13569

13570 13571 13572 13573 13574 13575 13576 13577 13578 13579 13580 13581 13582 13583 13584 13585 13586 13587 13588 13589 13590 13591 13592 13593 13594 13595 13596 13597 13598 13599 13600 13601 13602 13603
    def contain_var(one_list):
        for ele in one_list:
            if isinstance(ele, Variable):
                return True
        return False

    def get_new_shape_tensor(list_shape):
        new_shape_tensor = []
        for dim in list_shape:
            if isinstance(dim, Variable):
                dim.stop_gradient = True
                new_shape_tensor.append(dim)
            else:
                assert (isinstance(dim, int))
                temp_out = helper.create_variable_for_type_inference('int64')
                fill_constant([1], 'int64', dim, force_cpu=True, out=temp_out)
                new_shape_tensor.append(temp_out)
        return new_shape_tensor

    def get_attr_shape(list_shape):
        unk_dim_idx = -1
        attrs_shape = []
        for dim_idx, dim_size in enumerate(list_shape):
            if isinstance(dim_size, Variable):
                attrs_shape.append(-1)
            else:
                attrs_shape.append(dim_size)
                assert dim_size > 0, (
                    "Each dimension size given in shape must not be negtive "
                    "except one unknown dimension.")
        return attrs_shape

    helper = LayerHelper("uniform_random", **locals())
    inputs = dict()
13604
    attrs = {'seed': seed, 'min': min, 'max': max}
13605
    if in_dygraph_mode():
H
hong 已提交
13606
        attrs['shape'] = shape
13607 13608 13609 13610 13611 13612 13613 13614 13615 13616 13617 13618 13619 13620 13621 13622 13623
    else:
        if isinstance(shape, Variable):
            shape.stop_gradient = True
            inputs["ShapeTensor"] = shape
        elif isinstance(shape, (list, tuple)):
            assert len(shape) > 0, (
                "The size of argument(shape) can't be zero.")
            attrs["shape"] = get_attr_shape(shape)
            if contain_var(shape):
                inputs['ShapeTensorList'] = get_new_shape_tensor(shape)

    out = helper.create_variable_for_type_inference(dtype)
    helper.append_op(
        type="uniform_random", inputs=inputs, attrs=attrs,
        outputs={"Out": out})

    return helper.append_activation(out)