nn.py 508.9 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
Y
Yu Yang 已提交
37 38

__all__ = [
X
Xin Pan 已提交
39
    'fc',
H
HaoRen 已提交
40
    'center_loss',
X
Xin Pan 已提交
41 42 43 44 45 46 47 48 49
    'embedding',
    'dynamic_lstm',
    'dynamic_lstmp',
    'dynamic_gru',
    'gru_unit',
    'linear_chain_crf',
    'crf_decoding',
    'cos_sim',
    'cross_entropy',
50
    'bpr_loss',
X
Xin Pan 已提交
51 52 53 54 55 56 57 58 59 60
    'square_error_cost',
    'chunk_eval',
    'sequence_conv',
    'conv2d',
    'conv3d',
    'sequence_pool',
    'sequence_softmax',
    'softmax',
    'pool2d',
    'pool3d',
61 62
    'adaptive_pool2d',
    'adaptive_pool3d',
X
Xin Pan 已提交
63
    'batch_norm',
H
heqiaozhi 已提交
64
    'data_norm',
X
Xin Pan 已提交
65 66 67 68 69 70
    'beam_search_decode',
    'conv2d_transpose',
    'conv3d_transpose',
    'sequence_expand',
    'sequence_expand_as',
    'sequence_pad',
Y
Yibing Liu 已提交
71
    'sequence_unpad',
X
Xin Pan 已提交
72 73 74 75 76 77
    'lstm_unit',
    'reduce_sum',
    'reduce_mean',
    'reduce_max',
    'reduce_min',
    'reduce_prod',
Z
zhoukunsheng 已提交
78 79
    'reduce_all',
    'reduce_any',
X
Xin Pan 已提交
80 81
    'sequence_first_step',
    'sequence_last_step',
Y
Yibing Liu 已提交
82
    'sequence_slice',
X
Xin Pan 已提交
83 84 85 86 87 88 89 90 91 92 93 94
    'dropout',
    'split',
    'ctc_greedy_decoder',
    'edit_distance',
    'l2_normalize',
    'matmul',
    'topk',
    'warpctc',
    'sequence_reshape',
    'transpose',
    'im2sequence',
    'nce',
95
    'sampled_softmax_with_cross_entropy',
X
Xin Pan 已提交
96 97 98 99 100
    'hsigmoid',
    'beam_search',
    'row_conv',
    'multiplex',
    'layer_norm',
D
Dun 已提交
101
    'group_norm',
D
dengkaipeng 已提交
102
    'spectral_norm',
X
Xin Pan 已提交
103 104 105 106 107 108 109 110
    'softmax_with_cross_entropy',
    'smooth_l1',
    'one_hot',
    'autoincreased_step_counter',
    'reshape',
    'squeeze',
    'unsqueeze',
    'lod_reset',
111
    'lod_append',
X
Xin Pan 已提交
112 113 114 115 116
    'lrn',
    'pad',
    'pad_constant_like',
    'label_smooth',
    'roi_pool',
J
jerrywgz 已提交
117
    'roi_align',
X
Xin Pan 已提交
118 119 120 121
    'dice_loss',
    'image_resize',
    'image_resize_short',
    'resize_bilinear',
K
Kaipeng Deng 已提交
122
    'resize_trilinear',
123
    'resize_nearest',
X
Xin Pan 已提交
124
    'gather',
125
    'gather_nd',
X
Xin Pan 已提交
126
    'scatter',
127 128
    'scatter_nd_add',
    'scatter_nd',
X
Xin Pan 已提交
129 130 131 132
    'sequence_scatter',
    'random_crop',
    'mean_iou',
    'relu',
C
chengduo 已提交
133
    'selu',
X
Xin Pan 已提交
134 135 136
    'log',
    'crop',
    'rank_loss',
M
minqiyang 已提交
137
    'margin_rank_loss',
X
Xin Pan 已提交
138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153
    'elu',
    'relu6',
    'pow',
    'stanh',
    'hard_sigmoid',
    'swish',
    'prelu',
    'brelu',
    'leaky_relu',
    'soft_relu',
    'flatten',
    'sequence_mask',
    'stack',
    'pad2d',
    'unstack',
    'sequence_enumerate',
Z
zhoukunsheng 已提交
154
    'unique',
155
    'unique_with_counts',
X
Xin Pan 已提交
156 157 158 159 160 161 162 163 164 165
    'expand',
    'sequence_concat',
    'scale',
    'elementwise_add',
    'elementwise_div',
    'elementwise_sub',
    'elementwise_mul',
    'elementwise_max',
    'elementwise_min',
    'elementwise_pow',
Z
zhoukunsheng 已提交
166 167
    'elementwise_mod',
    'elementwise_floordiv',
X
Xin Pan 已提交
168 169 170 171 172 173 174
    'uniform_random_batch_size_like',
    'gaussian_random',
    'sampling_id',
    'gaussian_random_batch_size_like',
    'sum',
    'slice',
    'shape',
Z
zhoukunsheng 已提交
175
    'rank',
Z
zhoukunsheng 已提交
176
    'size',
X
Xin Pan 已提交
177 178 179 180 181 182 183 184 185 186
    'logical_and',
    'logical_or',
    'logical_xor',
    'logical_not',
    'clip',
    'clip_by_norm',
    'mean',
    'mul',
    'sigmoid_cross_entropy_with_logits',
    'maxout',
J
JiabinYang 已提交
187
    'space_to_depth',
W
whs 已提交
188
    'affine_grid',
S
sneaxiy 已提交
189
    'sequence_reverse',
190
    'sequence_topk_avg_pooling',
191
    'affine_channel',
B
barrierye 已提交
192
    'similarity_focus',
M
minqiyang 已提交
193
    'hash',
D
dengkaipeng 已提交
194
    'grid_sampler',
G
gmcather 已提交
195 196
    'log_loss',
    'add_position_encoding',
Q
Qiao Longfei 已提交
197
    'bilinear_tensor_product',
C
chengduo 已提交
198 199
    'merge_selected_rows',
    'get_tensor_from_selected_rows',
P
phlrain 已提交
200
    'lstm',
S
shippingwang 已提交
201
    'shuffle_channel',
202
    'temporal_shift',
S
sneaxiy 已提交
203
    'py_func',
204
    'psroi_pool',
H
heqiaozhi 已提交
205
    'teacher_student_sigmoid_loss',
M
minqiyang 已提交
206
    'huber_loss',
D
dengkaipeng 已提交
207
    'kldiv_loss',
Z
zhaozhehao 已提交
208
    'tree_conv',
C
ceci3 已提交
209
    'npair_loss',
R
ruri 已提交
210
    'pixel_shuffle',
211
    'fsp_matrix',
H
heqiaozhi 已提交
212
    'continuous_value_model',
Z
zhoukunsheng 已提交
213
    'where',
Z
zhoukunsheng 已提交
214
    'sign',
215
    'deformable_conv',
216
    'unfold',
C
cjt222 已提交
217
    'deformable_roi_pooling',
A
Aurelius84 已提交
218
    'match_matrix_tensor',
J
Jiawei Wang 已提交
219
    'filter_by_instag',
K
Kevin 已提交
220
    'var_conv_2d',
221
    'shard_index',
H
huangjun12 已提交
222
    'hard_swish',
Y
Yu Yang 已提交
223 224
]

J
jerrywgz 已提交
225 226
kIgnoreIndex = -100

Y
Yu Yang 已提交
227 228 229 230 231 232 233

def fc(input,
       size,
       num_flatten_dims=1,
       param_attr=None,
       bias_attr=None,
       act=None,
234
       name=None):
Y
Yu Yang 已提交
235
    """
236
    **Fully Connected Layer**
Y
Yu Yang 已提交
237

238
    This function creates a fully connected layer in the network. It can take
239
    one or multiple tensors as its inputs(input can be a list of Variable, see
A
Aurelius84 已提交
240
    Args in detail). It creates a variable called weights for each input tensor,
241 242 243 244
    which represents a fully connected weight matrix from each input unit to
    each output unit. The fully connected layer multiplies each input tensor
    with its corresponding weight to produce an output Tensor with shape [M, `size`],
    where M is batch size. If multiple input tensors are given, the results of
A
Aurelius84 已提交
245
    multiple output tensors with shape [M, `size`] will be summed up. If bias_attr
246 247
    is not None, a bias variable will be created and added to the output.
    Finally, if activation is not None, it will be applied to the output as well.
C
caoying03 已提交
248

249
    When the input is single tensor:
C
caoying03 已提交
250

251 252 253 254 255
    .. math::

        Out = Act({XW + b})

    When the input are multiple tensors:
256 257 258

    .. math::

259
        Out = Act({\sum_{i=0}^{N-1}X_iW_i + b})
260 261 262

    In the above equation:

263 264 265
    * :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
266
    * :math:`b`: The bias parameter created by this layer (if needed).
267
    * :math:`Act`: The activation function.
C
caoying03 已提交
268
    * :math:`Out`: The output tensor.
Y
Yu Yang 已提交
269

270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287
    See below for an example.

    .. code-block:: text

        Given:
            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
288
    Args:
R
ranqiu 已提交
289 290 291 292 293 294 295 296 297 298
        input (Variable|list of Variable): The input tensor(s) of this layer, and the dimension of
            the input tensor(s) is at least 2.
        size(int): The number of output units in this layer.
        num_flatten_dims (int, default 1): The fc layer can accept an input tensor with more than
            two dimensions. If this happens, the multidimensional tensor will first be flattened
            into a 2-dimensional matrix. The parameter `num_flatten_dims` determines how the input
            tensor is flattened: the first `num_flatten_dims` (inclusive, index starts from 1)
            dimensions will be flatten to form the first dimension of the final matrix (height of
            the matrix), and the rest `rank(X) - num_flatten_dims` dimensions are flattened to
            form the second dimension of the final matrix (width of the matrix). For example, suppose
H
haowang101779990 已提交
299
            `X` is a 5-dimensional tensor with a shape [2, 3, 4, 5, 6], and `num_flatten_dims` = 3.
R
ranqiu 已提交
300 301 302 303
            Then, the flattened matrix will have a shape [2 x 3 x 4, 5 x 6] = [24, 30].
        param_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for learnable
            parameters/weights of this layer.
        bias_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for the bias
304 305
            of this layer. If it is set to False, no bias will be added to the output units.
            If it is set to None, the bias is initialized zero. Default: None.
R
ranqiu 已提交
306 307
        act (str, default None): Activation to be applied to the output of this layer.
        name (str, default None): The name of this layer.
Y
Yu Yang 已提交
308

309
    Returns:
F
fengjiayi 已提交
310
        Variable: The transformation result.
311 312

    Raises:
C
caoying03 已提交
313
        ValueError: If rank of the input tensor is less than 2.
314 315 316 317

    Examples:
        .. code-block:: python

318
          import paddle.fluid as fluid
319
          # when input is single tensor
F
fengjiayi 已提交
320
          data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
321
          fc = fluid.layers.fc(input=data, size=1000, act="tanh")
322 323 324 325 326

          # when input are multiple tensors
          data_1 = fluid.layers.data(name="data_1", shape=[32, 32], dtype="float32")
          data_2 = fluid.layers.data(name="data_2", shape=[24, 36], dtype="float32")
          fc = fluid.layers.fc(input=[data_1, data_2], size=1000, act="tanh")
Y
Yu Yang 已提交
327
    """
C
caoying03 已提交
328
    helper = LayerHelper("fc", **locals())
Y
Yu Yang 已提交
329 330 331 332

    dtype = helper.input_dtype()

    mul_results = []
333 334
    for input_var, param_attr in helper.iter_inputs_and_params():
        input_shape = input_var.shape
Y
Yu Yang 已提交
335 336 337
        param_shape = [
            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
        ] + [size]
Y
ying 已提交
338

Y
Yu Yang 已提交
339
        w = helper.create_parameter(
340
            attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False)
X
Xin Pan 已提交
341
        tmp = helper.create_variable_for_type_inference(dtype)
342
        helper.append_op(
343 344 345
            type="mul",
            inputs={"X": input_var,
                    "Y": w},
346
            outputs={"Out": tmp},
M
mozga-intel 已提交
347 348
            attrs={"x_num_col_dims": num_flatten_dims,
                   "y_num_col_dims": 1})
349 350 351 352
        mul_results.append(tmp)

    if len(mul_results) == 1:
        pre_bias = mul_results[0]
353
    else:
X
Xin Pan 已提交
354
        pre_bias = helper.create_variable_for_type_inference(dtype)
355
        helper.append_op(
356 357 358
            type="sum",
            inputs={"X": mul_results},
            outputs={"Out": pre_bias},
X
Xin Pan 已提交
359
            attrs={"use_mkldnn": False})
360 361 362 363
    # 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
364 365


H
HaoRen 已提交
366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451
def center_loss(input,
                label,
                num_classes,
                alpha,
                param_attr,
                update_center=True):
    """
    **Center loss Cost layer**
    
    This layer accepts input (deep features,the output of the last hidden layer)
    and target label and return the center loss cost
    
    For deep features, :math:`X`, and target labels, :math:`Y`, the equation is:
    
    .. math::

        Out = \\frac{1}{2}(X - Y)^2

    Args:
        input (Variable): a 2-D tensor with shape[N x M].
        label (Variable): the groud truth which is a 2-D tensor
                         with shape[N x 1],where N is the batch size.
        num_classes (int): the number of classification categories.
        alpha (float|Variable): learning rate of centers.
        param_attr (ParamAttr): Attribute initializer of centers. 
        update_center (bool): whether to update value of center.

    Returns:
        Variable: 2-D tensor with shape [N * 1] 

    Examples:
        .. code-block:: python

          import paddle.fluid as fluid 

          input = fluid.layers.data(name='x',shape=[20,30],dtype='float32')
          label = fluid.layers.data(name='y',shape=[20,1],dtype='int64')
          num_classes = 1000
          alpha = 0.01
          param_attr = fluid.initializer.Xavier(uniform=False)
          center_loss=fluid.layers.center_loss(input=input,
                 label=label,
                 num_classes=1000,
                 alpha=alpha,
                 param_attr=fluid.initializer.Xavier(uniform=False),
                 update_center=True)
    """
    helper = LayerHelper('center_loss', **locals())
    dtype = helper.input_dtype()
    centers_shape = [num_classes, input.shape[1]]
    centers_param = helper.create_parameter(
        attr=param_attr, shape=centers_shape, dtype=dtype)
    centers_param.stop_gradient = True
    if isinstance(alpha, Variable):
        alpha_param = alpha
    else:
        assert isinstance(alpha, float)
        alpha_param = helper.create_variable(
            name="centerloss_alpha",
            shape=[1],
            dtype="float32",
            type=core.VarDesc.VarType.LOD_TENSOR,
            persistable=True,
            stop_gradient=True,
            initializer=Constant(alpha))

    centersdiff = helper.create_variable_for_type_inference(dtype=input.dtype)
    loss = helper.create_variable_for_type_inference(dtype=input.dtype)
    helper.append_op(
        type='center_loss',
        inputs={
            'X': [input],
            'Label': [label],
            'Centers': [centers_param],
            'CenterUpdateRate': [alpha_param]
        },
        outputs={
            'SampleCenterDiff': [centersdiff],
            'Loss': [loss],
            'CentersOut': [centers_param]
        },
        attrs={'cluster_num': num_classes,
               'need_update': update_center})
    return loss


452 453 454
def embedding(input,
              size,
              is_sparse=False,
455
              is_distributed=False,
456 457 458
              padding_idx=None,
              param_attr=None,
              dtype='float32'):
Y
Yu Yang 已提交
459
    """
460 461
    **Embedding Layer**

462
    This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in
463 464
    a lookup table. The result of this lookup is the embedding of each ID in the
    :attr:`input`.
465 466 467

    All the input variables are passed in as local variables to the LayerHelper
    constructor.
Y
Yu Yang 已提交
468 469

    Args:
470
        input(Variable): Input is a Tensor<int64> Variable, which contains the IDs information.
K
Kevin 已提交
471
            The value of the input IDs should satisfy :math:`0<= id < size[0]`.
472 473 474 475
        size(tuple|list): The shape of the look up table parameter. It should
            have two elements which indicate 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.
476
        is_distributed(bool): Whether to run lookup table from remote parameter server.
K
Kevin 已提交
477 478 479 480 481 482 483 484
        padding_idx(int|long|None): It will output all-zero padding data whenever
            lookup encounters :math:`padding\_idx` in Ids. If set :attr:`None`, it makes
            no effect to output. If :math:`padding\_idx < 0`, the :math:`padding\_idx`
            will automatically be converted to :math:`size[0] + padding\_idx` to use.
            Default: None.
        param_attr(ParamAttr): Parameters for this layer.
        dtype(np.dtype|core.VarDesc.VarType|str): The dtype refers to the data type of output
            tensor. It can be float32, float_16, int etc.
Y
Yu Yang 已提交
485

486 487 488
    Returns:
        Variable: The tensor variable storing the embeddings of the \
                  supplied inputs.
Y
Yu Yang 已提交
489

490 491
    Examples:
        .. code-block:: python
Y
Yu Yang 已提交
492

B
bdzhuxiaoning 已提交
493 494 495
          import paddle.fluid as fluid
          data = fluid.layers.data(name='sequence', shape=[1], dtype='int64', lod_level=1)
          emb = fluid.layers.embedding(input=data, size=[128, 64])    
Y
Yu Yang 已提交
496 497 498
    """

    helper = LayerHelper('embedding', **locals())
499
    remote_prefetch = is_sparse and (not is_distributed)
Q
Qiao Longfei 已提交
500 501
    if remote_prefetch:
        assert is_sparse is True and is_distributed is False
Y
Yu Yang 已提交
502 503
    w = helper.create_parameter(
        attr=helper.param_attr, shape=size, dtype=dtype, is_bias=False)
X
Xin Pan 已提交
504
    tmp = helper.create_variable_for_type_inference(dtype)
505 506
    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
507 508 509 510 511
    helper.append_op(
        type='lookup_table',
        inputs={'Ids': input,
                'W': w},
        outputs={'Out': tmp},
512 513 514
        attrs={
            'is_sparse': is_sparse,
            'is_distributed': is_distributed,
Q
Qiao Longfei 已提交
515
            'remote_prefetch': remote_prefetch,
516 517
            'padding_idx': padding_idx
        })
Y
Yu Yang 已提交
518 519 520
    return tmp


H
hutuxian 已提交
521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568
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


W
wopeizl 已提交
569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584
@templatedoc(op_type="lstm")
def dynamic_lstm(input,
                 size,
                 h_0=None,
                 c_0=None,
                 param_attr=None,
                 bias_attr=None,
                 use_peepholes=True,
                 is_reverse=False,
                 gate_activation='sigmoid',
                 cell_activation='tanh',
                 candidate_activation='tanh',
                 dtype='float32',
                 name=None):
    """
    ${comment}
Y
Yibing Liu 已提交
585

W
wopeizl 已提交
586 587 588 589 590 591 592 593 594 595 596
    Args:
        input (Variable): ${input_comment}
        size (int): 4 * hidden size.
        h_0(Variable): The initial hidden state is an optional input, default is zero.
                       This is a tensor with shape (N x D), where N is the
                       batch size and D is the hidden size.
        c_0(Variable): The initial cell state is an optional input, default is zero.
                       This is a tensor with shape (N x D), where N is the
                       batch size. `h_0` and `c_0` can be NULL but only at the same time.
        param_attr(ParamAttr|None): The parameter attribute for the learnable
                               hidden-hidden weights.
Y
Yu Yang 已提交
597

W
wopeizl 已提交
598 599 600 601
                               - Weights = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}
                               - The shape is (D x 4D), where D is the hidden
                                 size.
Y
Yu Yang 已提交
602

W
wopeizl 已提交
603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638
                               If it is set to None or one attribute of ParamAttr,
                               dynamic_lstm will create ParamAttr as param_attr.
                               If the Initializer of the param_attr is not set, the
                               parameter is initialized with Xavier. Default: None.
        bias_attr (ParamAttr|None): The bias attribute for the learnable bias
                              weights, which contains two parts, input-hidden
                              bias weights and peephole connections weights if
                              setting `use_peepholes` to `True`.

                              1. `use_peepholes = False`
                                 - Biases = {:math:`b_c, b_i, b_f, b_o`}.
                                 - The shape is (1 x 4D).
                              2. `use_peepholes = True`
                                 - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
                                                 W_{fc}, W_{oc}`}.
                                 - The shape is (1 x 7D).

                              If it is set to None or one attribute of ParamAttr,
                              dynamic_lstm will create ParamAttr as bias_attr.
                              If the Initializer of the bias_attr is not set,
                              the bias is initialized zero. Default: None.
        use_peepholes (bool): ${use_peepholes_comment}
        is_reverse (bool): ${is_reverse_comment}
        gate_activation (str): ${gate_activation_comment}
        cell_activation (str): ${cell_activation_comment}
        candidate_activation (str): ${candidate_activation_comment}
        dtype (str): Data type. Choices = ["float32", "float64"], default "float32".
        name (str|None): A name for this layer(optional). If set None, the layer
                         will be named automatically.

    Returns:
        tuple: The hidden state, and cell state of LSTM. The shape of both \
        is (T x D), and lod is the same with the `input`.

    Examples:
        .. code-block:: python
639
            
640
            import paddle.fluid as fluid
641 642
            emb_dim = 256
            vocab_size = 10000
W
wopeizl 已提交
643
            hidden_dim = 512
644 645 646 647 648 649
            
            data = fluid.layers.data(name='x', shape=[1],
                         dtype='int32', lod_level=1)
            emb = fluid.layers.embedding(input=data, size=[vocab_size, emb_dim], is_sparse=True)

            forward_proj = fluid.layers.fc(input=emb, size=hidden_dim * 4,
W
wopeizl 已提交
650
                                           bias_attr=False)
651

W
wopeizl 已提交
652 653 654
            forward, _ = fluid.layers.dynamic_lstm(
                input=forward_proj, size=hidden_dim * 4, use_peepholes=False)
    """
L
lujun 已提交
655
    assert in_dygraph_mode(
656
    ) is not True, "please use lstm instead of dynamic_lstm in dygraph mode!"
W
wopeizl 已提交
657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699
    assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp."
    helper = LayerHelper('lstm', **locals())
    size = size // 4
    weight = helper.create_parameter(
        attr=helper.param_attr, shape=[size, 4 * size], dtype=dtype)
    bias_size = [1, 7 * size]
    if not use_peepholes:
        bias_size[1] = 4 * size
    bias = helper.create_parameter(
        attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)

    hidden = helper.create_variable_for_type_inference(dtype)
    cell = helper.create_variable_for_type_inference(dtype)
    batch_gate = helper.create_variable_for_type_inference(dtype)
    batch_cell_pre_act = helper.create_variable_for_type_inference(dtype)
    inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
    batch_size = input.shape[0]
    if h_0:
        assert h_0.shape == (batch_size, size), \
            'The shape of h0 should be (batch_size, %d)' % size
        inputs['H0'] = h_0
    if c_0:
        assert c_0.shape == (batch_size, size), \
            'The shape of c0 should be (batch_size, %d)' % size
        inputs['C0'] = c_0

    helper.append_op(
        type='lstm',
        inputs=inputs,
        outputs={
            'Hidden': hidden,
            'Cell': cell,
            'BatchGate': batch_gate,
            'BatchCellPreAct': batch_cell_pre_act
        },
        attrs={
            'use_peepholes': use_peepholes,
            'is_reverse': is_reverse,
            'gate_activation': gate_activation,
            'cell_activation': cell_activation,
            'candidate_activation': candidate_activation
        })
    return hidden, cell
Y
Yu Yang 已提交
700 701


P
phlrain 已提交
702 703 704 705 706 707
def lstm(input,
         init_h,
         init_c,
         max_len,
         hidden_size,
         num_layers,
P
phlrain 已提交
708
         dropout_prob=0.0,
P
phlrain 已提交
709 710 711 712 713
         is_bidirec=False,
         is_test=False,
         name=None,
         default_initializer=None,
         seed=-1):
L
liuhongyu 已提交
714
    """
P
phlrain 已提交
715
    If Device is GPU, This op will use cudnn LSTM implementation
L
liuhongyu 已提交
716 717

    A four-gate Long Short-Term Memory network with no peephole connections.
M
minqiyang 已提交
718
    In the forward pass the output ht and cell output ct for a given iteration can be computed from the recurrent input ht-1,
L
liuhongyu 已提交
719 720
    the cell input ct-1 and the previous layer input xt given matrices W, R and biases bW, bR from the following equations:

H
haowang101779990 已提交
721
    .. math::
M
minqiyang 已提交
722 723 724 725 726 727 728

       i_t &= \sigma(W_{ix}x_{t} + W_{ih}h_{t-1} + bx_i + bh_i)

       f_t &= \sigma(W_{fx}x_{t} + W_{fh}h_{t-1} + bx_f + bh_f)

       o_t &= \sigma(W_{ox}x_{t} + W_{oh}h_{t-1} + bx_o + bh_o)

H
haowang101779990 已提交
729
       \\tilde{c_t} &= tanh(W_{cx}x_t + W_{ch}h_{t-1} + bx_c + bh_c)
M
minqiyang 已提交
730 731 732 733

       c_t &= f_t \odot c_{t-1} + i_t \odot \\tilde{c_t}

       h_t &= o_t \odot tanh(c_t)
H
haowang101779990 已提交
734 735

    - $W$ terms denote weight matrices (e.g. $W_{ix}$ is the matrix
P
phlrain 已提交
736 737 738 739 740 741
      of weights from the input gate to the input)
    - The b terms denote bias vectors ($bx_i$ and $bh_i$ are the input gate bias vector).
    - sigmoid is the logistic sigmoid function.
    - $i, f, o$ and $c$ are the input gate, forget gate, output gate,
      and cell activation vectors, respectively, all of which have the same size as
      the cell output activation vector $h$.
H
haowang101779990 已提交
742 743 744
    - The :math:`\odot` is the element-wise product of the vectors.
    - :math:`tanh` is the activation functions.
    - :math:`\\tilde{c_t}` is also called candidate hidden state,
P
phlrain 已提交
745
      which is computed based on the current input and the previous hidden state.
L
liuhongyu 已提交
746

M
minqiyang 已提交
747
    Where sigmoid is the sigmoid operator: :math:`sigmoid(x) = 1 / (1 + e^{-x})` , * represents a point-wise multiplication,
L
liuhongyu 已提交
748 749 750 751 752
    X represensts a matrix multiplication


    Args:
        input (Variable): LSTM input tensor, shape MUST be ( seq_len x batch_size x input_size )
M
minqiyang 已提交
753
        init_h(Variable): The initial hidden state of the LSTM
L
liuhongyu 已提交
754 755 756 757 758
                       This is a tensor with shape ( num_layers x batch_size x hidden_size)
                       if is_bidirec = True, shape should be ( num_layers*2 x batch_size x hidden_size)
        init_c(Variable): The initial cell state of the LSTM.
                       This is a tensor with shape ( num_layers x batch_size x hidden_size )
                       if is_bidirec = True, shape should be ( num_layers*2 x batch_size x hidden_size)
M
minqiyang 已提交
759
        max_len (int): max length of LSTM. the first dim of input tensor CAN NOT greater than max_len
L
liuhongyu 已提交
760 761
        hidden_size (int): hidden size of the LSTM
        num_layers (int): total layers number of the LSTM
P
phlrain 已提交
762 763
        dropout_prob(float|0.0): dropout prob, dropout ONLY work between rnn layers, NOT between time steps
                             There is NO dropout work on rnn output of the last RNN layers
L
liuhongyu 已提交
764 765 766 767 768 769
        is_bidirec (bool): If it is bidirectional
        is_test (bool): If it is in test phrase
        name (str|None): A name for this layer(optional). If set None, the layer
                         will be named automatically.
        default_initializer(Initialize|None): Where use initializer to initialize the Weight
                         If set None, defaule initializer will be used
P
phlrain 已提交
770
        seed(int): Seed for dropout in LSTM, If it's -1, dropout will use random seed
P
phlrain 已提交
771

L
liuhongyu 已提交
772 773

    Returns:
M
minqiyang 已提交
774 775
        rnn_out(Tensor),last_h(Tensor),last_c(Tensor):

H
haowang101779990 已提交
776
                        Three tensors, rnn_out, last_h, last_c:
M
minqiyang 已提交
777

H
haowang101779990 已提交
778 779 780 781
                        - rnn_out is result of LSTM hidden, shape is (seq_len x batch_size x hidden_size) \
                          if is_bidirec set to True, shape will be ( seq_len x batch_sze x hidden_size*2)
                        - last_h is the hidden state of the last step of LSTM \
                          shape is ( num_layers x batch_size x hidden_size ) \
M
minqiyang 已提交
782
                          if is_bidirec set to True, shape will be ( num_layers*2 x batch_size x hidden_size)
H
haowang101779990 已提交
783 784
                        - last_c(Tensor): the cell state of the last step of LSTM \
                          shape is ( num_layers x batch_size x hidden_size ) \
M
minqiyang 已提交
785
                          if is_bidirec set to True, shape will be ( num_layers*2 x batch_size x hidden_size)
L
liuhongyu 已提交
786 787 788 789


    Examples:
        .. code-block:: python
790
            
791 792 793
            import paddle.fluid as fluid
            import paddle.fluid.layers as layers

794 795 796 797 798
            emb_dim = 256
            vocab_size = 10000
            data = fluid.layers.data(name='x', shape=[-1, 100, 1],
                         dtype='int32')
            emb = fluid.layers.embedding(input=data, size=[vocab_size, emb_dim], is_sparse=True)
L
liuhongyu 已提交
799 800 801 802 803 804
            batch_size = 20
            max_len = 100
            dropout_prob = 0.2
            input_size = 100
            hidden_size = 150
            num_layers = 1
805 806 807 808 809
            init_h = layers.fill_constant( [num_layers, batch_size, hidden_size], 'float32', 0.0 )
            init_c = layers.fill_constant( [num_layers, batch_size, hidden_size], 'float32', 0.0 )
            rnn_out, last_h, last_c = layers.lstm( emb, init_h, init_c, \
                    max_len, hidden_size, num_layers, \
                    dropout_prob=dropout_prob)
L
liuhongyu 已提交
810 811 812 813
    """

    helper = LayerHelper('cudnn_lstm', **locals())

P
phlrain 已提交
814 815 816
    dtype = input.dtype
    input_shape = list(input.shape)
    input_size = input_shape[-1]
L
liuhongyu 已提交
817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875
    weight_size = 0
    for i in range(num_layers):
        if i == 0:
            input_weight_size = (input_size * hidden_size) * 4
        else:
            if is_bidirec:
                input_weight_size = (hidden_size * 2 * hidden_size) * 4
            else:
                input_weight_size = (hidden_size * hidden_size) * 4

        hidden_weight_size = (hidden_size * hidden_size) * 4

        if is_bidirec:
            weight_size += (input_weight_size + hidden_weight_size) * 2
            weight_size += hidden_size * 8 * 2
        else:
            weight_size += input_weight_size + hidden_weight_size
            weight_size += hidden_size * 8

    weight = helper.create_parameter(
        attr=helper.param_attr,
        shape=[weight_size],
        dtype=dtype,
        default_initializer=default_initializer)

    out = helper.create_variable_for_type_inference(dtype)
    last_h = helper.create_variable_for_type_inference(dtype)
    last_c = helper.create_variable_for_type_inference(dtype)

    cache = helper.create_variable(
        persistable=True, type=core.VarDesc.VarType.RAW, stop_gradient=True)

    helper.append_op(
        type='cudnn_lstm',
        inputs={
            'Input': input,
            'InitH': init_h,
            'InitC': init_c,
            'W': weight,
            'Cache': cache,
        },
        outputs={
            'Out': out,
            'last_h': last_h,
            'last_c': last_c,
        },
        attrs={
            'max_len': max_len,
            'is_bidirec': is_bidirec,
            'input_size': input_size,
            'hidden_size': hidden_size,
            'num_layers': num_layers,
            'is_test': is_test,
            'dropout_prob': dropout_prob,
            'seed': seed,
        })
    return out, last_h, last_c


Y
Yibing Liu 已提交
876 877 878 879 880 881 882 883 884 885
def dynamic_lstmp(input,
                  size,
                  proj_size,
                  param_attr=None,
                  bias_attr=None,
                  use_peepholes=True,
                  is_reverse=False,
                  gate_activation='sigmoid',
                  cell_activation='tanh',
                  candidate_activation='tanh',
X
xuezhong 已提交
886
                  proj_activation='tanh',
887
                  dtype='float32',
X
xuezhong 已提交
888 889 890 891 892
                  name=None,
                  h_0=None,
                  c_0=None,
                  cell_clip=None,
                  proj_clip=None):
Y
Yibing Liu 已提交
893 894 895
    """
    **Dynamic LSTMP Layer**

896 897 898 899 900 901
    LSTMP (LSTM with recurrent projection) layer has a separate projection
    layer after the LSTM layer, projecting the original hidden state to a
    lower-dimensional one, which is proposed to reduce the number of total
    parameters and furthermore computational complexity for the LSTM,
    espeacially for the case that the size of output units is relative
    large (https://research.google.com/pubs/archive/43905.pdf).
Y
Yibing Liu 已提交
902 903 904 905 906

    The formula is as follows:

    .. math::

907
        i_t & = \sigma(W_{ix}x_{t} + W_{ir}r_{t-1} + W_{ic}c_{t-1} + b_i)
Y
Yibing Liu 已提交
908

909
        f_t & = \sigma(W_{fx}x_{t} + W_{fr}r_{t-1} + W_{fc}c_{t-1} + b_f)
Y
Yibing Liu 已提交
910

911
        \\tilde{c_t} & = act_g(W_{cx}x_t + W_{cr}r_{t-1} + b_c)
Y
Yibing Liu 已提交
912

913
        o_t & = \sigma(W_{ox}x_{t} + W_{or}r_{t-1} + W_{oc}c_t + b_o)
Y
Yibing Liu 已提交
914

915
        c_t & = f_t \odot c_{t-1} + i_t \odot \\tilde{c_t}
Y
Yibing Liu 已提交
916

917
        h_t & = o_t \odot act_h(c_t)
Y
Yibing Liu 已提交
918

919
        r_t & = \overline{act_h}(W_{rh}h_t)
Y
Yibing Liu 已提交
920

Y
Yibing Liu 已提交
921 922 923 924 925 926
    In the above formula:

    * :math:`W`: Denotes weight matrices (e.g. :math:`W_{xi}` is \
          the matrix of weights from the input gate to the input).
    * :math:`W_{ic}`, :math:`W_{fc}`, :math:`W_{oc}`: Diagonal weight \
          matrices for peephole connections. In our implementation, \
翟飞跃 已提交
927
          we use vectors to represent these diagonal weight matrices.
Y
Yibing Liu 已提交
928
    * :math:`b`: Denotes bias vectors (e.g. :math:`b_i` is the input gate \
929
          bias vector).
Y
Yibing Liu 已提交
930 931 932
    * :math:`\sigma`: The activation, such as logistic sigmoid function.
    * :math:`i, f, o` and :math:`c`: The input gate, forget gate, output \
          gate, and cell activation vectors, respectively, all of which have \
933
          the same size as the cell output activation vector :math:`h`.
Y
Yibing Liu 已提交
934
    * :math:`h`: The hidden state.
935
    * :math:`r`: The recurrent projection of the hidden state.
Y
Yibing Liu 已提交
936 937
    * :math:`\\tilde{c_t}`: The candidate hidden state, whose \
          computation is based on the current input and previous hidden state.
938
    * :math:`\odot`: The element-wise product of the vectors.
Y
Yibing Liu 已提交
939
    * :math:`act_g` and :math:`act_h`: The cell input and cell output \
940
          activation functions and `tanh` is usually used for them.
Y
Yibing Liu 已提交
941 942
    * :math:`\overline{act_h}`: The activation function for the projection \
          output, usually using `identity` or same as :math:`act_h`.
Y
Yibing Liu 已提交
943 944 945 946

    Set `use_peepholes` to `False` to disable peephole connection. The formula
    is omitted here, please refer to the paper
    http://www.bioinf.jku.at/publications/older/2604.pdf for details.
947

Y
Yibing Liu 已提交
948 949 950 951 952 953 954 955 956 957 958 959
    Note that these :math:`W_{xi}x_{t}, W_{xf}x_{t}, W_{xc}x_{t}, W_{xo}x_{t}`
    operations on the input :math:`x_{t}` are NOT included in this operator.
    Users can choose to use fully-connected layer before LSTMP layer.

    Args:
        input(Variable): The input of dynamic_lstmp layer, which supports
                         variable-time length input sequence. The underlying
                         tensor in this Variable is a matrix with shape
                         (T X 4D), where T is the total time steps in this
                         mini-batch, D is the hidden size.
        size(int): 4 * hidden size.
        proj_size(int): The size of projection output.
960
        param_attr(ParamAttr|None): The parameter attribute for the learnable
Y
Yibing Liu 已提交
961 962
                               hidden-hidden weight and projection weight.

963 964
                               - Hidden-hidden weight = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}.
965 966
                               - The shape of hidden-hidden weight is (P x 4D),
                                 where P is the projection size and D the hidden
Y
Yibing Liu 已提交
967 968
                                 size.
                               - Projection weight = {:math:`W_{rh}`}.
969
                               - The shape of projection weight is (D x P).
C
chengduo 已提交
970 971 972 973 974

                               If it is set to None or one attribute of ParamAttr,
                               dynamic_lstm will create ParamAttr as param_attr.
                               If the Initializer of the param_attr is not set, the
                               parameter is initialized with Xavier. Default: None.
975
        bias_attr(ParamAttr|None): The bias attribute for the learnable bias
Y
Yibing Liu 已提交
976 977 978 979 980 981
                              weights, which contains two parts, input-hidden
                              bias weights and peephole connections weights if
                              setting `use_peepholes` to `True`.

                              1. `use_peepholes = False`
                                - Biases = {:math:`b_c, b_i, b_f, b_o`}.
982
                                - The shape is (1 x 4D).
Y
Yibing Liu 已提交
983 984 985
                              2. `use_peepholes = True`
                                - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
                                                 W_{fc}, W_{oc}`}.
986
                                - The shape is (1 x 7D).
C
chengduo 已提交
987 988 989 990 991

                              If it is set to None or one attribute of ParamAttr,
                              dynamic_lstm will create ParamAttr as bias_attr.
                              If the Initializer of the bias_attr is not set,
                              the bias is initialized zero. Default: None.
Y
Yibing Liu 已提交
992 993 994 995 996 997 998 999 1000
        use_peepholes(bool): Whether to enable diagonal/peephole connections,
                             default `True`.
        is_reverse(bool): Whether to compute reversed LSTM, default `False`.
        gate_activation(str): The activation for input gate, forget gate and
                              output gate. Choices = ["sigmoid", "tanh", "relu",
                              "identity"], default "sigmoid".
        cell_activation(str): The activation for cell output. Choices = ["sigmoid",
                              "tanh", "relu", "identity"], default "tanh".
        candidate_activation(str): The activation for candidate hidden state.
1001
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
Y
Yibing Liu 已提交
1002 1003
                              default "tanh".
        proj_activation(str): The activation for projection output.
1004
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
X
xuezhong 已提交
1005
                              default "tanh".
Y
Yibing Liu 已提交
1006
        dtype(str): Data type. Choices = ["float32", "float64"], default "float32".
1007 1008
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
X
xuezhong 已提交
1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019
        h_0(Variable): The initial hidden state is an optional input, default is zero.
                       This is a tensor with shape (N x D), where N is the
                       batch size and D is the projection size.
        c_0(Variable): The initial cell state is an optional input, default is zero.
                       This is a tensor with shape (N x D), where N is the
                       batch size. `h_0` and `c_0` can be NULL but only at the same time.
        cell_clip(float): If provided the cell state is clipped
                             by this value prior to the cell output activation.
        proj_clip(float): If `num_proj > 0` and `proj_clip` is
                            provided, then the projected values are clipped elementwise to within
                            `[-proj_clip, proj_clip]`.
Y
Yibing Liu 已提交
1020 1021

    Returns:
1022 1023 1024 1025
        tuple: A tuple of two output variable: the projection of hidden state, \
               and cell state of LSTMP. The shape of projection is (T x P), \
               for the cell state which is (T x D), and both LoD is the same \
               with the `input`.
Y
Yibing Liu 已提交
1026 1027

    Examples:
1028

Y
Yibing Liu 已提交
1029 1030
        .. code-block:: python

1031
            import paddle.fluid as fluid
1032 1033 1034 1035
            dict_dim, emb_dim = 128, 64
            data = fluid.layers.data(name='sequence', shape=[1],
                                     dtype='int32', lod_level=1)
            emb = fluid.layers.embedding(input=data, size=[dict_dim, emb_dim])
Y
Yibing Liu 已提交
1036
            hidden_dim, proj_dim = 512, 256
1037
            fc_out = fluid.layers.fc(input=emb, size=hidden_dim * 4,
Y
Yibing Liu 已提交
1038
                                     act=None, bias_attr=None)
1039 1040 1041
            proj_out, _ = fluid.layers.dynamic_lstmp(input=fc_out,
                                                     size=hidden_dim * 4,
                                                     proj_size=proj_dim,
Y
Yibing Liu 已提交
1042 1043 1044 1045
                                                     use_peepholes=False,
                                                     is_reverse=True,
                                                     cell_activation="tanh",
                                                     proj_activation="tanh")
Y
Yibing Liu 已提交
1046
    """
1047

L
lujun 已提交
1048
    assert in_dygraph_mode(
1049 1050
    ) is not True, "please use lstm instead of dynamic_lstmp in dygraph mode!"

C
chengduo 已提交
1051
    assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp."
Y
Yibing Liu 已提交
1052
    helper = LayerHelper('lstmp', **locals())
M
minqiyang 已提交
1053
    size = size // 4
Y
Yibing Liu 已提交
1054 1055 1056 1057 1058 1059 1060 1061 1062 1063
    weight = helper.create_parameter(
        attr=helper.param_attr, shape=[proj_size, 4 * size], dtype=dtype)
    proj_weight = helper.create_parameter(
        attr=helper.param_attr, shape=[size, proj_size], dtype=dtype)
    bias_size = [1, 7 * size]
    if not use_peepholes:
        bias_size[1] = 4 * size
    bias = helper.create_parameter(
        attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)

X
Xin Pan 已提交
1064 1065 1066 1067 1068 1069
    projection = helper.create_variable_for_type_inference(dtype)
    cell = helper.create_variable_for_type_inference(dtype)
    ordered_proj0 = helper.create_variable_for_type_inference(dtype)
    batch_hidden = helper.create_variable_for_type_inference(dtype)
    batch_gate = helper.create_variable_for_type_inference(dtype)
    batch_cell_pre_act = helper.create_variable_for_type_inference(dtype)
1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084
    inputs = {
        'Input': input,
        'Weight': weight,
        'ProjWeight': proj_weight,
        'Bias': bias
    }
    batch_size = input.shape[0]
    if h_0:
        assert h_0.shape == (batch_size, proj_size), \
            'The shape of h0 should be (batch_size, %d)' % proj_size
        inputs['H0'] = h_0
    if c_0:
        assert c_0.shape == (batch_size, size), \
            'The shape of c0 should be (batch_size, %d)' % size
        inputs['C0'] = c_0
Y
Yibing Liu 已提交
1085

X
xuezhong 已提交
1086 1087 1088 1089 1090
    if cell_clip:
        assert cell_clip >= 0, "cell_clip should not be negtive."
    if proj_clip:
        assert proj_clip >= 0, "proj_clip should not be negtive."

Y
Yibing Liu 已提交
1091 1092
    helper.append_op(
        type='lstmp',
1093
        inputs=inputs,
Y
Yibing Liu 已提交
1094 1095 1096 1097 1098 1099 1100 1101 1102
        outputs={
            'Projection': projection,
            'Cell': cell,
            'BatchHidden': batch_hidden,
            'BatchGate': batch_gate,
            'BatchCellPreAct': batch_cell_pre_act
        },
        attrs={
            'use_peepholes': use_peepholes,
1103 1104
            'cell_clip': cell_clip,
            'proj_clip': proj_clip,
Y
Yibing Liu 已提交
1105 1106 1107 1108 1109 1110 1111 1112 1113
            'is_reverse': is_reverse,
            'gate_activation': gate_activation,
            'cell_activation': cell_activation,
            'candidate_activation': candidate_activation,
            'proj_activation': proj_activation
        })
    return projection, cell


G
guosheng 已提交
1114 1115 1116 1117 1118 1119 1120
def dynamic_gru(input,
                size,
                param_attr=None,
                bias_attr=None,
                is_reverse=False,
                gate_activation='sigmoid',
                candidate_activation='tanh',
1121 1122
                h_0=None,
                origin_mode=False):
G
guosheng 已提交
1123
    """
1124
    **Gated Recurrent Unit (GRU) Layer**
G
guosheng 已提交
1125

1126 1127 1128
    if origin_mode is False, then the equation of a gru step is from paper
    `Empirical Evaluation of Gated Recurrent Neural Networks on Sequence
    Modeling <https://arxiv.org/pdf/1412.3555.pdf>`_ .
1129

G
guosheng 已提交
1130 1131 1132 1133 1134 1135 1136 1137 1138
    The formula is as follows:

    .. math::

        u_t & = act_g(W_{ux}x_{t} + W_{uh}h_{t-1} + b_u)

        r_t & = act_g(W_{rx}x_{t} + W_{rh}h_{t-1} + b_r)

        \\tilde{h_t} & = act_c(W_{cx}x_{t} + W_{ch}(r_t \odot h_{t-1}) + b_c)
1139

G
guosheng 已提交
1140
        h_t & = (1-u_t) \odot h_{t-1} + u_t \odot \\tilde{h_t}
1141

Q
Qiao Longfei 已提交
1142 1143 1144

    if origin_mode is True then the equation is from paper
    Learning Phrase Representations using RNN Encoder-Decoder for Statistical
1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156
    Machine Translation <https://arxiv.org/pdf/1406.1078.pdf>`_

    .. math::

        u_t & = act_g(W_{ux}x_{t} + W_{uh}h_{t-1} + b_u)

        r_t & = act_g(W_{rx}x_{t} + W_{rh}h_{t-1} + b_r)

        \\tilde{h_t} & = act_c(W_{cx}x_{t} + W_{ch}(r_t \odot h_{t-1}) + b_c)

        h_t & = u_t \odot h_{t-1} + (1-u_t) \odot \\tilde{h_t}

G
guosheng 已提交
1157
    The :math:`\odot` is the element-wise product of the vectors. :math:`act_g`
1158 1159
    is the update gate and reset gate activation function and :math:`sigmoid`
    is usually used for it. :math:`act_c` is the activation function for
G
guosheng 已提交
1160 1161 1162 1163
    candidate hidden state and :math:`tanh` is usually used for it.

    Note that these :math:`W_{ux}x_{t}, W_{rx}x_{t}, W_{cx}x_{t}` operations on
    the input :math:`x_{t}` are NOT included in this operator. Users can choose
1164
    to use fully-connect layer before GRU layer.
G
guosheng 已提交
1165 1166

    Args:
1167 1168
        input(Variable): The input of dynamic_gru layer, which supports
            variable-time length input sequence. The underlying tensor in this
G
guosheng 已提交
1169
            Variable is a matrix with shape :math:`(T \\times 3D)`, where
1170
            :math:`T` is the total time steps in this mini-batch, :math:`D`
G
guosheng 已提交
1171 1172
            is the hidden size.
        size(int): The dimension of the gru cell.
1173
        param_attr(ParamAttr|None): The parameter attribute for the learnable
G
guosheng 已提交
1174 1175
            hidden-hidden weight matrix. Note:

1176
            - The shape of the weight matrix is :math:`(T \\times 3D)`, where
G
guosheng 已提交
1177
              :math:`D` is the hidden size.
1178
            - All elements in the weight matrix can be divided into two parts.
G
guosheng 已提交
1179
              The first part are weights of the update gate and reset gate with
1180
              shape :math:`(D \\times 2D)`, and the second part are weights for
G
guosheng 已提交
1181
              candidate hidden state with shape :math:`(D \\times D)`.
1182 1183 1184 1185 1186

            If it is set to None or one attribute of ParamAttr, dynamic_gru will
            create ParamAttr as param_attr. If the Initializer of the param_attr
            is not set, the parameter is initialized with Xavier. Default: None.
        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias
1187
            of GRU.Note that the bias with :math:`(1 \\times 3D)` concatenates
1188
            the bias in the update gate, reset gate and candidate calculations.
1189 1190 1191
            If it is set to False, no bias will be applied to the update gate,
            reset gate and candidate calculations. If it is set to None or one
            attribute of ParamAttr, dynamic_gru will create ParamAttr as
1192 1193
            bias_attr. If the Initializer of the bias_attr is not set, the bias
            is initialized zero. Default: None.
1194
        is_reverse(bool): Whether to compute reversed GRU, default
G
guosheng 已提交
1195 1196 1197
            :attr:`False`.
        gate_activation(str): The activation for update gate and reset gate.
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "sigmoid".
1198
        candidate_activation(str): The activation for candidate hidden state.
G
guosheng 已提交
1199
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh".
1200 1201 1202 1203
        h_0 (Variable): This is initial hidden state. If not set, default is
            zero. This is a tensor with shape (N x D), where N is the number of
            total time steps of input mini-batch feature and D is the hidden
            size.
G
guosheng 已提交
1204 1205

    Returns:
G
guosheng 已提交
1206
        Variable: The hidden state of GRU. The shape is :math:`(T \\times D)`, \
1207
            and sequence length is the same with the input.
1208

G
guosheng 已提交
1209
    Examples:
1210

G
guosheng 已提交
1211 1212
        .. code-block:: python

1213 1214
            import paddle.fluid as fluid

1215 1216 1217 1218
            dict_dim, emb_dim = 128, 64
            data = fluid.layers.data(name='sequence', shape=[1],
                                     dtype='int32', lod_level=1)
            emb = fluid.layers.embedding(input=data, size=[dict_dim, emb_dim])
G
guosheng 已提交
1219
            hidden_dim = 512
1220
            x = fluid.layers.fc(input=emb, size=hidden_dim * 3)
T
Tink_Y 已提交
1221
            hidden = fluid.layers.dynamic_gru(input=x, size=hidden_dim)
G
guosheng 已提交
1222 1223
    """

L
lujun 已提交
1224
    assert in_dygraph_mode(
1225 1226
    ) is not True, "please use gru instead of dynamic_gru in dygraph mode!"

G
guosheng 已提交
1227 1228 1229 1230 1231 1232 1233
    helper = LayerHelper('gru', **locals())
    dtype = helper.input_dtype()

    weight = helper.create_parameter(
        attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype)
    bias = helper.create_parameter(
        attr=helper.bias_attr, shape=[1, 3 * size], dtype=dtype, is_bias=True)
Y
Yancey 已提交
1234
    batch_size = input.shape[0]
G
guosheng 已提交
1235
    inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
S
sneaxiy 已提交
1236
    if h_0:
G
guosheng 已提交
1237
        assert h_0.shape == (
Y
Yancey 已提交
1238 1239 1240
            batch_size, size
        ), 'The shape of h0 should be(batch_size, %d)' % size
        inputs['H0'] = h_0
G
guosheng 已提交
1241

X
Xin Pan 已提交
1242 1243 1244 1245
    hidden = helper.create_variable_for_type_inference(dtype)
    batch_gate = helper.create_variable_for_type_inference(dtype)
    batch_reset_hidden_prev = helper.create_variable_for_type_inference(dtype)
    batch_hidden = helper.create_variable_for_type_inference(dtype)
G
guosheng 已提交
1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258

    helper.append_op(
        type='gru',
        inputs=inputs,
        outputs={
            'Hidden': hidden,
            'BatchGate': batch_gate,
            'BatchResetHiddenPrev': batch_reset_hidden_prev,
            'BatchHidden': batch_hidden
        },
        attrs={
            'is_reverse': is_reverse,
            'gate_activation': gate_activation,
1259 1260
            'activation': candidate_activation,
            'origin_mode': origin_mode
G
guosheng 已提交
1261 1262 1263 1264
        })
    return hidden


Y
Yu Yang 已提交
1265 1266 1267
def gru_unit(input,
             hidden,
             size,
1268 1269
             param_attr=None,
             bias_attr=None,
Y
Yu Yang 已提交
1270
             activation='tanh',
Q
Qiao Longfei 已提交
1271 1272
             gate_activation='sigmoid',
             origin_mode=False):
Y
Yu Yang 已提交
1273
    """
1274 1275 1276
    **GRU unit layer**

    if origin_mode is True, then the equation of a gru step is from paper
Q
Qiao Longfei 已提交
1277
    `Learning Phrase Representations using RNN Encoder-Decoder for Statistical
1278
    Machine Translation <https://arxiv.org/pdf/1406.1078.pdf>`_
Y
Yu Yang 已提交
1279

1280 1281
        .. math::
            u_t & = actGate(xu_{t} + W_u h_{t-1} + b_u)
Y
Yu Yang 已提交
1282

1283
            r_t & = actGate(xr_{t} + W_r h_{t-1} + b_r)
Y
Yu Yang 已提交
1284

1285
            m_t & = actNode(xm_t + W_c dot(r_t, h_{t-1}) + b_m)
1286

1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301
            h_t & = dot(u_t, h_{t-1}) + dot((1-u_t), m_t)

    if origin_mode is False, then the equation of a gru step is from paper
    `Empirical Evaluation of Gated Recurrent Neural Networks on Sequence
    Modeling <https://arxiv.org/pdf/1412.3555.pdf>`_

        .. math::
            u_t & = actGate(xu_{t} + W_u h_{t-1} + b_u)

            r_t & = actGate(xr_{t} + W_r h_{t-1} + b_r)

            m_t & = actNode(xm_t + W_c dot(r_t, h_{t-1}) + b_m)

            h_t & = dot((1-u_t), h_{t-1}) + dot(u_t, m_t)

1302 1303

    The inputs of gru unit includes :math:`z_t`, :math:`h_{t-1}`. In terms
1304 1305 1306
    of the equation above, the :math:`z_t` is split into 3 parts -
    :math:`xu_t`, :math:`xr_t` and :math:`xm_t`. This means that in order to
    implement a full GRU unit operator for an input, a fully
1307 1308
    connected layer has to be applied, such that :math:`z_t = W_{fc}x_t`.

1309 1310
    The terms :math:`u_t` and :math:`r_t` represent the update and reset gates
    of the GRU cell. Unlike LSTM, GRU has one lesser gate. However, there is
1311 1312 1313
    an intermediate candidate hidden output, which is denoted by :math:`m_t`.
    This layer has three outputs :math:`h_t`, :math:`dot(r_t, h_{t-1})`
    and concatenation of :math:`u_t`, :math:`r_t` and :math:`m_t`.
1314 1315 1316

    Args:
        input (Variable): The fc transformed input value of current step.
1317
        hidden (Variable): The hidden value of gru unit from previous step.
1318
        size (integer): The input dimension value.
1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332
        param_attr(ParamAttr|None): The parameter attribute for the learnable
            hidden-hidden weight matrix. Note:

            - The shape of the weight matrix is :math:`(T \\times 3D)`, where
              :math:`D` is the hidden size.
            - All elements in the weight matrix can be divided into two parts.
              The first part are weights of the update gate and reset gate with
              shape :math:`(D \\times 2D)`, and the second part are weights for
              candidate hidden state with shape :math:`(D \\times D)`.

            If it is set to None or one attribute of ParamAttr, gru_unit will
            create ParamAttr as param_attr. If the Initializer of the param_attr
            is not set, the parameter is initialized with Xavier. Default: None.
        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias
1333
            of GRU.Note that the bias with :math:`(1 \\times 3D)` concatenates
1334
            the bias in the update gate, reset gate and candidate calculations.
1335 1336 1337
            If it is set to False, no bias will be applied to the update gate,
            reset gate and candidate calculations. If it is set to None or one
            attribute of ParamAttr, gru_unit will create ParamAttr as
1338 1339
            bias_attr. If the Initializer of the bias_attr is not set, the bias
            is initialized zero. Default: None.
1340 1341 1342 1343
        activation (string): The activation type for cell (actNode).
                             Default: 'tanh'
        gate_activation (string): The activation type for gates (actGate).
                                  Default: 'sigmoid'
Y
Yu Yang 已提交
1344

1345 1346 1347 1348 1349 1350
    Returns:
        tuple: The hidden value, reset-hidden value and gate values.

    Examples:

        .. code-block:: python
Y
Yu Yang 已提交
1351

1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362
            import paddle.fluid as fluid

            dict_dim, emb_dim = 128, 64
            data = fluid.layers.data(name='step_data', shape=[1], dtype='int32')
            emb = fluid.layers.embedding(input=data, size=[dict_dim, emb_dim])
            hidden_dim = 512
            x = fluid.layers.fc(input=emb, size=hidden_dim * 3)
            pre_hidden = fluid.layers.data(
                name='pre_hidden', shape=[hidden_dim], dtype='float32')
            hidden = fluid.layers.gru_unit(
                input=x, hidden=pre_hidden, size=hidden_dim * 3)
Y
Yu Yang 已提交
1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374

    """
    activation_dict = dict(
        identity=0,
        sigmoid=1,
        tanh=2,
        relu=3, )
    activation = activation_dict[activation]
    gate_activation = activation_dict[gate_activation]

    helper = LayerHelper('gru_unit', **locals())
    dtype = helper.input_dtype()
M
minqiyang 已提交
1375
    size = size // 3
Y
Yu Yang 已提交
1376 1377

    # create weight
1378 1379
    weight = helper.create_parameter(
        attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype)
Y
Yu Yang 已提交
1380

X
Xin Pan 已提交
1381 1382 1383
    gate = helper.create_variable_for_type_inference(dtype)
    reset_hidden_pre = helper.create_variable_for_type_inference(dtype)
    updated_hidden = helper.create_variable_for_type_inference(dtype)
1384
    inputs = {'Input': input, 'HiddenPrev': hidden, 'Weight': weight}
Y
Yu Yang 已提交
1385
    # create bias
1386
    if helper.bias_attr:
Y
Yu Yang 已提交
1387 1388 1389
        bias_size = [1, 3 * size]
        bias = helper.create_parameter(
            attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
1390
        inputs['Bias'] = bias
Y
Yu Yang 已提交
1391 1392 1393

    helper.append_op(
        type='gru_unit',
1394
        inputs=inputs,
Y
Yu Yang 已提交
1395 1396 1397 1398 1399 1400
        outputs={
            'Gate': gate,
            'ResetHiddenPrev': reset_hidden_pre,
            'Hidden': updated_hidden,
        },
        attrs={
1401 1402
            'activation': 2,  # tanh
            'gate_activation': 1,  # sigmoid
Y
Yu Yang 已提交
1403 1404 1405 1406 1407
        })

    return updated_hidden, reset_hidden_pre, gate


Y
yuyang18 已提交
1408
@templatedoc()
1409
def linear_chain_crf(input, label, param_attr=None, length=None):
Y
yuyang18 已提交
1410 1411 1412 1413 1414 1415 1416 1417
    """
    Linear Chain CRF.

    ${comment}

    Args:
        input(${emission_type}): ${emission_comment}
        label(${label_type}): ${label_comment}
1418
        Length(${length_type}): ${length_comment}
1419
        param_attr(ParamAttr): The attribute of the learnable parameter for transition parameter.
Y
yuyang18 已提交
1420 1421

    Returns:
D
dzhwinter 已提交
1422 1423 1424
        output(${emission_exps_type}): ${emission_exps_comment} \n
        output(${transition_exps_type}): ${transition_exps_comment} \n
        output(${log_likelihood_type}): ${log_likelihood_comment}
Y
yuyang18 已提交
1425

J
JesseyXujin 已提交
1426 1427 1428
    Examples:
        .. code-block:: python

1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468
            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):
                input_data = fluid.layers.data(name='input_data', shape=[10], dtype='float32', lod_level=1)
                label = fluid.layers.data(name='label', shape=[1], dtype='int', lod_level=1)
                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):
                input_data2 = fluid.layers.data(name='input_data2', shape=[10,10], dtype='float32')
                label2 = fluid.layers.data(name='label2', shape=[10,1], dtype='int')
                label_length = fluid.layers.data(name='length', shape=[1], dtype='int')
                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
1469
                     name='crfw',
1470 1471 1472 1473 1474 1475
                     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
1476

1477 1478 1479 1480 1481 1482 1483 1484
            #define data, using padding
            cc=np.random.rand(4,10,10).astype('float32')
            dd=np.random.rand(4,10,1).astype('int64')
            ll=np.array([[3,3,4,2]])
            feed2 = {'input_data2':cc,'label2':dd,'length':ll}

            loss2= exe.run(train_program,feed=feed2, fetch_list=[crf_cost2])
            print(loss2) 
1485 1486 1487 1488
            
            #you can use find_var to get transition parameter.
            transition=np.array(fluid.global_scope().find_var('crfw').get_tensor())
            print(transition)
Y
yuyang18 已提交
1489
    """
Y
Yu Yang 已提交
1490 1491 1492 1493 1494 1495
    helper = LayerHelper('linear_chain_crf', **locals())
    size = input.shape[1]
    transition = helper.create_parameter(
        attr=helper.param_attr,
        shape=[size + 2, size],
        dtype=helper.input_dtype())
X
Xin Pan 已提交
1496 1497 1498 1499 1500 1501 1502 1503
    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())
1504 1505 1506 1507 1508 1509 1510
    this_inputs = {
        "Emission": [input],
        "Transition": transition,
        "Label": [label]
    }
    if length:
        this_inputs['length'] = [length]
Y
Yu Yang 已提交
1511 1512
    helper.append_op(
        type='linear_chain_crf',
1513
        inputs=this_inputs,
Y
Yu Yang 已提交
1514 1515 1516 1517 1518 1519 1520 1521 1522 1523
        outputs={
            "Alpha": [alpha],
            "EmissionExps": [emission_exps],
            "TransitionExps": transition_exps,
            "LogLikelihood": log_likelihood
        })

    return log_likelihood


W
wopeizl 已提交
1524 1525 1526 1527
@templatedoc()
def crf_decoding(input, param_attr, label=None):
    """
    ${comment}
Y
yi.wu 已提交
1528

W
wopeizl 已提交
1529 1530
    Args:
        input(${emission_type}): ${emission_comment}
Y
yi.wu 已提交
1531

W
wopeizl 已提交
1532
        param_attr(ParamAttr): The parameter attribute for training.
Y
yuyang18 已提交
1533

W
wopeizl 已提交
1534
        label(${label_type}): ${label_comment}
1535

W
wopeizl 已提交
1536 1537
    Returns:
        Variable: ${viterbi_path_comment}
Y
yi.wu 已提交
1538

W
wopeizl 已提交
1539 1540
    Examples:
        .. code-block:: python
Y
yi.wu 已提交
1541

1542
           import paddle.fluid as fluid
Y
Yibing Liu 已提交
1543 1544 1545 1546 1547 1548 1549
           images = fluid.layers.data(name='pixel', shape=[784], dtype='float32')
           label = fluid.layers.data(name='label', shape=[1], dtype='int32')
           hidden = fluid.layers.fc(input=images, size=2)
           crf = fluid.layers.linear_chain_crf(input=hidden, label=label, 
                     param_attr=fluid.ParamAttr(name="crfw"))
           crf_decode = fluid.layers.crf_decoding(input=hidden, 
                     param_attr=fluid.ParamAttr(name="crfw"))
W
wopeizl 已提交
1550 1551 1552 1553 1554 1555 1556 1557
    """
    helper = LayerHelper('crf_decoding', **locals())
    transition = helper.get_parameter(param_attr.name)
    viterbi_path = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype())
    helper.append_op(
        type='crf_decoding',
        inputs={"Emission": [input],
Y
Yu Yang 已提交
1558
                "Transition": transition,
W
wopeizl 已提交
1559 1560
                "Label": label},
        outputs={"ViterbiPath": [viterbi_path]})
Y
Yu Yang 已提交
1561

W
wopeizl 已提交
1562
    return viterbi_path
Y
Yu Yang 已提交
1563 1564


Y
yi.wu 已提交
1565
@templatedoc()
F
fengjiayi 已提交
1566
def cos_sim(X, Y):
Y
Yu Yang 已提交
1567
    """
Y
yi.wu 已提交
1568 1569 1570
    ${comment}

    Args:
1571 1572
        X (Variable): ${x_comment}.
        Y (Variable): ${y_comment}.
F
fengjiayi 已提交
1573

Y
yi.wu 已提交
1574
    Returns:
1575
        Variable: the output of cosine(X, Y).
L
lvmengsi 已提交
1576 1577 1578 1579

    Examples:
        .. code-block:: python

1580
            import paddle.fluid as fluid
L
lvmengsi 已提交
1581 1582 1583
            x = fluid.layers.data(name='x', shape=[3, 7], dtype='float32', append_batch_size=False)
            y = fluid.layers.data(name='y', shape=[1, 7], dtype='float32', append_batch_size=False)
            out = fluid.layers.cos_sim(x, y)
Y
Yu Yang 已提交
1584
    """
F
fengjiayi 已提交
1585
    helper = LayerHelper('cos_sim', **locals())
X
Xin Pan 已提交
1586 1587 1588
    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
1589 1590 1591 1592 1593 1594 1595 1596 1597 1598
    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
1599 1600 1601 1602 1603
def dropout(x,
            dropout_prob,
            is_test=False,
            seed=None,
            name=None,
P
phlrain 已提交
1604
            dropout_implementation="downgrade_in_infer"):
1605 1606 1607 1608 1609
    """
    Computes dropout.

    Drop or keep each element of `x` independently. Dropout is a regularization
    technique for reducing overfitting by preventing neuron co-adaption during
1610
    training. The dropout operator randomly sets (according to the given dropout
1611 1612 1613
    probability) the outputs of some units to zero, while others are remain
    unchanged.

H
haowang101779990 已提交
1614 1615
    dropout op can be removed from the program to make the program more efficient.

1616
    Args:
1617 1618
        x (Variable): The input tensor variable.
        dropout_prob (float): Probability of setting units to zero.
1619 1620 1621 1622 1623 1624 1625
        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
                    units will be dropped. DO NOT use a fixed seed in training.
        name (str|None): A name for this layer(optional). If set None, the layer
                         will be named automatically.
H
haowang101779990 已提交
1626 1627
        dropout_implementation(string): ['downgrade_in_infer'(default)|'upscale_in_train']

P
phlrain 已提交
1628
                                        1. downgrade_in_infer(default), downgrade the outcome at inference
H
haowang101779990 已提交
1629 1630

                                           - train: out = input * mask
C
ceci3 已提交
1631
                                           - inference: out = input * (1.0 - dropout_prob)
H
haowang101779990 已提交
1632 1633 1634

                                           (mask is a tensor same shape with input, value is 0 or 1
                                           ratio of 0 is dropout_prob)
P
phlrain 已提交
1635
                                        2. upscale_in_train, upscale the outcome at training time
1636

H
haowang101779990 已提交
1637 1638
                                           - train: out = input * mask / ( 1.0 - dropout_prob )
                                           - inference: out = input
P
phlrain 已提交
1639

H
haowang101779990 已提交
1640 1641
                                           (mask is a tensor same shape with input, value is 0 or 1
                                           ratio of 0 is dropout_prob)
1642

M
minqiyang 已提交
1643

1644
    Returns:
1645
        Variable: A tensor variable is the shape with `x`.
1646 1647

    Examples:
1648

1649 1650
        .. code-block:: python

1651
            import paddle.fluid as fluid
1652 1653
            x = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
            droped = fluid.layers.dropout(x, dropout_prob=0.5)
1654 1655
    """

F
fengjiayi 已提交
1656
    helper = LayerHelper('dropout', **locals())
X
Xin Pan 已提交
1657 1658
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    mask = helper.create_variable_for_type_inference(
Z
Zeng Jinle 已提交
1659
        dtype=core.VarDesc.VarType.UINT8, stop_gradient=True)
C
chengduo 已提交
1660 1661 1662 1663

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

1664 1665 1666 1667 1668
    helper.append_op(
        type='dropout',
        inputs={'X': [x]},
        outputs={'Out': [out],
                 'Mask': [mask]},
1669 1670 1671 1672
        attrs={
            'dropout_prob': dropout_prob,
            'is_test': is_test,
            'fix_seed': seed is not None,
L
lvmengsi 已提交
1673
            'seed': seed if seed is not None else 0,
P
phlrain 已提交
1674
            'dropout_implementation': dropout_implementation,
1675
        })
1676 1677 1678
    return out


J
jerrywgz 已提交
1679
def cross_entropy(input, label, soft_label=False, ignore_index=kIgnoreIndex):
Y
Yu Yang 已提交
1680
    """
Y
Yibing Liu 已提交
1681 1682
    **Cross Entropy Layer**

1683 1684 1685
    This layer computes the cross entropy between `input` and `label`. It
    supports both standard cross-entropy and soft-label cross-entropy loss
    computation.
Y
Yibing Liu 已提交
1686 1687

    1) One-hot cross-entropy:
F
fengjiayi 已提交
1688
        `soft_label = False`, `Label[i, 0]` indicates the class index for sample i:
Y
yangyaming 已提交
1689

Y
Yibing Liu 已提交
1690
        .. math::
Y
yangyaming 已提交
1691

Y
Yibing Liu 已提交
1692 1693 1694
            Y[i] = -\log(X[i, Label[i]])

    2) Soft-label cross-entropy:
F
fengjiayi 已提交
1695 1696
        `soft_label = True`, `Label[i, j]` indicates the soft label of class j
        for sample i:
Y
Yibing Liu 已提交
1697 1698 1699 1700 1701

        .. math::

            Y[i] = \sum_j{-Label[i, j] * log(X[i, j])}

Y
Yibing Liu 已提交
1702
       Please make sure that in this case the summation of each row of `label`
Y
Yibing Liu 已提交
1703 1704 1705
       equals one.

    3) One-hot cross-entropy with vecterized `label`:
F
fengjiayi 已提交
1706 1707
         As a special case of 2), when each row of 'label' has only one
         non-zero element which is equal to 1, soft-label cross-entropy degenerates
Y
Yibing Liu 已提交
1708
         to a one-hot cross-entropy with one-hot label representation.
Y
yangyaming 已提交
1709

Y
Yibing Liu 已提交
1710
    Args:
Y
yangyaming 已提交
1711
        input (Variable|list):  a 2-D tensor with shape [N x D], where N is the
1712 1713 1714 1715
                                batch size and D is the number of classes. This
                                input is a probability computed by the previous
                                operator, which is almost always the result of
                                a softmax operator.
Y
yangyaming 已提交
1716
        label (Variable|list): the ground truth which is a 2-D tensor. When
1717 1718 1719 1720
                               `soft_label` is set to `False`, `label` is a
                               tensor<int64> with shape [N x 1]. When
                               `soft_label` is set to `True`, `label` is a
                               tensor<float/double> with shape [N x D].
F
fengjiayi 已提交
1721
        soft_label (bool): a flag indicating whether to
1722
                                           interpretate the given labels as soft
1723
                                           labels. Default: `False`.
M
minqiyang 已提交
1724 1725
        ignore_index (int): Specifies a target value that is ignored and does
                            not contribute to the input gradient. Only valid
J
jerrywgz 已提交
1726
                            if soft_label is set to False. Default: kIgnoreIndex
Y
Yibing Liu 已提交
1727 1728 1729 1730 1731

    Returns:
         A 2-D tensor with shape [N x 1], the cross entropy loss.

    Raises:
H
haowang101779990 已提交
1732 1733 1734
         ValueError:

                      1. the 1st dimension of ``input`` and ``label`` are not equal.
M
minqiyang 已提交
1735

H
haowang101779990 已提交
1736 1737
                      2. when ``soft_label == True``, and the 2nd dimension of
                         ``input`` and ``label`` are not equal.
M
minqiyang 已提交
1738

H
haowang101779990 已提交
1739 1740
                      3. when ``soft_label == False``, and the 2nd dimension of
                         ``label`` is not 1.
Y
Yibing Liu 已提交
1741 1742 1743 1744

    Examples:
        .. code-block:: python

1745
          import paddle.fluid as fluid
L
lvmengsi 已提交
1746 1747 1748 1749
          classdim = 7
          x = fluid.layers.data(name='x', shape=[3, 7], dtype='float32', append_batch_size=False)
          label = fluid.layers.data(name='label', shape=[3, 1], dtype='float32', append_batch_size=False)
          predict = fluid.layers.fc(input=x, size=classdim, act='softmax')
Y
Yibing Liu 已提交
1750
          cost = fluid.layers.cross_entropy(input=predict, label=label)
Y
Yu Yang 已提交
1751
    """
S
sneaxiy 已提交
1752 1753
    if not soft_label:
        return cross_entropy2(input, label, ignore_index)
F
fengjiayi 已提交
1754
    helper = LayerHelper('cross_entropy', **locals())
X
Xin Pan 已提交
1755
    out = helper.create_variable_for_type_inference(dtype=input.dtype)
Y
Yu Yang 已提交
1756 1757 1758 1759 1760
    helper.append_op(
        type='cross_entropy',
        inputs={'X': [input],
                'Label': [label]},
        outputs={'Y': [out]},
1761 1762
        attrs={"soft_label": soft_label,
               "ignore_index": ignore_index})
Y
Yu Yang 已提交
1763 1764 1765
    return out


S
sneaxiy 已提交
1766 1767 1768 1769
def cross_entropy2(input, label, ignore_index=kIgnoreIndex):
    helper = LayerHelper('cross_entropy2', **locals())
    out = helper.create_variable_for_type_inference(dtype=input.dtype)
    xshape = helper.create_variable_for_type_inference(dtype=input.dtype)
S
sneaxiy 已提交
1770
    match_x = helper.create_variable_for_type_inference(dtype=input.dtype)
S
sneaxiy 已提交
1771 1772 1773 1774 1775
    helper.append_op(
        type='cross_entropy2',
        inputs={'X': [input],
                'Label': [label]},
        outputs={'Y': [out],
S
sneaxiy 已提交
1776
                 'MatchX': [match_x],
S
sneaxiy 已提交
1777 1778 1779 1780 1781
                 'XShape': [xshape]},
        attrs={'ignore_index': ignore_index})
    return out


F
frankwhzhang 已提交
1782
def bpr_loss(input, label, name=None):
F
frankwhzhang 已提交
1783
    """
1784
    **Bayesian Personalized Ranking Loss Operator**
F
frankwhzhang 已提交
1785

1786
    This operator belongs to pairwise ranking loss. Label is the desired item.
F
frankwhzhang 已提交
1787
    The loss at a given point in one session is defined as:
1788 1789 1790

    .. math::
        Y[i] = 1/(N[i] - 1) * \sum_j{\log(\sigma(X[i, Label[i]]-X[i, j]))}
F
frankwhzhang 已提交
1791 1792

    Learn more details by reading paper <session-based recommendations with recurrent
1793
    neural networks>.
F
frankwhzhang 已提交
1794

1795 1796 1797 1798 1799 1800
    Args:
        input (Variable|list):  a 2-D tensor with shape [N x D], where N is the
                                batch size and D is the number of classes.
                                This input is not probability but logits.
        label (Variable|list):  the ground truth which is a 2-D tensor.  `label`
                                is a tensor<int64> with shape [N x 1].
F
frankwhzhang 已提交
1801 1802
        name (str|None):        A name for this layer(optional). If set None, the
                                layer will be named automatically. Default: None.
1803 1804 1805
    Returns:
        A 2-D tensor with shape [N x 1], the bpr loss.

F
frankwhzhang 已提交
1806 1807 1808
    Examples:
        .. code-block:: python

1809 1810 1811 1812 1813 1814 1815
          import paddle.fluid as fluid

          neg_size = 10
          label = fluid.layers.data(
                    name="label", shape=[1], dtype="int64")
          predict = fluid.layers.data(
                    name="predict", shape=[neg_size + 1], dtype="float32")
1816
          cost = fluid.layers.bpr_loss(input=predict, label=label)
F
frankwhzhang 已提交
1817
    """
1818 1819 1820 1821 1822
    helper = LayerHelper('bpr_loss', **locals())
    out = helper.create_variable_for_type_inference(dtype=input.dtype)
    helper.append_op(
        type='bpr_loss',
        inputs={'X': [input],
1823
                'Label': [label]},
1824 1825 1826 1827
        outputs={'Y': [out]})
    return out


F
fengjiayi 已提交
1828
def square_error_cost(input, label):
Y
Yu Yang 已提交
1829
    """
1830 1831
    **Square error cost layer**

1832 1833
    This layer accepts input predictions and target label and returns the
    squared error cost.
Y
ying 已提交
1834

1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847
    For predictions, :math:`X`, and target labels, :math:`Y`, the equation is:

    .. math::

        Out = (X - Y)^2

    In the above equation:

        * :math:`X`: Input predictions, a tensor.
        * :math:`Y`: Input labels, a tensor.
        * :math:`Out`: Output value, same shape with :math:`X`.

    Args:
1848 1849
        input (Variable): Input tensor, has predictions.
        label (Variable): Label tensor, has target labels.
1850 1851

    Returns:
G
guosheng 已提交
1852
        Variable: The tensor variable storing the element-wise squared error \
1853
                  difference of input and label.
1854 1855 1856 1857

    Examples:
        .. code-block:: python

1858
          import paddle.fluid as fluid
R
ruri 已提交
1859 1860 1861
          y = fluid.layers.data(name='y', shape=[1], dtype='float32')
          y_predict = fluid.layers.data(name='y_predict', shape=[1], dtype='float32')
          cost = fluid.layers.square_error_cost(input=y_predict, label=y)
1862

Y
Yu Yang 已提交
1863
    """
F
fengjiayi 已提交
1864
    helper = LayerHelper('square_error_cost', **locals())
X
Xin Pan 已提交
1865
    minus_out = helper.create_variable_for_type_inference(dtype=input.dtype)
Y
Yu Yang 已提交
1866 1867 1868 1869 1870 1871
    helper.append_op(
        type='elementwise_sub',
        inputs={'X': [input],
                'Y': [label]},
        outputs={'Out': [minus_out]})

X
Xin Pan 已提交
1872
    square_out = helper.create_variable_for_type_inference(dtype=input.dtype)
Y
Yu Yang 已提交
1873
    helper.append_op(
F
fengjiayi 已提交
1874 1875
        type='square', inputs={'X': [minus_out]},
        outputs={'Out': [square_out]})
Y
Yu Yang 已提交
1876 1877 1878
    return square_out


Y
yi.wu 已提交
1879
@templatedoc()
Y
Yu Yang 已提交
1880 1881 1882 1883
def chunk_eval(input,
               label,
               chunk_scheme,
               num_chunk_types,
1884 1885
               excluded_chunk_types=None,
               seq_length=None):
Y
Yu Yang 已提交
1886
    """
Y
yi.wu 已提交
1887
    **Chunk Evaluator**
Y
yi.wu 已提交
1888

Y
yangyaming 已提交
1889
    This function computes and outputs the precision, recall and
1890
    F1-score of chunk detection.
Y
yi.wu 已提交
1891

M
minqiyang 已提交
1892
    For some basics of chunking, please refer to
H
haowang101779990 已提交
1893
    `Chunking with Support Vector Machines <https://aclanthology.info/pdf/N/N01/N01-1025.pdf>`_ .
Y
yi.wu 已提交
1894 1895 1896 1897 1898 1899

    ChunkEvalOp computes the precision, recall, and F1-score of chunk detection,
    and supports IOB, IOE, IOBES and IO (also known as plain) tagging schemes.
    Here is a NER example of labeling for these tagging schemes:

    .. code-block:: python
1900

Y
yi.wu 已提交
1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925
       ====== ====== ======  =====  ==  ============   =====  ===== =====  ==  =========
              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)
    and LOC(LOCATION), and we can see that the labels have the form <tag type>-<chunk type>.

    Since the calculations actually use label ids rather than labels, extra attention
    should be paid when mapping labels to ids to make CheckEvalOp work. The key point
    is that the listed equations are satisfied by ids.

    .. 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
1926

Y
yi.wu 已提交
1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950
       Scheme Begin Inside End   Single
        plain   0     -      -     -
        IOB     0     1      -     -
        IOE     -     0      1     -
        IOBES   0     1      2     3

    Still use NER as example, assuming the tagging scheme is IOB while chunk types are ORG,
    PER and LOC. To satisfy the above equations, the label map can be like this:

    .. code-block:: python

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

    It's not hard to verify the equations noting that the num of chunk types
    is 3 and the num of tag types in IOB scheme is 2. For example, the label
    id of I-LOC is 5, the tag type id of I-LOC is 1, and the chunk type id of
    I-LOC is 2, which consistent with the results from the equations.

Y
yi.wu 已提交
1951
    Args:
1952 1953 1954 1955 1956
        input (Variable): prediction output of the network.
        label (Variable): label of the test data set.
        chunk_scheme (str): ${chunk_scheme_comment}
        num_chunk_types (int): ${num_chunk_types_comment}
        excluded_chunk_types (list): ${excluded_chunk_types_comment}
1957
        seq_length(Variable): 1-D Tensor specifying sequence length when input and label are Tensor type.
F
fengjiayi 已提交
1958

Y
yi.wu 已提交
1959
    Returns:
Y
update  
yi.wu 已提交
1960 1961 1962
        tuple: tuple containing: precision, recall, f1_score,
        num_infer_chunks, num_label_chunks,
        num_correct_chunks
1963

Y
yi.wu 已提交
1964 1965 1966
    Examples:
        .. code-block:: python

1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977
            import paddle.fluid as fluid

            dict_size = 10000
            label_dict_len = 7
            sequence = fluid.layers.data(
                name='id', shape=[1], lod_level=1, dtype='int64')
            embedding = fluid.layers.embedding(
                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
1978
            crf = fluid.layers.linear_chain_crf(
1979
                input=hidden, label=label, param_attr=fluid.ParamAttr(name="crfw"))
Y
yi.wu 已提交
1980
            crf_decode = fluid.layers.crf_decoding(
1981
                input=hidden, param_attr=fluid.ParamAttr(name="crfw"))
Y
yi.wu 已提交
1982 1983 1984 1985 1986
            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
1987
    """
F
fengjiayi 已提交
1988
    helper = LayerHelper("chunk_eval", **locals())
Y
Yu Yang 已提交
1989 1990

    # prepare output
X
Xin Pan 已提交
1991 1992 1993 1994 1995 1996 1997
    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
1998

1999 2000 2001 2002 2003
    this_input = {"Inference": [input], "Label": [label]}

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

Y
Yu Yang 已提交
2004 2005
    helper.append_op(
        type="chunk_eval",
2006
        inputs=this_input,
Y
Yu Yang 已提交
2007 2008 2009
        outputs={
            "Precision": [precision],
            "Recall": [recall],
2010 2011 2012 2013
            "F1-Score": [f1_score],
            "NumInferChunks": [num_infer_chunks],
            "NumLabelChunks": [num_label_chunks],
            "NumCorrectChunks": [num_correct_chunks]
Y
Yu Yang 已提交
2014 2015 2016
        },
        attrs={
            "num_chunk_types": num_chunk_types,
G
guosheng 已提交
2017 2018
            "chunk_scheme": chunk_scheme,
            "excluded_chunk_types": excluded_chunk_types or []
Y
Yu Yang 已提交
2019
        })
2020 2021
    return (precision, recall, f1_score, num_infer_chunks, num_label_chunks,
            num_correct_chunks)
Y
Yu Yang 已提交
2022 2023


2024
@templatedoc()
Y
Yu Yang 已提交
2025 2026 2027 2028
def sequence_conv(input,
                  num_filters,
                  filter_size=3,
                  filter_stride=1,
2029 2030
                  padding=True,
                  padding_start=None,
Y
Yu Yang 已提交
2031 2032
                  bias_attr=None,
                  param_attr=None,
C
chengduo 已提交
2033 2034
                  act=None,
                  name=None):
Y
Yu Yang 已提交
2035
    """
2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071
    The sequence_conv receives input sequences with variable length and other convolutional
    configuration parameters for the filter and stride to apply the convolution operation.
    It fills all-zero padding data on both sides of the sequence by default to ensure that
    the output is the same length as the input. You can customize the padding behavior by
    configuring the parameter :attr:`padding\_start`.
    
    **Warning:** the parameter :attr:`padding` take no effect and will be deprecated in the future.

    .. code-block:: text

            Here we'll illustrate the details of the padding operation:
            For a mini-batch of 2 variable lengths sentences, containing 3, and 1 time-steps:
            Assumed input (X) is a [4, M, N] float LoDTensor, and X->lod()[0] = [0, 3, 4].
            Besides, for the sake of simplicity, we assume M=1 and N=2.
            X = [[a1, a2;
                  b1, b2;
                  c1, c2]
                 [d1, d2]]

            This is to say that input (X) has 4 words and the dimension of each word
            representation is 2.

            * Case1:

                If padding_start is -1 and filter_size is 3.
                The length of padding data is calculated as follows:
                up_pad_len = max(0, -padding_start) = 1
                down_pad_len = max(0, filter_size + padding_start - 1) = 1

                The output of the input sequence after padding is:
                data_aftet_padding = [[0,  0,  a1, a2, b1, b2;
                                       a1, a2, b1, b2, c1, c2;
                                       b1, b2, c1, c2, 0,  0 ]
                                      [0,  0,  d1, d2, 0,  0 ]]

                It will be multiplied by the filter weight to get the final output.
2072 2073 2074

    Args:
        input (Variable): ${x_comment}
2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092
        num_filters (int): the number of filters.
        filter_size (int): the height of filter, the width is hidden size by default.
        filter_stride (int): stride of the filter. Currently only supports :attr:`stride` = 1.
        padding (bool): the parameter :attr:`padding` take no effect and will be discarded in the
            future. Currently, it will always pad input to make sure the length of the output is
            the same as input whether :attr:`padding` is set true or false. Because the length of
            input sequence may be shorter than :attr:`filter\_size`, which will cause the convolution
            result to not be computed correctly. These padding data will not be trainable or updated
            while trainnig. 
        padding_start (int|None): It is used to indicate the start index for padding the input
            sequence, which can be negative. The negative number means to pad
            :attr:`|padding_start|` time-steps of all-zero data at the beginning of each instance.
            The positive number means to skip :attr:`padding_start` time-steps of each instance,
            and it will pad :math:`filter\_size + padding\_start - 1` time-steps of all-zero data
            at the end of the sequence to ensure that the output is the same length as the input.
            If set None, the same length :math:`\\frac{filter\_size}{2}` of data will be filled
            on both sides of the sequence. If set 0, the length of :math:`filter\_size - 1` data
            is padded at the end of each input sequence.
C
chengduo 已提交
2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105
        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of sequence_conv.
            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, sequence_conv
            will create ParamAttr as bias_attr. If the Initializer of the bias_attr
            is not set, the bias is initialized zero. Default: None.
        param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
            of sequence_conv. If it is set to None or one attribute of ParamAttr, sequence_conv
            will create ParamAttr as param_attr. If the Initializer of the param_attr
            is not set, the parameter is initialized with Xavier. Default: None.
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None.
        name (str|None): A name for this layer(optional). If set None, the layer
            will be named automatically. Default: None.
F
fengjiayi 已提交
2106

2107 2108
    Returns:
        Variable: output of sequence_conv
B
bdzhuxiaoning 已提交
2109 2110

    Examples:
2111

B
bdzhuxiaoning 已提交
2112 2113 2114
        .. code-block:: python

             import paddle.fluid as fluid
2115

B
bdzhuxiaoning 已提交
2116
             x = fluid.layers.data(name='x', shape=[10,10], append_batch_size=False, dtype='float32')
2117
             x_conved = fluid.layers.sequence_conv(input=x, num_filters=2, filter_size=3, padding_start=-1)
Y
Yu Yang 已提交
2118 2119
    """

L
lujun 已提交
2120
    assert not in_dygraph_mode(), (
2121
        "sequence layer is not supported in dygraph mode yet.")
Y
Yu Yang 已提交
2122 2123 2124 2125 2126
    helper = LayerHelper('sequence_conv', **locals())
    dtype = helper.input_dtype()
    filter_shape = [filter_size * input.shape[1], num_filters]
    filter_param = helper.create_parameter(
        attr=helper.param_attr, shape=filter_shape, dtype=dtype)
X
Xin Pan 已提交
2127
    pre_bias = helper.create_variable_for_type_inference(dtype)
2128 2129
    if padding_start is None:
        padding_start = -int(filter_size // 2)
Y
Yu Yang 已提交
2130 2131 2132 2133 2134 2135 2136 2137 2138 2139

    helper.append_op(
        type='sequence_conv',
        inputs={
            'X': [input],
            'Filter': [filter_param],
        },
        outputs={"Out": pre_bias},
        attrs={
            'contextStride': filter_stride,
2140 2141
            'contextStart': padding_start,
            'contextLength': filter_size,
Y
Yu Yang 已提交
2142 2143 2144 2145 2146
        })
    pre_act = helper.append_bias_op(pre_bias)
    return helper.append_activation(pre_act)


C
chengduo 已提交
2147
def sequence_softmax(input, use_cudnn=False, name=None):
2148 2149 2150
    """
    This function computes the softmax activation among all time-steps for each
    sequence. The dimension of each time-step should be 1. Thus, the shape of
2151
    input Tensor can be either :math:`[N, 1]` or :math:`[N]`, where :math:`N`
2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167
    is the sum of the length of all sequences.

    For i-th sequence in a mini-batch:

    .. math::

        Out(X[lod[i]:lod[i+1]], :) = \\frac{\exp(X[lod[i]:lod[i+1], :])}{\sum(\exp(X[lod[i]:lod[i+1], :]))}

    For example, for a mini-batch of 3 sequences with variable-length,
    each containing 2, 3, 2 time-steps, the lod of which is [0, 2, 5, 7],
    then softmax will be computed among :math:`X[0:2, :]`, :math:`X[2:5, :]`,
    :math:`X[5:7, :]`, and :math:`N` turns out to be 7.

    Args:
        input (Variable): The input variable which is a LoDTensor.
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \
C
chengduo 已提交
2168 2169 2170
            library is installed. Default: False.
        name (str|None): A name for this layer(optional). If set None, the layer
            will be named automatically. Default: None.
2171

2172 2173 2174 2175 2176 2177 2178
    Returns:
        Variable: output of sequence_softmax

    Examples:

        .. code-block:: python

2179
             import paddle.fluid as fluid
2180 2181 2182 2183
             x = fluid.layers.data(name='x', shape=[7, 1],
                              dtype='float32', lod_level=1)
             x_sequence_softmax = fluid.layers.sequence_softmax(input=x)
    """
L
lujun 已提交
2184
    assert not in_dygraph_mode(), (
2185
        "sequence layer is not supported in dygraph mode yet.")
2186 2187
    helper = LayerHelper('sequence_softmax', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
2188
    softmax_out = helper.create_variable_for_type_inference(dtype)
2189 2190 2191 2192 2193 2194 2195 2196
    helper.append_op(
        type="sequence_softmax",
        inputs={"X": input},
        outputs={"Out": softmax_out},
        attrs={"use_cudnn": use_cudnn})
    return softmax_out


D
dengkaipeng 已提交
2197
def softmax(input, use_cudnn=False, name=None, axis=-1):
Q
qiaolongfei 已提交
2198
    """
2199
    The input of the softmax operator is a tensor of any rank. The output tensor
F
fengjiayi 已提交
2200
    has the same shape as the input.
Q
qiaolongfei 已提交
2201

D
dengkaipeng 已提交
2202
    The dimension :attr:`axis` of the input tensor will be permuted to the last.
D
dengkaipeng 已提交
2203
    Then the input tensor will be logically flattened to a 2-D matrix. The matrix's
D
dengkaipeng 已提交
2204
    second dimension(row length) is the same as the dimension :attr:`axis` of the input
2205 2206 2207
    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
D
dengkaipeng 已提交
2208
    of the input tensor's dimension :attr:`axis`) vector of arbitrary real values to a
F
fengjiayi 已提交
2209
    K-dimensional vector of real values in the range [0, 1] that add up to 1.
Q
qiaolongfei 已提交
2210 2211 2212 2213 2214 2215 2216

    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.

F
fengjiayi 已提交
2217
    For each row :math:`i` and each column :math:`j` in the matrix, we have:
Q
qiaolongfei 已提交
2218 2219 2220 2221 2222 2223 2224 2225

    .. math::

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

    Args:
        input (Variable): The input variable.
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \
J
jerrywgz 已提交
2226 2227
            library is installed. To improve numerical stablity, set use_cudnn to \
            False by default. Default: False
C
chengduo 已提交
2228 2229
        name (str|None): A name for this layer(optional). If set None, the layer
            will be named automatically. Default: None.
D
dengkaipeng 已提交
2230 2231 2232
        axis (int): The index of dimension to perform softmax calculations, it should
            be in range :math:`[-1, rank - 1]`, while :math:`rank` is the rank of
            input variable. Default: -1.
Q
qiaolongfei 已提交
2233 2234 2235 2236 2237 2238 2239 2240

    Returns:
        Variable: output of softmax

    Examples:

        .. code-block:: python

J
JesseyXujin 已提交
2241 2242
             import paddle.fluid as fluid
             x = fluid.layers.data(name='x', shape=[2], dtype='float32')
Q
qiaolongfei 已提交
2243
             fc = fluid.layers.fc(input=x, size=10)
D
dengkaipeng 已提交
2244
             # perform softmax in the second dimension
D
dengkaipeng 已提交
2245
             softmax = fluid.layers.softmax(input=fc, axis=1)
D
dengkaipeng 已提交
2246 2247
             # perform softmax in the last dimension
             softmax = fluid.layers.softmax(input=fc, axis=-1)
Q
qiaolongfei 已提交
2248 2249

    """
2250 2251
    helper = LayerHelper('softmax', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
2252
    softmax_out = helper.create_variable_for_type_inference(dtype)
2253 2254 2255 2256
    helper.append_op(
        type="softmax",
        inputs={"X": input},
        outputs={"Out": softmax_out},
D
dengkaipeng 已提交
2257 2258
        attrs={"axis": axis,
               "use_cudnn": use_cudnn})
2259 2260 2261
    return softmax_out


Y
Yu Yang 已提交
2262 2263 2264
def conv2d(input,
           num_filters,
           filter_size,
C
chengduoZH 已提交
2265 2266
           stride=1,
           padding=0,
2267
           dilation=1,
Y
Yu Yang 已提交
2268 2269 2270
           groups=None,
           param_attr=None,
           bias_attr=None,
C
chengduoZH 已提交
2271
           use_cudnn=True,
2272 2273
           act=None,
           name=None):
Y
Yu Yang 已提交
2274
    """
C
chengduoZH 已提交
2275
    The convolution2D layer calculates the output based on the input, filter
T
tensor-tang 已提交
2276 2277
    and strides, paddings, dilations, groups parameters. Input and
    Output are in NCHW format, where N is batch size, C is the number of
2278
    channels, H is the height of the feature, and W is the width of the feature.
T
tensor-tang 已提交
2279 2280 2281 2282 2283 2284
    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/>`_
2285
    for more details.
2286 2287 2288
    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
2289

2290
    For each input :math:`X`, the equation is:
C
refine  
chengduoZH 已提交
2291

C
chengduoZH 已提交
2292 2293
    .. math::

C
refine  
chengduoZH 已提交
2294
        Out = \sigma (W \\ast X + b)
C
chengduoZH 已提交
2295

T
tensor-tang 已提交
2296
    Where:
C
chengduoZH 已提交
2297

2298 2299 2300 2301 2302
    * :math:`X`: Input value, a tensor with NCHW format.
    * :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
2303
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
2304 2305 2306

    Example:

2307 2308
        - Input:

W
weixing02 已提交
2309
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
C
refine  
chengduoZH 已提交
2310

W
weixing02 已提交
2311
          Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)`
C
refine  
chengduoZH 已提交
2312

2313
        - Output:
T
tensor-tang 已提交
2314

W
weixing02 已提交
2315
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
C
refine  
chengduoZH 已提交
2316

C
chengduoZH 已提交
2317
        Where
2318 2319

        .. math::
C
chengduoZH 已提交
2320

W
weixing02 已提交
2321 2322
            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
2323 2324

    Args:
2325
        input (Variable): The input image with [N, C, H, W] format.
T
tensor-tang 已提交
2326
        num_filters(int): The number of filter. It is as same as the output
2327
            image channel.
2328
        filter_size (int|tuple): The filter size. If filter_size is a tuple,
2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343
            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 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
2344 2345 2346 2347 2348
            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
2349
            and the :math:`std` is :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None.
C
chengduo 已提交
2350 2351 2352 2353 2354
        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.
2355 2356
        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
2357 2358
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None
2359
        name (str|None): A name for this layer(optional). If set None, the layer
C
chengduo 已提交
2360
            will be named automatically. Default: None
C
chengduoZH 已提交
2361 2362

    Returns:
G
guosheng 已提交
2363
        Variable: The tensor variable storing the convolution and \
C
chengduoZH 已提交
2364 2365
                  non-linearity activation result.

C
refine  
chengduoZH 已提交
2366
    Raises:
2367 2368
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
C
refine  
chengduoZH 已提交
2369

C
chengduoZH 已提交
2370 2371 2372
    Examples:
        .. code-block:: python

2373
          import paddle.fluid as fluid
2374 2375
          data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32')
          conv2d = fluid.layers.conv2d(input=data, num_filters=2, filter_size=3, act="relu")
Y
Yu Yang 已提交
2376 2377 2378
    """

    num_channels = input.shape[1]
C
chengduo 已提交
2379
    assert param_attr is not False, "param_attr should not be False here."
2380
    l_type = 'conv2d'
X
xzl 已提交
2381 2382
    if (num_channels == groups and num_filters % num_channels == 0 and
            not use_cudnn):
2383
        l_type = 'depthwise_conv2d'
2384 2385 2386 2387

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

Y
Yu Yang 已提交
2388 2389 2390 2391 2392
    if groups is None:
        num_filter_channels = num_channels
    else:
        if num_channels % groups != 0:
            raise ValueError("num_channels must be divisible by groups.")
M
minqiyang 已提交
2393
        num_filter_channels = num_channels // groups
Y
Yu Yang 已提交
2394

C
chengduoZH 已提交
2395 2396 2397
    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')
2398
    dilation = utils.convert_to_list(dilation, 2, 'dilation')
C
chengduoZH 已提交
2399

C
chengduoZH 已提交
2400 2401
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
2402 2403

    input_shape = input.shape
M
minqiyang 已提交
2404
    filter_shape = [num_filters, int(num_filter_channels)] + filter_size
Y
Yu Yang 已提交
2405 2406

    def _get_default_param_initializer():
C
chengduo 已提交
2407 2408
        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
2409 2410 2411 2412 2413 2414 2415 2416
        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
2417
    pre_bias = helper.create_variable_for_type_inference(dtype)
Y
Yu Yang 已提交
2418 2419

    helper.append_op(
2420
        type=l_type,
Y
Yu Yang 已提交
2421 2422 2423 2424 2425
        inputs={
            'Input': input,
            'Filter': filter_param,
        },
        outputs={"Output": pre_bias},
C
chengduoZH 已提交
2426 2427 2428
        attrs={
            'strides': stride,
            'paddings': padding,
2429
            'dilations': dilation,
C
chengduoZH 已提交
2430
            'groups': groups,
2431
            'use_cudnn': use_cudnn,
2432
            'use_mkldnn': False,
2433
            'fuse_relu_before_depthwise_conv': False
C
chengduoZH 已提交
2434
        })
Y
Yu Yang 已提交
2435 2436 2437 2438 2439 2440

    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)

    return helper.append_activation(pre_act)


C
chengduoZH 已提交
2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457
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,
           name=None):
    """
    **Convlution3D Layer**

    The convolution3D layer calculates the output based on the input, filter
    and strides, paddings, dilations, groups parameters. Input(Input) and
2458 2459 2460 2461 2462 2463
    Output(Output) 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. 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
2464 2465 2466 2467 2468 2469 2470 2471 2472

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

    .. math::

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

    In the above equation:

2473 2474
    * :math:`X`: Input value, a tensor with NCDHW format.
    * :math:`W`: Filter value, a tensor with MCDHW format.
C
chengduoZH 已提交
2475 2476 2477
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
2478
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500

    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:
        input (Variable): The input image with [N, C, D, H, W] format.
2501
        num_filters(int): The number of filter. It is as same as the output
C
chengduoZH 已提交
2502 2503
            image channel.
        filter_size (int|tuple|None): The filter size. If filter_size is a tuple,
2504
            it must contain three integers, (filter_size_D, filter_size_H, filter_size_W).
C
chengduoZH 已提交
2505 2506
            Otherwise, the filter will be a square.
        stride (int|tuple): The stride size. If stride is a tuple, it must
2507
            contain three integers, (stride_D, stride_H, stride_W). Otherwise, the
C
chengduoZH 已提交
2508 2509
            stride_D = stride_H = stride_W = stride. Default: stride = 1.
        padding (int|tuple): The padding size. If padding is a tuple, it must
2510
            contain three integers, (padding_D, padding_H, padding_W). Otherwise, the
C
chengduoZH 已提交
2511 2512
            padding_D = padding_H = padding_W = padding. Default: padding = 0.
        dilation (int|tuple): The dilation size. If dilation is a tuple, it must
2513
            contain three integers, (dilation_D, dilation_H, dilation_W). Otherwise, the
C
chengduoZH 已提交
2514 2515 2516 2517 2518 2519
            dilation_D = dilation_H = dilation_W = dilation. Default: dilation = 1.
        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
2520 2521 2522 2523 2524 2525 2526 2527 2528 2529
        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
2530 2531
        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
2532 2533
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None.
C
chengduoZH 已提交
2534
        name (str|None): A name for this layer(optional). If set None, the layer
C
chengduo 已提交
2535
            will be named automatically. Default: None.
C
chengduoZH 已提交
2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547

    Returns:
        Variable: The tensor variable storing the convolution and \
                  non-linearity activation result.

    Raises:
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.

    Examples:
        .. code-block:: python

2548
          import paddle.fluid as fluid
2549 2550
          data = fluid.layers.data(name='data', shape=[3, 12, 32, 32], dtype='float32')
          conv3d = fluid.layers.conv3d(input=data, num_filters=2, filter_size=3, act="relu")
C
chengduoZH 已提交
2551 2552 2553
    """

    l_type = 'conv3d'
C
chengduo 已提交
2554
    assert param_attr is not False, "param_attr should not be False here."
C
chengduoZH 已提交
2555 2556 2557 2558 2559 2560 2561 2562 2563 2564
    helper = LayerHelper(l_type, **locals())
    dtype = helper.input_dtype()

    num_channels = input.shape[1]

    if groups is None:
        num_filter_channels = num_channels
    else:
        if num_channels % groups != 0:
            raise ValueError("num_channels must be divisible by groups.")
M
minqiyang 已提交
2565
        num_filter_channels = num_channels // groups
C
chengduoZH 已提交
2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578

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

    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")

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

    def _get_default_param_initializer():
C
chengduo 已提交
2579 2580 2581
        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
2582 2583 2584 2585 2586 2587 2588 2589
        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
2590
    pre_bias = helper.create_variable_for_type_inference(dtype)
C
chengduoZH 已提交
2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604

    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,
X
Xin Pan 已提交
2605
            'use_mkldnn': False
C
chengduoZH 已提交
2606 2607
        })

2608
    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
C
chengduoZH 已提交
2609 2610 2611 2612

    return helper.append_activation(pre_act)


2613
def sequence_pool(input, pool_type, is_test=False, pad_value=0.0):
Y
Yu Yang 已提交
2614
    """
Y
yangyaming 已提交
2615 2616 2617
    This function add the operator for sequence pooling.
    It pools features of all time-steps of each instance, and is applied
    on top of the input using pool_type mentioned in the parameters.
L
Luo Tao 已提交
2618 2619 2620 2621 2622 2623 2624 2625 2626 2627

    It supports four pool_type:

    - average: :math:`Out[i] = \\frac{\sum_i X_i}{N}`
    - sum:     :math:`Out[i] = \sum_jX_{ij}`
    - sqrt:    :math:`Out[i] = \\frac{\sum_jX_{ij}}{\sqrt{len(X_i)}}`
    - max:     :math:`Out[i] = max(X_i)`

    .. code-block:: text

2628 2629
       x is a 1-level LoDTensor and **pad_value** = 0.0:
         x.lod = [[2, 3, 2, 0]]
L
Luo Tao 已提交
2630 2631 2632 2633
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
2634
         out.dim = [4, 1]
2635
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
2636 2637

       for different pool_type:
2638 2639 2640
         average: out.data = [2, 4, 3, 0.0], where 2=(1+3)/2, 4=(2+4+6)/3, 3=(5+1)/2
         sum    : out.data = [4, 12, 6, 0.0], where 4=1+3, 12=2+4+6, 6=5+1
         sqrt   : out.data = [2.82, 6.93, 4.24, 0.0], where 2.82=(1+3)/sqrt(2),
L
Luo Tao 已提交
2641
                    6.93=(2+4+6)/sqrt(3), 4.24=(5+1)/sqrt(2)
2642 2643 2644 2645 2646
         max    : out.data = [3, 6, 5, 0.0], where 3=max(1,3), 6=max(2,4,6), 5=max(5,1)
         last   : out.data = [3, 6, 1, 0.0], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1)
         first  : out.data = [1, 2, 5, 0.0], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1)

         and all above 0.0 = **pad_value**.
F
fengjiayi 已提交
2647

L
Luo Tao 已提交
2648
    Args:
2649
        input (variable): The input variable which is a LoDTensor.
Y
yangyaming 已提交
2650
        pool_type (string): The pooling type of sequence_pool.
L
Luo Tao 已提交
2651
            It supports average, sum, sqrt and max.
2652 2653
        is_test (bool): Used to distinguish training from scoring mode. Default False.
        pad_value (float): Used to pad the pooling result for empty input sequence.
L
Luo Tao 已提交
2654 2655 2656 2657 2658 2659 2660

    Returns:
        The sequence pooling variable which is a Tensor.

    Examples:

        .. code-block:: python
F
fengjiayi 已提交
2661

2662 2663
             import paddle.fluid as fluid

Y
yangyaming 已提交
2664
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
2665 2666 2667 2668 2669
                              dtype='float32', lod_level=1)
             avg_x = fluid.layers.sequence_pool(input=x, pool_type='average')
             sum_x = fluid.layers.sequence_pool(input=x, pool_type='sum')
             sqrt_x = fluid.layers.sequence_pool(input=x, pool_type='sqrt')
             max_x = fluid.layers.sequence_pool(input=x, pool_type='max')
2670 2671
             last_x = fluid.layers.sequence_pool(input=x, pool_type='last')
             first_x = fluid.layers.sequence_pool(input=x, pool_type='first')
Y
Yu Yang 已提交
2672
    """
L
lujun 已提交
2673
    assert not in_dygraph_mode(), (
2674
        "sequence layer is not supported in dygraph mode yet.")
F
fengjiayi 已提交
2675
    helper = LayerHelper('sequence_pool', **locals())
Y
Yu Yang 已提交
2676
    dtype = helper.input_dtype()
X
Xin Pan 已提交
2677 2678
    pool_out = helper.create_variable_for_type_inference(dtype)
    max_index = helper.create_variable_for_type_inference(dtype)
Y
Yu Yang 已提交
2679 2680 2681 2682 2683 2684

    helper.append_op(
        type="sequence_pool",
        inputs={"X": input},
        outputs={"Out": pool_out,
                 "MaxIndex": max_index},
2685 2686 2687 2688 2689
        attrs={
            "pooltype": pool_type.upper(),
            "is_test": is_test,
            "pad_value": pad_value
        })
Y
Yu Yang 已提交
2690

Y
yangyaming 已提交
2691 2692 2693 2694 2695
    # when pool_type is max, variable max_index is initialized,
    # so we stop the gradient explicitly here
    if pool_type == 'max':
        max_index.stop_gradient = True

Y
Yu Yang 已提交
2696 2697 2698
    return pool_out


C
add doc  
chengduoZH 已提交
2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714
@templatedoc()
def sequence_concat(input, name=None):
    """
    ${comment}

    Args:
        input(list): List of Variables to be concatenated.
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.

    Returns:
        Variable: Output variable of the concatenation.

    Examples:
        .. code-block:: python

B
bdzhuxiaoning 已提交
2715 2716 2717 2718
           import paddle.fluid as fluid
           x = fluid.layers.data(name='x', shape=[10], dtype='float32')
           y = fluid.layers.data(name='y', shape=[10], dtype='float32')
           out = fluid.layers.sequence_concat(input=[x, y])
C
add doc  
chengduoZH 已提交
2719
    """
L
lujun 已提交
2720
    assert not in_dygraph_mode(), (
2721
        "sequence layer is not supported in dygraph mode yet.")
C
add doc  
chengduoZH 已提交
2722
    helper = LayerHelper('sequence_concat', **locals())
X
Xin Pan 已提交
2723
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
C
add doc  
chengduoZH 已提交
2724 2725 2726 2727 2728
    helper.append_op(
        type='sequence_concat', inputs={'X': input}, outputs={'Out': [out]})
    return out


F
fengjiayi 已提交
2729
def sequence_first_step(input):
L
Luo Tao 已提交
2730
    """
L
Luo Tao 已提交
2731
    This function gets the first step of sequence.
L
Luo Tao 已提交
2732 2733 2734 2735

    .. code-block:: text

       x is a 1-level LoDTensor:
2736
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
2737 2738 2739 2740 2741
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
2742
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
2743
         out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1)
F
fengjiayi 已提交
2744

L
Luo Tao 已提交
2745 2746 2747 2748 2749 2750 2751 2752 2753
    Args:
        input(variable): The input variable which is a LoDTensor.

    Returns:
        The sequence's first step variable which is a Tensor.

    Examples:

        .. code-block:: python
F
fengjiayi 已提交
2754

2755
             import paddle.fluid as fluid
Y
yangyaming 已提交
2756
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
2757 2758 2759
                              dtype='float32', lod_level=1)
             x_first_step = fluid.layers.sequence_first_step(input=x)
    """
2760 2761 2762
    return sequence_pool(input=input, pool_type="first")


F
fengjiayi 已提交
2763
def sequence_last_step(input):
L
Luo Tao 已提交
2764
    """
L
Luo Tao 已提交
2765
    This function gets the last step of sequence.
L
Luo Tao 已提交
2766 2767 2768 2769

    .. code-block:: text

       x is a 1-level LoDTensor:
2770
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
2771 2772 2773 2774 2775
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
2776
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
2777
         out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1)
F
fengjiayi 已提交
2778

L
Luo Tao 已提交
2779 2780 2781 2782 2783 2784 2785 2786 2787
    Args:
        input(variable): The input variable which is a LoDTensor.

    Returns:
        The sequence's last step variable which is a Tensor.

    Examples:

        .. code-block:: python
F
fengjiayi 已提交
2788

2789
             import paddle.fluid as fluid
Y
yangyaming 已提交
2790
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
2791 2792 2793
                              dtype='float32', lod_level=1)
             x_last_step = fluid.layers.sequence_last_step(input=x)
    """
2794 2795 2796
    return sequence_pool(input=input, pool_type="last")


Y
Yibing Liu 已提交
2797 2798 2799 2800
def sequence_slice(input, offset, length, name=None):
    """
    **Sequence Slice Layer**

2801
    The layer crops a subsequence from given sequence with given start
Y
Yibing Liu 已提交
2802 2803 2804 2805 2806
    offset and subsequence length.

    It only supports sequence data (LoDTensor with lod_level equal to 1).

    .. code-block:: text
2807

H
haowang101779990 已提交
2808
              - Case:
Y
Yibing Liu 已提交
2809

2810
            Given the input Variable **input**:
2811

2812 2813 2814
                input.data = [[a1, a2], [b1, b2], [c1, c2], [d1, d2], [e1, e2]],
                input.lod = [[3, 2]],
                input.dims = (5, 2),
Y
Yibing Liu 已提交
2815

2816
            with offset.data = [[0], [1]] and length.data = [[2], [1]],
Y
Yibing Liu 已提交
2817

2818
            the output Variable will be
2819

2820 2821 2822
                out.data = [[a1, a2], [b1, b2], [e1, e2]],
                out.lod = [[2, 1]],
                out.dims = (3, 2).
2823

M
minqiyang 已提交
2824
    Note:
H
haowang101779990 已提交
2825
          The first dimension size of **input**, **offset** and **length**
2826
          should be equal. The **offset** should start from 0.
2827

Y
Yibing Liu 已提交
2828
    Args:
2829
        input(Variable): The input Variable which consists of the complete
Y
Yibing Liu 已提交
2830
                         sequences.
Y
Yibing Liu 已提交
2831 2832 2833 2834 2835 2836
        offset(Variable): The offset to slice each sequence.
        length(Variable): The length of each subsequence.
        name(str|None): A name for this layer(optional). If set None, the
                        layer will be named automatically.

    Returns:
Y
Yibing Liu 已提交
2837
        Variable: The output subsequences.
Y
Yibing Liu 已提交
2838 2839 2840 2841 2842

    Examples:

        .. code-block:: python

2843
             import paddle.fluid as fluid
Y
Yibing Liu 已提交
2844 2845 2846 2847 2848
             import numpy as np
             seqs = fluid.layers.data(name='x', shape=[10, 5],
                              dtype='float32', lod_level=1)
             offset = fluid.layers.assign(input=np.array([[0, 1]]).astype("int32"))
             length = fluid.layers.assign(input=np.array([[2, 1]]).astype("int32"))
2849
             subseqs = fluid.layers.sequence_slice(input=seqs, offset=offset,
Y
Yibing Liu 已提交
2850 2851
                                                   length=length)
    """
L
lujun 已提交
2852
    assert not in_dygraph_mode(), (
2853
        "sequence layer is not supported in dygraph mode yet.")
Y
Yibing Liu 已提交
2854 2855
    helper = LayerHelper("sequence_slice", **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
2856
    out = helper.create_variable_for_type_inference(dtype)
Y
Yibing Liu 已提交
2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870

    offset.stop_gradient = True
    length.stop_gradient = True

    helper.append_op(
        type="sequence_slice",
        inputs={"X": input,
                "Offset": offset,
                "Length": length},
        outputs={"Out": out})

    return out


F
fengjiayi 已提交
2871
@templatedoc()
Y
Yu Yang 已提交
2872
def pool2d(input,
C
chengduoZH 已提交
2873 2874
           pool_size=-1,
           pool_type="max",
C
chengduoZH 已提交
2875 2876
           pool_stride=1,
           pool_padding=0,
C
caoying03 已提交
2877
           global_pooling=False,
C
chengduoZH 已提交
2878
           use_cudnn=True,
2879
           ceil_mode=False,
2880 2881
           name=None,
           exclusive=True):
Y
Yu Yang 已提交
2882
    """
F
fengjiayi 已提交
2883
    ${comment}
2884 2885

    Args:
2886 2887 2888
        input (Variable): The input tensor of pooling operator. The format of
                          input tensor is NCHW, where N is batch size, C is
                          the number of channels, H is the height of the
F
fengjiayi 已提交
2889
                          feature, and W is the width of the feature.
J
JiabinYang 已提交
2890
        pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list,
J
JiabinYang 已提交
2891 2892
            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
2893
        pool_type: ${pooling_type_comment}
J
JiabinYang 已提交
2894 2895 2896 2897 2898 2899
        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.
        pool_padding (int|list|tuple): The pool padding size. If pool padding size is a tuple,
            it must contain two integers, (pool_padding_on_Height, pool_padding_on_Width).
            Otherwise, the pool padding size will be a square of an int.
2900 2901 2902
        global_pooling (bool): ${global_pooling_comment}
        use_cudnn (bool): ${use_cudnn_comment}
        ceil_mode (bool): ${ceil_mode_comment}
2903
        name (str|None): A name for this layer(optional). If set None, the
F
fengjiayi 已提交
2904
                        layer will be named automatically.
2905
        exclusive (bool): Whether to exclude padding points in average pooling
2906
                          mode, default is true
F
fengjiayi 已提交
2907

2908
    Returns:
F
fengjiayi 已提交
2909
        Variable: The pooling result.
F
fengjiayi 已提交
2910 2911 2912 2913 2914 2915 2916 2917 2918 2919

    Raises:
        ValueError: If 'pool_type' is not "max" nor "avg"
        ValueError: If 'global_pooling' is False and 'pool_size' is -1
        ValueError: If 'use_cudnn' is not a bool value.

    Examples:

        .. code-block:: python

2920
          import paddle.fluid as fluid
F
fengjiayi 已提交
2921 2922
          data = fluid.layers.data(
              name='data', shape=[3, 32, 32], dtype='float32')
D
dengkaipeng 已提交
2923
          pool2d = fluid.layers.pool2d(
2924 2925 2926 2927
                            input=data,
                            pool_size=2,
                            pool_type='max',
                            pool_stride=1,
F
fengjiayi 已提交
2928
                            global_pooling=False)
Y
Yu Yang 已提交
2929 2930 2931 2932 2933
    """
    if pool_type not in ["max", "avg"]:
        raise ValueError(
            "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",
            str(pool_type))
C
chengduoZH 已提交
2934

C
chengduoZH 已提交
2935 2936 2937 2938 2939
    if global_pooling is False and pool_size == -1:
        raise ValueError(
            "When the global_pooling is False, pool_size must be passed "
            "and be a valid value. Received pool_size: " + str(pool_size))

C
chengduoZH 已提交
2940 2941 2942 2943
    pool_size = utils.convert_to_list(pool_size, 2, 'pool_size')
    pool_padding = utils.convert_to_list(pool_padding, 2, 'pool_padding')
    pool_stride = utils.convert_to_list(pool_stride, 2, 'pool_stride')

C
chengduoZH 已提交
2944 2945
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
2946

C
Add doc  
chengduoZH 已提交
2947
    l_type = 'pool2d'
2948 2949

    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
2950
    dtype = helper.input_dtype()
X
Xin Pan 已提交
2951
    pool_out = helper.create_variable_for_type_inference(dtype)
Y
Yu Yang 已提交
2952 2953

    helper.append_op(
2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964
        type=l_type,
        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,
            "use_cudnn": use_cudnn,
            "ceil_mode": ceil_mode,
2965 2966
            "use_mkldnn": False,
            "exclusive": exclusive,
2967 2968 2969 2970 2971
        })

    return pool_out


D
dengkaipeng 已提交
2972
@templatedoc()
2973 2974 2975 2976 2977 2978 2979 2980
def pool3d(input,
           pool_size=-1,
           pool_type="max",
           pool_stride=1,
           pool_padding=0,
           global_pooling=False,
           use_cudnn=True,
           ceil_mode=False,
2981 2982
           name=None,
           exclusive=True):
2983
    """
2984
    ${comment}
2985 2986

    Args:
D
dengkaipeng 已提交
2987 2988 2989 2990 2991
        input (Variable): The input tensor of pooling operator. 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,
                          H is the height of the feature, and W is the width
                          of the feature.
D
dengkaipeng 已提交
2992 2993 2994 2995 2996
        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}
2997 2998 2999 3000 3001 3002 3003
        pool_stride (int): stride of the pooling layer.
        pool_padding (int): padding size.
        global_pooling (bool): ${global_pooling_comment}
        use_cudnn (bool): ${use_cudnn_comment}
        ceil_mode (bool): ${ceil_mode_comment}
        name (str): A name for this layer(optional). If set None, the layer
            will be named automatically.
3004
        exclusive (bool): Whether to exclude padding points in average pooling
3005
                          mode, default is true
3006

3007
    Returns:
3008
        Variable: output of pool3d layer.
D
dengkaipeng 已提交
3009 3010 3011 3012 3013

    Examples:

        .. code-block:: python

3014
          import paddle.fluid as fluid
D
dengkaipeng 已提交
3015 3016 3017 3018 3019 3020 3021 3022
          data = fluid.layers.data(
              name='data', shape=[3, 32, 32, 32], dtype='float32')
          pool3d = fluid.layers.pool3d(
                            input=data,
                            pool_size=2,
                            pool_type='max',
                            pool_stride=1,
                            global_pooling=False)
Y
Yu Yang 已提交
3023 3024 3025 3026 3027
    """
    if pool_type not in ["max", "avg"]:
        raise ValueError(
            "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",
            str(pool_type))
C
chengduoZH 已提交
3028

C
chengduoZH 已提交
3029 3030 3031 3032 3033
    if global_pooling is False and pool_size == -1:
        raise ValueError(
            "When the global_pooling is False, pool_size must be passed "
            "and be a valid value. Received pool_size: " + str(pool_size))

3034 3035 3036
    pool_size = utils.convert_to_list(pool_size, 3, 'pool_size')
    pool_padding = utils.convert_to_list(pool_padding, 3, 'pool_padding')
    pool_stride = utils.convert_to_list(pool_stride, 3, 'pool_stride')
C
chengduoZH 已提交
3037

C
chengduoZH 已提交
3038 3039
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
3040

3041 3042
    l_type = "pool3d"
    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
3043
    dtype = helper.input_dtype()
X
Xin Pan 已提交
3044
    pool_out = helper.create_variable_for_type_inference(dtype)
Y
Yu Yang 已提交
3045 3046

    helper.append_op(
3047
        type=l_type,
Y
Yu Yang 已提交
3048 3049 3050 3051 3052 3053 3054
        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
3055
            "paddings": pool_padding,
3056
            "use_cudnn": use_cudnn,
3057
            "ceil_mode": ceil_mode,
3058 3059
            "use_mkldnn": False,
            "exclusive": exclusive,
Y
Yu Yang 已提交
3060 3061 3062 3063 3064
        })

    return pool_out


3065 3066 3067 3068 3069 3070 3071
@templatedoc(op_type="pool2d")
def adaptive_pool2d(input,
                    pool_size,
                    pool_type="max",
                    require_index=False,
                    name=None):
    """
D
dengkaipeng 已提交
3072 3073 3074 3075 3076 3077 3078
    **Adaptive Pool2d Operator**
    The adaptive_pool2d operation calculates the output based on the input, pool_size,
    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)
    is same as Parameter(pool_size).
3079

3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092
    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)}
3093 3094 3095 3096 3097 3098 3099 3100 3101

    Args:
        input (Variable): The input tensor of pooling operator. 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.
        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
3102 3103
        require_index (bool): If true, the index of max pooling point will be returned along
            with outputs. It cannot be set in average pooling type.
3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117
        name (str|None): A name for this layer(optional). If set None, the
                        layer will be named automatically.

    Returns:
        Variable: The pooling result.

    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

M
minqiyang 已提交
3118
          # suppose input data in shape of [N, C, H, W], `pool_size` is [m, n],
3119
          # output shape is [N, C, m, n], adaptive pool divide H and W dimentions
M
minqiyang 已提交
3120
          # of input data into m * n grids averagely and performs poolings in each
3121 3122
          # grid to get output.
          # adaptive average pool performs calculations as follow:
M
minqiyang 已提交
3123
          #
3124 3125 3126 3127 3128 3129 3130 3131
          #     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])
          #
3132
          import paddle.fluid as fluid
3133 3134
          data = fluid.layers.data(
              name='data', shape=[3, 32, 32], dtype='float32')
D
dengkaipeng 已提交
3135
          pool_out = fluid.layers.adaptive_pool2d(
3136 3137
                            input=data,
                            pool_size=[3, 3],
3138
                            pool_type='avg')
3139 3140 3141 3142 3143 3144 3145 3146 3147 3148
    """
    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'.")

3149
    pool_size = utils.convert_to_list(pool_size, 2, 'pool_size')
3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174

    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
3175
    return (pool_out, mask) if require_index else pool_out
3176 3177 3178 3179 3180 3181 3182 3183 3184


@templatedoc(op_type="pool3d")
def adaptive_pool3d(input,
                    pool_size,
                    pool_type="max",
                    require_index=False,
                    name=None):
    """
D
dengkaipeng 已提交
3185 3186 3187 3188 3189 3190 3191
    **Adaptive Pool3d Operator**
    The adaptive_pool3d operation calculates the output based on the input, pool_size,
    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
    dimensions of output(Out) is same as Parameter(pool_size).
3192

3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209
    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)}
3210 3211 3212

    Args:
        input (Variable): The input tensor of pooling operator. The format of
D
dengkaipeng 已提交
3213 3214 3215
                          input tensor is NCDHW, 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.
3216
        pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list,
D
dengkaipeng 已提交
3217
            it must contain three integers, (Depth, Height, Width).
3218
        pool_type: ${pooling_type_comment}
D
dengkaipeng 已提交
3219 3220
        require_index (bool): If true, the index of max pooling point will be returned along
            with outputs. It cannot be set in average pooling type.
3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234
        name (str|None): A name for this layer(optional). If set None, the
                        layer will be named automatically.

    Returns:
        Variable: The pooling result.

    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

3235 3236
          # 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
3237
          # of input data into l * m * n grids averagely and performs poolings in each
3238 3239
          # grid to get output.
          # adaptive average pool performs calculations as follow:
M
minqiyang 已提交
3240
          #
3241 3242 3243 3244 3245 3246 3247 3248 3249
          #     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
3250
          #                 output[:, :, i, j, k] =
3251 3252
          #                     avg(input[:, :, dstart:dend, hstart: hend, wstart: wend])
          #
K
Kaipeng Deng 已提交
3253 3254 3255

          import paddle.fluid as fluid

3256
          data = fluid.layers.data(
K
Kaipeng Deng 已提交
3257 3258
              name='data', shape=[3, 32, 32, 32], dtype='float32')
          pool_out = fluid.layers.adaptive_pool3d(
3259
                            input=data,
D
dengkaipeng 已提交
3260
                            pool_size=[3, 3, 3],
3261
                            pool_type='avg')
3262 3263 3264 3265 3266 3267 3268 3269 3270 3271
    """
    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'.")

3272
    pool_size = utils.convert_to_list(pool_size, 3, 'pool_size')
3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297

    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
3298
    return (pool_out, mask) if require_index else pool_out
3299 3300


Y
Yu Yang 已提交
3301 3302 3303 3304 3305 3306 3307
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
3308
               data_layout='NCHW',
Y
Yang Yang 已提交
3309
               in_place=False,
3310 3311
               name=None,
               moving_mean_name=None,
W
wanghaoshuang 已提交
3312
               moving_variance_name=None,
3313
               do_model_average_for_mean_and_var=False,
3314 3315
               fuse_with_relu=False,
               use_global_stats=False):
Y
Yu Yang 已提交
3316
    """
Q
qiaolongfei 已提交
3317 3318 3319 3320
    **Batch Normalization Layer**

    Can be used as a normalizer function for conv2d and fully_connected operations.
    The required data format for this layer is one of the following:
Q
qiaolongfei 已提交
3321

Q
qiaolongfei 已提交
3322
    1. NHWC `[batch, in_height, in_width, in_channels]`
Q
qiaolongfei 已提交
3323

Q
qiaolongfei 已提交
3324 3325
    2. NCHW `[batch, in_channels, in_height, in_width]`

Q
qiaolongfei 已提交
3326 3327 3328
    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
3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340

    :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
3341

3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354

    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
3355 3356 3357 3358
    Note:
        if build_strategy.sync_batch_norm=True, the batch_norm in network will use 
        sync_batch_norm automatically.

3359
    Args:
Q
qingqing01 已提交
3360
        input(variable): The rank of input variable can be 2, 3, 4, 5.
Q
qiaolongfei 已提交
3361
        act(string, Default None): Activation type, linear|relu|prelu|...
Q
qingqing01 已提交
3362 3363 3364 3365 3366 3367 3368 3369 3370
        is_test (bool, Default False): A flag indicating whether it is in
            test phrase or not.
        momentum(float, Default 0.9): The value used for the moving_mean and
            moving_var computation. The updated formula is:
            :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
3371 3372
        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
3373 3374 3375
	     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
3376 3377
        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
3378 3379 3380
	     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.
Q
qiaolongfei 已提交
3381
        data_layout(string, default NCHW): NCHW|NHWC
3382
        in_place(bool, Default False): Make the input and output of batch norm reuse memory.
Q
qiaolongfei 已提交
3383 3384
        name(string, Default None): A name for this layer(optional). If set None, the layer
            will be named automatically.
3385 3386 3387
        moving_mean_name(string, Default None): The name of moving_mean which store the global Mean. If it 
            is set to None, batch_norm will save global mean with a random name, otherwise, batch_norm 
            will save global mean with the string.
Q
qiaolongfei 已提交
3388
        moving_variance_name(string, Default None): The name of the moving_variance which store the global Variance.
3389 3390
            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.
Q
qiaolongfei 已提交
3391
        do_model_average_for_mean_and_var(bool, Default False): Do model average for mean and variance or not.
3392
        fuse_with_relu (bool): if True, this OP performs relu after batch norm.
3393 3394 3395 3396 3397
        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.
3398 3399

    Returns:
Q
qiaolongfei 已提交
3400
        Variable: A tensor variable which is the result after applying batch normalization on the input.
Q
qiaolongfei 已提交
3401 3402 3403 3404 3405

    Examples:

        .. code-block:: python

3406
            import paddle.fluid as fluid
L
lvmengsi 已提交
3407
            x = fluid.layers.data(name='x', shape=[3, 7, 3, 7], dtype='float32', append_batch_size=False)
Q
qiaolongfei 已提交
3408 3409
            hidden1 = fluid.layers.fc(input=x, size=200, param_attr='fc1.w')
            hidden2 = fluid.layers.batch_norm(input=hidden1)
Y
Yu Yang 已提交
3410
    """
C
chengduo 已提交
3411
    assert bias_attr is not False, "bias_attr should not be False in batch_norm."
Y
Yu Yang 已提交
3412 3413 3414
    helper = LayerHelper('batch_norm', **locals())
    dtype = helper.input_dtype()

W
Wu Yi 已提交
3415 3416 3417 3418
    # use fp32 for bn parameter
    if dtype == core.VarDesc.VarType.FP16:
        dtype = core.VarDesc.VarType.FP32

Y
Yu Yang 已提交
3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436
    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(
3437
        attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
Y
Yu Yang 已提交
3438

3439 3440
    mean = helper.create_parameter(
        attr=ParamAttr(
W
wanghaoshuang 已提交
3441 3442 3443
            name=moving_mean_name,
            initializer=Constant(0.0),
            trainable=False,
W
wanghaoshuang 已提交
3444
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
3445
        shape=param_shape,
W
Wu Yi 已提交
3446
        dtype=dtype)
3447 3448 3449 3450 3451 3452
    mean.stop_gradient = True

    variance = helper.create_parameter(
        attr=ParamAttr(
            name=moving_variance_name,
            initializer=Constant(1.0),
W
wanghaoshuang 已提交
3453
            trainable=False,
W
wanghaoshuang 已提交
3454
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
3455
        shape=param_shape,
W
Wu Yi 已提交
3456
        dtype=dtype)
3457
    variance.stop_gradient = True
Y
Yu Yang 已提交
3458 3459 3460 3461 3462 3463

    # 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
3464 3465 3466 3467
    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
3468

X
Xin Pan 已提交
3469 3470
    batch_norm_out = input if in_place else helper.create_variable_for_type_inference(
        dtype)
Y
Yu Yang 已提交
3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487

    helper.append_op(
        type="batch_norm",
        inputs={
            "X": input,
            "Scale": scale,
            "Bias": bias,
            "Mean": mean,
            "Variance": variance
        },
        outputs={
            "Y": batch_norm_out,
            "MeanOut": mean_out,
            "VarianceOut": variance_out,
            "SavedMean": saved_mean,
            "SavedVariance": saved_variance
        },
3488 3489 3490 3491
        attrs={
            "momentum": momentum,
            "epsilon": epsilon,
            "is_test": is_test,
3492
            "data_layout": data_layout,
X
Xin Pan 已提交
3493
            "use_mkldnn": False,
3494 3495
            "fuse_with_relu": fuse_with_relu,
            "use_global_stats": use_global_stats
3496
        })
Y
Yu Yang 已提交
3497 3498 3499 3500

    return helper.append_activation(batch_norm_out)


H
heqiaozhi 已提交
3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551
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,
              do_model_average_for_mean_and_var=False):
    """
    **Data Normalization Layer**

    Can be used as a normalizer function for conv2d and fully_connected operations.
    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`.
        data_layout(string, default NCHW): NCHW|NHWC
        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.
        do_model_average_for_mean_and_var(bool, Default False): Do model average for mean and variance or not.

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

    Examples:

        .. code-block:: python
3552 3553
            
            import paddle.fluid as fluid
H
heqiaozhi 已提交
3554

3555 3556
            hidden1 = fluid.layers.data(name="hidden1", shape=[200])
            hidden2 = fluid.layers.data_norm(name="hidden2", input=hidden1)
H
heqiaozhi 已提交
3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621
    """
    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
        },
        outputs={"Y": data_norm_out,
                 "Means": means,
                 "Scales": scales},
H
heqiaozhi 已提交
3622
        attrs={"epsilon": epsilon})
H
heqiaozhi 已提交
3623 3624 3625 3626

    return helper.append_activation(data_norm_out)


Y
yuyang18 已提交
3627
@templatedoc()
G
guosheng 已提交
3628 3629 3630 3631 3632 3633 3634 3635 3636 3637
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):
    """
Y
yuyang18 已提交
3638
    ${comment}
G
guosheng 已提交
3639 3640 3641

    The formula is as follows:

Y
yuyang18 已提交
3642
    ..  math::
G
guosheng 已提交
3643 3644 3645 3646 3647 3648 3649

        \\mu & = \\frac{1}{H}\\sum_{i=1}^{H} a_i

        \\sigma & = \\sqrt{\\frac{1}{H}\sum_{i=1}^{H}(a_i - \\mu)^2}

        h & = f(\\frac{g}{\\sigma}(a - \\mu) + b)

Y
yuyang18 已提交
3650 3651 3652 3653 3654 3655 3656 3657
    * :math:`a`: the vector representation of the summed inputs to the neurons
    in that layer.

    * :math:`H`: the number of hidden units in a layers

    * :math:`g`: the trainable scale parameter.

    * :math:`b`: the trainable bias parameter.
Y
yuyang18 已提交
3658

G
guosheng 已提交
3659 3660
    Args:
        input(Variable): The input tensor variable.
3661
        scale(bool): Whether to learn the adaptive gain :math:`g` after
S
sneaxiy 已提交
3662
            normalization. Default True.
3663
        shift(bool): Whether to learn the adaptive bias :math:`b` after
S
sneaxiy 已提交
3664 3665
            normalization. Default True.
        begin_norm_axis(int): The normalization will be performed along
G
guosheng 已提交
3666
            dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`.
S
sneaxiy 已提交
3667
            Default 1.
3668
        epsilon(float): The small value added to the variance to prevent
S
sneaxiy 已提交
3669
            division by zero. Default 1e-05.
G
guosheng 已提交
3670
        param_attr(ParamAttr|None): The parameter attribute for the learnable
S
sneaxiy 已提交
3671 3672
            gain :math:`g`. If :attr:`scale` is False, :attr:`param_attr` is
            omitted. If :attr:`scale` is True and :attr:`param_attr` is None,
3673 3674
            a default :code:`ParamAttr` would be added as scale. The
            :attr:`param_attr` is initialized as 1 if it is added. Default None.
G
guosheng 已提交
3675
        bias_attr(ParamAttr|None): The parameter attribute for the learnable
S
sneaxiy 已提交
3676 3677
            bias :math:`b`. If :attr:`shift` is False, :attr:`bias_attr` is
            omitted. If :attr:`shift` is True and :attr:`param_attr` is None,
3678
            a default :code:`ParamAttr` would be added as bias. The
S
sneaxiy 已提交
3679
            :attr:`bias_attr` is initialized as 0 if it is added. Default None.
G
guosheng 已提交
3680
        act(str): Activation to be applied to the output of layer normalizaiton.
S
sneaxiy 已提交
3681 3682 3683
                  Default None.
        name(str): The name of this layer. It is optional. Default None, and a
                   unique name would be generated automatically.
G
guosheng 已提交
3684 3685

    Returns:
Y
yuyang18 已提交
3686
        ${y_comment}
G
guosheng 已提交
3687 3688 3689

    Examples:

3690
        >>> import paddle.fluid as fluid
Y
yuyang18 已提交
3691 3692 3693
        >>> data = fluid.layers.data(name='data', shape=[3, 32, 32],
        >>>                          dtype='float32')
        >>> x = fluid.layers.layer_norm(input=data, begin_norm_axis=1)
G
guosheng 已提交
3694
    """
L
lujun 已提交
3695
    assert in_dygraph_mode(
L
lujun 已提交
3696
    ) is not True, "please use FC instead of fc in dygraph mode!"
G
guosheng 已提交
3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710
    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:
        scale = helper.create_parameter(
            attr=helper.param_attr,
            shape=param_shape,
            dtype=dtype,
            default_initializer=Constant(1.0))
        inputs['Scale'] = scale
G
guosheng 已提交
3711
    if shift:
G
guosheng 已提交
3712 3713 3714 3715 3716 3717
        assert bias_attr is not False
        bias = helper.create_parameter(
            attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
        inputs['Bias'] = bias

    # create output
X
Xin Pan 已提交
3718 3719 3720 3721 3722
    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
3723 3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737

    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
3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748 3749
@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
3750
    Refer to `Group Normalization <https://arxiv.org/abs/1803.08494>`_ .
D
Dun 已提交
3751 3752 3753 3754 3755 3756 3757 3758 3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771

    Args:
        input(Variable): The input tensor variable.
        groups(int): The number of groups that divided from channels.
        epsilon(float): The small value added to the variance to prevent
            division by zero.
        param_attr(ParamAttr|None): The parameter attribute for the learnable
            scale :math:`g`. If it is set to False, no scale will be added to the output units.
            If it is set to None, the bias is initialized one. Default: None.
        bias_attr(ParamAttr|None): The parameter attribute for the learnable
            bias :math:`b`. If it is set to False, no bias will be added to the output units.
            If it is set to None, the bias is initialized zero. Default: None.
        act(str): Activation to be applied to the output of group normalizaiton.
        data_layout(string|NCHW): Only NCHW is supported.
        name (str): The name of this layer. It is optional.

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

    Examples:

3772
        >>> import paddle.fluid as fluid
D
Dun 已提交
3773 3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798
        >>> data = fluid.layers.data(name='data', shape=[8, 32, 32],
        >>>                          dtype='float32')
        >>> x = fluid.layers.group_norm(input=data, groups=4)
    """
    helper = LayerHelper('group_norm', **locals())
    dtype = helper.input_dtype()

    # create intput and parameters
    inputs = {'X': input}
    input_shape = input.shape
    if data_layout != 'NCHW':
        raise ValueError("unsupported data layout:" + data_layout)
    param_shape = [input_shape[1]]
    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
3799 3800
    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
3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817
    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,
        },
        attrs={"epsilon": epsilon,
               "groups": groups})

    return helper.append_activation(group_norm_out)


@templatedoc()
3818
def spectral_norm(weight, dim=0, power_iters=1, eps=1e-12, name=None):
D
dengkaipeng 已提交
3819 3820 3821
    """
    **Spectral Normalization Layer**

D
dengkaipeng 已提交
3822
    This layer calculates the spectral normalization value of weight parameters of
3823
    fc, conv1d, conv2d, conv3d layers which should be 2-D, 3-D, 4-D, 5-D
D
dengkaipeng 已提交
3824
    Parameters. Calculations are showed as follows.
3825

D
dengkaipeng 已提交
3826 3827 3828
    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
3829
    and W is the product result of remaining dimensions.
D
dengkaipeng 已提交
3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841

    Step 2:
    :attr:`power_iters` shoule be a positive interger, do following
    calculations with U and V for :attr:`power_iters` rounds.

    .. 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
3842
    Calculate :math:`\sigma(\mathbf{W})` and normalize weight values.
D
dengkaipeng 已提交
3843 3844 3845 3846

    .. math::

        \sigma(\mathbf{W}) = \mathbf{u}^{T} \mathbf{W} \mathbf{v}
3847

D
dengkaipeng 已提交
3848
        \mathbf{W} = \\frac{\mathbf{W}}{\sigma(\mathbf{W})}
3849 3850
                

D
dengkaipeng 已提交
3851 3852 3853 3854
    Refer to `Spectral Normalization <https://arxiv.org/abs/1802.05957>`_ .

    Args:
        weight(${weight_type}): ${weight_comment}
D
dengkaipeng 已提交
3855 3856 3857
        dim(int): ${dim_comment}
        power_iters(int): ${power_iters_comment}
        eps(float): ${eps_comment}
D
dengkaipeng 已提交
3858 3859 3860
        name (str): The name of this layer. It is optional.

    Returns:
D
dengkaipeng 已提交
3861
        Variable: A tensor variable of weight parameters after spectral normalization.
D
dengkaipeng 已提交
3862 3863

    Examples:
K
Kaipeng Deng 已提交
3864
       .. code-block:: python
D
dengkaipeng 已提交
3865

K
Kaipeng Deng 已提交
3866 3867 3868 3869 3870
            import paddle.fluid as fluid

            weight = fluid.layers.data(name='weight', shape=[2, 8, 32, 32], 
                                       append_batch_size=False, dtype='float32')
            x = fluid.layers.spectral_norm(weight=weight, dim=1, power_iters=2)
D
dengkaipeng 已提交
3871 3872
    """
    helper = LayerHelper('spectral_norm', **locals())
3873
    dtype = weight.dtype
D
dengkaipeng 已提交
3874 3875 3876

    # create intput and parameters
    inputs = {'Weight': weight}
3877 3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892 3893 3894
    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
3895 3896

    # create output
3897
    out = helper.create_variable(dtype=dtype)
D
Dun 已提交
3898 3899

    helper.append_op(
3900
        type="spectral_norm",
D
Dun 已提交
3901
        inputs=inputs,
3902 3903 3904 3905 3906 3907
        outputs={"Out": out, },
        attrs={
            "dim": dim,
            "power_iters": power_iters,
            "eps": eps,
        })
D
Dun 已提交
3908

3909
    return out
D
Dun 已提交
3910 3911


Y
Yu Yang 已提交
3912 3913 3914 3915
def conv2d_transpose(input,
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
3916 3917 3918
                     padding=0,
                     stride=1,
                     dilation=1,
3919
                     groups=None,
C
caoying03 已提交
3920
                     param_attr=None,
3921
                     bias_attr=None,
C
chengduoZH 已提交
3922
                     use_cudnn=True,
3923
                     act=None,
C
caoying03 已提交
3924
                     name=None):
Y
Yu Yang 已提交
3925
    """
3926 3927 3928 3929 3930 3931 3932 3933
    **Convlution2D transpose layer**

    The convolution2D transpose layer calculates the output based on the input,
    filter, and dilations, strides, paddings. Input(Input) and output(Output)
    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(dilations, strides, paddings) are two elements. These two elements
    represent height and width, respectively. The details of convolution transpose
3934
    layer, please refer to the following explanation and references
L
lvmengsi 已提交
3935
    `therein <https://arxiv.org/pdf/1603.07285.pdf>`_.
3936 3937 3938
    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.
3939 3940 3941 3942 3943

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

    .. math::

3944
        Out = \sigma (W \\ast X + b)
3945

3946
    Where:
3947 3948 3949

    * :math:`X`: Input value, a tensor with NCHW format.
    * :math:`W`: Filter value, a tensor with MCHW format.
3950 3951 3952 3953
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :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
3954

3955 3956 3957 3958
    Example:

        - Input:

3959
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
3960

3961
          Filter shape: :math:`(C_{in}, C_{out}, H_f, W_f)`
3962 3963 3964

        - Output:

3965
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
3966 3967

        Where
Y
Yu Yang 已提交
3968

3969 3970
        .. math::

3971 3972
           H^\prime_{out} &= (H_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (H_f - 1) + 1 \\\\
           W^\prime_{out} &= (W_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (W_f - 1) + 1 \\\\
L
lvmengsi 已提交
3973 3974 3975 3976 3977 3978 3979 3980 3981
           H_{out} &\in [ H^\prime_{out}, H^\prime_{out} + strides[0] ] \\\\
           W_{out} &\in [ W^\prime_{out}, W^\prime_{out} + strides[1] ] 

    Note:
          if output_size is None, :math:`H_{out} = H^\prime_{out}, W_{out} = W^\prime_{out}`; 
          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
3982 3983

    Args:
3984 3985 3986 3987
        input(Variable): The input image with [N, C, H, W] format.
        num_filters(int): The number of the filter. It is as same as the output
            image channel.
        output_size(int|tuple|None): The output image size. If output size is a
3988 3989 3990 3991
            tuple, it must contain two integers, (image_H, image_W). None if use
            filter_size, padding, and stride to calculate output_size.
            if output_size and filter_size are specified at the same time, They
            should follow the formula above.
3992 3993 3994 3995 3996 3997 3998 3999 4000 4001 4002 4003 4004 4005 4006 4007 4008 4009
        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
            it must contain two integers, (filter_size_H, filter_size_W).
            Otherwise, the filter will be a square. None if use output size to
            calculate filter_size.
        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.
        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.
        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 Conv2d transpose layer. Inspired by
            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
4010 4011 4012 4013 4014 4015 4016 4017 4018 4019
            Default: groups = 1.
        param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
            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.
        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv2d_transpose.
            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.
4020
        use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
C
chengduo 已提交
4021 4022 4023
            library is installed. Default: True.
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None.
4024
        name(str|None): A name for this layer(optional). If set None, the layer
C
chengduo 已提交
4025
            will be named automatically. Default: True.
Y
Yu Yang 已提交
4026 4027

    Returns:
4028
        Variable: The tensor variable storing the convolution transpose result.
4029 4030

    Raises:
4031 4032
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
4033 4034 4035 4036

    Examples:
       .. code-block:: python

4037
          import paddle.fluid as fluid
4038 4039
          data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32')
          conv2d_transpose = fluid.layers.conv2d_transpose(input=data, num_filters=2, filter_size=3)
Y
Yu Yang 已提交
4040
    """
C
chengduo 已提交
4041
    assert param_attr is not False, "param_attr should not be False in conv2d_transpose."
4042 4043 4044 4045 4046 4047 4048 4049
    input_channel = input.shape[1]

    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
4050 4051 4052
    if not isinstance(input, Variable):
        raise TypeError("Input of conv2d_transpose must be Variable")

C
chengduoZH 已提交
4053 4054 4055
    padding = utils.convert_to_list(padding, 2, 'padding')
    stride = utils.convert_to_list(stride, 2, 'stride')
    dilation = utils.convert_to_list(dilation, 2, 'dilation')
G
guosheng 已提交
4056

C
chengduoZH 已提交
4057 4058
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
G
guosheng 已提交
4059

Y
Yu Yang 已提交
4060 4061 4062 4063 4064
    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
4065

Y
Yu Yang 已提交
4066 4067
        h_in = input.shape[2]
        w_in = input.shape[3]
G
guosheng 已提交
4068

C
chengduoZH 已提交
4069
        filter_size_h = (output_size[0] - (h_in - 1) * stride[0] + 2 *
M
minqiyang 已提交
4070
                         padding[0] - 1) // dilation[0] + 1
C
chengduoZH 已提交
4071
        filter_size_w = (output_size[1] - (w_in - 1) * stride[1] + 2 *
M
minqiyang 已提交
4072
                         padding[1] - 1) // dilation[1] + 1
Y
Yu Yang 已提交
4073
        filter_size = [filter_size_h, filter_size_w]
C
chengduoZH 已提交
4074 4075 4076
    else:
        filter_size = utils.convert_to_list(filter_size, 2,
                                            'conv2d_transpose.filter_size')
C
chengduo 已提交
4077

4078 4079 4080 4081 4082 4083 4084
    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")
    padding = utils.convert_to_list(padding, 2, 'padding')
4085
    groups = 1 if groups is None else groups
M
minqiyang 已提交
4086
    filter_shape = [input_channel, num_filters // groups] + filter_size
C
chengduo 已提交
4087

Y
Yu Yang 已提交
4088 4089 4090
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

X
Xin Pan 已提交
4091
    pre_bias = helper.create_variable_for_type_inference(dtype=input.dtype)
Y
Yu Yang 已提交
4092
    helper.append_op(
4093
        type=op_type,
Y
Yu Yang 已提交
4094 4095
        inputs={'Input': [input],
                'Filter': [img_filter]},
4096
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
4097
        attrs={
4098
            'output_size': output_size,
4099 4100 4101 4102 4103
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn
Y
Yu Yang 已提交
4104 4105
        })

4106 4107 4108
    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
    out = helper.append_activation(pre_act)
    return out
Y
Yu Yang 已提交
4109 4110


4111
def conv3d_transpose(input,
Y
Yu Yang 已提交
4112 4113 4114
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
4115 4116 4117
                     padding=0,
                     stride=1,
                     dilation=1,
4118
                     groups=None,
C
caoying03 已提交
4119
                     param_attr=None,
4120
                     bias_attr=None,
C
chengduoZH 已提交
4121
                     use_cudnn=True,
4122
                     act=None,
C
caoying03 已提交
4123
                     name=None):
Y
Yu Yang 已提交
4124
    """
4125
    **Convlution3D transpose layer**
4126

4127
    The convolution3D transpose layer calculates the output based on the input,
4128
    filter, and dilations, strides, paddings. Input(Input) and output(Output)
4129 4130 4131 4132 4133
    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(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
4134
    explanation and references `therein <https://arxiv.org/pdf/1603.07285.pdf>`_.
4135 4136 4137
    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.
4138 4139 4140 4141 4142

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

    .. math::

4143
        Out = \sigma (W \\ast X + b)
4144 4145 4146

    In the above equation:

4147 4148
    * :math:`X`: Input value, a tensor with NCDHW format.
    * :math:`W`: Filter value, a tensor with MCDHW format.
4149 4150 4151 4152
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :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
4153

4154 4155 4156 4157
    Example:

        - Input:

4158
          Input shape: :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})`
4159

4160
          Filter shape: :math:`(C_{in}, C_{out}, D_f, H_f, W_f)`
4161 4162 4163

        - Output:

4164
          Output shape: :math:`(N, C_{out}, D_{out}, H_{out}, W_{out})`
4165 4166

        Where
Y
Yu Yang 已提交
4167

4168 4169
        .. math::

4170 4171 4172
           D_{out} &= (D_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (D_f - 1) + 1 \\\\
           H_{out} &= (H_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (H_f - 1) + 1 \\\\
           W_{out} &= (W_{in} - 1) * strides[2] - 2 * paddings[2] + dilations[2] * (W_f - 1) + 1
Y
Yu Yang 已提交
4173 4174

    Args:
4175
        input(Variable): The input image with [N, C, D, H, W] format.
4176 4177 4178
        num_filters(int): The number of the filter. It is as same as the output
            image channel.
        output_size(int|tuple|None): The output image size. If output size is a
4179
            tuple, it must contain three integers, (image_D, image_H, image_W). This
4180 4181
            parameter only works when filter_size is None.
        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
4182
            it must contain three integers, (filter_size_D, filter_size_H, filter_size_W).
4183 4184 4185
            Otherwise, the filter will be a square. None if use output size to
            calculate filter_size.
        padding(int|tuple): The padding size. If padding is a tuple, it must
4186 4187
            contain three integers, (padding_D, padding_H, padding_W). Otherwise, the
            padding_D = padding_H = padding_W = padding. Default: padding = 0.
4188
        stride(int|tuple): The stride size. If stride is a tuple, it must
4189 4190
            contain three integers, (stride_D, stride_H, stride_W). Otherwise, the
            stride_D = stride_H = stride_W = stride. Default: stride = 1.
4191
        dilation(int|tuple): The dilation size. If dilation is a tuple, it must
4192 4193 4194
            contain three integers, (dilation_D, dilation_H, dilation_W). Otherwise, the
            dilation_D = dilation_H = dilation_W = dilation. Default: dilation = 1.
        groups(int): The groups number of the Conv3d transpose layer. Inspired by
4195 4196 4197 4198 4199
            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
C
chengduo 已提交
4200 4201 4202 4203 4204 4205 4206 4207 4208
        param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
            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.
        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv3d_transpose.
            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.
4209 4210
        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
4211 4212
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None.
4213 4214
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.
Y
Yu Yang 已提交
4215 4216

    Returns:
4217
        Variable: The tensor variable storing the convolution transpose result.
4218 4219

    Raises:
4220 4221
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
4222 4223 4224 4225

    Examples:
       .. code-block:: python

4226
          import paddle.fluid as fluid
4227 4228
          data = fluid.layers.data(name='data', shape=[3, 12, 32, 32], dtype='float32')
          conv3d_transpose = fluid.layers.conv3d_transpose(input=data, num_filters=2, filter_size=3)
Y
Yu Yang 已提交
4229
    """
C
chengduo 已提交
4230
    assert param_attr is not False, "param_attr should not be False in conv3d_transpose."
4231 4232
    l_type = "conv3d_transpose"
    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
4233
    if not isinstance(input, Variable):
4234
        raise TypeError("Input of conv3d_transpose must be Variable")
Y
Yu Yang 已提交
4235 4236
    input_channel = input.shape[1]

4237 4238 4239
    padding = utils.convert_to_list(padding, 3, 'padding')
    stride = utils.convert_to_list(stride, 3, 'stride')
    dilation = utils.convert_to_list(dilation, 3, 'dilation')
C
chengduoZH 已提交
4240

C
chengduoZH 已提交
4241 4242 4243
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")

Y
Yu Yang 已提交
4244 4245 4246 4247 4248 4249
    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]

4250 4251 4252
        d_in = input.shape[2]
        h_in = input.shape[3]
        w_in = input.shape[4]
C
chengduoZH 已提交
4253

4254
        filter_size_d = (output_size[0] - (d_in - 1) * stride[0] + 2 *
M
minqiyang 已提交
4255
                         padding[0] - 1) // dilation[0] + 1
4256
        filter_size_h = (output_size[1] - (h_in - 1) * stride[1] + 2 *
M
minqiyang 已提交
4257
                         padding[1] - 1) // dilation[1] + 1
4258
        filter_size_w = (output_size[2] - (w_in - 1) * stride[2] + 2 *
M
minqiyang 已提交
4259
                         padding[2] - 1) // dilation[2] + 1
4260
        filter_size = [filter_size_d, filter_size_h, filter_size_w]
C
chengduoZH 已提交
4261
    else:
4262 4263
        filter_size = utils.convert_to_list(filter_size, 3,
                                            'conv3d_transpose.filter_size')
Y
Yu Yang 已提交
4264

4265
    groups = 1 if groups is None else groups
M
minqiyang 已提交
4266
    filter_shape = [input_channel, num_filters // groups] + filter_size
Y
Yu Yang 已提交
4267 4268 4269
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

X
Xin Pan 已提交
4270
    pre_bias = helper.create_variable_for_type_inference(dtype=input.dtype)
Y
Yu Yang 已提交
4271
    helper.append_op(
4272
        type=l_type,
Y
Yu Yang 已提交
4273 4274
        inputs={'Input': [input],
                'Filter': [img_filter]},
4275
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
4276 4277 4278 4279
        attrs={
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
4280
            'groups': groups,
C
chengduoZH 已提交
4281 4282
            'use_cudnn': use_cudnn
        })
Y
Yu Yang 已提交
4283

4284 4285
    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
    out = helper.append_activation(pre_act)
Y
Yu Yang 已提交
4286
    return out
Y
yangyaming 已提交
4287 4288


Y
yangyaming 已提交
4289
def sequence_expand(x, y, ref_level=-1, name=None):
4290
    """Sequence Expand Layer. This layer will expand the input variable **x**
Y
yangyaming 已提交
4291 4292 4293 4294
    according to specified level lod of **y**. Please note that lod level of
    **x** is at most 1 and rank of **x** is at least 2. When rank of **x**
    is greater than 2, then it would be viewed as a 2-D tensor.
    Following examples will explain how sequence_expand works:
4295 4296 4297 4298 4299

    .. code-block:: text

        * Case 1
            x is a LoDTensor:
4300
                x.lod  = [[2,        2]]
Y
yangyaming 已提交
4301
                x.data = [[a], [b], [c], [d]]
4302 4303 4304
                x.dims = [4, 1]

            y is a LoDTensor:
4305 4306
                y.lod = [[2,    2],
                         [3, 3, 1, 1]]
4307

Y
yangyaming 已提交
4308
            ref_level: 0
4309

Y
yangyaming 已提交
4310
            then output is a 1-level LoDTensor:
4311
                out.lod =  [[2,        2,        2,        2]]
Y
yangyaming 已提交
4312
                out.data = [[a], [b], [a], [b], [c], [d], [c], [d]]
4313 4314 4315 4316
                out.dims = [8, 1]

        * Case 2
            x is a Tensor:
Y
yangyaming 已提交
4317
                x.data = [[a], [b], [c]]
4318 4319 4320
                x.dims = [3, 1]

            y is a LoDTensor:
4321
                y.lod = [[2, 0, 3]]
4322

Y
yangyaming 已提交
4323
            ref_level: -1
4324

Y
yangyaming 已提交
4325 4326 4327
            then output is a Tensor:
                out.data = [[a], [a], [c], [c], [c]]
                out.dims = [5, 1]
4328 4329 4330
    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
        y (Variable): The input variable which is a LoDTensor.
Y
yangyaming 已提交
4331 4332
        ref_level (int): Lod level of `y` to be referred by `x`. If set to -1,
                         refer the last level of lod.
C
caoying03 已提交
4333
        name(str|None): A name for this layer(optional). If set None, the layer
Y
yangyaming 已提交
4334
                        will be named automatically.
4335 4336 4337 4338 4339 4340

    Returns:
        Variable: The expanded variable which is a LoDTensor.

    Examples:
        .. code-block:: python
4341
	
4342
            import paddle.fluid as fluid
4343
            import paddle.fluid.layers as layers
4344 4345 4346
            x = fluid.layers.data(name='x', shape=[10], dtype='float32')
            y = fluid.layers.data(name='y', shape=[10, 20],
                             dtype='float32', lod_level=1)
Y
yangyaming 已提交
4347
            out = layers.sequence_expand(x=x, y=y, ref_level=0)
4348
    """
L
lujun 已提交
4349
    assert not in_dygraph_mode(), (
4350
        "sequence layer is not supported in dygraph mode yet.")
Y
yangyaming 已提交
4351
    helper = LayerHelper('sequence_expand', input=x, **locals())
4352
    dtype = helper.input_dtype()
X
Xin Pan 已提交
4353
    tmp = helper.create_variable_for_type_inference(dtype)
4354
    helper.append_op(
Y
yangyaming 已提交
4355 4356 4357 4358 4359
        type='sequence_expand',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': tmp},
        attrs={'ref_level': ref_level})
4360
    return tmp
4361 4362


C
chengduo 已提交
4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410
def sequence_expand_as(x, y, name=None):
    """Sequence Expand As Layer. This layer will expand the input variable **x**
    according to the zeroth level lod of **y**. Current implementation requires
    the level number of Input(Y)'s lod must be 1, and the first dimension of
    Input(X) should be equal to the size of Input(Y)'s zeroth level lod, and
    lod of Input(X) is not considered.

    Following examples will explain how sequence_expand_as works:

    .. code-block:: text

        * Case 1:

            Given a 1-level LoDTensor input(X)
                X.data = [[a], [b], [c], [d]]
                X.dims = [4, 1]
            and input(Y)
                Y.lod = [[0, 3, 6, 7, 8]]
            ref_level: 0
            then we get 1-level LoDTensor
                Out.lod =  [[0,            3,              6,  7,  8]]
                Out.data = [[a], [a], [a], [b], [b], [b], [c], [d]]
                Out.dims = [8, 1]

        * Case 2:

            Given a common Tensor input(X)
                X.data = [[a, b], [c, d], [e, f]]
                X.dims = [3, 2]
            and input(Y)
                Y.lod = [[0, 2, 3, 6]]
            ref_level: 0
            then we get a common LoDTensor
                Out.lod =  [[0,             2,     3,                    6]]
                Out.data = [[a, b], [a, b] [c, d], [e, f], [e, f], [e, f]]
                Out.dims = [6, 2]

    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
        y (Variable): The input variable which is a LoDTensor.
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        Variable: The expanded variable which is a LoDTensor.

    Examples:
        .. code-block:: python
4411 4412
            
            import paddle.fluid as fluid
4413
            import paddle.fluid.layers as layers
C
chengduo 已提交
4414 4415 4416 4417 4418 4419

            x = fluid.layers.data(name='x', shape=[10], dtype='float32')
            y = fluid.layers.data(name='y', shape=[10, 20],
                             dtype='float32', lod_level=1)
            out = layers.sequence_expand_as(x=x, y=y)
    """
L
lujun 已提交
4420
    assert not in_dygraph_mode(), (
4421
        "sequence layer is not supported in dygraph mode yet.")
C
chengduo 已提交
4422 4423
    helper = LayerHelper('sequence_expand_as', input=x, **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
4424
    tmp = helper.create_variable_for_type_inference(dtype)
C
chengduo 已提交
4425 4426 4427 4428 4429 4430 4431 4432
    helper.append_op(
        type='sequence_expand_as',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': tmp})
    return tmp


F
fengjiayi 已提交
4433
@templatedoc()
4434
def sequence_pad(x, pad_value, maxlen=None, name=None):
F
fengjiayi 已提交
4435 4436 4437 4438 4439
    """
    ${comment}

    Args:
        x(Variable): Input variable which should contain lod information.
M
minqiyang 已提交
4440 4441 4442
        pad_value(Variable): The Variable that holds values that will be fill
            into padded steps. It can be a scalar or a tensor whose shape
            equals to time steps in sequences. If it's a scalar, it will be
F
fengjiayi 已提交
4443
            automatically broadcasted to the shape of time step.
M
minqiyang 已提交
4444 4445 4446 4447
        maxlen(int, default None): The length of padded sequences. It can be
            None or any positive int. When it is None, all sequences will be
            padded up to the length of the longest one among them; when it a
            certain positive value, it must be greater than the length of the
4448 4449 4450
            longest original sequence.
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.
M
minqiyang 已提交
4451

F
fengjiayi 已提交
4452
    Returns:
M
minqiyang 已提交
4453
        Variable: The padded sequence batch and the original lengths before
4454
                  padding. All sequences has the same length.
M
minqiyang 已提交
4455

F
fengjiayi 已提交
4456 4457 4458
    Examples:
        .. code-block:: python

4459
            import paddle.fluid as fluid
F
fengjiayi 已提交
4460 4461 4462 4463
            import numpy

            x = fluid.layers.data(name='y', shape=[10, 5],
                             dtype='float32', lod_level=1)
G
gmcather 已提交
4464
            pad_value = fluid.layers.assign(
D
dongzhihong 已提交
4465
                input=numpy.array([0.0], dtype=numpy.float32))
F
fengjiayi 已提交
4466 4467 4468
            out = fluid.layers.sequence_pad(x=x, pad_value=pad_value)
    """

L
lujun 已提交
4469
    assert not in_dygraph_mode(), (
4470
        "sequence layer is not supported in dygraph mode yet.")
F
fengjiayi 已提交
4471 4472
    helper = LayerHelper('sequence_pad', input=x, **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
4473 4474
    out = helper.create_variable_for_type_inference(dtype)
    length = helper.create_variable_for_type_inference(dtype)
4475 4476 4477 4478

    pad_value.stop_gradient = True
    length.stop_gradient = True

F
fengjiayi 已提交
4479 4480 4481 4482 4483 4484
    if maxlen is None:
        maxlen = -1
    helper.append_op(
        type='sequence_pad',
        inputs={'X': x,
                'PadValue': pad_value},
4485 4486
        outputs={'Out': out,
                 'Length': length},
F
fengjiayi 已提交
4487
        attrs={'padded_length': maxlen})
4488
    return out, length
F
fengjiayi 已提交
4489 4490


4491
def sequence_unpad(x, length, name=None):
Y
Yibing Liu 已提交
4492
    """
4493
    **Sequence Unpad Layer**
Y
Yibing Liu 已提交
4494

4495 4496
    This layer removes the padding data in the input sequences and convert
    them into sequences with actual length as output, identitied by lod
Y
Yibing Liu 已提交
4497 4498 4499 4500 4501 4502 4503 4504 4505
    information.

    .. code-block:: text

	Example:

	Given input Variable **x**:
	    x.data = [[ 1.0,  2.0,  3.0,  4.0,  5.0],
		      [ 6.0,  7.0,  8.0,  9.0, 10.0],
4506 4507 4508
		      [11.0, 12.0, 13.0, 14.0, 15.0]],

	in which there are 3 sequences padded to length 5, and the acutal length
4509
	specified by input Variable **length**:
Y
Yibing Liu 已提交
4510 4511 4512 4513 4514 4515

	    length.data = [[2], [3], [4]],

	after unpadding, the output Variable will be:

	    out.data = [[1.0, 2.0, 6.0, 7.0, 8.0, 11.0, 12.0, 13.0, 14.0]]
4516
	    out.lod = [[2, 3, 4]]
Y
Yibing Liu 已提交
4517 4518 4519 4520 4521 4522

    Args:
        x(Variable): Input Variable which contains the padded sequences with
            equal length.
        length(Variable): The Variable that specifies the actual ength of
            sequences after unpadding.
4523 4524
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.
Y
Yibing Liu 已提交
4525 4526 4527 4528 4529 4530 4531

    Returns:
        Variable: The Variable contains the unpadded sequences.

    Examples:
        .. code-block:: python

4532
            import paddle.fluid as fluid
Y
Yibing Liu 已提交
4533 4534 4535 4536 4537
            x = fluid.layers.data(name='x', shape=[10, 5], dtype='float32')
            len = fluid.layers.data(name='length', shape=[1], dtype='int64')
            out = fluid.layers.sequence_unpad(x=x, length=len)
    """

L
lujun 已提交
4538
    assert not in_dygraph_mode(), (
4539
        "sequence layer is not supported in dygraph mode yet.")
Y
Yibing Liu 已提交
4540 4541
    helper = LayerHelper('sequence_unpad', input=x, **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
4542
    out = helper.create_variable_for_type_inference(dtype)
Y
Yibing Liu 已提交
4543 4544 4545 4546 4547 4548 4549 4550 4551 4552 4553

    length.stop_gradient = True

    helper.append_op(
        type='sequence_unpad',
        inputs={'X': x,
                'Length': length},
        outputs={'Out': out})
    return out


4554 4555 4556 4557 4558 4559 4560
def beam_search(pre_ids,
                pre_scores,
                ids,
                scores,
                beam_size,
                end_id,
                level=0,
4561
                is_accumulated=True,
4562 4563
                name=None,
                return_parent_idx=False):
4564
    """
4565 4566
    Beam search is a classical algorithm for selecting candidate words in a
    machine translation task.
Y
Yan Chunwei 已提交
4567 4568 4569

    Refer to `Beam search <https://en.wikipedia.org/wiki/Beam_search>`_
    for more details.
M
minqiyang 已提交
4570 4571

    This layer does the search in beams for one time step. Specifically, it
4572 4573 4574
    selects the top-K candidate word ids of current step from :attr:`ids`
    according to their :attr:`scores` for all source sentences, where K is
    :attr:`beam_size` and :attr:`ids, scores` are predicted results from the
4575 4576 4577 4578 4579 4580 4581 4582 4583 4584 4585
    computation cell. If :attr:`ids` is not set, it will be calculated out
    according to :attr:`scores`. Additionally, :attr:`pre_ids` and
    :attr:`pre_scores` are the output of beam_search at previous step, they
    are needed for special use to handle ended candidate translations.

    Note that if :attr:`is_accumulated` is :attr:`True`, the :attr:`scores`
    passed in should be accumulated scores. Else, the :attr:`scores` are
    considered as the straightforward scores and will be transformed to the
    log field and accumulated the :attr:`pre_scores` in this operator.
    Length penalty should be done with extra operators before calculating the
    accumulated scores if needed.
4586 4587 4588 4589

    Please see the following demo for a fully beam search usage example:

        fluid/tests/book/test_machine_translation.py
Y
Yan Chunwei 已提交
4590

4591
    Args:
4592 4593 4594 4595 4596 4597 4598 4599 4600 4601 4602 4603 4604 4605 4606 4607 4608 4609 4610 4611 4612 4613 4614
        pre_ids(Variable): The LodTensor variable which is the output of
            beam_search at previous step. It should be a LodTensor with shape
            :math:`(batch_size, 1)` and lod
            :math:`[[0, 1, ... , batch_size], [0, 1, ..., batch_size]]` at the
            first step.
        pre_scores(Variable): The LodTensor variable which is the output of
            beam_search at previous step.
        ids(Variable): The LodTensor variable containing the candidates ids.
            Its shape should be :math:`(batch_size \\times beam_size, K)`,
            where :math:`K` supposed to be :attr:`beam_size`.
        scores(Variable): The LodTensor variable containing the accumulated
            scores corresponding to :attr:`ids` and its shape is the same as
            the shape of :attr:`ids`.
        beam_size(int): The beam width used in beam search.
        end_id(int): The id of end token.
        level(int, default 0): It can be ignored and mustn't change currently.
            It means the source level of lod, which is explained as following.
            The lod level of :attr:`ids` should be 2. The first level is source
            level which describes how many prefixes (branchs) for each source
            sentece (beam), and the second level is sentence level which
            describes how these candidates belong to the prefix. The paths
            linking prefixes and selected candidates are organized and reserved
            in lod.
4615 4616
        is_accumulated(bool, default True): Whether the input :attr:`score` is
             accumulated scores.
4617 4618
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
4619 4620 4621 4622
        return_parent_idx(bool): Whether to return an extra Tensor variable 
                        preserving the selected_ids' parent indice in pre_ids
                        in output, which can be used to gather cell states at
                        the next time step.
F
fengjiayi 已提交
4623

4624
    Returns:
4625 4626 4627 4628
        Variable: The LodTensor tuple containing the selected ids and the \
            corresponding scores. If :attr:`return_parent_idx` is :attr:`True`, \
            an extra Tensor variable preserving the selected_ids' parent indice \
            is included.
Y
Yan Chunwei 已提交
4629 4630 4631 4632

    Examples:
        .. code-block:: python

4633 4634
            import paddle.fluid as fluid

4635 4636 4637
            # Suppose `probs` contains predicted results from the computation
            # cell and `pre_ids` and `pre_scores` is the output of beam_search
            # at previous step.
4638 4639 4640 4641 4642 4643 4644 4645 4646 4647 4648 4649
            beam_size = 4
            end_id = 1
            pre_ids = fluid.layers.data(
                name='pre_id', shape=[1], lod_level=2, dtype='int64')
            pre_scores = fluid.layers.data(
                name='pre_scores', shape=[1], lod_level=2, dtype='float32')
            probs = fluid.layers.data(
                name='probs', shape=[10000], dtype='float32')
            topk_scores, topk_indices = fluid.layers.topk(probs, k=beam_size)
            accu_scores = fluid.layers.elementwise_add(
                x=fluid.layers.log(x=topk_scores),
                y=fluid.layers.reshape(pre_scores, shape=[-1]),
4650
                axis=0)
4651
            selected_ids, selected_scores = fluid.layers.beam_search(
4652 4653 4654 4655 4656 4657 4658
                pre_ids=pre_ids,
                pre_scores=pre_scores,
                ids=topk_indices,
                scores=accu_scores,
                beam_size=beam_size,
                end_id=end_id)
    """
Q
Qiao Longfei 已提交
4659
    helper = LayerHelper('beam_search', **locals())
4660 4661 4662 4663 4664 4665
    score_type = pre_scores.dtype
    id_type = pre_ids.dtype

    inputs = {"pre_ids": pre_ids, "pre_scores": pre_scores, "scores": scores}
    if ids is not None:
        inputs["ids"] = ids
Q
Qiao Longfei 已提交
4666

X
Xin Pan 已提交
4667 4668 4669
    selected_scores = helper.create_variable_for_type_inference(
        dtype=score_type)
    selected_ids = helper.create_variable_for_type_inference(dtype=id_type)
4670 4671 4672 4673 4674
    # parent_idx is a tensor used to gather cell states at the next time
    # step. Though lod in selected_ids can also be used to gather by
    # sequence_expand, it is not efficient.
    # gather_op's index input only supports int32 dtype currently
    parent_idx = helper.create_variable_for_type_inference(dtype="int32")
Q
Qiao Longfei 已提交
4675 4676 4677

    helper.append_op(
        type='beam_search',
4678
        inputs=inputs,
Q
Qiao Longfei 已提交
4679 4680 4681
        outputs={
            'selected_ids': selected_ids,
            'selected_scores': selected_scores,
4682
            'parent_idx': parent_idx
Q
Qiao Longfei 已提交
4683 4684 4685 4686 4687 4688
        },
        attrs={
            # TODO(ChunweiYan) to assure other value support
            'level': level,
            'beam_size': beam_size,
            'end_id': end_id,
4689
            'is_accumulated': is_accumulated,
Q
Qiao Longfei 已提交
4690
        })
4691 4692 4693 4694
    if return_parent_idx:
        return selected_ids, selected_scores, parent_idx
    else:
        return selected_ids, selected_scores
Q
Qiao Longfei 已提交
4695 4696


4697 4698 4699 4700 4701 4702 4703
def beam_search_decode(ids, scores, beam_size, end_id, name=None):
    """
    Beam Search Decode Layer. This layer constructs the full hypotheses for
    each source sentence by walking back along the LoDTensorArray :attr:`ids`
    whose lods can be used to restore the path in the beam search tree.
    Please see the following demo for a fully beam search usage example:
        fluid/tests/book/test_machine_translation.py
G
guosheng 已提交
4704

4705 4706 4707 4708 4709 4710 4711 4712 4713
    Args:
        ids(Variable): The LodTensorArray variable containing the selected ids
            of all steps.
        scores(Variable): The LodTensorArray variable containing the selected
            scores of all steps.
        beam_size(int): The beam width used in beam search.
        end_id(int): The id of end token.
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
G
guosheng 已提交
4714

4715 4716 4717 4718 4719 4720
    Returns:
        Variable: The LodTensor pair containing the generated id sequences \
            and the corresponding scores. The shapes and lods of the two \
            LodTensor are same. The lod level is 2 and the two levels \
            separately indicate how many hypotheses each source sentence has \
            and how many ids each hypothesis has.
G
guosheng 已提交
4721

4722 4723
    Examples:
        .. code-block:: python
T
Tink_Y 已提交
4724

4725 4726
            import paddle.fluid as fluid

4727 4728
            # Suppose `ids` and `scores` are LodTensorArray variables reserving
            # the selected ids and scores of all steps
4729 4730 4731
            ids = fluid.layers.create_array(dtype='int64')
            scores = fluid.layers.create_array(dtype='float32')
            finished_ids, finished_scores = fluid.layers.beam_search_decode(
4732 4733 4734
                ids, scores, beam_size=5, end_id=0)
    """
    helper = LayerHelper('beam_search_decode', **locals())
X
Xin Pan 已提交
4735 4736
    sentence_ids = helper.create_variable_for_type_inference(dtype=ids.dtype)
    sentence_scores = helper.create_variable_for_type_inference(dtype=ids.dtype)
4737 4738 4739 4740 4741 4742 4743 4744 4745 4746 4747 4748 4749 4750 4751

    helper.append_op(
        type="beam_search_decode",
        inputs={"Ids": ids,
                "Scores": scores},
        outputs={
            "SentenceIds": sentence_ids,
            "SentenceScores": sentence_scores
        },
        attrs={"beam_size": beam_size,
               "end_id": end_id})

    return sentence_ids, sentence_scores


Y
yangyaming 已提交
4752 4753 4754 4755
def lstm_unit(x_t,
              hidden_t_prev,
              cell_t_prev,
              forget_bias=0.0,
Y
yangyaming 已提交
4756
              param_attr=None,
C
caoying03 已提交
4757 4758
              bias_attr=None,
              name=None):
Y
yangyaming 已提交
4759 4760 4761 4762
    """Lstm unit layer. The equation of a lstm step is:

        .. math::

4763
            i_t & = \sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i)
Y
yangyaming 已提交
4764

4765
            f_t & = \sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + b_f)
Y
yangyaming 已提交
4766

4767
            c_t & = f_tc_{t-1} + i_t tanh (W_{x_c}x_t + W_{h_c}h_{t-1} + b_c)
Y
yangyaming 已提交
4768

4769
            o_t & = \sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + b_o)
Y
yangyaming 已提交
4770 4771 4772

            h_t & = o_t tanh(c_t)

4773 4774 4775 4776 4777 4778
    The inputs of lstm unit include :math:`x_t`, :math:`h_{t-1}` and
    :math:`c_{t-1}`. The 2nd dimensions of :math:`h_{t-1}` and :math:`c_{t-1}`
    should be same. The implementation separates the linear transformation and
    non-linear transformation apart. Here, we take :math:`i_t` as an example.
    The linear transformation is applied by calling a `fc` layer and the
    equation is:
Y
yangyaming 已提交
4779 4780 4781

        .. math::

4782
            L_{i_t} = W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i
Y
yangyaming 已提交
4783 4784 4785 4786 4787 4788 4789 4790

    The non-linear transformation is applied by calling `lstm_unit_op` and the
    equation is:

        .. math::

            i_t = \sigma(L_{i_t})

4791
    This layer has two outputs including :math:`h_t` and :math:`c_t`.
Y
yangyaming 已提交
4792 4793

    Args:
Y
yangyaming 已提交
4794 4795 4796 4797 4798 4799
        x_t (Variable): The input value of current step, a 2-D tensor with shape
            M x N, M for batch size and N for input size.
        hidden_t_prev (Variable): The hidden value of lstm unit, a 2-D tensor
            with shape M x S, M for batch size and S for size of lstm unit.
        cell_t_prev (Variable): The cell value of lstm unit, a 2-D tensor with
            shape M x S, M for batch size and S for size of lstm unit.
Y
yangyaming 已提交
4800
        forget_bias (float): The forget bias of lstm unit.
C
chengduo 已提交
4801 4802 4803 4804 4805 4806 4807 4808 4809 4810 4811 4812
        param_attr(ParamAttr|None): The parameter attribute for the learnable
                               hidden-hidden weights.
                               If it is set to None or one attribute of ParamAttr,
                               lstm_unit will create ParamAttr as param_attr.
                               If the Initializer of the param_attr is not set, the
                               parameter is initialized with Xavier. Default: None.
        bias_attr (ParamAttr|None): The bias attribute for the learnable bias
                              weights. 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,
                              lstm_unit will create ParamAttr as bias_attr.
                              If the Initializer of the bias_attr is not set,
                              the bias is initialized zero. Default: None.
C
caoying03 已提交
4813 4814
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
Y
yangyaming 已提交
4815 4816

    Returns:
Y
yangyaming 已提交
4817
        tuple: The hidden value and cell value of lstm unit.
Y
yangyaming 已提交
4818 4819

    Raises:
4820 4821 4822 4823
        ValueError: The ranks of **x_t**, **hidden_t_prev** and **cell_t_prev**
                    not be 2 or the 1st dimensions of **x_t**, **hidden_t_prev**
                    and **cell_t_prev** not be the same or the 2nd dimensions of
                    **hidden_t_prev** and **cell_t_prev** not be the same.
Y
yangyaming 已提交
4824 4825 4826 4827 4828

    Examples:

        .. code-block:: python

4829 4830 4831 4832 4833 4834 4835 4836 4837 4838 4839 4840 4841
            import paddle.fluid as fluid

            dict_dim, emb_dim, hidden_dim = 128, 64, 512
            data = fluid.layers.data(name='step_data', shape=[1], dtype='int32')
            x = fluid.layers.embedding(input=data, size=[dict_dim, emb_dim])
            pre_hidden = fluid.layers.data(
                name='pre_hidden', shape=[hidden_dim], dtype='float32')
            pre_cell = fluid.layers.data(
                name='pre_cell', shape=[hidden_dim], dtype='float32')
            hidden = fluid.layers.lstm_unit(
                x_t=x,
                hidden_t_prev=pre_hidden,
                cell_t_prev=pre_cell)
Y
yangyaming 已提交
4842 4843 4844 4845 4846 4847 4848 4849 4850 4851 4852 4853 4854 4855
    """
    helper = LayerHelper('lstm_unit', **locals())

    if len(x_t.shape) != 2:
        raise ValueError("Rank of x_t must be 2.")

    if len(hidden_t_prev.shape) != 2:
        raise ValueError("Rank of hidden_t_prev must be 2.")

    if len(cell_t_prev.shape) != 2:
        raise ValueError("Rank of cell_t_prev must be 2.")

    if x_t.shape[0] != hidden_t_prev.shape[0] or x_t.shape[
            0] != cell_t_prev.shape[0]:
Y
yangyaming 已提交
4856
        raise ValueError("The 1st dimensions of x_t, hidden_t_prev and "
4857 4858 4859 4860
                         "cell_t_prev must be the same.")

    if hidden_t_prev.shape[1] != cell_t_prev.shape[1]:
        raise ValueError("The 2nd dimensions of hidden_t_prev and "
Y
yangyaming 已提交
4861 4862
                         "cell_t_prev must be the same.")

Y
yangyaming 已提交
4863 4864 4865
    if bias_attr is None:
        bias_attr = ParamAttr()

Y
yangyaming 已提交
4866
    size = cell_t_prev.shape[1]
4867
    concat_out = concat(input=[x_t, hidden_t_prev], axis=1)
Y
yangyaming 已提交
4868 4869
    fc_out = fc(input=concat_out,
                size=4 * size,
Y
yangyaming 已提交
4870
                param_attr=param_attr,
4871
                bias_attr=bias_attr)
Y
yangyaming 已提交
4872
    dtype = x_t.dtype
X
Xin Pan 已提交
4873 4874
    c = helper.create_variable_for_type_inference(dtype)
    h = helper.create_variable_for_type_inference(dtype)
Y
yangyaming 已提交
4875 4876 4877 4878 4879 4880 4881 4882 4883

    helper.append_op(
        type='lstm_unit',
        inputs={"X": fc_out,
                "C_prev": cell_t_prev},
        outputs={"C": c,
                 "H": h},
        attrs={"forget_bias": forget_bias})

Y
yangyaming 已提交
4884
    return h, c
G
guosheng 已提交
4885 4886


C
caoying03 已提交
4887
def reduce_sum(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
4888
    """
Y
yangyaming 已提交
4889
    Computes the sum of tensor elements over the given dimension.
G
guosheng 已提交
4890 4891 4892

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
4893
        dim (list|int|None): The dimensions along which the sum is performed. If
Y
yangyaming 已提交
4894 4895
            :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
4896 4897
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
4898
        keep_dim (bool|False): Whether to reserve the reduced dimension in the
Y
yangyaming 已提交
4899
            output Tensor. The result tensor will have one fewer dimension
G
guosheng 已提交
4900
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
4901 4902
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
4903 4904 4905

    Returns:
        Variable: The reduced Tensor variable.
F
fengjiayi 已提交
4906

G
guosheng 已提交
4907 4908 4909
    Examples:
        .. code-block:: python

4910
            import paddle.fluid as fluid
G
guosheng 已提交
4911 4912 4913
            # 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
4914
            # Each example is followed by the corresponding output tensor.
4915
            x = fluid.layers.data(name='x', shape=[4, 2], dtype='float32')
G
guosheng 已提交
4916 4917 4918 4919
            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
4920

4921
            # y is a Tensor variable with shape [2, 2, 2] and elements as below:
W
whs 已提交
4922 4923
            #      [[[1, 2], [3, 4]],
            #      [[5, 6], [7, 8]]]
Q
qiaolongfei 已提交
4924
            # Each example is followed by the corresponding output tensor.
4925 4926 4927
            y = fluid.layers.data(name='y', shape=[2, 2, 2], dtype='float32')
            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
4928

G
guosheng 已提交
4929 4930
    """
    helper = LayerHelper('reduce_sum', **locals())
X
Xin Pan 已提交
4931
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
4932 4933
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
G
guosheng 已提交
4934 4935 4936 4937 4938
    helper.append_op(
        type='reduce_sum',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
4939
            'dim': dim if dim != None else [0],
G
guosheng 已提交
4940 4941 4942 4943
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
4944 4945


C
caoying03 已提交
4946
def reduce_mean(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
4947
    """
Y
Yibing Liu 已提交
4948
    Computes the mean of the input tensor's elements along the given dimension.
G
guosheng 已提交
4949 4950 4951

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
Yibing Liu 已提交
4952 4953 4954
        dim (list|int|None): The dimension along which the mean is computed. If
            `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
4955
            must be in the range :math:`[-rank(input), rank(input))`. If
4956
            :math:`dim[i] < 0`, the dimension to reduce is
Y
Yibing Liu 已提交
4957
            :math:`rank(input) + dim[i]`.
Y
yangyaming 已提交
4958 4959
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
G
guosheng 已提交
4960
            than the :attr:`input` unless :attr:`keep_dim` is true.
Y
Yibing Liu 已提交
4961
        name(str|None): A name for this layer(optional). If set `None`, the layer
C
caoying03 已提交
4962
                       will be named automatically.
G
guosheng 已提交
4963 4964

    Returns:
Y
Yibing Liu 已提交
4965
        Variable: The reduced mean Variable.
F
fengjiayi 已提交
4966

G
guosheng 已提交
4967 4968 4969
    Examples:
        .. code-block:: python

4970
            import paddle.fluid as fluid
G
guosheng 已提交
4971 4972 4973 4974
            # 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.
4975
            x = fluid.layers.data(name='x', shape=[4, 2], dtype='float32')
G
guosheng 已提交
4976 4977 4978
            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]
4979
            fluid.layers.reduce_mean(x, dim=1, keep_dim=True)  # [[0.475], [0.4]]
W
whs 已提交
4980

4981
            # y is a Tensor variable with shape [2, 2, 2] and elements as below:
W
whs 已提交
4982 4983 4984
            #      [[[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.
4985 4986 4987
            y = fluid.layers.data(name='y', shape=[2, 2, 2], dtype='float32')
            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
4988 4989
    """
    helper = LayerHelper('reduce_mean', **locals())
X
Xin Pan 已提交
4990
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
4991 4992
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
G
guosheng 已提交
4993 4994 4995 4996 4997
    helper.append_op(
        type='reduce_mean',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
4998
            'dim': dim if dim != None else [0],
G
guosheng 已提交
4999 5000 5001 5002
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
5003 5004


C
caoying03 已提交
5005
def reduce_max(input, dim=None, keep_dim=False, name=None):
5006
    """
Y
yangyaming 已提交
5007
    Computes the maximum of tensor elements over the given dimension.
5008 5009 5010

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
5011
        dim (list|int|None): The dimension along which the maximum is computed.
Y
yangyaming 已提交
5012 5013 5014
            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
5015
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
Y
yangyaming 已提交
5016 5017
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
5018
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
5019 5020
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
5021 5022 5023

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
5024

5025 5026 5027
    Examples:
        .. code-block:: python

5028
            import paddle.fluid as fluid
5029 5030 5031 5032
            # 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.
5033
            x = fluid.layers.data(name='x', shape=[4, 2], dtype='float32')
5034 5035 5036 5037
            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
5038

5039
            # y is a Tensor variable with shape [2, 2, 2] and elements as below:
W
whs 已提交
5040 5041 5042
            #      [[[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.
5043 5044 5045
            y = fluid.layers.data(name='y', shape=[2, 2, 2], dtype='float32')
            fluid.layers.reduce_max(y, dim=[1, 2]) # [4.0, 8.0]
            fluid.layers.reduce_max(y, dim=[0, 1]) # [7.0, 8.0]
5046 5047
    """
    helper = LayerHelper('reduce_max', **locals())
X
Xin Pan 已提交
5048
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
5049 5050
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
5051 5052 5053 5054 5055
    helper.append_op(
        type='reduce_max',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
5056
            'dim': dim if dim != None else [0],
5057 5058 5059 5060 5061 5062
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


C
caoying03 已提交
5063
def reduce_min(input, dim=None, keep_dim=False, name=None):
5064
    """
Y
yangyaming 已提交
5065
    Computes the minimum of tensor elements over the given dimension.
5066 5067 5068

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
5069
        dim (list|int|None): The dimensions along which the minimum is computed.
Y
yangyaming 已提交
5070 5071 5072
            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
5073
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
Y
yangyaming 已提交
5074 5075
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
5076
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
5077 5078
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
5079 5080 5081

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
5082

5083 5084 5085
    Examples:
        .. code-block:: python

5086
            import paddle.fluid as fluid
5087 5088 5089 5090
            # 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.
5091
            x = fluid.layers.data(name='x', shape=[4, 2], dtype='float32')
5092 5093 5094 5095
            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
5096

5097
            # y is a Tensor variable with shape [2, 2, 2] and elements as below:
W
whs 已提交
5098 5099 5100
            #      [[[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.
5101 5102 5103
            y = fluid.layers.data(name='y', shape=[2, 2, 2], dtype='float32')
            fluid.layers.reduce_min(y, dim=[1, 2]) # [1.0, 5.0]
            fluid.layers.reduce_min(y, dim=[0, 1]) # [1.0, 2.0]
5104 5105
    """
    helper = LayerHelper('reduce_min', **locals())
X
Xin Pan 已提交
5106
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
5107 5108
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
5109 5110 5111 5112 5113
    helper.append_op(
        type='reduce_min',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
5114
            'dim': dim if dim != None else [0],
5115 5116 5117 5118
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
5119 5120


5121 5122 5123 5124 5125 5126
def reduce_prod(input, dim=None, keep_dim=False, name=None):
    """
    Computes the product of tensor elements over the given dimension.

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
5127
        dim (list|int|None): The dimensions along which the product is performed. If
5128 5129
            :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
5130 5131
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
5132 5133 5134
        keep_dim (bool|False): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
            than the :attr:`input` unless :attr:`keep_dim` is true.
Y
yangyaming 已提交
5135
        name(str|None): A name for this layer(optional). If set None, the
Z
zhouhanqing 已提交
5136
            layer will be named automatically.
5137 5138 5139 5140 5141 5142 5143

    Returns:
        Variable: The reduced Tensor variable.

    Examples:
        .. code-block:: python

5144
            import paddle.fluid as fluid
5145 5146 5147 5148
            # 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.
5149
            x = fluid.layers.data(name='x', shape=[4, 2], dtype='float32')
5150 5151 5152
            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
5153
            fluid.layers.reduce_prod(x, dim=1,
Z
zhouhanqing 已提交
5154
                                     keep_dim=True)  # [[0.027], [0.0084]]
W
whs 已提交
5155

5156
            # y is a Tensor variable with shape [2, 2, 2] and elements as below:
W
whs 已提交
5157 5158 5159
            #      [[[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.
5160 5161 5162
            y = fluid.layers.data(name='y', shape=[2, 2, 2], dtype='float32')
            fluid.layers.reduce_prod(y, dim=[1, 2]) # [24.0, 1680.0]
            fluid.layers.reduce_prod(y, dim=[0, 1]) # [105.0, 384.0]
5163 5164
    """
    helper = LayerHelper('reduce_prod', **locals())
X
Xin Pan 已提交
5165
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
5166 5167
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
5168 5169 5170 5171 5172
    helper.append_op(
        type='reduce_prod',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
5173
            'dim': dim if dim != None else [0],
5174 5175 5176 5177 5178 5179
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


Z
zhoukunsheng 已提交
5180 5181
def reduce_all(input, dim=None, keep_dim=False, name=None):
    """
Z
zhoukunsheng 已提交
5182
    Computes the ``logical and`` of tensor elements over the given dimension.
Z
zhoukunsheng 已提交
5183 5184 5185 5186 5187 5188 5189 5190 5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
        dim (list|int|None): The dimension along which the logical and is computed.
            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))`.
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
            than the :attr:`input` unless :attr:`keep_dim` is true.
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.

    Returns:
        Variable: The reduced Tensor variable.

    Examples:
        .. code-block:: python
Z
zhoukunsheng 已提交
5202
        
5203
            import paddle.fluid as fluid
5204 5205 5206
            import paddle.fluid.layers as layers
            import numpy as np

Z
zhoukunsheng 已提交
5207 5208 5209
            # x is a bool Tensor variable with following elements:
            #    [[True, False]
            #     [True, True]]
5210 5211 5212 5213 5214 5215 5216
            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]
            out = layers.reduce_all(x, dim=1, keep_dim=True)  # [[False], [True]]
Z
zhoukunsheng 已提交
5217 5218 5219 5220 5221 5222 5223 5224 5225 5226 5227 5228 5229 5230 5231 5232 5233 5234 5235 5236

    """
    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):
    """
Z
zhoukunsheng 已提交
5237
    Computes the ``logical or`` of tensor elements over the given dimension.
Z
zhoukunsheng 已提交
5238 5239 5240 5241 5242 5243 5244 5245 5246 5247 5248 5249 5250 5251 5252 5253 5254 5255 5256

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
        dim (list|int|None): The dimension along which the logical or is computed.
            If :attr:`None`, compute the logical or 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))`.
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
            than the :attr:`input` unless :attr:`keep_dim` is true.
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.

    Returns:
        Variable: The reduced Tensor variable.

    Examples:
        .. code-block:: python
Z
zhoukunsheng 已提交
5257

5258
            import paddle.fluid as fluid
5259 5260 5261
            import paddle.fluid.layers as layers
            import numpy as np

Z
zhoukunsheng 已提交
5262 5263 5264
            # x is a bool Tensor variable with following elements:
            #    [[True, False]
            #     [False, False]]
5265 5266 5267 5268 5269 5270 5271
            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]
            out = layers.reduce_any(x, dim=1,
Z
zhoukunsheng 已提交
5272 5273 5274 5275 5276 5277 5278 5279 5280 5281 5282 5283 5284 5285
                                     keep_dim=True)  # [[True], [False]]

    """
    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,
5286 5287 5288 5289 5290
            'reduce_all': True if dim == None else False
        })
    return out


C
caoying03 已提交
5291
def split(input, num_or_sections, dim=-1, name=None):
G
guosheng 已提交
5292
    """
C
caoying03 已提交
5293
    Split the input tensor into multiple sub-tensors.
G
guosheng 已提交
5294 5295 5296

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
C
caoying03 已提交
5297 5298 5299 5300 5301
        num_or_sections (int|list): If :attr:`num_or_sections` is an integer,
            then the integer indicates the number of equal sized sub-tensors
            that the tensor will be divided into. If :attr:`num_or_sections`
            is a list of integers, the length of list indicates the number of
            sub-tensors and the integers indicate the sizes of sub-tensors'
G
guosheng 已提交
5302
            :attr:`dim` dimension orderly.
C
caoying03 已提交
5303
        dim (int): The dimension along which to split. If :math:`dim < 0`, the
G
guosheng 已提交
5304
            dimension to split along is :math:`rank(input) + dim`.
C
caoying03 已提交
5305 5306
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
5307 5308

    Returns:
D
dzhwinter 已提交
5309
        list(Variable): The list of segmented tensor variables.
G
guosheng 已提交
5310 5311 5312 5313

    Examples:
        .. code-block:: python

5314 5315 5316 5317 5318 5319
            import paddle.fluid as fluid

            # input is a variable which shape is [-1, 3, 9, 5]
            input = fluid.layers.data(
                 name="input", shape=[3, 9, 5], dtype="float32")

5320
            x0, x1, x2 = fluid.layers.split(input, num_or_sections=3, dim=2)
5321 5322 5323 5324 5325 5326 5327 5328
            # x0.shape [-1, 3, 3, 5]
            # x1.shape [-1, 3, 3, 5]
            # x2.shape [-1, 3, 3, 5]

            x0, x1, x2 = fluid.layers.split(input, num_or_sections=3, dim=2)
            # x0.shape [-1, 3, 2, 5]
            # x1.shape [-1, 3, 3, 5]
            # x2.shape [-1, 3, 4, 5]
G
guosheng 已提交
5329 5330 5331 5332 5333 5334 5335 5336
    """
    helper = LayerHelper('split', **locals())
    input_shape = input.shape
    dim = (len(input_shape) + dim) if dim < 0 else dim
    if isinstance(num_or_sections, int):
        assert num_or_sections > 1, 'num_or_sections must be more than 1.'
        num = num_or_sections
    else:
T
tink2123 已提交
5337
        assert len(num_or_sections) <= input_shape[
G
guosheng 已提交
5338 5339 5340
            dim], 'len(num_or_sections) must not be more than input.shape[dim].'
        num = len(num_or_sections)
    outs = [
X
Xin Pan 已提交
5341
        helper.create_variable_for_type_inference(dtype=helper.input_dtype())
G
guosheng 已提交
5342 5343 5344 5345 5346 5347 5348 5349 5350 5351 5352 5353 5354
        for i in range(num)
    ]
    helper.append_op(
        type='split',
        inputs={'X': input},
        outputs={'Out': outs},
        attrs={
            'num': num_or_sections if isinstance(num_or_sections, int) else 0,
            'sections': num_or_sections
            if isinstance(num_or_sections, list) else [],
            'axis': dim
        })
    return outs
C
caoying03 已提交
5355 5356 5357 5358 5359 5360 5361 5362 5363


def l2_normalize(x, axis, epsilon=1e-12, name=None):
    """
    **L2 normalize Layer**

    The l2 normalize layer normalizes `x` along dimension `axis` using an L2
    norm. For a 1-D tensor (`dim` is fixed to 0), this layer computes

5364
    .. math::
5365 5366

        y = \\frac{x}{ \sqrt{\sum {x^2} + epsion }}
C
caoying03 已提交
5367 5368 5369 5370 5371

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

    Args:
5372
        x(Variable|list): The input tensor to l2_normalize layer.
5373
        axis(int): The axis on which to apply normalization. If `axis < 0`, \
5374 5375
            the dimension to normalization is rank(X) + axis. -1 is the
            last dimension.
5376
        epsilon(float): The epsilon value is used to avoid division by zero, \
翟飞跃 已提交
5377
            the default value is 1e-12.
5378
        name(str|None): A name for this layer(optional). If set None, the layer \
5379
            will be named automatically.
C
caoying03 已提交
5380 5381

    Returns:
5382
        Variable: The output tensor variable is the same shape with `x`.
C
caoying03 已提交
5383 5384

    Examples:
5385

C
caoying03 已提交
5386 5387
        .. code-block:: python

5388
            import paddle.fluid as fluid
5389 5390 5391 5392
            data = fluid.layers.data(name="data",
                                     shape=(3, 17, 13),
                                     dtype="float32")
            normed = fluid.layers.l2_normalize(x=data, axis=1)
C
caoying03 已提交
5393 5394
    """

F
fengjiayi 已提交
5395 5396
    if len(x.shape) == 1:
        axis = 0
C
caoying03 已提交
5397 5398
    helper = LayerHelper("l2_normalize", **locals())

X
Xin Pan 已提交
5399 5400
    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
5401
    helper.append_op(
5402 5403 5404 5405
        type="norm",
        inputs={"X": x},
        outputs={"Out": out,
                 "Norm": norm},
C
caoying03 已提交
5406
        attrs={
5407 5408
            "axis": 1 if axis is None else axis,
            "epsilon": epsilon,
C
caoying03 已提交
5409 5410
        })
    return out
5411 5412


S
sneaxiy 已提交
5413
def matmul(x, y, transpose_x=False, transpose_y=False, alpha=1.0, name=None):
G
guosheng 已提交
5414
    """
Y
ying 已提交
5415 5416 5417 5418
    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
5419

C
chengduoZH 已提交
5420
    The actual behavior depends on the shapes of :math:`x`, :math:`y` and the
5421
    flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically:
G
guosheng 已提交
5422

5423 5424 5425 5426 5427
    - 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
5428
      :math:`[1, D]` in transposed form.
G
guosheng 已提交
5429

C
chengduoZH 已提交
5430
    - After transpose, the two tensors are 2-D or n-D and matrix multiplication
5431
      performs in the following way.
G
guosheng 已提交
5432

5433
      - If both are 2-D, they are multiplied like conventional matrices.
C
chengduoZH 已提交
5434
      - If either is n-D, it is treated as a stack of matrices residing in the
Y
ying 已提交
5435
        last two dimensions and a batched matrix multiply supporting broadcast
5436
        applies on the two tensors.
G
guosheng 已提交
5437

Y
ying 已提交
5438 5439
    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
5440
    removed after matrix multiplication.
G
guosheng 已提交
5441 5442 5443

    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
5444 5445 5446
        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
5447
        alpha (float): The scale of output. Default 1.0.
5448
        name(str|None): A name for this layer(optional). If set None, the layer
5449
            will be named automatically.
G
guosheng 已提交
5450 5451

    Returns:
石晓伟 已提交
5452
        Variable: The product Tensor (or LoDTensor) variable.
G
guosheng 已提交
5453

G
guosheng 已提交
5454 5455 5456
    Examples:
        .. code-block:: python

5457
            # Examples to clarify shapes of the inputs and output
C
chengduoZH 已提交
5458
            # x: [B, ..., M, K], y: [B, ..., K, N]
5459
            # fluid.layers.matmul(x, y)  # out: [B, ..., M, N]
Y
ying 已提交
5460

5461
            # x: [B, M, K], y: [B, K, N]
5462
            # fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
5463

5464
            # x: [B, M, K], y: [K, N]
5465
            # fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
5466

5467
            # x: [M, K], y: [K, N]
5468
            # fluid.layers.matmul(x, y)  # out: [M, N]
Y
ying 已提交
5469 5470

            # x: [B, M, K], y: [K]
5471
            # fluid.layers.matmul(x, y)  # out: [B, M]
Y
ying 已提交
5472

5473
            # x: [K], y: [K]
5474
            # fluid.layers.matmul(x, y)  # out: [1]
5475

Y
ying 已提交
5476
            # x: [M], y: [N]
5477 5478
            # fluid.layers.matmul(x, y, True, True)  # out: [M, N]

5479
            import paddle.fluid as fluid
5480 5481 5482
            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
5483
    """
Y
ying 已提交
5484 5485 5486 5487 5488 5489 5490

    def __check_input(x, y):
        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
5491
            y_shape = y_shape + [1]
Y
ying 已提交
5492 5493 5494 5495 5496 5497 5498

        # 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]:
5499 5500
            raise ValueError("Invalid inputs for matmul. x: %s, y: %s\n" %
                             (x_shape, y_shape))
Y
ying 已提交
5501

C
chengduo 已提交
5502
        if len(y_shape) > 2 and len(x_shape) > 2:
Y
ying 已提交
5503
            for i, dim_x in enumerate(x_shape[:-2]):
P
phlrain 已提交
5504 5505 5506
                # don't check neg shape
                if dim_x < 0 or y_shape[i] < 0:
                    continue
Y
ying 已提交
5507
                if dim_x != y_shape[i]:
C
chengduo 已提交
5508 5509
                    raise ValueError("Invalid inputs for matmul. x(%s), y(%s)" %
                                     (x.shape, y.shape))
Y
ying 已提交
5510 5511 5512

    __check_input(x, y)

5513
    helper = LayerHelper('matmul', **locals())
X
Xin Pan 已提交
5514
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
G
guosheng 已提交
5515
    helper.append_op(
5516 5517 5518 5519
        type='matmul',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
S
sneaxiy 已提交
5520 5521 5522
        attrs={
            'transpose_X': transpose_x,
            'transpose_Y': transpose_y,
S
sneaxiy 已提交
5523
            'alpha': float(alpha),
S
sneaxiy 已提交
5524
        })
5525
    return out
5526 5527


5528
def topk(input, k, name=None):
Q
qingqing01 已提交
5529 5530 5531 5532
    """
    This operator is used to find values and indices of the k largest entries
    for the last dimension.

F
fengjiayi 已提交
5533
    If the input is a vector (1-D Tensor), finds the k largest entries in the vector
Q
qingqing01 已提交
5534 5535 5536 5537 5538 5539
    and outputs their values and indices as vectors. Thus values[j] is the j-th
    largest entry in input, and its index is indices[j].

    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
5540 5541 5542 5543 5544 5545 5546 5547 5548 5549 5550 5551 5552 5553 5554 5555 5556 5557 5558 5559 5560
    For example:

    .. code-block:: text

        If:
            input = [[5, 4, 2, 3],
                     [9, 7, 10, 25],
                     [6, 2, 10, 1]]
            k = 2

        Then:
            The first output:
            values = [[5, 4],
                      [10, 25],
                      [6, 10]]

            The second output:
            indices = [[0, 1],
                       [2, 3],
                       [0, 2]]

Q
qingqing01 已提交
5561 5562 5563
    Args:
        input(Variable): The input variable which can be a vector or Tensor with
            higher rank.
W
whs 已提交
5564
        k(int | Variable):  The number of top elements to look for along the last dimension
F
fengjiayi 已提交
5565
                 of input.
5566
        name(str|None): A name for this layer(optional). If set None, the layer
5567
                       will be named automatically.
F
fengjiayi 已提交
5568
                       Default: None
Q
qingqing01 已提交
5569 5570

    Returns:
5571 5572 5573
        Tuple[Variable]: A tuple with two elements. Each element is a Variable.
        The first one is k largest elements along each last
        dimensional slice. The second one is indices of values
F
fengjiayi 已提交
5574
        within the last dimension of input.
Q
qingqing01 已提交
5575

F
fengjiayi 已提交
5576 5577
    Raises:
        ValueError: If k < 1 or k is not less than the last dimension of input
Q
qingqing01 已提交
5578 5579 5580 5581

    Examples:
        .. code-block:: python

5582
            import paddle.fluid as fluid
5583 5584
            import paddle.fluid.layers as layers
            input = layers.data(name="input", shape=[13, 11], dtype='float32')
Q
qingqing01 已提交
5585 5586 5587
            top5_values, top5_indices = layers.topk(input, k=5)
    """
    helper = LayerHelper("top_k", **locals())
X
Xin Pan 已提交
5588 5589
    values = helper.create_variable_for_type_inference(dtype=input.dtype)
    indices = helper.create_variable_for_type_inference(dtype="int64")
W
whs 已提交
5590 5591 5592 5593 5594 5595
    inputs = {"X": [input]}
    attrs = None
    if isinstance(k, Variable):
        inputs['K'] = k
    else:
        attrs = {'k': k}
Q
qingqing01 已提交
5596 5597
    helper.append_op(
        type="top_k",
W
whs 已提交
5598
        inputs=inputs,
Q
qingqing01 已提交
5599 5600
        outputs={"Out": [values],
                 "Indices": [indices]},
W
whs 已提交
5601
        attrs=attrs)
Q
qingqing01 已提交
5602 5603 5604 5605 5606
    values.stop_gradient = True
    indices.stop_gradient = True
    return values, indices


5607 5608 5609 5610 5611 5612
def edit_distance(input,
                  label,
                  normalized=True,
                  ignored_tokens=None,
                  input_length=None,
                  label_length=None):
5613
    """
R
ruri 已提交
5614
    Edit distance operator computes the edit distances between a batch of
Y
ying 已提交
5615 5616 5617 5618 5619 5620 5621 5622
    hypothesis strings and their references. Edit distance, also called
    Levenshtein distance, measures how dissimilar two strings are by counting
    the minimum number of operations to transform one string into anthor.
    Here the operations include insertion, deletion, and substitution.

    For example, given hypothesis string A = "kitten" and reference
    B = "sitting", the edit distance is 3 for A will be transformed into B
    at least after two substitutions and one insertion:
W
wanghaoshuang 已提交
5623

Y
ying 已提交
5624
    "kitten" -> "sitten" -> "sittin" -> "sitting"
W
wanghaoshuang 已提交
5625

5626
    The input is a LoDTensor/Tensor consisting of all the hypothesis strings with
Y
ying 已提交
5627
    the total number denoted by `batch_size`, and the separation is specified
5628 5629
    by the LoD information or input_length. And the `batch_size` reference strings are arranged
    in order in the same way as `input`.
W
wanghaoshuang 已提交
5630

5631
    The output contains the `batch_size` results and each stands for the edit
Y
ying 已提交
5632 5633
    distance for a pair of strings respectively. If Attr(normalized) is true,
    the edit distance will be divided by the length of reference string.
W
wanghaoshuang 已提交
5634

5635
    Args:
5636 5637
        input(Variable): The indices for hypothesis strings, it should have rank 2 and dtype int64.
        label(Variable): The indices for reference strings, it should have rank 2 and dtype int64.
5638
        normalized(bool, default True): Indicated whether to normalize the edit distance by
Y
ying 已提交
5639
                          the length of reference string.
5640
        ignored_tokens(list<int>, default None): Tokens that should be removed before
Y
ying 已提交
5641
                                     calculating edit distance.
5642 5643
        input_length(Variable): The length for each sequence in `input` if it's of Tensor type, it should have shape `[batch_size]` and dtype int64.
        label_length(Variable): The length for each sequence in `label` if it's of Tensor type, it should have shape `[batch_size]` and dtype int64.
5644

W
wanghaoshuang 已提交
5645
    Returns:
5646 5647 5648
        edit_distance_out(Variable): edit distance result in shape [batch_size, 1]. \n
        sequence_num(Variable): sequence number in shape [].
        
W
wanghaoshuang 已提交
5649 5650 5651

    Examples:
        .. code-block:: python
5652
            
R
ruri 已提交
5653 5654
            import paddle.fluid as fluid

5655 5656 5657 5658
            # using LoDTensor
            x_lod = fluid.layers.data(name='x_lod', shape=[1], dtype='int64', lod_level=1)
            y_lod = fluid.layers.data(name='y_lod', shape=[1], dtype='int64', lod_level=1)
            distance_lod, seq_num_lod = fluid.layers.edit_distance(input=x_lod, label=y_lod)
R
ruri 已提交
5659

5660 5661 5662 5663 5664 5665 5666 5667
            # using Tensor
            x_seq_len = 5
            y_seq_len = 6
            x_pad = fluid.layers.data(name='x_pad', shape=[x_seq_len], dtype='int64')
            y_pad = fluid.layers.data(name='y_pad', shape=[y_seq_len], dtype='int64')
            x_len = fluid.layers.data(name='x_len', shape=[], dtype='int64')
            y_len = fluid.layers.data(name='y_len', shape=[], dtype='int64')
            distance_pad, seq_num_pad = fluid.layers.edit_distance(input=x_pad, label=y_pad, input_length=x_len, label_length=y_len)
R
ruri 已提交
5668

5669
    """
5670
    helper = LayerHelper("edit_distance", **locals())
5671

5672
    # remove some tokens from input and labels
W
wanghaoshuang 已提交
5673
    if ignored_tokens is not None and len(ignored_tokens) > 0:
X
Xin Pan 已提交
5674 5675
        erased_input = helper.create_variable_for_type_inference(dtype="int64")
        erased_label = helper.create_variable_for_type_inference(dtype="int64")
5676 5677 5678 5679 5680

        helper.append_op(
            type="sequence_erase",
            inputs={"X": [input]},
            outputs={"Out": [erased_input]},
W
wanghaoshuang 已提交
5681
            attrs={"tokens": ignored_tokens})
5682 5683 5684 5685 5686
        input = erased_input

        helper.append_op(
            type="sequence_erase",
            inputs={"X": [label]},
W
whs 已提交
5687
            outputs={"Out": [erased_label]},
W
wanghaoshuang 已提交
5688
            attrs={"tokens": ignored_tokens})
5689 5690
        label = erased_label

5691 5692 5693 5694 5695
    this_inputs = {"Hyps": [input], "Refs": [label]}
    if input_length and label_length:
        this_inputs['HypsLength'] = [input_length]
        this_inputs['RefsLength'] = [label_length]

5696
    # edit distance op
X
Xin Pan 已提交
5697 5698
    edit_distance_out = helper.create_variable_for_type_inference(dtype="int64")
    sequence_num = helper.create_variable_for_type_inference(dtype="int64")
5699 5700
    helper.append_op(
        type="edit_distance",
5701
        inputs=this_inputs,
5702 5703
        outputs={"Out": [edit_distance_out],
                 "SequenceNum": [sequence_num]},
5704 5705
        attrs={"normalized": normalized})

5706
    return edit_distance_out, sequence_num
5707 5708 5709 5710 5711


def ctc_greedy_decoder(input, blank, name=None):
    """
    This op is used to decode sequences by greedy policy by below steps:
Y
yi.wu 已提交
5712

Y
ying 已提交
5713 5714 5715 5716
    1. Get the indexes of max value for each row in input. a.k.a.
       numpy.argmax(input, axis=0).
    2. For each sequence in result of step1, merge repeated tokens between two
       blanks and delete all blanks.
5717 5718 5719 5720 5721 5722 5723 5724 5725 5726 5727 5728 5729 5730 5731 5732 5733

    A simple example as below:

    .. code-block:: text

        Given:

        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]]

5734
        input.lod = [[4, 4]]
M
minqiyang 已提交
5735

W
whs 已提交
5736
        Computation:
5737

W
whs 已提交
5738 5739 5740 5741 5742 5743
        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:
5744 5745 5746 5747 5748

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

5749
        output.lod = [[2, 1]]
5750

W
whs 已提交
5751

5752 5753
    Args:

Y
ying 已提交
5754 5755 5756 5757 5758 5759 5760 5761 5762
        input(Variable): (LoDTensor<float>), the probabilities of
                         variable-length sequences, which is a 2-D Tensor with
                         LoD information. It's shape is [Lp, num_classes + 1],
                         where Lp is the sum of all input sequences' length and
                         num_classes is the true number of classes. (not
                         including the blank label).
        blank(int): the blank label index of Connectionist Temporal
                    Classification (CTC) loss, which is in thehalf-opened
                    interval [0, num_classes + 1).
5763
        name (str): The name of this layer. It is optional.
5764 5765

    Returns:
H
haowang101779990 已提交
5766 5767 5768
        Variable: CTC greedy decode result which is a 2-D tensor with shape [Lp, 1]. \
                  'Lp' is the sum if all output sequences' length. If all the sequences \
                  in result were empty, the result LoDTensor will be [-1] with  \
M
minqiyang 已提交
5769
                  LoD [[]] and dims [1, 1].
5770 5771 5772 5773

    Examples:
        .. code-block:: python

S
SunGaofeng 已提交
5774
            import paddle.fluid as fluid
5775 5776
            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
            cost = fluid.layers.ctc_greedy_decoder(input=x, blank=0)
W
wanghaoshuang 已提交
5777
    """
5778
    helper = LayerHelper("ctc_greedy_decoder", **locals())
Q
qingqing01 已提交
5779
    _, topk_indices = topk(input, k=1)
5780 5781

    # ctc align op
X
Xin Pan 已提交
5782
    ctc_out = helper.create_variable_for_type_inference(dtype="int64")
5783 5784 5785
    helper.append_op(
        type="ctc_align",
        inputs={"Input": [topk_indices]},
W
wanghaoshuang 已提交
5786
        outputs={"Output": [ctc_out]},
5787 5788
        attrs={"merge_repeated": True,
               "blank": blank})
5789
    return ctc_out
5790 5791


5792 5793 5794 5795 5796 5797
def warpctc(input,
            label,
            blank=0,
            norm_by_times=False,
            input_length=None,
            label_length=None):
W
wanghaoshuang 已提交
5798
    """
5799 5800
    An operator integrating the open source Warp-CTC library
    (https://github.com/baidu-research/warp-ctc)
W
wanghaoshuang 已提交
5801
    to compute Connectionist Temporal Classification (CTC) loss.
5802 5803
    It can be aliased as softmax with CTC, since a native softmax activation is
    interated to the Warp-CTC library, to to normlize values for each row of the
W
wanghaoshuang 已提交
5804 5805 5806
    input tensor.

    Args:
5807
       input (Variable): The unscaled probabilities of variable-length sequences,
5808 5809 5810
         which is a 2-D Tensor with LoD information, or a 3-D Tensor without Lod
         information. When it is a 2-D LodTensor, it's shape is 
         [Lp, num_classes + 1], where Lp is the sum of all input
W
wanghaoshuang 已提交
5811
         sequences' length and num_classes is the true number of classes.
5812 5813 5814 5815
         (not including the blank label). When it is a 3-D Tensor, it's shape 
         is [max_logit_length, batch_size, num_classes + 1],
         where max_logit_length is the length of the longest
         input logit sequence.
5816
       label (Variable): The ground truth of variable-length sequence,
5817 5818 5819
         which is a 2-D Tensor with LoD information or a 2-D Tensor without
         LoD information. When it is a 2-D LoDTensor or 2-D Tensor, 
         it is of the shape [Lg, 1], where Lg is th sum of all labels' length.
5820
       blank (int, default 0): The blank label index of Connectionist
W
wanghaoshuang 已提交
5821 5822
         Temporal Classification (CTC) loss, which is in the
         half-opened interval [0, num_classes + 1).
5823 5824 5825
       norm_by_times(bool, default false): Whether to normalize the gradients
         by the number of time-step, which is also the sequence's length.
         There is no need to normalize the gradients if warpctc layer was
5826
         follewed by a mean_op.
5827 5828 5829 5830
       input_length(Variable): The length for each input sequence if it is 
         of Tensor type, it should have shape `[batch_size]` and dtype int64.
       label_length(Variable): The length for each label sequence if it is
         of Tensor type, it should have shape `[batch_size]` and dtype int64.
W
wanghaoshuang 已提交
5831 5832

    Returns:
5833 5834
        Variable: The Connectionist Temporal Classification (CTC) loss,
        which is a 2-D Tensor of the shape [batch_size, 1].
W
wanghaoshuang 已提交
5835 5836 5837

    Examples:
        .. code-block:: python
5838

5839
            # using LoDTensor
B
Bai Yifan 已提交
5840
            import paddle.fluid as fluid
5841 5842 5843
            import numpy as np
            
            label = fluid.layers.data(name='label', shape=[12, 1],
B
Bai Yifan 已提交
5844
                                      dtype='float32', lod_level=1)
5845 5846 5847
            predict = fluid.layers.data(name='predict', 
                                        shape=[11, 8],
                                        dtype='float32',lod_level=1)
5848
            cost = fluid.layers.warpctc(input=predict, label=label)
W
wanghaoshuang 已提交
5849

5850 5851 5852 5853 5854 5855 5856 5857 5858 5859 5860 5861 5862 5863 5864 5865 5866 5867
            # using Tensor
            input_length = fluid.layers.data(name='logits_length', shape=[11],
                                         dtype='int64')
            label_length = fluid.layers.data(name='labels_length', shape=[12],
                                         dtype='int64')
            target = fluid.layers.data(name='target', shape=[12, 1],
                                       dtype='int32')
            # length of the longest logit sequence
            max_seq_length = 4
            # number of logit sequences
            batch_size = 4
            output = fluid.layers.data(name='output', 
                                       shape=[max_seq_length, batch_size, 8],
                                       dtype='float32')
            loss = fluid.layers.warpctc(input=output,label=target,
                                        input_length=input_length,
                                        label_length=label_length)

W
wanghaoshuang 已提交
5868
    """
F
fengjiayi 已提交
5869
    helper = LayerHelper('warpctc', **locals())
5870 5871 5872 5873 5874
    this_inputs = {'Logits': [input], 'Label': [label]}
    if input_length and label_length:
        this_inputs['LogitsLength'] = [input_length]
        this_inputs['LabelLength'] = [label_length]

X
Xin Pan 已提交
5875 5876
    loss_out = helper.create_variable_for_type_inference(dtype=input.dtype)
    grad_out = helper.create_variable_for_type_inference(dtype=input.dtype)
5877

W
wanghaoshuang 已提交
5878 5879
    helper.append_op(
        type='warpctc',
5880
        inputs=this_inputs,
W
wanghaoshuang 已提交
5881 5882
        outputs={'WarpCTCGrad': [grad_out],
                 'Loss': [loss_out]},
W
Wu Yi 已提交
5883 5884 5885 5886
        attrs={
            'blank': blank,
            'norm_by_times': norm_by_times,
        })
W
wanghaoshuang 已提交
5887
    return loss_out
5888 5889 5890 5891 5892 5893 5894 5895 5896 5897 5898 5899 5900 5901 5902


def sequence_reshape(input, new_dim):
    """
    **Sequence Reshape Layer**

    This layer will rearrange the input sequences. The new dimension is set by
    user. Length of each sequence is computed according to original length,
    original dimension and new dimension. The following example will help to
    illustrate the function of this layer:

    .. code-block:: text

        x is a LoDTensor:
            x.lod  = [[0, 2, 6]]
5903 5904 5905
            x.data = [[1,  2], [3,  4],
                      [5,  6], [7,  8],
                      [9, 10], [11, 12]]
5906 5907 5908 5909 5910
            x.dims = [6, 2]

        set new_dim = 4

        then out is a LoDTensor:
5911

5912
            out.lod  = [[0, 1, 3]]
5913 5914 5915 5916

            out.data = [[1,  2,  3,  4],
                        [5,  6,  7,  8],
                        [9, 10, 11, 12]]
5917 5918 5919 5920 5921 5922 5923
            out.dims = [3, 4]

    Currently, only 1-level LoDTensor is supported and please make sure
    (original length * original dimension) can be divided by new dimension with
    no remainder for each sequence.

    Args:
5924 5925 5926

       input (Variable): A 2-D LoDTensor with shape being [N, M] where M for dimension.
       new_dim (int): New dimension that the input LoDTensor is reshaped to.
5927 5928

    Returns:
5929

5930 5931 5932 5933 5934
        Variable: Reshaped LoDTensor according to new dimension.

    Examples:
        .. code-block:: python

B
bdzhuxiaoning 已提交
5935 5936 5937
            import paddle.fluid as fluid
            x = fluid.layers.data(name='x', shape=[2, 6], append_batch_size=False, dtype='float32', lod_level=1)
            x_reshaped = fluid.layers.sequence_reshape(input=x, new_dim=4)
5938
    """
L
lujun 已提交
5939
    assert not in_dygraph_mode(), (
5940
        "sequence layer is not supported in dygraph mode yet.")
5941
    helper = LayerHelper('sequence_reshape', **locals())
X
Xin Pan 已提交
5942
    out = helper.create_variable_for_type_inference(helper.input_dtype())
5943 5944 5945 5946 5947 5948
    helper.append_op(
        type='sequence_reshape',
        inputs={'X': [input]},
        outputs={'Out': [out]},
        attrs={'new_dim': new_dim})
    return out
Y
ying 已提交
5949 5950


5951 5952 5953 5954
# FIXME(wuyi): let docstring_checker.py understand @autodoc.
# For now, the comments in c++ use types like Tensor, but in python side
# the type is often "Variable", and arguments may vary.
@templatedoc(op_type="nce")
Y
Yang Yu 已提交
5955 5956 5957 5958 5959 5960
def nce(input,
        label,
        num_total_classes,
        sample_weight=None,
        param_attr=None,
        bias_attr=None,
C
chengduo 已提交
5961
        num_neg_samples=None,
5962 5963 5964
        name=None,
        sampler="uniform",
        custom_dist=None,
5965 5966
        seed=0,
        is_sparse=False):
5967 5968 5969 5970 5971 5972 5973
    """
    ${comment}

    Args:
        input (Variable): input variable.
        label (Variable): label.
        num_total_classes (int):${num_total_classes_comment}
5974 5975
        sample_weight (Variable|None): A Variable of shape [batch_size, 1]
            storing a weight for each sample. The default weight for each
5976
            sample is 1.0.
C
chengduo 已提交
5977 5978 5979 5980 5981 5982 5983 5984 5985
        param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
             of nce. If it is set to None or one attribute of ParamAttr, nce
             will create ParamAttr as param_attr. If the Initializer of the param_attr
             is not set, the parameter is initialized with Xavier. Default: None.
        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of nce.
             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, nce
             will create ParamAttr as bias_attr. If the Initializer of the bias_attr
             is not set, the bias is initialized zero. Default: None.
5986
        num_neg_samples (int): ${num_neg_samples_comment}
C
chengduo 已提交
5987 5988
        name (str|None): A name for this layer(optional). If set None, the layer
             will be named automatically. Default: None.
5989 5990 5991
        sampler (str): The sampler used to sample class from negtive classes.
                       It can be 'uniform', 'log_uniform' or 'custom_dist'.
                       default: 'uniform'.
5992
        custom_dist (float[]): A float[] with size=num_total_classes.
5993 5994 5995 5996
                       It is used when sampler is set to 'custom_dist'.
                       custom_dist[i] is the probsbility of i-th class to be sampled.
                       default: None.
        seed (int): The seed used in sampler. default: 0.
5997
        is_sparse(bool): The flag indicating whether to use sparse update, the weight@GRAD and bias@GRAD will be changed to SelectedRows.
F
fengjiayi 已提交
5998

5999
    Returns:
Y
Yibing Liu 已提交
6000 6001 6002 6003 6004 6005
        Variable: The output nce loss.

    Examples:
        .. code-block:: python


X
xsrobin 已提交
6006 6007 6008 6009 6010 6011 6012 6013 6014 6015 6016 6017 6018 6019 6020 6021 6022 6023 6024 6025 6026 6027 6028 6029 6030 6031 6032 6033 6034 6035 6036 6037 6038 6039
            import paddle.fluid as fluid
            import numpy as np

            window_size = 5
            words = []
            for i in xrange(window_size):
                words.append(fluid.layers.data(
                    name='word_{0}'.format(i), shape=[1], dtype='int64'))

            dict_size = 10000
            label_word = int(window_size / 2) + 1

            embs = []
            for i in xrange(window_size):
                if i == label_word:
                    continue

                emb = fluid.layers.embedding(input=words[i], size=[dict_size, 32],
                                   param_attr='embed', is_sparse=True)
                embs.append(emb)

            embs = fluid.layers.concat(input=embs, axis=1)
            loss = fluid.layers.nce(input=embs, label=words[label_word],
                      num_total_classes=dict_size, param_attr='nce.w_0',
                      bias_attr='nce.b_0')

             #or use custom distribution
             dist = np.array([0.05,0.5,0.1,0.3,0.05])
             loss = fluid.layers.nce(input=embs, label=words[label_word],
                       num_total_classes=5, param_attr='nce.w_1',
                       bias_attr='nce.b_1',
                       num_neg_samples=3,
                       sampler="custom_dist",
                       custom_dist=dist)
6040
    """
Y
Yang Yu 已提交
6041 6042 6043
    helper = LayerHelper('nce', **locals())
    assert isinstance(input, Variable)
    assert isinstance(label, Variable)
C
chengduo 已提交
6044 6045

    dim = input.shape[1]
Y
Yang Yu 已提交
6046 6047 6048 6049 6050 6051
    num_true_class = label.shape[1]
    w = helper.create_parameter(
        attr=helper.param_attr,
        shape=[num_total_classes, dim],
        is_bias=False,
        dtype=input.dtype)
6052
    inputs = {}
C
chengduo 已提交
6053 6054 6055 6056 6057 6058 6059
    if helper.bias_attr:
        b = helper.create_parameter(
            attr=helper.bias_attr,
            shape=[num_total_classes, 1],
            is_bias=True,
            dtype=input.dtype)
        inputs['Bias'] = b
X
Xin Pan 已提交
6060 6061 6062
    cost = helper.create_variable_for_type_inference(dtype=input.dtype)
    sample_logits = helper.create_variable_for_type_inference(dtype=input.dtype)
    sample_labels = helper.create_variable_for_type_inference(dtype=label.dtype)
Y
Yang Yu 已提交
6063

6064 6065 6066 6067
    inputs['Input'] = input
    inputs['Label'] = label
    inputs['Weight'] = w
    inputs['SampleWeight'] = sample_weight if sample_weight is not None else []
6068 6069 6070 6071 6072 6073 6074

    if sampler == "uniform":
        sampler = 0
    elif sampler == "log_uniform":
        sampler = 1
    elif sampler == "custom_dist":
        assert custom_dist is not None
6075 6076
        # assert isinstance(custom_dist, Variable)

Y
Yibing Liu 已提交
6077
        custom_dist_len = num_total_classes
6078 6079 6080 6081 6082 6083
        alias_probs_ = [0] * custom_dist_len
        alias_ = [0] * custom_dist_len
        bigs = []
        littles = []
        for i in range(custom_dist_len):
            normal_prob = custom_dist[i] * custom_dist_len
6084
            if normal_prob - 1.0 > 0:
6085
                bigs.append((i, normal_prob))
6086
            elif 1.0 - normal_prob > 0:
6087 6088 6089 6090 6091 6092 6093 6094 6095 6096 6097 6098 6099 6100 6101
                littles.append((i, normal_prob))
            else:
                alias_probs_[i] = normal_prob
                alias_[i] = -1

        while len(bigs) and len(littles):
            big = bigs.pop(0)
            little = littles.pop(0)

            big_idx = big[0]
            big_prob = big[1]

            alias_probs_[little[0]] = little[1]
            alias_[little[0]] = big_idx
            big_left = big[1] + little[1] - 1
6102
            if big_left - 1.0 > 0:
6103
                bigs.append((big_idx, big_left))
6104
            elif 1.0 - big_left > 0:
6105 6106 6107 6108 6109 6110 6111 6112 6113 6114 6115 6116 6117 6118
                littles.append((big_idx, big_left))
            else:
                alias_probs_[big_idx] = big_left
                alias_[big_idx] = -1

        if len(bigs):
            big = bigs.pop(0)
            alias_probs_[big[0]] = 1.0
            alias_[big[0]] = -1
        if len(littles):
            little = littles.pop(0)
            alias_probs_[little[0]] = 1.0
            alias_[little[0]] = -1

6119 6120 6121 6122 6123 6124 6125 6126 6127 6128 6129 6130 6131 6132 6133
        def _init_by_numpy_array(numpy_array):
            ret = helper.create_parameter(
                attr=ParamAttr(),
                shape=numpy_array.shape,
                dtype=numpy_array.dtype,
                default_initializer=NumpyArrayInitializer(numpy_array))
            ret.stop_gradient = True
            return ret

        inputs['CustomDistProbs'] = _init_by_numpy_array(
            np.array(custom_dist).astype('float32'))
        inputs['CustomDistAlias'] = _init_by_numpy_array(
            np.array(alias_).astype('int32'))
        inputs['CustomDistAliasProbs'] = _init_by_numpy_array(
            np.array(alias_probs_).astype('float32'))
6134 6135 6136 6137
        sampler = 2
    else:
        raise Exception("Unsupported sampler type.")

6138 6139 6140 6141 6142
    if num_neg_samples is None:
        num_neg_samples = 10
    else:
        num_neg_samples = int(num_neg_samples)

6143 6144 6145 6146
    remote_prefetch = is_sparse
    print(
        "With sparse mode, if your models has only small parameter prefetch may cause speed down"
    )
6147

Y
Yang Yu 已提交
6148 6149
    attrs = {
        'num_total_classes': int(num_total_classes),
6150 6151
        'num_neg_samples': num_neg_samples,
        'seed': seed,
6152
        'sampler': sampler,
6153 6154
        'is_sparse': is_sparse,
        'remote_prefetch': remote_prefetch
Y
Yang Yu 已提交
6155
    }
Y
Yang Yu 已提交
6156 6157 6158

    helper.append_op(
        type='nce',
C
chengduo 已提交
6159
        inputs=inputs,
Y
Yang Yu 已提交
6160 6161 6162 6163 6164 6165
        outputs={
            'Cost': cost,
            'SampleLogits': sample_logits,
            'SampleLabels': sample_labels
        },
        attrs=attrs)
Y
Yang Yu 已提交
6166
    return cost / (num_neg_samples + 1)
6167 6168


C
chengduo 已提交
6169 6170
def hsigmoid(input,
             label,
6171
             num_classes,
C
chengduo 已提交
6172 6173
             param_attr=None,
             bias_attr=None,
J
JiabinYang 已提交
6174
             name=None,
6175 6176 6177
             path_table=None,
             path_code=None,
             is_custom=False,
J
JiabinYang 已提交
6178
             is_sparse=False):
W
weixing02 已提交
6179 6180
    """
    The hierarchical sigmoid operator is used to accelerate the training
M
minqiyang 已提交
6181
    process of language model. This operator organizes the classes into a
M
minqiyang 已提交
6182
    complete binary tree, or you can use is_custom to pass your own tree to
6183
    implement hierarchical. Each leaf node represents a class(a word) and each
G
guosheng 已提交
6184 6185 6186 6187 6188 6189
    internal node acts as a binary classifier. For each word there's a unique
    path from root to it's leaf node, hsigmoid calculate the cost for each
    internal node on the path, and sum them to get a total cost. hsigmoid can
    achive a acceleration from :math:`O(N)` to :math:`O(logN)`, where :math:`N`
    represents the size of word dict.

6190
    Using default tree you can Refer to `Hierarchical Probabilistic Neural Network Language Model
G
guosheng 已提交
6191
    <http://www.iro.umontreal.ca/~lisa/pointeurs/hierarchical-nnlm-aistats05.pdf>`_
M
minqiyang 已提交
6192

6193 6194
    And if you want to use the costumed tree by set 'is_custom' as true you may need to do following things first:

H
haowang101779990 已提交
6195 6196 6197 6198
    1. using your word dict to build a binary tree, each leaf node should be an word of your word dict
    2. build a dict to store word_id -> word's leaf to root path, we call it path_table.
    3. build a dict to store word_id -> code of word's leaf to root path, we call it path_code. Code
       means label of each binary classification, using 1 indicate true, 0 indicate false.
M
minqiyang 已提交
6199
    4. now, each word should has its path and code along the path, you can pass a batch of path and code
H
haowang101779990 已提交
6200
       related to the same batch of inputs.
6201

W
weixing02 已提交
6202
    Args:
M
minqiyang 已提交
6203
        input (Variable): The input tensor variable with shape
G
guosheng 已提交
6204 6205 6206 6207
            :math:`[N \\times D]`, where :math:`N` is the size of mini-batch,
            and :math:`D` is the feature size.
        label (Variable): The tensor variable contains labels of training data.
            It's a tensor with shape is :math:`[N \\times 1]`.
M
minqiyang 已提交
6208 6209
        num_classes: (int), The number of classes, must not be less than 2. with default tree this has to be set,
            it should never be None under is_custom=False, but while is_custom is true, it should be non leaf num
6210
            which indicates the num of classes using by binary classify.
C
chengduo 已提交
6211 6212 6213 6214 6215 6216 6217 6218 6219 6220 6221
        param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
             of hsigmoid. If it is set to None or one attribute of ParamAttr, hsigmoid
             will create ParamAttr as param_attr. If the Initializer of the param_attr
             is not set, the parameter is initialized with Xavier. Default: None.
        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of hsigmoid.
             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, hsigmoid
             will create ParamAttr as bias_attr. If the Initializer of the bias_attr
             is not set, the bias is initialized zero. Default: None.
        name (str|None): A name for this layer(optional). If set None, the layer
             will be named automatically. Default: None.
M
minqiyang 已提交
6222
        path_table: (Variable|None) this variable can store each batch of samples' path to root,
6223
            it should be in leaf -> root order
M
minqiyang 已提交
6224 6225 6226
            path_table should have the same shape with path_code, and for each sample i path_table[i] indicates a np.array like
            structure and each element in this array is indexes in parent nodes' Weight Matrix.
        path_code:  (Variable|None) this variable can store each batch of samples' code,
6227
            each code consist with every code of parent nodes. it should be in leaf -> root order
M
minqiyang 已提交
6228
        is_custom: (bool|False)using user defined binary tree instead of default complete binary tree, if costum is
6229
             set you need to set path_table/path_code/num_classes, otherwise num_classes should be set
M
minqiyang 已提交
6230
        is_sparse: (bool|False)using sparse update instead of dense update, if set, the gradient
6231
             of W and input will be sparse.
W
weixing02 已提交
6232 6233

    Returns:
J
JiabinYang 已提交
6234
        Out: (LodTensor) The cost of hierarchical sigmoid operator. the shape is [N, 1]
W
weixing02 已提交
6235 6236 6237 6238 6239

    Examples:

        .. code-block:: python

6240
            import paddle.fluid as fluid
G
guosheng 已提交
6241 6242 6243
            x = fluid.layers.data(name='x', shape=[2], dtype='float32')
            y = fluid.layers.data(name='y', shape=[1], dtype='int64')
            out = fluid.layers.hsigmoid(input=x, label=y, num_classes=6)
W
weixing02 已提交
6244 6245 6246 6247
    """

    helper = LayerHelper('hierarchical_sigmoid', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
6248 6249
    out = helper.create_variable_for_type_inference(dtype)
    pre_out = helper.create_variable_for_type_inference(dtype)
W
weixing02 已提交
6250
    dim = input.shape[1]
6251
    if ((num_classes is None) or (num_classes < 2)) and (not is_custom):
J
JiabinYang 已提交
6252 6253 6254
        raise ValueError(
            "num_classes must not be less than 2 with default tree")

6255 6256 6257 6258 6259 6260 6261 6262 6263
    if (not is_custom) and (is_sparse):
        print("Sparse mode should not be used without custom tree")
        is_sparse = False

    if (not is_custom) and ((path_table is not None) or
                            (path_code is not None)):
        raise ValueError(
            "only num_classes should be passed without custom tree")

6264
    if (is_custom) and (path_code is None):
6265
        raise ValueError("path_code should not be None with custom tree")
6266
    elif (is_custom) and (path_table is None):
6267
        raise ValueError("path_table should not be None with custom tree")
6268
    elif (is_custom) and (num_classes is None):
6269
        raise ValueError("num_classes should not be None with custom tree")
6270 6271 6272
    else:
        pass

J
JiabinYang 已提交
6273
    weights = None
6274 6275 6276 6277
    remote_prefetch = is_sparse
    print(
        "With sparse mode, if your models has only small parameter prefetch may cause speed down"
    )
6278
    if not is_custom:
J
JiabinYang 已提交
6279 6280 6281 6282 6283 6284 6285 6286
        weights = helper.create_parameter(
            attr=helper.param_attr,
            shape=[num_classes - 1, dim],
            is_bias=False,
            dtype=input.dtype)
    else:
        weights = helper.create_parameter(
            attr=helper.param_attr,
6287
            shape=[num_classes, dim],
J
JiabinYang 已提交
6288 6289
            is_bias=False,
            dtype=input.dtype)
6290 6291 6292
    inputs = {
        "X": input,
        "W": weights,
6293
        "PathTable": path_table,
6294
        "PathCode": path_code,
6295 6296
        "Label": label
    }
W
weixing02 已提交
6297
    if helper.bias_attr:
6298
        if not is_custom:
J
JiabinYang 已提交
6299 6300
            bias = helper.create_parameter(
                attr=helper.bias_attr,
J
JiabinYang 已提交
6301
                shape=[num_classes - 1, 1],
J
JiabinYang 已提交
6302 6303 6304 6305 6306 6307
                is_bias=True,
                dtype=input.dtype)
            inputs['Bias'] = bias
        else:
            bias = helper.create_parameter(
                attr=helper.bias_attr,
6308
                shape=[num_classes, 1],
J
JiabinYang 已提交
6309 6310 6311
                is_bias=True,
                dtype=input.dtype)
            inputs['Bias'] = bias
W
weixing02 已提交
6312 6313
    helper.append_op(
        type="hierarchical_sigmoid",
W
weixing02 已提交
6314
        inputs=inputs,
W
weixing02 已提交
6315
        outputs={"Out": out,
6316 6317 6318 6319 6320 6321 6322
                 "PreOut": pre_out,
                 "W_Out": weights},
        attrs={
            "num_classes": num_classes,
            "is_sparse": is_sparse,
            "remote_prefetch": remote_prefetch
        })
W
weixing02 已提交
6323 6324 6325
    return out


Y
fix ci.  
ying 已提交
6326
def transpose(x, perm, name=None):
Y
ying 已提交
6327 6328 6329 6330 6331 6332 6333
    """
    Permute the dimensions of `input` according to `perm`.

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

    Args:
6334 6335 6336
        x (Variable): The input Tensor.
        perm (list): A permutation of the dimensions of `input`.
        name (str): The name of this layer. It is optional.
Y
ying 已提交
6337 6338 6339 6340 6341 6342 6343

    Returns:
        Variable: A transposed Tensor.

    Examples:
        .. code-block:: python

6344
            # use append_batch_size=False to avoid prepending extra
6345
            # batch size in shape
6346
            import paddle.fluid as fluid
6347
            x = fluid.layers.data(name='x', shape=[5, 10, 15],
6348
                            dtype='float32', append_batch_size=False)
6349
            x_transposed = fluid.layers.transpose(x, perm=[1, 0, 2])
Y
ying 已提交
6350 6351
    """

Y
fix ci.  
ying 已提交
6352
    if len(perm) != len(x.shape):
Y
ying 已提交
6353 6354
        raise ValueError(
            "Input(perm) is the permutation of dimensions of Input(input). "
6355
            "Its length should be equal to Input(input)'s rank.")
Y
ying 已提交
6356 6357 6358 6359 6360 6361
    for idx, dim in enumerate(perm):
        if dim >= len(x.shape):
            raise ValueError(
                "Each element in perm should be less than x's rank. "
                "%d-th element in perm is %d which accesses x's rank %d." %
                (idx, perm[idx], len(x.shape)))
Y
ying 已提交
6362 6363

    helper = LayerHelper('transpose', **locals())
X
Xin Pan 已提交
6364 6365
    out = helper.create_variable_for_type_inference(x.dtype)
    x_shape = helper.create_variable_for_type_inference(x.dtype)
Y
ying 已提交
6366
    helper.append_op(
6367
        type='transpose2',
Y
fix ci.  
ying 已提交
6368
        inputs={'X': [x]},
6369 6370
        outputs={'Out': [out],
                 'XShape': [x_shape]},
Y
ying 已提交
6371 6372
        attrs={'axis': perm})
    return out
6373 6374


6375 6376 6377 6378 6379 6380 6381
def im2sequence(input,
                filter_size=1,
                stride=1,
                padding=0,
                input_image_size=None,
                out_stride=1,
                name=None):
6382
    """
6383 6384 6385 6386 6387 6388 6389
    Extracts image patches from the input tensor to form a tensor of shape
    {input.batch_size * output_height * output_width, filter_size_H *
    filter_size_W * input.channels} which is similar with im2col.
    This op use filter / kernel to scan images and convert these images to
    sequences. After expanding, the number of time step are
    output_height * output_width for an image, in which output_height and
    output_width are calculated by below equation:
6390 6391 6392 6393 6394 6395 6396 6397 6398 6399

    .. math::

        output\_size = 1 + \
            (2 * padding + img\_size - block\_size + stride - 1) / stride

    And the dimension of each time step is block_y * block_x * input.channels.

    Args:
        input (Variable): The input should be a tensor in NCHW format.
W
wanghaoshuang 已提交
6400 6401 6402 6403 6404 6405 6406 6407 6408 6409 6410 6411 6412 6413 6414 6415 6416 6417

        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
            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 can
            contain two integers like (padding_H, padding_W) which means
            padding_up = padding_down = padding_H and
            padding_left = padding_right = padding_W. Or it can use
            (padding_up, padding_left, padding_down, padding_right) to indicate
            paddings of four direction. Otherwise, a scalar padding means
            padding_up = padding_down = padding_left = padding_right = padding
            Default: padding = 0.

6418 6419 6420 6421 6422 6423 6424 6425 6426
        input_image_size(Variable): the input contains image real size.It's dim
            is [batchsize, 2]. It is dispensable.It is just for batch inference.

        out_stride(int|tuple): The scaling of image through CNN. It is
            dispensable. It is valid only when input_image_size is not null.
            If out_stride is tuple,  it must contain two intergers,
            (out_stride_H, out_stride_W). Otherwise,
            the out_stride_H = out_stride_W = out_stride.

6427 6428 6429
        name (int): The name of this layer. It is optional.

    Returns:
W
wanghaoshuang 已提交
6430 6431 6432 6433 6434
        output: The output is a LoDTensor with shape
        {input.batch_size * output_height * output_width,
        filter_size_H * filter_size_W * input.channels}.
        If we regard output as a matrix, each row of this matrix is
        a step of a sequence.
6435 6436 6437 6438 6439 6440 6441 6442 6443 6444 6445 6446 6447 6448 6449 6450 6451 6452 6453 6454 6455 6456 6457 6458 6459 6460 6461

    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
6462 6463 6464
            filter = [2, 2]
            stride = [1, 1]
            padding = [0, 0]
6465 6466 6467 6468 6469 6470 6471 6472 6473 6474 6475 6476

            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.]]

6477
            output.dims = {8, 8}
6478

6479
            output.lod = [[4, 4]]
6480

T
Tink_Y 已提交
6481
    Examples:
6482 6483 6484

        .. code-block:: python

B
Bai Yifan 已提交
6485 6486 6487
            import paddle.fluid as fluid
            data = fluid.layers.data(name='data', shape=[3, 32, 32],
                                     dtype='float32')
6488
            output = fluid.layers.im2sequence(
B
Bai Yifan 已提交
6489 6490
                input=data, stride=[1, 1], filter_size=[2, 2])

6491 6492

    """
L
lujun 已提交
6493
    assert not in_dygraph_mode(), (
6494
        "sequence layer is not supported in dygraph mode yet.")
W
wanghaoshuang 已提交
6495 6496 6497 6498 6499 6500 6501 6502 6503 6504

    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])
6505
    inputs = {"X": input}
6506
    attrs = {"kernels": filter_size, "strides": stride, "paddings": padding}
6507 6508 6509 6510 6511
    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
6512
    helper = LayerHelper('im2sequence', **locals())
X
Xin Pan 已提交
6513
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
6514
    helper.append_op(
6515
        type='im2sequence', inputs=inputs, outputs={'Out': out}, attrs=attrs)
6516
    return out
6517 6518


Y
yuyang18 已提交
6519
@templatedoc()
6520
def row_conv(input, future_context_size, param_attr=None, act=None):
Y
yuyang18 已提交
6521 6522
    """
    ${comment}
6523 6524

    Args:
Y
yuyang18 已提交
6525
        input (${x_type}): ${x_comment}.
Y
yangyaming 已提交
6526 6527
        future_context_size (int): Future context size. Please note, the shape
            of convolution kernel is [future_context_size + 1, D].
6528 6529 6530 6531 6532
        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
6533
        ${out_comment}.
6534 6535

    Examples:
Y
yuyang18 已提交
6536 6537 6538 6539
        >>> import paddle.fluid as fluid
        >>> x = fluid.layers.data(name='x', shape=[16],
        >>>                        dtype='float32', lod_level=1)
        >>> out = fluid.layers.row_conv(input=x, future_context_size=2)
6540 6541 6542 6543 6544 6545
    """
    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
6546
    out = helper.create_variable_for_type_inference(dtype)
6547 6548 6549 6550 6551
    helper.append_op(
        type='row_conv',
        inputs={'X': [input],
                'Filter': [filter_param]},
        outputs={'Out': [out]})
Y
yangyaming 已提交
6552
    return helper.append_activation(out)
6553 6554


Y
yuyang18 已提交
6555
@templatedoc()
6556 6557
def multiplex(inputs, index):
    """
Y
yuyang18 已提交
6558 6559
    ${comment}

L
lujun 已提交
6560 6561 6562 6563 6564 6565 6566 6567 6568 6569 6570 6571 6572 6573 6574 6575 6576 6577 6578 6579 6580 6581 6582 6583 6584 6585 6586 6587 6588 6589 6590 6591 6592 6593 6594 6595 6596 6597 6598 6599 6600 6601 6602
    For Example:

    .. code-block:: text

        case 1:

        Given:

        X = [[[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]]]

        index = [3,0,1,2]

        out:[[3 0 3 4]    // X[3,0] (3 = index[i], 0 = i); i=0
             [0 1 3 4]    // X[0,1] (0 = index[i], 1 = i); i=1
             [1 2 4 2]    // X[1,2] (0 = index[i], 2 = i); i=2
             [2 3 3 4]]   // X[2,3] (0 = index[i], 3 = i); i=3

        case 2:

        Given:

        X = [[[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]]]

        index = [1,0]

        out:[[1 0 3 4]    // X[1,0] (3 = index[0], 0 = i); i=1
             [0 1 3 4]    // X[0,1] (0 = index[1], 1 = i); i=2
             [0 2 4 4]    // X[0,2] (0 = 0, 2 = i); i=3
             [0 3 3 4]]   // X[0,3] (0 = 0, 3 = i); i=4

    Examples:

    .. code-block:: python

        import paddle.fluid as fluid
        x1 = fluid.layers.data(name='x1', shape=[4], dtype='float32')
        x2 = fluid.layers.data(name='x2', shape=[4], dtype='float32')
        index = fluid.layers.data(name='index', shape=[1], dtype='int32')
        out = fluid.layers.multiplex(inputs=[x1, x2], index=index)
6603 6604

    Args:
Y
yuyang18 已提交
6605 6606
       inputs (list): ${x_comment}.
       index (${ids_type}): ${ids_comment}.
6607 6608

    Returns:
Y
yuyang18 已提交
6609
        ${out_comment}.
6610 6611
    """
    helper = LayerHelper('multiplex', **locals())
Y
yangyaming 已提交
6612 6613 6614 6615 6616

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

X
Xin Pan 已提交
6617
    out = helper.create_variable_for_type_inference(inputs[0].dtype)
6618 6619 6620 6621 6622 6623
    helper.append_op(
        type='multiplex',
        inputs={'X': inputs,
                'Ids': index},
        outputs={'Out': [out]})
    return out
6624 6625


6626 6627 6628
def softmax_with_cross_entropy(logits,
                               label,
                               soft_label=False,
J
jerrywgz 已提交
6629
                               ignore_index=kIgnoreIndex,
6630
                               numeric_stable_mode=True,
6631 6632
                               return_softmax=False,
                               axis=-1):
6633 6634
    """
    **Softmax With Cross Entropy Operator.**
6635

6636
    Cross entropy loss with softmax is used as the output layer extensively. This
6637 6638 6639
    operator computes the softmax normalized values for dimension :attr:`axis` of 
    the input tensor, after which cross-entropy loss is computed. This provides 
    a more numerically stable gradient.
6640

6641 6642 6643
    Because this operator performs a softmax on logits internally, it expects
    unscaled logits. This operator should not be used with the output of
    softmax operator since that would produce incorrect results.
6644

6645 6646 6647 6648
    When the attribute :attr:`soft_label` is set :attr:`False`, this operators 
    expects mutually exclusive hard labels, each sample in a batch is in exactly 
    one class with a probability of 1.0. Each sample in the batch will have a 
    single label.
6649

6650
    The equation is as follows:
6651

6652
    1) Hard label (one-hot label, so every sample has exactly one class)
6653

6654 6655 6656 6657
    .. math::

        loss_j =  -\\text{logit}_{label_j} +
        \\log\\left(\\sum_{i=0}^{K}\\exp(\\text{logit}_i)\\right), j = 1,..., K
6658

6659 6660 6661
    2) Soft label (each sample can have a distribution over all classes)

    .. math::
6662

6663 6664 6665 6666
        loss_j =  -\\sum_{i=0}^{K}\\text{label}_i
        \\left(\\text{logit}_i - \\log\\left(\\sum_{i=0}^{K}
        \\exp(\\text{logit}_i)\\right)\\right), j = 1,...,K

6667 6668
    3) If :attr:`numeric_stable_mode` is :attr:`True`, softmax is calculated 
    first by:
S
sneaxiy 已提交
6669 6670

    .. math::
6671

H
haowang101779990 已提交
6672
        max_j &= \\max_{i=0}^{K}{\\text{logit}_i}
S
sneaxiy 已提交
6673

H
haowang101779990 已提交
6674
        log\\_max\\_sum_j &= \\log\\sum_{i=0}^{K}\\exp(logit_i - max_j)
S
sneaxiy 已提交
6675

H
haowang101779990 已提交
6676
        softmax_j &= \\exp(logit_j - max_j - {log\\_max\\_sum}_j)
S
sneaxiy 已提交
6677 6678 6679

    and then cross entropy loss is calculated by softmax and label.

6680
    Args:
6681 6682 6683 6684 6685 6686
        logits (Variable): The input tensor of unscaled log probabilities.
        label (Variable): The ground truth  tensor. If :attr:`soft_label`
            is set to :attr:`True`, Label is a Tensor<float/double> in the 
            same shape with :attr:`logits`. If :attr:`soft_label` is set to 
            :attr:`True`, Label is a Tensor<int64> in the same shape with 
            :attr:`logits` expect shape in dimension :attr:`axis` as 1.
6687
        soft_label (bool): A flag to indicate whether to interpretate the given
6688
            labels as soft labels. Default False.
M
minqiyang 已提交
6689 6690
        ignore_index (int): Specifies a target value that is ignored and does
                            not contribute to the input gradient. Only valid
6691 6692
                            if :attr:`soft_label` is set to :attr:`False`. 
                            Default: kIgnoreIndex
S
sneaxiy 已提交
6693 6694
        numeric_stable_mode (bool): A flag to indicate whether to use a more
                                    numerically stable algorithm. Only valid
6695 6696 6697 6698
                                    when :attr:`soft_label` is :attr:`False` 
                                    and GPU is used. When :attr:`soft_label` 
                                    is :attr:`True` or CPU is used, the 
                                    algorithm is always numerically stable.
6699
                                    Note that the speed may be slower when use
6700
                                    stable algorithm. Default: True
6701
        return_softmax (bool): A flag indicating whether to return the softmax
6702
                               along with the cross entropy loss. Default: False
6703 6704 6705
        axis (int): The index of dimension to perform softmax calculations. It 
                    should be in range :math:`[-1, rank - 1]`, while :math:`rank`
                    is the rank of input :attr:`logits`. Default: -1.
6706

6707
    Returns:
H
haowang101779990 已提交
6708 6709
        Variable or Tuple of two Variables: Return the cross entropy loss if \
                                            `return_softmax` is False, otherwise the tuple \
6710 6711 6712 6713
                                            (loss, softmax), softmax is in the same shape \
                                            with input logits and cross entropy loss is in \
                                            the same shape with input logits except shape \
                                            in dimension :attr:`axis` as 1.
6714 6715 6716 6717

    Examples:
        .. code-block:: python

6718 6719
            import paddle.fluid as fluid

6720 6721 6722
            data = fluid.layers.data(name='data', shape=[128], dtype='float32')
            label = fluid.layers.data(name='label', shape=[1], dtype='int64')
            fc = fluid.layers.fc(input=data, size=100)
F
stash  
fengjiayi 已提交
6723 6724
            out = fluid.layers.softmax_with_cross_entropy(
                logits=fc, label=label)
6725 6726
    """
    helper = LayerHelper('softmax_with_cross_entropy', **locals())
X
Xin Pan 已提交
6727 6728
    softmax = helper.create_variable_for_type_inference(dtype=logits.dtype)
    loss = helper.create_variable_for_type_inference(dtype=logits.dtype)
6729 6730 6731 6732 6733 6734
    helper.append_op(
        type='softmax_with_cross_entropy',
        inputs={'Logits': logits,
                'Label': label},
        outputs={'Softmax': softmax,
                 'Loss': loss},
S
sneaxiy 已提交
6735 6736 6737
        attrs={
            'soft_label': soft_label,
            'ignore_index': ignore_index,
6738 6739
            'numeric_stable_mode': numeric_stable_mode,
            'axis': axis
S
sneaxiy 已提交
6740
        })
6741 6742 6743 6744

    if return_softmax:
        return loss, softmax

6745 6746 6747
    return loss


6748 6749 6750
def sampled_softmax_with_cross_entropy(logits,
                                       label,
                                       num_samples,
X
xuezhong 已提交
6751
                                       num_true=1,
6752
                                       remove_accidental_hits=True,
X
xuezhong 已提交
6753 6754 6755
                                       use_customized_samples=False,
                                       customized_samples=None,
                                       customized_probabilities=None,
6756
                                       seed=0):
X
xuezhong 已提交
6757 6758 6759 6760 6761
    """
    **Sampled Softmax With Cross Entropy Operator.**

    Cross entropy loss with sampled softmax is used as the output layer for 
    larger output classes extensively. This operator samples a number of samples
6762
    for all examples, and computes the softmax normalized values for each 
X
xuezhong 已提交
6763 6764 6765 6766 6767 6768 6769 6770
    row of the sampled tensor, after which cross-entropy loss is computed. 

    Because this operator performs a softmax on logits internally, it expects
    unscaled logits. This operator should not be used with the output of
    softmax operator since that would produce incorrect results.
    
    For examples with T true labels (T >= 1), we assume that each true label has
    a probability of 1/T. For each sample, S samples are generated using a
X
xuezhong 已提交
6771
    log uniform distribution. True labels are concatenated with these samples to
X
xuezhong 已提交
6772 6773 6774 6775 6776 6777 6778 6779
    form T + S samples for each example. So, assume the shape of logits is
    [N x K], the shape for samples is [N x (T+S)]. For each sampled label, a 
    probability is calculated, which corresponds to the Q(y|x) in 
    [Jean et al., 2014](http://arxiv.org/abs/1412.2007).
    
    Logits are sampled according to the sampled labels. Then if 
    remove_accidental_hits is True, if a sample[i, j] accidentally hits true 
    labels, then the corresponding sampled_logits[i, j] is minus by 1e20 to 
X
xuezhong 已提交
6780
    make its softmax result close to zero. Then sampled logits are subtracted by
X
xuezhong 已提交
6781 6782 6783 6784 6785 6786 6787 6788 6789 6790 6791
    logQ(y|x), these sampled logits and re-indexed labels are used to compute 
    a softmax with cross entropy.

    Args:
        logits (Variable): The unscaled log probabilities, which is a 2-D tensor
            with shape [N x K]. N is the batch_size, and K is the class number.
        label (Variable): The ground truth which is a 2-D tensor. Label is a 
            Tensor<int64> with shape [N x T], where T is the number of true 
            labels per example. 
        num_samples (int): The number for each example, num_samples should be 
            less than the number of class.
6792
        num_true(int): The number of target classes per training example.
X
xuezhong 已提交
6793 6794 6795 6796 6797
        remove_accidental_hits (bool): A flag indicating whether to remove 
            accidental hits when sampling. If True and if a sample[i, j] 
            accidentally hits true labels, then the corresponding 
            sampled_logits[i, j] is minus by 1e20 to make its softmax result 
            close to zero. Default is True.
X
xuezhong 已提交
6798
        use_customized_samples (bool): Whether to use custom samples and probabities to sample
6799
            logits.
X
xuezhong 已提交
6800 6801 6802 6803 6804
        customized_samples (Variable): User defined samples, which is a 2-D tensor
            with shape [N, T + S]. S is the num_samples, and T is the number of true 
            labels per example. 
        customized_probabilities (Variable): User defined probabilities of samples, 
            a 2-D tensor which has the same shape with customized_samples.
6805 6806 6807
        seed (int): The random seed for generating random number, which is used
            in the process of sampling. Default is 0.

X
xuezhong 已提交
6808 6809 6810 6811 6812 6813 6814
    Returns:
        Variable: Return the cross entropy loss which is a 2-D tensor with shape
                  [N x 1].

    Examples:
        .. code-block:: python

6815 6816 6817
            import paddle.fluid as fluid

            input = fluid.layers.data(name='data', shape=[256], dtype='float32')
6818
            label = fluid.layers.data(name='label', shape=[1], dtype='int64')
6819
            fc = fluid.layers.fc(input=input, size=100)
X
xuezhong 已提交
6820
            out = fluid.layers.sampled_softmax_with_cross_entropy(
6821
                      logits=fc, label=label, num_samples=25)
X
xuezhong 已提交
6822 6823 6824 6825 6826 6827 6828 6829
    """
    helper = LayerHelper('sample_logits', **locals())
    samples = helper.create_variable_for_type_inference(dtype='int64')
    probabilities = helper.create_variable_for_type_inference(
        dtype=logits.dtype)
    sampled_logits \
        = helper.create_variable_for_type_inference(dtype=logits.dtype)
    sampled_label = helper.create_variable_for_type_inference(dtype='int64')
X
xuezhong 已提交
6830 6831
    sampled_softlabel = helper.create_variable_for_type_inference(
        dtype=logits.dtype)
6832 6833
    logits_dim = helper.create_variable_for_type_inference(dtype=logits.dtype)
    labels_dim = helper.create_variable_for_type_inference(dtype=label.type)
X
xuezhong 已提交
6834 6835 6836 6837 6838

    helper.append_op(
        type='sample_logits',
        inputs={
            'Logits': logits,
X
xuezhong 已提交
6839
            'Labels': label,
X
xuezhong 已提交
6840 6841
            'CustomizedSamples': customized_samples,
            'CustomizedProbabilities': customized_probabilities
X
xuezhong 已提交
6842 6843 6844 6845
        },
        outputs={
            'Samples': samples,
            'Probabilities': probabilities,
X
xuezhong 已提交
6846
            'SampledLabels': sampled_label,
6847 6848 6849
            'SampledLogits': sampled_logits,
            'LogitsDim': logits_dim,
            'LabelsDim': labels_dim
X
xuezhong 已提交
6850 6851
        },
        attrs={
X
xuezhong 已提交
6852
            'use_customized_samples': use_customized_samples,
6853
            'uniq': True,
X
xuezhong 已提交
6854 6855 6856 6857
            'remove_accidental_hits': remove_accidental_hits,
            'num_samples': num_samples,
            'seed': seed
        })
X
xuezhong 已提交
6858 6859
    loss = helper.create_variable_for_type_inference(dtype=logits.dtype)
    softmax = helper.create_variable_for_type_inference(dtype=logits.dtype)
X
xuezhong 已提交
6860 6861 6862 6863 6864 6865
    helper.append_op(
        type='one_hot',
        inputs={'X': sampled_label},
        attrs={'depth': num_samples + 1},
        outputs={'Out': sampled_softlabel})

6866 6867
    helper.append_op(
        type='softmax_with_cross_entropy',
X
xuezhong 已提交
6868
        inputs={'Logits': sampled_logits,
X
xuezhong 已提交
6869
                'Label': sampled_softlabel},
X
xuezhong 已提交
6870 6871 6872
        outputs={'Softmax': softmax,
                 'Loss': loss},
        attrs={
X
xuezhong 已提交
6873
            'soft_label': True,
X
xuezhong 已提交
6874 6875 6876
            'ignore_index': False,
            'numeric_stable_mode': False
        })
X
xuezhong 已提交
6877
    return loss / num_true
X
xuezhong 已提交
6878 6879


6880 6881
def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None):
    """
Y
Yibing Liu 已提交
6882 6883
    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
6884
    For each instance, it computes the smooth L1 loss element by element first
6885
    and then sums all the losses. So the shape of ouput Variable is
6886
    [batch_size, 1].
6887

6888 6889
    Args:
        x (Variable): A tensor with rank at least 2. The input value of smooth
Q
qingqing01 已提交
6890
            L1 loss op with shape [batch_size, dim1, ..., dimN].
6891
        y (Variable): A tensor with rank at least 2. The target value of smooth
Y
Yibing Liu 已提交
6892
            L1 loss op with same shape as :attr:`x`.
6893
        inside_weight (Variable|None):  A tensor with rank at least 2. This
6894 6895
            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
6896
            by this tensor element by element.
6897
        outside_weight (Variable|None): A tensor with rank at least 2. This
6898 6899
            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
6900
            element by element.
6901
        sigma (float|None): Hyper parameter of smooth L1 loss layer. A float
6902 6903
           scalar with default value 1.0.

6904
    Returns:
6905
        Variable: The output smooth L1 loss with shape [batch_size, 1].
6906 6907 6908 6909

    Examples:
        .. code-block:: python

6910
            import paddle.fluid as fluid
6911
            data = fluid.layers.data(name='data', shape=[128], dtype='float32')
F
stash  
fengjiayi 已提交
6912 6913
            label = fluid.layers.data(
                name='label', shape=[100], dtype='float32')
6914
            fc = fluid.layers.fc(input=data, size=100)
F
fengjiayi 已提交
6915
            out = fluid.layers.smooth_l1(x=fc, y=label)
6916
    """
6917

6918
    helper = LayerHelper('smooth_l1_loss', **locals())
X
Xin Pan 已提交
6919 6920
    diff = helper.create_variable_for_type_inference(dtype=x.dtype)
    loss = helper.create_variable_for_type_inference(dtype=x.dtype)
6921 6922 6923 6924 6925 6926 6927 6928 6929 6930
    helper.append_op(
        type='smooth_l1_loss',
        inputs={
            'X': x,
            'Y': y,
            'InsideWeight': inside_weight,
            'OutsideWeight': outside_weight
        },
        outputs={'Diff': diff,
                 'Out': loss},
6931
        attrs={'sigma': sigma if sigma is not None else 1.0})
6932
    return loss
6933 6934


6935
def one_hot(input, depth, allow_out_of_range=False):
6936
    """
Y
Yibing Liu 已提交
6937
    This layer creates the one-hot representations for input indices.
6938 6939

    Args:
Y
Yibing Liu 已提交
6940 6941
        input(Variable): Input indices, last dimension must be 1.
        depth(scalar): An interger defining the depth of the one-hot dimension.
6942 6943 6944 6945
        allow_out_of_range(bool): A bool value indicating whether the input
            indices could be out of range [0, depth). When input indices are
            out of range, exceptions is raised if allow_out_of_range is False,
            or zero-filling representations is created if it is set True
6946 6947

    Returns:
Y
Yibing Liu 已提交
6948
        Variable: The one-hot representations of input.
6949 6950

    Examples:
C
caoying03 已提交
6951
        .. code-block:: python
6952

6953
            import paddle.fluid as fluid
Y
Yibing Liu 已提交
6954 6955
            label = fluid.layers.data(name="label", shape=[1], dtype="int64")
            one_hot_label = fluid.layers.one_hot(input=label, depth=10)
6956 6957
    """
    helper = LayerHelper("one_hot", **locals())
6958

X
Xin Pan 已提交
6959
    one_hot_out = helper.create_variable_for_type_inference(dtype='float32')
6960 6961 6962 6963 6964 6965 6966 6967 6968 6969

    if in_dygraph_mode():
        inputs = {'X': input}
        attrs = {'depth': depth}
    else:
        if not isinstance(depth, Variable):
            # user attribute 
            inputs = {'X': input}
            attrs = {'depth': depth}
        else:
H
Hongyu Liu 已提交
6970
            depth.stop_gradient = True
6971 6972
            inputs = {'X': input, 'depth_tensor': depth}
            attrs = {}
6973 6974
    helper.append_op(
        type="one_hot",
6975 6976
        inputs=inputs,
        attrs=attrs,
6977 6978
        outputs={'Out': one_hot_out},
        stop_gradient=True)
6979
    return one_hot_out
Y
Yu Yang 已提交
6980 6981


Y
Yu Yang 已提交
6982
def autoincreased_step_counter(counter_name=None, begin=1, step=1):
Y
Yu Yang 已提交
6983
    """
Y
yi.wu 已提交
6984 6985 6986
    Create an auto-increase variable
    which will be automatically increased by 1 every mini-batch
    Return the run counter of the main program, default is started from 1.
Y
Yu Yang 已提交
6987 6988 6989 6990 6991 6992

    Args:
        counter_name(str): The counter name, default is '@STEP_COUNTER@'.
        begin(int): The first value of this counter.
        step(int): The increment step between each execution.

6993 6994
    Returns:
        Variable: The global run counter.
Y
yi.wu 已提交
6995 6996 6997 6998

    Examples:
        .. code-block:: python

6999
           import paddle.fluid as fluid
Y
yi.wu 已提交
7000
           global_step = fluid.layers.autoincreased_step_counter(
Y
Yibing Liu 已提交
7001
               counter_name='@LR_DECAY_COUNTER@', begin=0, step=1)
Y
Yu Yang 已提交
7002 7003
    """
    helper = LayerHelper('global_step_counter')
Y
Yu Yang 已提交
7004 7005
    if counter_name is None:
        counter_name = '@STEP_COUNTER@'
Y
Yu Yang 已提交
7006 7007 7008 7009 7010
    counter, is_new_var = helper.create_or_get_global_variable(
        name=counter_name, dtype='int64', shape=[1], persistable=True)
    if is_new_var:
        helper.set_variable_initializer(
            counter, initializer=Constant(
Y
Yu Yang 已提交
7011
                value=begin - 1, force_cpu=True))
W
Wu Yi 已提交
7012
        helper.main_program.global_block()._prepend_op(
Y
Yu Yang 已提交
7013 7014
            type='increment',
            inputs={'X': [counter]},
Y
Yu Yang 已提交
7015
            outputs={'Out': [counter]},
M
minqiyang 已提交
7016 7017
            attrs={'step': float(step)},
            stop_gradient=True)
Y
Yu Yang 已提交
7018 7019 7020
        counter.stop_gradient = True

    return counter
Y
yangyaming 已提交
7021 7022


7023
def reshape(x, shape, actual_shape=None, act=None, inplace=False, name=None):
C
caoying03 已提交
7024
    """
C
caoying03 已提交
7025 7026
    Gives a new shape to the input Tensor without changing its data.

7027
    The target shape can be given by :attr:`shape` or :attr:`actual_shape`.
7028
    :attr:`shape` is a list of integer or tensor variable while :attr:`actual_shape` is a tensor
7029
    variable. :attr:`actual_shape` has a higher priority than :attr:`shape`
7030
    if it is provided and it only contains integer, while :attr:`shape` still should be set correctly to
7031
    gurantee shape inference in compile-time.
C
caoying03 已提交
7032

7033
    Some tricks exist when specifying the target shape.
C
caoying03 已提交
7034

7035 7036 7037 7038
    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.

7039
    2. 0 means the actual dimension value is going to be copied from the
7040 7041 7042 7043
    corresponding dimension of x. The indice of 0s in shape can not exceed
    Rank(X).

    Here are some examples to explain it.
C
caoying03 已提交
7044 7045

    1. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
W
wanghaoshuang 已提交
7046
    is [6, 8], the reshape operator will transform x into a 2-D tensor with
7047
    shape [6, 8] and leaving x's data unchanged.
C
caoying03 已提交
7048

7049
    2. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
7050 7051
    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
7052 7053
    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
7054
    dimensions.
C
caoying03 已提交
7055

7056
    3. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
7057 7058 7059 7060
    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
7061

7062 7063
    **Warning:** the parameter :attr:`actual_shape` will be deprecated in the future and only use :attr:`shape` instead.

C
caoying03 已提交
7064
    Args:
7065
        x(variable): The input tensor.
7066 7067 7068 7069
        shape(list|tuple|Variable): The new shape. At most one dimension of the new shape can
                     be -1. If :attr:`shape` is a list or tuple, it can contain Variable or not and
                     the shape of Variable must be [1].

7070 7071 7072 7073
        actual_shape(variable): An optional input. If provided, reshape
                                according to this given shape rather than
                                :attr:`shape` specifying shape. That is to
                                say :attr:`actual_shape` has a higher priority
7074 7075 7076 7077
                                than :attr:`shape(list|tuple)` but not :attr:`shape(Variable)`. \
                                This argument :attr:`actual_shape` will be removed in a future version. \
                                Instructions for updating: :attr:`actual_shape` is deprecated,
                                only use :attr:`shape` instead.
7078 7079
        act (str): The non-linear activation to be applied to the reshaped tensor
                   variable.
C
chengduozh 已提交
7080 7081 7082
        inplace(bool): 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 variables. Note that if :attr:`x`
C
chengduozh 已提交
7083
                       is more than one layer's input, ``inplace`` must be :attr:`False`.
7084
        name (str): The name of this layer. It is optional.
C
caoying03 已提交
7085

7086
    Returns:
G
guosheng 已提交
7087 7088 7089 7090
        Variable: The reshaped tensor variable if :attr:`act` is None. It is a \
                  new tensor variable if :attr:`inplace` is :attr:`False`, \
                  otherwise it is :attr:`x`. If :attr:`act` is not None, return \
                  the activated tensor variable.
C
caoying03 已提交
7091

X
Xin Pan 已提交
7092 7093 7094
    Raises:
        TypeError: if actual_shape is neither Variable nor None.

C
caoying03 已提交
7095 7096
    Examples:
        .. code-block:: python
G
guosheng 已提交
7097

7098
            import paddle.fluid as fluid
7099 7100 7101 7102 7103 7104 7105 7106 7107 7108 7109 7110 7111

            # example 1:
            # attr shape is a list which doesn't contain tensor Variable.
            data_1 = fluid.layers.data(
                name='data_1', shape=[2, 4, 6], dtype='float32')
            reshaped_1 = fluid.layers.reshape(
                x=data_1, shape=[-1, 0, 3, 2], inplace=True)

            # 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])
C
caoying03 已提交
7112 7113
    """

7114 7115 7116
    if not isinstance(shape, (list, tuple, Variable)):
        raise TypeError(
            "Input shape must be an Variable or python list or tuple.")
7117

7118 7119
    if not isinstance(actual_shape, Variable) and (actual_shape is not None):
        raise TypeError("actual_shape should either be Variable or None.")
7120

7121
    helper = LayerHelper("reshape2", **locals())
7122 7123 7124 7125 7126 7127 7128 7129 7130 7131 7132 7133 7134 7135 7136 7137 7138 7139 7140 7141 7142 7143 7144 7145 7146 7147 7148 7149 7150 7151 7152 7153 7154 7155 7156 7157 7158 7159 7160 7161 7162 7163 7164
    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, (
                        "Only one dimension in shape can be unknown.")
                    unk_dim_idx = dim_idx
                elif dim_size == 0:
                    assert dim_idx < len(x.shape), (
                        "The indice of 0s in shape can not exceed Rank(X).")
                else:
                    assert dim_size > 0, (
                        "Each dimension size given in shape must not be negtive "
                        "except one unknown dimension.")
        return attrs_shape

7165 7166 7167 7168
    if in_dygraph_mode():
        inputs = {'X': x}
        attrs = {'shape': shape}
    else:
7169 7170 7171 7172 7173 7174 7175 7176 7177 7178 7179 7180
        if isinstance(shape, Variable):
            shape.stop_gradient = True
            inputs["Shape"] = 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['ShapeTensor'] = get_new_shape_tensor(shape)
            elif isinstance(actual_shape, Variable):
                actual_shape.stop_gradient = True
                inputs["Shape"] = actual_shape
7181

7182 7183
    out = x if inplace else helper.create_variable_for_type_inference(
        dtype=x.dtype)
X
Xin Pan 已提交
7184
    x_shape = helper.create_variable_for_type_inference(dtype=x.dtype)
C
caoying03 已提交
7185
    helper.append_op(
7186
        type="reshape2",
X
Xin Pan 已提交
7187
        inputs=inputs,
7188
        attrs=attrs,
7189 7190
        outputs={"Out": out,
                 "XShape": x_shape})
C
caoying03 已提交
7191

D
dzhwinter 已提交
7192
    return helper.append_activation(out)
7193

7194

7195
def squeeze(input, axes, name=None):
Y
Yibing Liu 已提交
7196
    """
M
minqiyang 已提交
7197 7198 7199
    Remove single-dimensional entries from the shape of a tensor. Takes a
    parameter axes with a list of axes to squeeze. If axes is not provided, all
    the single dimensions will be removed from the shape. If an axis is
Y
Yibing Liu 已提交
7200
    selected with shape entry not equal to one, an error is raised.
M
minqiyang 已提交
7201

H
haowang101779990 已提交
7202 7203 7204 7205 7206 7207 7208 7209 7210 7211 7212 7213 7214 7215 7216 7217 7218 7219 7220 7221 7222
    For example:

    .. code-block:: text

        Case 1:

          Given
            X.shape = (1, 3, 1, 5)
          and
            axes = [0]
          we get:
            Out.shape = (3, 1, 5)

        Case 2:

          Given
            X.shape = (1, 3, 1, 5)
          and
            axes = []
          we get:
            Out.shape = (3, 5)
M
minqiyang 已提交
7223

Y
Yibing Liu 已提交
7224
    Args:
7225
        input (Variable): The input variable to be squeezed.
Y
Yibing Liu 已提交
7226
        axes (list): List of integers, indicating the dimensions to be squeezed.
7227
        name (str|None): Name for this layer.
Y
Yibing Liu 已提交
7228 7229 7230 7231 7232 7233 7234

    Returns:
        Variable: Output squeezed variable.

    Examples:
        .. code-block:: python

7235
            import paddle.fluid as fluid
7236
            import paddle.fluid.layers as layers
Y
Yibing Liu 已提交
7237
            x = layers.data(name='x', shape=[5, 1, 10])
7238
            y = layers.squeeze(input=x, axes=[1])
Y
Yibing Liu 已提交
7239
    """
L
lujun 已提交
7240
    assert not in_dygraph_mode(), (
L
lujun 已提交
7241
        "squeeze layer is not supported in dygraph mode yet.")
Y
Yibing Liu 已提交
7242
    helper = LayerHelper("squeeze", **locals())
X
Xin Pan 已提交
7243 7244
    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
7245
    helper.append_op(
7246
        type="squeeze2",
7247
        inputs={"X": input},
Y
Yibing Liu 已提交
7248
        attrs={"axes": axes},
7249 7250
        outputs={"Out": out,
                 "XShape": x_shape})
Y
Yibing Liu 已提交
7251

7252 7253 7254
    return out


7255
def unsqueeze(input, axes, name=None):
Y
Yibing Liu 已提交
7256
    """
M
minqiyang 已提交
7257 7258 7259
    Insert single-dimensional entries to the shape of a tensor. Takes one
    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
7260

M
minqiyang 已提交
7261
    For example:
H
haowang101779990 已提交
7262 7263 7264

    .. code-block:: text

M
minqiyang 已提交
7265
      Given a tensor such that tensor with shape [3, 4, 5],
Y
Yibing Liu 已提交
7266
      then Unsqueezed tensor with axes=[0, 4] has shape [1, 3, 4, 5, 1].
M
minqiyang 已提交
7267

Y
Yibing Liu 已提交
7268
    Args:
7269
        input (Variable): The input variable to be unsqueezed.
Y
Yibing Liu 已提交
7270
        axes (list): List of integers, indicating the dimensions to be inserted.
7271
        name (str|None): Name for this layer.
Y
Yibing Liu 已提交
7272 7273 7274 7275 7276 7277 7278

    Returns:
        Variable: Output unsqueezed variable.

    Examples:
        .. code-block:: python

7279 7280 7281
            import paddle.fluid as fluid
            x = fluid.layers.data(name='x', shape=[5, 10])
            y = fluid.layers.unsqueeze(input=x, axes=[1])
Y
Yibing Liu 已提交
7282 7283
    """
    helper = LayerHelper("unsqueeze", **locals())
X
Xin Pan 已提交
7284 7285
    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
7286
    helper.append_op(
7287
        type="unsqueeze2",
7288
        inputs={"X": input},
Y
Yibing Liu 已提交
7289
        attrs={"axes": axes},
7290 7291
        outputs={"Out": out,
                 "XShape": x_shape})
Y
Yibing Liu 已提交
7292

7293 7294
    return out

7295

Y
yangyaming 已提交
7296
def lod_reset(x, y=None, target_lod=None):
Y
yangyaming 已提交
7297
    """
Y
Yibing Liu 已提交
7298
    Set LoD of :attr:`x` to a new one specified by :attr:`y` or
7299 7300 7301 7302
    :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
7303
    :attr:`y.data` or :attr:`target_lod`, only one level LoD is supported.
Y
yangyaming 已提交
7304 7305 7306 7307 7308 7309

    .. code-block:: text

        * Example 1:

            Given a 1-level LoDTensor x:
7310
                x.lod =  [[ 2,           3,                   1 ]]
Y
yangyaming 已提交
7311 7312 7313
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

7314
            target_lod: [4, 2]
Y
yangyaming 已提交
7315 7316

            then we get a 1-level LoDTensor:
7317
                out.lod =  [[4,                          2]]
Y
yangyaming 已提交
7318 7319 7320 7321 7322 7323
                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:
7324
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
7325 7326 7327 7328
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a Tensor:
7329
                y.data = [[2, 4]]
Y
yangyaming 已提交
7330 7331 7332
                y.dims = [1, 3]

            then we get a 1-level LoDTensor:
7333
                out.lod =  [[2,            4]]
Y
yangyaming 已提交
7334 7335 7336 7337 7338 7339
                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:
7340
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
7341 7342 7343 7344
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a 2-level LoDTensor:
7345
                y.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
7346 7347 7348 7349
                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:
7350
                out.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
7351 7352 7353 7354
                out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                out.dims = [6, 1]

    Args:
7355
        x (Variable): Input variable which could be a Tensor or LoDTensor.
7356
        y (Variable|None): If provided, output's LoD would be derived
Y
Yibing Liu 已提交
7357
                           from :attr:`y`.
Y
yangyaming 已提交
7358
        target_lod (list|tuple|None): One level LoD which should be considered
Y
Yibing Liu 已提交
7359
                                      as target LoD when :attr:`y` not provided.
Y
yangyaming 已提交
7360 7361

    Returns:
Y
Yibing Liu 已提交
7362
        Variable: Output variable with LoD specified by this layer.
Y
yangyaming 已提交
7363 7364

    Raises:
Y
Yibing Liu 已提交
7365
        ValueError: If :attr:`y` and :attr:`target_lod` are both None.
Y
yangyaming 已提交
7366 7367 7368 7369

    Examples:
        .. code-block:: python

7370
            import paddle.fluid as fluid
7371 7372 7373
            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
7374 7375
    """
    helper = LayerHelper("lod_reset", **locals())
X
Xin Pan 已提交
7376
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
Y
yangyaming 已提交
7377 7378 7379 7380 7381 7382 7383 7384 7385 7386 7387
    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:
7388 7389 7390 7391 7392 7393 7394 7395 7396 7397 7398 7399 7400 7401 7402 7403 7404 7405 7406 7407 7408 7409 7410 7411 7412 7413
        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.
7414
        level (list|tuple|Variable): The LoD level to be appended into LoD of x.
7415 7416 7417 7418 7419 7420

    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
7421

7422 7423 7424 7425 7426 7427 7428 7429 7430 7431
    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.")
7432 7433 7434
    if (not isinstance(level, Iterable)) and (not isinstance(level, Variable)):
        raise ValueError("Input(level) must be list, tuple or Variable.")

7435 7436
    helper = LayerHelper("lod_append", **locals())
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
7437 7438 7439 7440 7441 7442 7443 7444

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

    if isinstance(level, Variable):
        inputs['Y'] = level
    else:
        attrs['target_lod'] = level
7445
    helper.append_op(
7446
        type="lod_reset", inputs=inputs, attrs=attrs, outputs={'Out': out})
Y
yangyaming 已提交
7447
    return out
D
dragonwarrior 已提交
7448 7449 7450 7451 7452 7453 7454 7455 7456 7457 7458


def lrn(input, n=5, k=1.0, alpha=1e-4, beta=0.75, name=None):
    """
    Local Response Normalization Layer. This layer performs a type of
    "lateral inhibition" by normalizing over local input regions.

    The formula is as follows:

    .. math::

X
xiaoting 已提交
7459
      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
7460 7461 7462 7463 7464 7465 7466 7467 7468 7469 7470 7471 7472 7473 7474 7475 7476 7477 7478 7479 7480 7481 7482 7483 7484 7485 7486 7487

    In the above equation:

    * :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.

    Refer to `ImageNet Classification with Deep Convolutional Neural Networks
    <https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf>`_

    Args:
        input (Variable): The input tensor of this layer, and the dimension of input tensor must be 4.
        n (int, default 5): The number of channels to sum over.
        k (float, default 1.0): An offset (usually positive to avoid dividing by 0).
        alpha (float, default 1e-4): The scaling parameter.
        beta (float, default 0.75): The exponent.
        name (str, default None): A name for this operation.

    Raises:
        ValueError: If rank of the input tensor is not 4.

    Returns:
        A tensor variable storing the transformation result.

    Examples:
        .. code-block:: python

7488
          import paddle.fluid as fluid
F
stash  
fengjiayi 已提交
7489 7490
          data = fluid.layers.data(
              name="data", shape=[3, 112, 112], dtype="float32")
D
dragonwarrior 已提交
7491 7492 7493 7494 7495 7496 7497 7498 7499 7500 7501 7502
          lrn = fluid.layers.lrn(input=data)
    """
    helper = LayerHelper('lrn', **locals())
    dtype = helper.input_dtype()
    input_shape = input.shape
    dims = len(input_shape)

    if dims != 4:
        raise ValueError(
            "dims of input must be 4(not %d), and it's order must be NCHW" %
            (dims))

X
Xin Pan 已提交
7503 7504 7505
    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
7506 7507 7508 7509 7510 7511 7512 7513 7514 7515 7516 7517 7518
    helper.append_op(
        type="lrn",
        inputs={"X": input},
        outputs={
            "Out": lrn_out,
            "MidOut": mid_out,
        },
        attrs={"n": n,
               "k": k,
               "alpha": alpha,
               "beta": beta})

    return lrn_out
G
guosheng 已提交
7519 7520 7521 7522


def pad(x, paddings, pad_value=0., name=None):
    """
G
guosheng 已提交
7523
    Pads a tensor with a constant value given by :attr:`pad_value`, and the
W
wanghaoshuang 已提交
7524
    padded width is specified by :attr:`paddings`.
G
guosheng 已提交
7525

G
guosheng 已提交
7526
    Specifically, the number of values padded before the contents of :attr:`x`
7527
    in dimension :attr:`i` is indicated by :attr:`paddings[2i]`, and the number
G
guosheng 已提交
7528
    of values padded after the contents of :attr:`x` in dimension :attr:`i` is
7529
    indicated by :attr:`paddings[2i+1]`.
G
guosheng 已提交
7530 7531 7532 7533 7534 7535 7536 7537 7538 7539 7540 7541 7542 7543 7544 7545 7546 7547 7548 7549 7550 7551

    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:
        x (Variable): The input tensor variable.
        paddings (list): A list of integers. Its elements specify the padded
                         width before and after for each dimension in turn.
W
wanghaoshuang 已提交
7552
                         The length of :attr:paddings must be
G
guosheng 已提交
7553 7554 7555 7556 7557 7558 7559 7560 7561 7562
                         :math:`rank(x) \\times 2`.
        pad_value (float): The constant value used to pad.
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        Variable: The padded tensor variable.

    Examples:
        .. code-block:: python
G
guosheng 已提交
7563

G
guosheng 已提交
7564
            # x is a rank 2 tensor variable.
S
SunGaofeng 已提交
7565 7566
            import paddle.fluid as fluid
            x = fluid.layers.data(name='data', shape=[224], dtype='float32')
G
guosheng 已提交
7567 7568 7569 7570 7571
            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
7572
    out = helper.create_variable_for_type_inference(dtype)
G
guosheng 已提交
7573 7574 7575 7576 7577 7578 7579
    helper.append_op(
        type='pad',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'paddings': paddings,
               'pad_value': float(pad_value)})
    return out
7580 7581


C
chengduo 已提交
7582 7583 7584 7585 7586 7587 7588 7589 7590 7591 7592 7593 7594 7595 7596 7597 7598 7599 7600 7601 7602 7603 7604 7605 7606 7607 7608 7609 7610 7611 7612
def pad_constant_like(x, y, pad_value=0., name=None):
    """
    Pad input(Y) with :attr:`pad_value`, the number of values padded to
    the edges of each axis is specified by the difference of the shape
    of X and Y. ((0, shape_x_0 - shape_y_0), ... (0, shape_x_n - shape_y_n))
    unique pad widths for each axis. The input should be a k-D
    tensor(k > 0 and k < 7).

    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
7613 7614
		And
            pad_value = -1,
C
chengduo 已提交
7615

T
Tink_Y 已提交
7616 7617 7618 7619 7620 7621 7622 7623 7624 7625 7626 7627 7628 7629
        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
7630 7631 7632 7633 7634 7635 7636 7637 7638 7639 7640 7641 7642 7643 7644 7645

    Args:
        x (Variable): The input tensor variable.
        y (Variable): The input tensor variable.
        pad_value (float): The constant value used to pad.
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        Variable: The padded tensor variable.

    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
7646 7647 7648
            import paddle.fluid as fluid
            x = fluid.layers.data(name='x', shape=[2,3,2,3], dtype='float32')
            y = fluid.layers.data(name='y', shape=[1,3,1,3], dtype='float32')
C
chengduo 已提交
7649 7650 7651 7652 7653
            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
7654
    out = helper.create_variable_for_type_inference(dtype)
C
chengduo 已提交
7655 7656 7657 7658 7659 7660 7661 7662 7663
    helper.append_op(
        type='pad_constant_like',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
        attrs={'pad_value': float(pad_value)})
    return out


7664 7665 7666 7667 7668 7669 7670
def label_smooth(label,
                 prior_dist=None,
                 epsilon=0.1,
                 dtype="float32",
                 name=None):
    """
    Label smoothing is a mechanism to regularize the classifier layer and is
7671 7672
    called label-smoothing regularization (LSR).

7673 7674 7675 7676 7677 7678 7679 7680 7681 7682 7683 7684 7685 7686 7687 7688 7689 7690 7691 7692 7693 7694 7695
    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.

    Args:
        label(Variable): The input variable containing the label data. The
                          label data should use one-hot representation.
        prior_dist(Variable): The prior distribution to be used to smooth
                              labels. If not provided, an uniform distribution
                              is used. The shape of :attr:`prior_dist` should
7696
                              be :math:`(1, class\_num)`.
7697 7698
        epsilon(float): The weight used to mix up the original ground-truth
                        distribution and the fixed distribution.
7699
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32,
7700 7701 7702 7703 7704 7705 7706 7707 7708
                                                  float_64, int etc.
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

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

    Examples:
        .. code-block:: python
7709
            
7710
            import paddle.fluid as fluid
7711
            import paddle.fluid.layers as layers
7712 7713 7714 7715 7716 7717 7718 7719 7720 7721

            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
7722
    smooth_label = helper.create_variable_for_type_inference(dtype)
7723 7724 7725 7726 7727 7728 7729
    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
7730 7731


W
wopeizl 已提交
7732 7733 7734 7735 7736 7737 7738
@templatedoc()
def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0):
    """
    ${comment}

    Args:
        input (Variable): ${x_comment}
7739 7740 7741 7742 7743
        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.
W
wopeizl 已提交
7744 7745 7746 7747 7748 7749 7750 7751 7752 7753
        pooled_height (integer): ${pooled_height_comment} Default: 1
        pooled_width (integer): ${pooled_width_comment} Default: 1
        spatial_scale (float): ${spatial_scale_comment} Default: 1.0

    Returns:
        Variable: ${out_comment}.

    Examples:
        .. code-block:: python

7754 7755 7756 7757 7758 7759 7760 7761 7762 7763 7764 7765 7766
            import paddle.fluid as fluid

            x = fluid.layers.data(
                name='x', shape=[8, 112, 112], dtype='float32')
            rois = fluid.layers.data(
                name='roi', shape=[4], lod_level=1, dtype='float32')
            pool_out = fluid.layers.roi_pool(
                input=x,
                rois=rois,
                pooled_height=7,
                pooled_width=7,
                spatial_scale=1.0)

W
wopeizl 已提交
7767 7768 7769 7770 7771 7772 7773 7774 7775 7776 7777 7778 7779 7780 7781 7782 7783
    """
    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
7784 7785


J
jerrywgz 已提交
7786 7787 7788 7789 7790 7791
@templatedoc()
def roi_align(input,
              rois,
              pooled_height=1,
              pooled_width=1,
              spatial_scale=1.0,
J
jerrywgz 已提交
7792 7793
              sampling_ratio=-1,
              name=None):
J
jerrywgz 已提交
7794 7795 7796 7797 7798
    """
    ${comment}

    Args:
        input (Variable): ${x_comment}
7799 7800 7801 7802 7803
        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. 
J
jerrywgz 已提交
7804 7805 7806 7807 7808 7809 7810 7811 7812 7813
        pooled_height (integer): ${pooled_height_comment} Default: 1
        pooled_width (integer): ${pooled_width_comment} Default: 1
        spatial_scale (float): ${spatial_scale_comment} Default: 1.0
        sampling_ratio(intger): ${sampling_ratio_comment} Default: -1

    Returns:
        Variable: ${out_comment}.
    Examples:
        .. code-block:: python

7814
            import paddle.fluid as fluid
J
jerrywgz 已提交
7815 7816 7817 7818
            x = fluid.layers.data(
                name='data', shape=[256, 32, 32], dtype='float32')
            rois = fluid.layers.data(
                name='rois', shape=[4], dtype='float32')
7819 7820 7821
            align_out = fluid.layers.roi_align(input=x,
                                               rois=rois,
                                               pooled_height=7,
J
jerrywgz 已提交
7822 7823 7824 7825 7826 7827
                                               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
7828
    align_out = helper.create_variable_for_type_inference(dtype)
J
jerrywgz 已提交
7829 7830 7831 7832 7833 7834 7835 7836 7837 7838 7839 7840 7841 7842
    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


W
whs 已提交
7843 7844 7845 7846 7847 7848 7849 7850 7851 7852 7853 7854 7855 7856 7857 7858 7859 7860 7861 7862 7863 7864 7865 7866 7867 7868
def dice_loss(input, label, epsilon=0.00001):
    """
    Dice loss for comparing the similarity of two batch of data,
    usually is used for binary image segmentation i.e. labels are binary.
    The dice loss can be defined as below equation:

    .. 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}


    Args:
        input (Variable): The predictions with rank>=2. The first dimension is batch size,
                          and the last dimension is class number.
        label (Variable): The groud truth with the same rank with input. The first dimension
                          is batch size, and the last dimension is 1.
        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

    Returns:
        dice_loss (Variable): The dice loss with shape [1].

    Examples:
7869 7870
        .. code-block:: python

S
SunGaofeng 已提交
7871 7872 7873
            import paddle.fluid as fluid
            x = fluid.layers.data(name='data', shape = [3, 224, 224, 2], dtype='float32')
            label = fluid.layers.data(name='label', shape=[3, 224, 224, 1], dtype='float32')
W
whs 已提交
7874
            predictions = fluid.layers.softmax(x)
S
SunGaofeng 已提交
7875
            loss = fluid.layers.dice_loss(input=predictions, label=label)
W
whs 已提交
7876 7877
    """
    label = one_hot(label, depth=input.shape[-1])
7878
    reduce_dim = list(range(1, len(input.shape)))
W
whs 已提交
7879 7880 7881 7882 7883 7884
    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)
7885 7886


7887 7888 7889 7890
def image_resize(input,
                 out_shape=None,
                 scale=None,
                 name=None,
7891
                 resample='BILINEAR',
7892 7893
                 actual_shape=None,
                 align_corners=True,
T
tink2123 已提交
7894
                 align_mode=1):
7895
    """
Q
qiaolongfei 已提交
7896
    **Resize a Batch of Images**
F
stash  
fengjiayi 已提交
7897

K
Kaipeng Deng 已提交
7898 7899 7900
    The input must be a tensor of the shape (num_batches, channels, in_h, in_w)
    or (num_batches, channels, in_d, in_h, in_w), and the resizing only applies 
    on the last two/three dimensions(depth, hight and width).
7901 7902

    Supporting resample methods:
Q
update  
qiaolongfei 已提交
7903

7904
        'BILINEAR' : Bilinear interpolation
T
Tink_Y 已提交
7905

K
Kaipeng Deng 已提交
7906 7907
        'TRILINEAR' : Trilinear interpolation

7908
        'NEAREST' : Nearest neighbor interpolation
F
stash  
fengjiayi 已提交
7909

7910 7911 7912 7913 7914 7915 7916 7917 7918 7919
    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
7920 7921 7922 7923 7924
    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
7925
    Align_corners and align_mode are optinal parameters,the calculation method 
7926 7927 7928 7929
    of interpolation can be selected by them.

    Example:

T
Tink_Y 已提交
7930
    .. code-block:: text
7931

T
Tink_Y 已提交
7932
        For scale:
7933
          
T
Tink_Y 已提交
7934
            if align_corners = True && out_size > 1 :
7935

T
Tink_Y 已提交
7936 7937 7938 7939 7940 7941 7942 7943 7944 7945 7946
              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
7947

T
Tink_Y 已提交
7948 7949
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
7950

T
Tink_Y 已提交
7951 7952
              H_out = floor (H_{in} * scale_{factor})
              W_out = floor (W_{in} * scale_{factor})
7953

T
Tink_Y 已提交
7954 7955
          else:
              align_corners = True
7956

T
Tink_Y 已提交
7957 7958
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
7959

T
Tink_Y 已提交
7960 7961
              H_out = round(H_{in} * scale_{factor})
              W_out = round(W_{in} * scale_{factor})
7962

T
Tink_Y 已提交
7963 7964 7965 7966 7967 7968 7969 7970 7971 7972
        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
7973

T
Tink_Y 已提交
7974 7975 7976 7977
          else:
           
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
7978

T
Tink_Y 已提交
7979 7980
              H_out = H_{in} * scale_{factor}
              W_out = W_{in} * scale_{factor}
7981

K
Kaipeng Deng 已提交
7982 7983 7984 7985 7986 7987 7988 7989 7990 7991 7992 7993 7994 7995 7996 7997 7998 7999 8000 8001 8002 8003
        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}
          
8004 8005 8006 8007 8008 8009
    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
8010 8011 8012
    For details of trilinear interpolation, please refer to Wikipedia: 
    https://en.wikipedia.org/wiki/Trilinear_interpolation.

8013 8014


8015
    Args:
8016
        input (Variable): The input tensor of image resize layer,
8017
                          This is a 4-D tensor of the shape
K
Kaipeng Deng 已提交
8018 8019 8020
                          (num_batches, channels, in_h, in_w) or a
                          5-D tensor of the shape
                          (num_batches, channls, in_d, in_h, in_w).
8021
        out_shape(list|tuple|Variable|None): Output shape of image resize
K
Kaipeng Deng 已提交
8022 8023 8024 8025
                                    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
D
dengkaipeng 已提交
8026
        scale(float|None): The multiplier for the input height or width. At
D
dengkaipeng 已提交
8027
             least one of :attr:`out_shape` or :attr:`scale` must be set. 
D
dengkaipeng 已提交
8028
             And :attr:`out_shape` has a higher priority than :attr:`scale`. 
D
dengkaipeng 已提交
8029
             Default: None.
8030 8031
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
K
Kaipeng Deng 已提交
8032 8033
        resample(str): The resample method. It supports 'BILINEAR', 'TRILINEAR'
                       and 'NEAREST' currently. Default: 'BILINEAR'
8034 8035 8036
        actual_shape(Variable): An optional input to specify output shape
                                dynamically. If provided, image resize
                                according to this given shape rather than
8037
                                :attr:`out_shape` and :attr:`scale` specifying
8038 8039 8040 8041 8042 8043 8044
                                shape. That is to say actual_shape has the
                                highest priority. It is recommended to use
                                actual_shape instead of :attr:`out_shape` if you
                                want to specify output shape dynamically. 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
8045 8046
                                constructing stage.
                                Default: None
8047 8048 8049 8050
        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
8051
        align_mode(int)  :  An optional for bilinear interpolation. can be \'0\' 
T
tink2123 已提交
8052 8053
                            for src_idx = scale*(dst_indx+0.5)-0.5 , can be \'1\' for 
                            src_idx = scale*dst_index .
8054 8055

    Returns:
Q
update  
qiaolongfei 已提交
8056
        Variable: The output is a 4-D tensor of the shape
K
Kaipeng Deng 已提交
8057 8058
        (num_batches, channls, out_h, out_w) or a 5-D tensor of the shape
        (num_batches, channels, out_d, out_h, out_w).
F
stash  
fengjiayi 已提交
8059

8060 8061 8062
    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
8063 8064 8065 8066
        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.
8067
        ValueError: One of out_shape and scale must not be None.
K
Kaipeng Deng 已提交
8068 8069
        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
8070
        ValueError: scale should be greater than zero.
8071 8072
        TypeError: align_corners shoule be a bool value
        ValueError: align_mode can only be '0' or '1'
8073

8074 8075 8076
    Examples:
        .. code-block:: python

8077
            import paddle.fluid as fluid
R
ruri 已提交
8078
            input = fluid.layers.data(name="input", shape=[3,6,9], dtype="float32")
8079
            out = fluid.layers.image_resize(input, out_shape=[12, 12], resample="NEAREST")
8080
    """
8081 8082
    resample_methods = {
        'BILINEAR': 'bilinear',
K
Kaipeng Deng 已提交
8083
        'TRILINEAR': 'trilinear',
8084 8085
        'NEAREST': 'nearest',
    }
8086 8087
    if resample not in resample_methods:
        raise ValueError(
K
Kaipeng Deng 已提交
8088 8089
            "The 'resample' of image_resize can only be 'BILINEAR', 'TRILINEAR' "
            "or 'NEAREST' currently.")
8090
    resample_type = resample_methods[resample]
8091

K
Kaipeng Deng 已提交
8092 8093 8094 8095 8096
    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.")

8097 8098 8099 8100 8101
    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")

8102
    if out_shape is None and scale is None:
8103
        raise ValueError("One of out_shape and scale must not be None.")
8104
    helper = LayerHelper('{}_interp'.format(resample_type), **locals())
8105
    dtype = helper.input_dtype()
8106 8107 8108 8109

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

8110
    inputs = {"X": input}
D
dengkaipeng 已提交
8111
    attrs = {
K
Kaipeng Deng 已提交
8112
        "out_d": 0,
D
dengkaipeng 已提交
8113 8114
        "out_h": 0,
        "out_w": 0,
D
dengkaipeng 已提交
8115 8116 8117 8118 8119
        "interp_method": resample_type,
        "align_corners": align_corners,
        "align_mode": align_mode
    }

8120
    if out_shape is not None:
8121 8122 8123 8124
        if isinstance(out_shape, Variable):
            warnings.warn("out_shape as Variable type is deprecated, \
                    it is recommended to use actual_shape instead of \
                    out_shape to specify output shape dynamically.")
8125
            inputs['OutSize'] = out_shape
8126 8127
        else:
            if not (_is_list_or_turple_(out_shape)):
D
dengkaipeng 已提交
8128 8129
                raise TypeError(
                    "out_shape should be a list or tuple or Variable.")
K
Kaipeng Deng 已提交
8130 8131 8132 8133 8134 8135 8136 8137 8138 8139 8140 8141 8142 8143 8144
            if len(input.shape) == 4:
                if len(out_shape) != 2:
                    raise ValueError("out_shape length should be 2 for "
                                     "input 4-D tensor.")
                out_shape = list(map(int, out_shape))
                attrs['out_h'] = out_shape[0]
                attrs['out_w'] = out_shape[1]
            if len(input.shape) == 5:
                if len(out_shape) != 3:
                    raise ValueError("out_shape length should be 3 for "
                                     "input 5-D tensor.")
                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]
8145

8146
    else:
D
dengkaipeng 已提交
8147 8148
        if scale <= 0:
            raise ValueError("scale should be greater than zero.")
D
dengkaipeng 已提交
8149
        attrs['scale'] = float(scale)
8150

8151 8152 8153 8154 8155
    if isinstance(actual_shape, Variable):
        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
8156
    out = helper.create_variable_for_type_inference(dtype)
8157
    helper.append_op(
8158
        type='{}_interp'.format(resample_type),
8159
        inputs=inputs,
8160
        outputs={"Out": out},
D
dengkaipeng 已提交
8161
        attrs=attrs)
8162
    return out
F
stash  
fengjiayi 已提交
8163 8164


8165
@templatedoc(op_type="bilinear_interp")
8166 8167 8168 8169
def resize_bilinear(input,
                    out_shape=None,
                    scale=None,
                    name=None,
8170 8171
                    actual_shape=None,
                    align_corners=True,
T
tink2123 已提交
8172
                    align_mode=1):
8173
    """
8174 8175
    Resize input by performing bilinear interpolation based on given
    output shape which specified by actual_shape, out_shape and scale
8176 8177
    in priority order.

8178 8179 8180 8181
    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
8182 8183
    again in the other direction.

8184
    For details of bilinear interpolation, please refer to Wikipedia:
8185
    https://en.wikipedia.org/wiki/Bilinear_interpolation
Y
yuyang18 已提交
8186

T
tink2123 已提交
8187
    Align_corners and align_mode are optinal parameters,the calculation 
8188 8189 8190 8191
    method of interpolation can be selected by them.

    Example:

T
Tink_Y 已提交
8192
    .. code-block:: text
8193

T
Tink_Y 已提交
8194
        For scale:
8195
          
T
Tink_Y 已提交
8196
            if align_corners = True && out_size > 1 :
8197

T
Tink_Y 已提交
8198 8199 8200 8201 8202
              scale_factor = (in_size-1.0)/(out_size-1.0)
            
            else:
              
              scale_factor = float(in_size/out_size)     
8203

T
Tink_Y 已提交
8204 8205 8206 8207 8208 8209 8210 8211 8212 8213
        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
8214 8215


T
Tink_Y 已提交
8216
          else:
T
tink2123 已提交
8217

T
Tink_Y 已提交
8218 8219
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
8220

T
Tink_Y 已提交
8221 8222
              H_out = H_{in} * scale_{factor}
              W_out = W_{in} * scale_{factor}
8223 8224 8225



Y
yuyang18 已提交
8226
    Args:
K
Kaipeng Deng 已提交
8227
        input(${x_type}): input should be a 4-D tensor.
Y
yuyang18 已提交
8228

D
dengkaipeng 已提交
8229 8230 8231
        out_shape(list|tuple|Variable|None): Output shape of resize bilinear
                                    layer, the shape is (out_h, out_w).
                                    Default: None
8232

Y
yuyang18 已提交
8233
        scale(float|None): The multiplier for the input height or width. At
D
dengkaipeng 已提交
8234
             least one of :attr:`out_shape` or :attr:`scale` must be set. 
D
dengkaipeng 已提交
8235
             And :attr:`out_shape` has a higher priority than :attr:`scale`. 
D
dengkaipeng 已提交
8236
             Default: None.
Y
yuyang18 已提交
8237 8238

        name(str|None): The output variable name.
8239 8240 8241
        actual_shape(Variable): An optional input to specify output shape
                                dynamically. If provided, image resize
                                according to this given shape rather than
8242
                                :attr:`out_shape` and :attr:`scale` specifying
8243 8244 8245 8246 8247 8248 8249
                                shape. That is to say actual_shape has the
                                highest priority. It is recommended to use
                                actual_shape instead of :attr:`out_shape` if you
                                want to specify output shape dynamically. 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
8250 8251
                                constructing stage.
                                Default: None
8252 8253
        align_corners(bool): ${align_corners_comment}
        align_mode(bool): ${align_mode_comment}
Y
yuyang18 已提交
8254 8255

    Returns:
K
Kaipeng Deng 已提交
8256
        A 4-D tensor in shape of (num_batches, channels, out_h, out_w)
8257 8258 8259 8260

    Examples:
        .. code-block:: python

8261
            import paddle.fluid as fluid
R
ruri 已提交
8262
            input = fluid.layers.data(name="input", shape=[3,6,9], dtype="float32")
8263
            out = fluid.layers.resize_bilinear(input, out_shape=[12, 12])
8264 8265
    """

8266 8267
    return image_resize(input, out_shape, scale, name, 'BILINEAR', actual_shape,
                        align_corners, align_mode)
8268 8269


K
Kaipeng Deng 已提交
8270 8271 8272 8273 8274 8275 8276 8277 8278 8279 8280 8281 8282 8283 8284 8285 8286 8287 8288 8289 8290 8291 8292 8293 8294 8295 8296 8297 8298 8299 8300 8301 8302 8303 8304 8305 8306 8307 8308 8309 8310 8311 8312 8313 8314 8315 8316 8317 8318 8319 8320 8321 8322 8323 8324 8325 8326 8327 8328 8329 8330 8331 8332 8333 8334 8335 8336 8337 8338 8339 8340 8341 8342 8343 8344 8345 8346 8347 8348 8349 8350 8351 8352 8353 8354 8355 8356 8357 8358 8359 8360 8361 8362 8363 8364 8365 8366 8367 8368 8369 8370 8371 8372 8373 8374 8375
@templatedoc(op_type="trilinear_interp")
def resize_trilinear(input,
                     out_shape=None,
                     scale=None,
                     name=None,
                     actual_shape=None,
                     align_corners=True,
                     align_mode=1):
    """
    Resize input by performing trilinear interpolation based on given
    output shape which specified by actual_shape, out_shape and scale
    in priority order.

    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:
              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}



    Args:
        input(${x_type}): input should be a 4-D tensor.

        out_shape(list|tuple|Variable|None): Output shape of resize bilinear
                                    layer, the shape is (out_d, out_h, out_w).
                                    Default: None

        scale(float|None): The multiplier for the input depth, 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`. 
             Default: None.

        name(str|None): The output variable name.
        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
                                actual_shape instead of :attr:`out_shape` if you
                                want to specify output shape dynamically. 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.
                                Default: None
        align_corners(bool): ${align_corners_comment}
        align_mode(bool): ${align_mode_comment}

    Returns:
        A 5-D tensor in shape (num_batches, channels, out_d, out_h, out_w)

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
            input = fluid.layers.data(name="input", shape=[3,6,9,11], dtype="float32")
            out = fluid.layers.resize_trilinear(input, out_shape=[12, 12, 12])
    """

    return image_resize(input, out_shape, scale, name, 'TRILINEAR',
                        actual_shape, align_corners, align_mode)


8376
@templatedoc(op_type="nearest_interp")
8377 8378 8379 8380
def resize_nearest(input,
                   out_shape=None,
                   scale=None,
                   name=None,
8381 8382
                   actual_shape=None,
                   align_corners=True):
8383
    """
8384
    Resize input by performing nearest neighbor interpolation in both the
T
Tink_Y 已提交
8385 8386
    3rd dimension(in height direction) and the 4th dimension(in width
    direction) based on given output shape which is specified by actual_shape,
8387 8388
    out_shape and scale in priority order.

8389 8390
    Example:

T
Tink_Y 已提交
8391 8392 8393 8394 8395
    .. code-block:: text

        For scale:
          
            if align_corners = True && out_size > 1 :
8396

T
Tink_Y 已提交
8397 8398 8399 8400 8401 8402 8403 8404
              scale_factor = (in_size-1.0)/(out_size-1.0)
            
            else:
              
              scale_factor = float(in_size/out_size)
            
          
        Nearest neighbor interpolation:
8405
          
T
Tink_Y 已提交
8406 8407
          if:
              align_corners = False
8408

T
Tink_Y 已提交
8409 8410
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
8411

T
Tink_Y 已提交
8412 8413
              H_out = floor(H_{in} * scale_{factor})
              W_out = floor(W_{in} * scale_{factor})
8414

T
Tink_Y 已提交
8415 8416
          else:
              align_corners = True
8417

T
Tink_Y 已提交
8418 8419
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
8420

T
Tink_Y 已提交
8421 8422
              H_out = round(H_{in} * scale_{factor})
              W_out = round(W_{in} * scale_{factor})
8423 8424


8425
    For details of nearest neighbor interpolation, please refer to Wikipedia:
8426
    https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation
Y
yuyang18 已提交
8427 8428

    Args:
K
Kaipeng Deng 已提交
8429
        input(${x_type}): input should be a 4-D tensor.
Y
yuyang18 已提交
8430

D
dengkaipeng 已提交
8431 8432 8433
        out_shape(list|tuple|Variable|None): Output shape of resize nearest
                                    layer, the shape is (out_h, out_w).
                                    Default: None
8434

Y
yuyang18 已提交
8435
        scale(float|None): The multiplier for the input height or width. At
D
dengkaipeng 已提交
8436
             least one of :attr:`out_shape` or :attr:`scale` must be set. 
D
dengkaipeng 已提交
8437
             And :attr:`out_shape` has a higher priority than :attr:`scale`. 
D
dengkaipeng 已提交
8438
             Default: None.
Y
yuyang18 已提交
8439 8440

        name(str|None): The output variable name.
8441 8442 8443
        actual_shape(Variable): An optional input to specify output shape
                                dynamically. If provided, image resize
                                according to this given shape rather than
8444
                                :attr:`out_shape` and :attr:`scale` specifying
8445 8446 8447 8448 8449 8450 8451
                                shape. That is to say actual_shape has the
                                highest priority. It is recommended to use
                                actual_shape instead of :attr:`out_shape` if you
                                want to specify output shape dynamically. 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
8452 8453
                                constructing stage.
                                Default: None
8454
        align_corners(bool): ${align_corners_comment}
Y
yuyang18 已提交
8455 8456

    Returns:
K
Kaipeng Deng 已提交
8457
        A 4-D tensor in shape of (num_batches, channels, out_h, out_w)
8458 8459 8460 8461

    Examples:
        .. code-block:: python

8462
            import paddle.fluid as fluid
R
ruri 已提交
8463
            input = fluid.layers.data(name="input", shape=[3,6,9], dtype="float32")
8464
            out = fluid.layers.resize_nearest(input, out_shape=[12, 12])
8465 8466
    """

8467 8468
    return image_resize(input, out_shape, scale, name, 'NEAREST', actual_shape,
                        align_corners)
8469 8470 8471 8472


def image_resize_short(input, out_short_len, resample='BILINEAR'):
    """
8473 8474 8475
    Resize a batch of images. The short edge of input images will be
    resized to the given 'out_short_len'. The long edge of input images
    will be resized proportionately to make images' length-width ratio
8476 8477 8478 8479 8480 8481 8482
    constant.

    Args:
        input (Variable): The input tensor of image resize layer,
                          This is a 4-D tensor of the shape
                          (num_batches, channels, in_h, in_w).
        out_short_len(int): The length of output images' short edge.
8483
        resample (str): resample method, default: BILINEAR.
F
fengjiayi 已提交
8484

8485
    Returns:
Q
update  
qiaolongfei 已提交
8486
        Variable: The output is a 4-D tensor of the shape
8487
        (num_batches, channls, out_h, out_w).
R
ruri 已提交
8488 8489 8490 8491

    Examples:
        .. code-block:: python

8492
            import paddle.fluid as fluid
R
ruri 已提交
8493 8494
            input = fluid.layers.data(name="input", shape=[3,6,9], dtype="float32")
            out = fluid.layers.image_resize_short(input, out_short_len=3)
8495 8496 8497 8498 8499 8500 8501 8502 8503 8504
    """
    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
8505 8506 8507
    out_shape[long_idx] = int(
        float(out_shape[long_idx]) * (float(out_short_len) / float(hw[
            short_idx])) + 0.5)
8508 8509 8510
    return image_resize(input=input, out_shape=out_shape, resample=resample)


8511
def gather(input, index, overwrite=True):
W
whs 已提交
8512
    """
Q
qiaolongfei 已提交
8513 8514
    **Gather Layer**

8515
    Output is obtained by gathering entries of the outer-most dimension
W
whs 已提交
8516 8517 8518 8519
    of X indexed by `index` and concatenate them together.

    .. math::

8520
        Out = X[Index]
W
whs 已提交
8521 8522 8523 8524 8525 8526 8527


    .. code-block:: text


                Given:

8528 8529
                X = [[1, 2],
                     [3, 4],
W
whs 已提交
8530 8531 8532 8533 8534 8535 8536 8537 8538 8539
                     [5, 6]]

                Index = [1, 2]

                Then:

                Out = [[3, 4],
                       [5, 6]]

    Args:
8540
        input (Variable): The source input with rank>=1.
W
whs 已提交
8541
        index (Variable): The index input with rank=1.
8542 8543 8544 8545 8546 8547
        overwrite (bool): The mode that updating the grad when has same index.
            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
8548 8549 8550 8551 8552

    Returns:
        output (Variable): The output is a tensor with the same rank as input.

    Examples:
W
whs 已提交
8553

W
whs 已提交
8554 8555
        .. code-block:: python

8556
            import paddle.fluid as fluid
Y
Yibing Liu 已提交
8557 8558
            x = fluid.layers.data(name='x', shape=[-1, 5], dtype='float32')
            index = fluid.layers.data(name='index', shape=[-1, 1], dtype='int32')
W
whs 已提交
8559 8560 8561 8562
            output = fluid.layers.gather(x, index)
    """
    helper = LayerHelper('gather', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
8563
    out = helper.create_variable_for_type_inference(dtype)
W
whs 已提交
8564 8565 8566 8567
    helper.append_op(
        type="gather",
        inputs={"X": input,
                "Index": index},
8568 8569
        outputs={"Out": out},
        attrs={'overwrite': overwrite})
W
whs 已提交
8570 8571 8572
    return out


8573 8574 8575 8576 8577 8578 8579 8580 8581 8582 8583 8584 8585 8586 8587 8588 8589 8590 8591 8592 8593 8594 8595 8596 8597 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 8635 8636 8637 8638 8639 8640 8641 8642 8643 8644 8645 8646 8647 8648 8649 8650 8651 8652 8653 8654 8655 8656 8657
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:
        input (Variable): The source input
        index (Variable): The index input with rank > 1, index.shape[-1] <= input.rank
        name (str|None): A name for this layer(optional). If set None, the
                         layer will be named automatically

    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
            x = fluid.layers.data(name='x', shape=[3, 4, 5], dtype='float32')
            index = fluid.layers.data(name='index', shape=[2, 2], dtype='int32')
            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


8658
def scatter(input, index, updates, name=None, overwrite=True):
8659 8660 8661 8662 8663 8664 8665 8666 8667 8668 8669 8670 8671 8672 8673 8674 8675
    """
    **Scatter Layer**

    Output is obtained by updating the input on selected indices on the first
    axis.

    .. math::

        Out = X
        Out[Ids] = Updates

    Args:
        input (Variable): The source input with rank>=1.
        index (Variable): The index input with rank=1. Its dtype should be
                          int32 or int64 as it is used as indexes.
        updates (Variable): The updated value of scatter op.
        name (str|None): The output variable name. Default None.
8676 8677 8678 8679
        overwrite (bool): The mode that updating the output when has same index.
            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. 
	    Default value is True.You can set overwrite=False to implement scatter_add.
8680 8681 8682 8683 8684 8685 8686 8687

    Returns:
        output (Variable): The output is a tensor with the same shape as input.

    Examples:

        .. code-block:: python

8688 8689 8690 8691 8692
            import paddle.fluid as fluid

            input = fluid.layers.data(name='data', shape=[3, 5, 9], dtype='float32', append_batch_size=False)
            index = fluid.layers.data(name='index', shape=[3], dtype='int64', append_batch_size=False)
            updates = fluid.layers.data(name='update', shape=[3, 5, 9], dtype='float32', append_batch_size=False)
8693

8694
            output = fluid.layers.scatter(input, index, updates)
8695 8696 8697
    """
    helper = LayerHelper('scatter', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
8698
    out = helper.create_variable_for_type_inference(dtype)
8699 8700 8701 8702 8703
    helper.append_op(
        type="scatter",
        inputs={"X": input,
                "Ids": index,
                "Updates": updates},
8704
        attrs={'overwrite': overwrite},
8705 8706 8707 8708
        outputs={"Out": out})
    return out


8709 8710 8711 8712 8713 8714 8715 8716 8717 8718 8719 8720 8721 8722 8723 8724 8725 8726 8727 8728 8729 8730 8731 8732 8733 8734 8735 8736 8737 8738 8739 8740 8741 8742 8743 8744 8745 8746 8747 8748 8749 8750 8751 8752 8753 8754 8755 8756 8757 8758 8759 8760 8761 8762 8763 8764 8765 8766 8767 8768 8769 8770 8771 8772 8773 8774 8775 8776 8777 8778 8779 8780 8781 8782 8783 8784 8785 8786 8787 8788 8789 8790 8791 8792 8793 8794 8795 8796 8797 8798 8799 8800 8801 8802 8803 8804 8805 8806 8807 8808 8809 8810 8811 8812 8813 8814 8815 8816 8817 8818 8819 8820 8821 8822 8823 8824 8825 8826 8827 8828 8829
def scatter_nd_add(ref, index, updates, name=None):
    """
    **Scatter_nd_add Layer**

    Output is obtained by applying sparse addition to a single value
    or slice in a Variable. :attr:`ref` is a Tensor with rank :math:`R` 
    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]:]` .
    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:
        ref (Variable): The ref input.
        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.
        updates (Variable): The updated value of scatter_nd_add op, and it must have the same type
                            as ref. It must have the shape index.shape[:-1] + ref.shape[index.shape[-1]:]
        name (str|None): The output variable name. Default None.

    Returns:
        output (Variable): The output is a tensor with the same shape and type as ref.

    Examples:

        .. code-block:: python

            import paddle.fluid as fluid

            ref = fluid.layers.data(name='ref', shape=[3, 5, 9, 10], dtype='float32', append_batch_size=False)
            index = fluid.layers.data(name='index', shape=[3, 2], dtype='int32', append_batch_size=False)
            updates = fluid.layers.data(name='update', shape=[3, 9, 10], dtype='float32', append_batch_size=False)

            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())
    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="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.
        updates (Variable): The updated value of scatter_nd op. 
                            It must have the shape index.shape[:-1] + shape[index.shape[-1]:]
        shape(tuple|list): Shape of output tensor.
        name (str|None): The output variable name. Default None.

    Returns:
        output (Variable): The output is a tensor with the same type as :attr:`updates` .

    Examples:

        .. code-block:: python

            import paddle.fluid as fluid

            index = fluid.layers.data(name='index', shape=[3, 2], dtype='int64', append_batch_size=False)
            updates = fluid.layers.data(name='update', shape=[3, 9, 10], dtype='float32', append_batch_size=False)
            shape = [3, 5, 9, 10]

            output = fluid.layers.scatter_nd(index, updates, shape)
    """
    return scatter_nd_add(zeros(shape, updates.dtype), index, updates, name)


Q
Qingsheng Li 已提交
8830 8831 8832 8833 8834 8835 8836 8837 8838
def sequence_scatter(input, index, updates, name=None):
    """
    **Sequence Scatter Layer**

    This operator scatters the Updates tensor to the input X. It uses the LoD
    information of Ids to select the rows to update, and use the values in Ids as
    the columns to update in each row of X.

    Here is an example:
H
haowang101779990 已提交
8839

Q
Qingsheng Li 已提交
8840
    Given the following input:
H
haowang101779990 已提交
8841

Q
Qingsheng Li 已提交
8842
    .. code-block:: text
H
haowang101779990 已提交
8843

Q
Qingsheng Li 已提交
8844 8845 8846 8847 8848 8849 8850 8851 8852 8853 8854 8855
        input.data = [[1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
                      [1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
                      [1.0, 1.0, 1.0, 1.0, 1.0, 1.0]]
        input.dims = [3, 6]

        index.data = [[0], [1], [2], [5], [4], [3], [2], [1], [3], [2], [5], [4]]
        index.lod =  [[0,        3,                       8,                 12]]

        updates.data = [[0.3], [0.3], [0.4], [0.1], [0.2], [0.3], [0.4], [0.0], [0.2], [0.3], [0.1], [0.4]]
        updates.lod =  [[  0,            3,                                 8,                         12]]

    Then we have the output:
H
haowang101779990 已提交
8856

Q
Qingsheng Li 已提交
8857
    .. code-block:: text
H
haowang101779990 已提交
8858

Q
Qingsheng Li 已提交
8859 8860 8861 8862 8863 8864 8865 8866 8867 8868 8869 8870 8871 8872 8873
        out.data = [[1.3, 1.3, 1.4, 1.0, 1.0, 1.0],
                    [1.0, 1.0, 1.4, 1.3, 1.2, 1.1],
                    [1.0, 1.0, 1.3, 1.2, 1.4, 1.1]]
        out.dims = X.dims = [3, 6]

    Args:
        input (Variable): The source input with rank>=1.
        index (Variable): A LoD Tensor. The index input of sequence scatter op
            where input will be  updated. The index input with rank=1. Its dtype
            should be int32 or int64 as it is used as indexes.
        updates (Variable): A LoD Tensor. The values to scatter to the input
            tensor X, must be a LoDTensor with the same LoD information as index.
        name (str|None): The output variable name. Default None.

    Returns:
H
haowang101779990 已提交
8874
        Variable: The output is a tensor with the same shape as input.
Q
Qingsheng Li 已提交
8875 8876 8877 8878

    Examples:

        .. code-block:: python
8879
	
8880
            import paddle.fluid as fluid
8881
            import paddle.fluid.layers as layers
Q
Qingsheng Li 已提交
8882

8883 8884 8885
            input = layers.data( name="x", shape=[3, 6], append_batch_size=False, dtype='float32' )
            index = layers.data( name='index', shape=[1], dtype='int32')
            updates = layers.data( name='updates', shape=[1], dtype='float32')
Q
Qingsheng Li 已提交
8886 8887 8888
            output = fluid.layers.sequence_scatter(input, index, updates)

    """
L
lujun 已提交
8889
    assert not in_dygraph_mode(), (
8890
        "sequence layer is not supported in dygraph mode yet.")
Q
Qingsheng Li 已提交
8891 8892
    helper = LayerHelper('sequence_scatter', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
8893
    out = helper.create_variable_for_type_inference(dtype)
Q
Qingsheng Li 已提交
8894 8895 8896 8897 8898 8899 8900 8901 8902
    helper.append_op(
        type="sequence_scatter",
        inputs={"X": input,
                "Ids": index,
                "Updates": updates},
        outputs={"Out": out})
    return out


Y
yuyang18 已提交
8903 8904 8905 8906 8907 8908 8909 8910 8911 8912 8913 8914 8915
@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}
8916

8917
    Examples:
8918
        >>> import paddle.fluid as fluid
8919 8920
        >>> img = fluid.layers.data("img", [3, 256, 256])
        >>> cropped_img = fluid.layers.random_crop(img, shape=[3, 224, 224])
Y
yuyang18 已提交
8921
    """
F
stash  
fengjiayi 已提交
8922
    helper = LayerHelper("random_crop", **locals())
F
fengjiayi 已提交
8923
    dtype = x.dtype
X
Xin Pan 已提交
8924
    out = helper.create_variable_for_type_inference(dtype)
Y
yuyang18 已提交
8925
    if seed is None:
8926
        seed = np.random.randint(-65536, 65536)
F
fengjiayi 已提交
8927
    op_attrs = {"shape": shape}
F
stash  
fengjiayi 已提交
8928
    if isinstance(seed, int):
F
fengjiayi 已提交
8929 8930 8931 8932 8933
        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
8934 8935 8936 8937
    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
8938
        inputs={"X": x,
F
stash  
fengjiayi 已提交
8939 8940
                "Seed": seed},
        outputs={"Out": out,
F
fengjiayi 已提交
8941 8942
                 "SeedOut": seed},
        attrs=op_attrs)
F
stash  
fengjiayi 已提交
8943
    return out
W
whs 已提交
8944 8945


8946
def log(x, name=None):
W
wanghaoshuang 已提交
8947 8948 8949 8950 8951
    """
    Calculates the natural log of the given input tensor, element-wise.

    .. math::

8952
        Out = \\ln(x)
W
wanghaoshuang 已提交
8953 8954

    Args:
8955
        x (Variable): Input tensor.
8956 8957
        name (str|None, default None): A name for this layer If set None,
            the layer will be named automatically.
W
wanghaoshuang 已提交
8958 8959 8960 8961 8962 8963 8964 8965

    Returns:
        Variable: The natural log of the input tensor computed element-wise.

    Examples:

        .. code-block:: python

8966
            import paddle.fluid as fluid
8967
            x = fluid.layers.data(name="x", shape=[3, 4], dtype="float32")
8968
            output = fluid.layers.log(x)
W
wanghaoshuang 已提交
8969 8970
    """
    helper = LayerHelper('log', **locals())
W
wanghaoshuang 已提交
8971
    dtype = helper.input_dtype(input_param_name='x')
X
Xin Pan 已提交
8972
    out = helper.create_variable_for_type_inference(dtype)
W
wanghaoshuang 已提交
8973
    helper.append_op(type="log", inputs={"X": x}, outputs={"Out": out})
W
wanghaoshuang 已提交
8974 8975 8976
    return out


8977
def relu(x, name=None):
W
wanghaoshuang 已提交
8978 8979
    """
    Relu takes one input data (Tensor) and produces one output data (Tensor)
8980
    where the rectified linear function, y = max(0, x), is applied to
W
wanghaoshuang 已提交
8981 8982 8983 8984
    the tensor elementwise.

    .. math::

8985
        Out = \\max(0, x)
W
wanghaoshuang 已提交
8986 8987

    Args:
8988
        x (Variable): The input tensor.
8989 8990
        name (str|None, default None): A name for this layer If set None,
            the layer will be named automatically.
W
wanghaoshuang 已提交
8991 8992 8993 8994 8995 8996 8997 8998

    Returns:
        Variable: The output tensor with the same shape as input.

    Examples:

        .. code-block:: python

8999
            import paddle.fluid as fluid
9000
            x = fluid.layers.data(name="x", shape=[3, 4], dtype="float32")
9001
            output = fluid.layers.relu(x)
W
wanghaoshuang 已提交
9002 9003
    """
    helper = LayerHelper('relu', **locals())
W
wanghaoshuang 已提交
9004
    dtype = helper.input_dtype(input_param_name='x')
X
Xin Pan 已提交
9005
    out = helper.create_variable_for_type_inference(dtype)
X
Xin Pan 已提交
9006 9007
    helper.append_op(
        type="relu", inputs={"X": helper.input('x')}, outputs={"Out": out})
W
wanghaoshuang 已提交
9008
    return out
9009 9010


C
chengduo 已提交
9011 9012 9013 9014 9015 9016 9017 9018 9019 9020 9021 9022 9023 9024 9025 9026 9027 9028 9029 9030 9031 9032 9033 9034
@templatedoc()
def selu(x, scale=None, alpha=None, name=None):
    """
    ${comment}

    Args:
        x (Variable): The input tensor.
        scale(float, None): If the scale is not set,
            the default value is 1.0507009873554804934193349852946.
            For more information about this value, please refer
            to: https://arxiv.org/abs/1706.02515.
        alpha(float, None): If the alpha is not set,
            the default value is 1.6732632423543772848170429916717.
            For more information about this value, please refer
            to: https://arxiv.org/abs/1706.02515.
        name (str|None, default None): A name for this layer If set None,
            the layer will be named automatically.

    Returns:
        Variable: The output tensor with the same shape as input.

    Examples:

        .. code-block:: python
9035 9036 9037 9038 9039 9040
             
            import paddle.fluid as fluid
          
            input = fluid.layers.data(
                 name="input", shape=[3, 9, 5], dtype="float32")
            output = fluid.layers.selu(input)
C
chengduo 已提交
9041 9042 9043 9044 9045 9046 9047 9048 9049 9050 9051 9052 9053 9054 9055
    """
    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
9056 9057 9058
def mean_iou(input, label, num_classes):
    """
    Mean Intersection-Over-Union is a common evaluation metric for
9059 9060 9061 9062
    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
9063
    .. math::
9064

H
haowang101779990 已提交
9065
        IOU = \\frac{true\_positive}{(true\_positive + false\_positive + false\_negative)}.
W
whs 已提交
9066

9067
    The predictions are accumulated in a confusion matrix and mean-IOU
W
whs 已提交
9068 9069 9070 9071 9072
    is then calculated from it.


    Args:
        input (Variable): A Tensor of prediction results for semantic labels with type int32 or int64.
9073
        label (Variable): A Tensor of ground truth labels with type int32 or int64.
W
whs 已提交
9074
                           Its shape should be the same as input.
9075
        num_classes (int): The possible number of labels.
W
whs 已提交
9076 9077

    Returns:
M
minqiyang 已提交
9078 9079
        mean_iou (Variable),out_wrong(Variable),out_correct(Variable):

H
haowang101779990 已提交
9080
                     Three variables:
M
minqiyang 已提交
9081

H
haowang101779990 已提交
9082 9083 9084
                     - mean_iou : A Tensor representing the mean intersection-over-union with shape [1].
                     - out_wrong: A Tensor with shape [num_classes]. The wrong numbers of each class.
                     - out_correct: A Tensor with shape [num_classes]. The correct numbers of each class.
W
whs 已提交
9085 9086 9087 9088

    Examples:

        .. code-block:: python
9089

B
Bai Yifan 已提交
9090
            import paddle.fluid as fluid
9091 9092 9093 9094
            iou_shape = [32, 32]
            num_classes = 5
            predict = fluid.layers.data(name='predict', shape=iou_shape)
            label = fluid.layers.data(name='label', shape=iou_shape)
B
Bai Yifan 已提交
9095
            iou, wrongs, corrects = fluid.layers.mean_iou(predict, label,
9096
                                                          num_classes)
W
whs 已提交
9097 9098 9099
    """
    helper = LayerHelper('mean_iou', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
9100 9101 9102
    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
9103 9104
    helper.append_op(
        type="mean_iou",
W
whs 已提交
9105 9106
        inputs={"Predictions": input,
                "Labels": label},
W
whs 已提交
9107
        outputs={
W
whs 已提交
9108 9109 9110
            "OutMeanIou": out_mean_iou,
            "OutWrong": out_wrong,
            "OutCorrect": out_correct
W
whs 已提交
9111 9112 9113
        },
        attrs={"num_classes": num_classes})
    return out_mean_iou, out_wrong, out_correct
9114 9115 9116 9117 9118 9119 9120 9121 9122 9123 9124 9125 9126 9127 9128 9129 9130 9131 9132 9133 9134 9135 9136 9137 9138 9139 9140 9141 9142 9143 9144 9145 9146 9147 9148 9149 9150 9151 9152 9153 9154 9155


def crop(x, shape=None, offsets=None, name=None):
    """
    Crop input into output, as specified by offsets and shape.

    .. 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]].

    Args:
        x (Variable): The input tensor variable.
        shape (Variable|list/tuple of integer): The output shape is specified
            by `shape`, which can a Variable or a list/tupe of integer.
            If a tensor Variable, it's rank must be the same as `x`. This way
            is suitable for the case that the output shape may be changed each
            iteration. If a list/tupe of integer, it's length must be the same
            as the rank of `x`
S
SunGaofeng 已提交
9156
        offsets (Variable|list/tuple of integer|None): Specifies the cropping
9157
            offsets at each dimension. It can be a Variable or or a list/tupe
S
SunGaofeng 已提交
9158
            of integers. If a tensor Variable, it's rank must be the same as `x`.
9159 9160 9161 9162 9163 9164 9165 9166 9167 9168 9169 9170 9171 9172 9173 9174 9175
            This way is suitable for the case that the offsets may be changed
            each iteration. If a list/tupe of integer, it's length must be the
            same as the rank of `x`. If None, the offsets are 0 at each
            dimension.
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        Variable: The cropped tensor variable.

    Raises:
        ValueError: If shape is not a list, tuple or Variable.

    Examples:

        .. code-block:: python

S
SunGaofeng 已提交
9176
            import paddle.fluid as fluid
9177 9178 9179 9180 9181 9182
            x = fluid.layers.data(name="x", shape=[3, 5], dtype="float32")
            y = fluid.layers.data(name="y", shape=[2, 3], dtype="float32")
            crop = fluid.layers.crop(x, shape=y)

            # or
            z = fluid.layers.data(name="z", shape=[3, 5], dtype="float32")
T
Tink_Y 已提交
9183
            crop = fluid.layers.crop(z, shape=[-1, 2, 3])
9184 9185 9186 9187 9188

    """
    helper = LayerHelper('crop', **locals())

    if not (isinstance(shape, list) or isinstance(shape, tuple) or \
9189
            isinstance(shape, Variable)):
9190 9191 9192 9193 9194
        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
9195
    out = helper.create_variable_for_type_inference(x.dtype)
9196 9197 9198 9199 9200 9201 9202 9203 9204 9205 9206 9207 9208 9209 9210 9211 9212
    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
9213 9214


W
whs 已提交
9215 9216 9217 9218 9219 9220 9221 9222 9223 9224 9225 9226 9227 9228 9229 9230 9231
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.

    .. code-block:: text

        * Case 1:

          Given:

              theta = [[[x_11, x_12, x_13]
                        [x_14, x_15, x_16]]
                       [[x_21, x_22, x_23]
                        [x_24, x_25, x_26]]]
9232

W
whs 已提交
9233
              out_shape = [2, 3, 5, 5]
9234

W
whs 已提交
9235
          Step 1:
9236

W
whs 已提交
9237 9238 9239
              Generate normalized coordinates according to out_shape.
              The values of the normalized coordinates are in the interval between -1 and 1.
              The shape of the normalized coordinates is [2, H, W] as below:
9240

W
whs 已提交
9241 9242 9243 9244 9245 9246 9247 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
              C = [[[-1.  -1.  -1.  -1.  -1. ]
                    [-0.5 -0.5 -0.5 -0.5 -0.5]
                    [ 0.   0.   0.   0.   0. ]
                    [ 0.5  0.5  0.5  0.5  0.5]
                    [ 1.   1.   1.   1.   1. ]]
                   [[-1.  -0.5  0.   0.5  1. ]
                    [-1.  -0.5  0.   0.5  1. ]
                    [-1.  -0.5  0.   0.5  1. ]
                    [-1.  -0.5  0.   0.5  1. ]
                    [-1.  -0.5  0.   0.5  1. ]]]
              C[0] is the coordinates in height axis and  C[1] is the coordinates in width axis.

          Step2:

              Tanspose and reshape C to shape [H * W, 2] and append ones to last dimension. The we get:
              C_ = [[-1.  -1.   1. ]
                    [-0.5 -1.   1. ]
                    [ 0.  -1.   1. ]
                    [ 0.5 -1.   1. ]
                    [ 1.  -1.   1. ]
                    [-1.  -0.5  1. ]
                    [-0.5 -0.5  1. ]
                    [ 0.  -0.5  1. ]
                    [ 0.5 -0.5  1. ]
                    [ 1.  -0.5  1. ]
                    [-1.   0.   1. ]
                    [-0.5  0.   1. ]
                    [ 0.   0.   1. ]
                    [ 0.5  0.   1. ]
                    [ 1.   0.   1. ]
                    [-1.   0.5  1. ]
                    [-0.5  0.5  1. ]
                    [ 0.   0.5  1. ]
                    [ 0.5  0.5  1. ]
                    [ 1.   0.5  1. ]
                    [-1.   1.   1. ]
                    [-0.5  1.   1. ]
                    [ 0.   1.   1. ]
                    [ 0.5  1.   1. ]
                    [ 1.   1.   1. ]]
          Step3:
              Compute output by equation $$Output[i] = C_ * Theta[i]^T$$

    Args:
        theta (Variable): A batch of affine transform parameters with shape [N, 2, 3].
M
minqiyang 已提交
9286
        out_shape (Variable | list | tuple): The shape of target output with format [N, C, H, W].
H
haowang101779990 已提交
9287
                                             ``out_shape`` can be a Variable or a list or tuple.
W
whs 已提交
9288 9289 9290 9291 9292 9293 9294 9295 9296 9297 9298 9299
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        Variable: The output with shape [N, H, W, 2].

    Raises:
        ValueError: If the type of arguments is not supported.

    Examples:

        .. code-block:: python
H
haowang101779990 已提交
9300

S
SunGaofeng 已提交
9301
            import paddle.fluid as fluid
W
whs 已提交
9302 9303 9304 9305 9306 9307 9308 9309 9310 9311 9312
            theta = fluid.layers.data(name="x", shape=[2, 3], dtype="float32")
            out_shape = fluid.layers.data(name="y", shape=[-1], dtype="float32")
            data = fluid.layers.affine_grid(theta, out_shape)

            # or
            data = fluid.layers.affine_grid(theta, [5, 3, 28, 28])

    """
    helper = LayerHelper('affine_grid')

    if not (isinstance(out_shape, list) or isinstance(out_shape, tuple) or \
9313
            isinstance(out_shape, Variable)):
W
whs 已提交
9314 9315 9316 9317 9318 9319 9320 9321 9322 9323 9324 9325 9326 9327 9328 9329 9330 9331 9332 9333 9334
        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


9335 9336
def rank_loss(label, left, right, name=None):
    """
H
haowang101779990 已提交
9337

9338 9339
    **Rank loss layer for RankNet**

H
haowang101779990 已提交
9340
    `RankNet <http://icml.cc/2015/wp-content/uploads/2015/06/icml_ranking.pdf>`_
9341 9342 9343
    is a pairwise ranking model with a training sample consisting of a pair
    of documents, A and B. Label P indicates whether A is ranked higher than B
    or not:
M
minqiyang 已提交
9344

9345 9346
    P = {0, 1} or {0, 0.5, 1}, where 0.5 means that there is no information
    about the rank of the input pair.
M
minqiyang 已提交
9347

H
haowang101779990 已提交
9348 9349
    Rank loss layer takes three inputs: left ( :math:`o_i` ), right ( :math:`o_j` ) and
    label ( :math:`P_{i,j}` ). The inputs respectively represent RankNet's output scores
9350 9351
    for documents A and B and the value of label P. The following equation
    computes rank loss C_{i,j} from the inputs:
M
minqiyang 已提交
9352

H
haowang101779990 已提交
9353 9354 9355 9356 9357 9358 9359 9360
    .. math::

      C_{i,j} &= -\\tilde{P_{ij}} * o_{i,j} + \log(1 + e^{o_{i,j}}) \\\\

      o_{i,j} &=  o_i - o_j  \\\\

      \\tilde{P_{i,j}} &= \\left \{0, 0.5, 1 \\right \} \ or \ \\left \{0, 1 \\right \}

M
minqiyang 已提交
9361 9362 9363

    Rank loss layer takes batch inputs with size batch_size (batch_size >= 1).

9364 9365 9366 9367 9368 9369 9370 9371 9372 9373 9374 9375 9376 9377 9378 9379 9380
    Args:
        label (Variable): Indicats whether A ranked higher than B or not.
        left (Variable): RankNet's output score for doc A.
        right (Variable): RankNet's output score for doc B.
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        list: The value of rank loss.

    Raises:
        ValueError: Any of label, left, and right is not a variable.

    Examples:

        .. code-block:: python

9381
            import paddle.fluid as fluid
9382 9383 9384
            label = fluid.layers.data(name="label", shape=[-1, 1], dtype="float32")
            left = fluid.layers.data(name="left", shape=[-1, 1], dtype="float32")
            right = fluid.layers.data(name="right", shape=[-1, 1], dtype="float32")
9385 9386 9387 9388 9389 9390 9391 9392 9393 9394 9395 9396 9397 9398
            out = fluid.layers.rank_loss(label, left, right)

    """
    helper = LayerHelper('rank_loss', **locals())

    if not (isinstance(label, Variable)):
        raise ValueError("The label should be a Variable")

    if not (isinstance(left, Variable)):
        raise ValueError("The left should be a Variable")

    if not (isinstance(right, Variable)):
        raise ValueError("The right should be a Variable")

X
Xin Pan 已提交
9399
    out = helper.create_variable_for_type_inference("float32")
9400 9401 9402 9403 9404 9405 9406 9407

    helper.append_op(
        type='rank_loss',
        inputs={"Label": label,
                "Left": left,
                "Right": right},
        outputs={'Out': out})
    return out
9408 9409


M
minqiyang 已提交
9410 9411
def margin_rank_loss(label, left, right, margin=0.1, name=None):
    """
M
minqiyang 已提交
9412
    Margin Ranking Loss Layer for ranking problem,
M
minqiyang 已提交
9413
    which compares left score and right score passed in.
M
minqiyang 已提交
9414
    The ranking loss can be defined as following equation:
M
minqiyang 已提交
9415 9416 9417

    .. math::

H
haowang101779990 已提交
9418
        rank\_loss = max(0, -label * (left - right) + margin)
M
minqiyang 已提交
9419 9420

    Args:
M
minqiyang 已提交
9421
       label (Variable): Indicates whether the left is ranked higher than the right or not.
M
minqiyang 已提交
9422 9423
       left (Variable): Ranking score for left.
       right (Variable): Ranking score for right.
M
minqiyang 已提交
9424
       margin (float): Indicates the given margin.
M
minqiyang 已提交
9425 9426
       name (str|None): A name for this layer (optional). If set None, the layer
                       will be named automatically.
H
haowang101779990 已提交
9427

M
minqiyang 已提交
9428
    Returns:
M
minqiyang 已提交
9429
       Variable: The ranking loss.
H
haowang101779990 已提交
9430

M
minqiyang 已提交
9431
    Raises:
M
minqiyang 已提交
9432
       ValueError: Any of label, left, and right is not a Variable.
H
haowang101779990 已提交
9433

M
minqiyang 已提交
9434
    Examples:
H
haowang101779990 已提交
9435

M
minqiyang 已提交
9436
        .. code-block:: python
H
haowang101779990 已提交
9437

9438
           import paddle.fluid as fluid
Y
Yibing Liu 已提交
9439 9440 9441
           label = fluid.layers.data(name="label", shape=[-1, 1], dtype="float32")
           left = fluid.layers.data(name="left", shape=[-1, 1], dtype="float32")
           right = fluid.layers.data(name="right", shape=[-1, 1], dtype="float32")
M
minqiyang 已提交
9442 9443
           out = fluid.layers.margin_rank_loss(label, left, right)
    """
M
minqiyang 已提交
9444
    helper = LayerHelper('margin_rank_loss', **locals())
M
minqiyang 已提交
9445 9446 9447 9448 9449 9450
    if not isinstance(label, Variable):
        raise ValueError("The label should be a Variable.")
    if not isinstance(left, Variable):
        raise ValueError("The left should be a Variable.")
    if not isinstance(right, Variable):
        raise ValueError("The right should be a Variable.")
X
Xin Pan 已提交
9451 9452
    out = helper.create_variable_for_type_inference(left.dtype)
    act = helper.create_variable_for_type_inference(left.dtype)
M
minqiyang 已提交
9453 9454 9455 9456 9457 9458 9459 9460 9461 9462 9463
    helper.append_op(
        type='margin_rank_loss',
        inputs={"Label": label,
                "X1": left,
                "X2": right},
        outputs={'Out': out,
                 'Activated': act},
        attrs={'margin': margin})
    return out


W
whs 已提交
9464 9465 9466 9467 9468 9469 9470 9471 9472 9473 9474 9475
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.

    Example:
T
Tink_Y 已提交
9476
        .. code-block:: text
W
whs 已提交
9477

T
Tink_Y 已提交
9478
	      Given that X is a channel of image from input:
M
minqiyang 已提交
9479

T
Tink_Y 已提交
9480 9481
	      X = [[1, 2, 3],
		   [4, 5, 6]]
M
minqiyang 已提交
9482

T
Tink_Y 已提交
9483
	      Case 0:
M
minqiyang 已提交
9484

T
Tink_Y 已提交
9485 9486 9487
		paddings = [0, 1, 2, 3],
		mode = 'constant'
		pad_value = 0
M
minqiyang 已提交
9488

T
Tink_Y 已提交
9489 9490 9491
		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
9492

T
Tink_Y 已提交
9493
	      Case 1:
M
minqiyang 已提交
9494

T
Tink_Y 已提交
9495 9496
		paddings = [0, 1, 2, 1],
		mode = 'reflect'
M
minqiyang 已提交
9497

T
Tink_Y 已提交
9498 9499 9500
		Out = [[3, 2, 1, 2, 3, 2]
		       [6, 5, 4, 5, 6, 5]
		       [3, 2, 1, 2, 3, 2]]
M
minqiyang 已提交
9501

T
Tink_Y 已提交
9502
	      Case 2:
M
minqiyang 已提交
9503

T
Tink_Y 已提交
9504 9505
		paddings = [0, 1, 2, 1],
		mode = 'edge'
M
minqiyang 已提交
9506

T
Tink_Y 已提交
9507 9508 9509
		Out = [[1, 1, 1, 2, 3, 3]
		       [4, 4, 4, 5, 6, 6]
		       [4, 4, 4, 5, 6, 6]]
M
minqiyang 已提交
9510 9511


W
whs 已提交
9512 9513
    Args:
        input (Variable): The input image with [N, C, H, W] format or [N, H, W, C] format.
9514
        paddings (tuple|list|Variable): The padding size. If padding is a tuple, it must
W
whs 已提交
9515 9516 9517 9518 9519 9520 9521 9522 9523 9524 9525 9526 9527 9528 9529 9530 9531
            contain four integers, (padding_top, padding_bottom, padding_left, padding_right).
            Default: padding = [0, 0, 0, 0].
        mode (str): Three modes: constant(default), reflect, edge. Default: constant
        pad_value (float32): The value to fill the padded areas in constant mode. Default: 0
        data_format (str): An optional string from: "NHWC", "NCHW". Specify the data format of
                           the input data.
                           Default: "NCHW"
        name (str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.

    Returns:
        Variable: The tensor variable padded accordding to paddings and mode.


    Examples:
        .. code-block:: python

B
Bai Yifan 已提交
9532 9533 9534 9535 9536
          import paddle.fluid as fluid
          data = fluid.layers.data(name='data', shape=[3, 32, 32],
                                   dtype='float32')
          result = fluid.layers.pad2d(input=data, paddings=[1, 2, 3, 4],
                                      mode='reflect')
W
whs 已提交
9537 9538 9539 9540
    """

    helper = LayerHelper('pad2d', **locals())
    dtype = helper.input_dtype(input_param_name='input')
X
Xin Pan 已提交
9541
    out = helper.create_variable_for_type_inference(dtype)
9542 9543 9544 9545 9546 9547 9548 9549 9550
    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
9551
    helper.append_op(
9552
        type='pad2d', inputs=inputs, outputs={"Out": out}, attrs=attrs)
W
whs 已提交
9553 9554 9555 9556

    return out


9557 9558 9559 9560 9561 9562 9563 9564 9565 9566 9567 9568
@templatedoc()
def elu(x, alpha=1.0, name=None):
    """
    ${comment}
    Args:
        x(${x_type}): ${x_comment}
        alpha(${alpha_type}|1.0): ${alpha_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        output(${out_type}): ${out_comment}
Z
ZhenWang 已提交
9569 9570 9571 9572 9573

    Examples:

        .. code-block:: python

9574
            import paddle.fluid as fluid
Z
ZhenWang 已提交
9575 9576
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
            y = fluid.layers.elu(x, alpha=0.2)
9577 9578
    """
    helper = LayerHelper('elu', **locals())
X
Xin Pan 已提交
9579
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
9580 9581 9582 9583 9584 9585 9586 9587 9588 9589 9590 9591 9592 9593 9594 9595 9596 9597 9598 9599
    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):
    """
    ${comment}
    Args:
        x(${x_type}): ${x_comment}
        threshold(${threshold_type}|6.0): ${threshold_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        output(${out_type}): ${out_comment}
Z
ZhenWang 已提交
9600 9601 9602 9603 9604

    Examples:

        .. code-block:: python

9605
            import paddle.fluid as fluid
Z
ZhenWang 已提交
9606 9607
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
            y = fluid.layers.relu6(x, threshold=6.0)
9608 9609
    """
    helper = LayerHelper('relu6', **locals())
X
Xin Pan 已提交
9610
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
9611 9612 9613 9614 9615 9616 9617 9618 9619 9620 9621 9622 9623 9624 9625 9626 9627 9628 9629 9630
    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):
    """
    ${comment}
    Args:
        x(${x_type}): ${x_comment}
        factor(${factor_type}|1.0): ${factor_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        output(${out_type}): ${out_comment}
Z
ZhenWang 已提交
9631 9632 9633 9634 9635

    Examples:

        .. code-block:: python

9636
            import paddle.fluid as fluid
Z
ZhenWang 已提交
9637 9638
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
            y = fluid.layers.pow(x, factor=2.0)
9639 9640
    """
    helper = LayerHelper('pow', **locals())
X
Xin Pan 已提交
9641
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
9642 9643 9644 9645 9646 9647 9648 9649 9650 9651 9652 9653 9654 9655 9656 9657 9658 9659 9660 9661 9662
    helper.append_op(
        type='pow',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'factor': factor})
    return out


@templatedoc()
def stanh(x, scale_a=2.0 / 3.0, scale_b=1.7159, name=None):
    """
    ${comment}
    Args:
        x(${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:
        output(${out_type}): ${out_comment}
Z
ZhenWang 已提交
9663 9664 9665 9666 9667

    Examples:

        .. code-block:: python

9668
            import paddle.fluid as fluid
Z
ZhenWang 已提交
9669
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
Z
ZhenWang 已提交
9670
            y = fluid.layers.stanh(x, scale_a=0.67, scale_b=1.72)
9671 9672
    """
    helper = LayerHelper('stanh', **locals())
X
Xin Pan 已提交
9673
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
9674 9675 9676 9677 9678 9679 9680 9681 9682 9683 9684 9685 9686 9687 9688 9689 9690 9691 9692 9693 9694 9695
    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):
    """
    ${comment}
    Args:
        x(${x_type}): ${x_comment}
        slope(${slope_type}|0.2): ${slope_comment}
        offset(${offset_type}|0.5): ${offset_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        output(${out_type}): ${out_comment}
Z
ZhenWang 已提交
9696 9697 9698 9699 9700

    Examples:

        .. code-block:: python

9701
            import paddle.fluid as fluid
Z
ZhenWang 已提交
9702 9703
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
            y = fluid.layers.hard_sigmoid(x, slope=0.3, offset=0.8)
9704 9705
    """
    helper = LayerHelper('hard_sigmoid', **locals())
X
Xin Pan 已提交
9706
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
9707 9708 9709 9710 9711 9712 9713 9714 9715 9716 9717 9718 9719 9720 9721 9722 9723 9724 9725 9726 9727
    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):
    """
    ${comment}
    Args:
        x(${x_type}): ${x_comment}
        beta(${beta_type}|1.0): ${beta_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        output(${out_type}): ${out_comment}
Z
ZhenWang 已提交
9728 9729 9730 9731 9732

    Examples:

        .. code-block:: python

9733
            import paddle.fluid as fluid
Z
ZhenWang 已提交
9734 9735
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
            y = fluid.layers.swish(x, beta=2.0)
9736 9737
    """
    helper = LayerHelper('swish', **locals())
X
Xin Pan 已提交
9738
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
9739 9740 9741 9742 9743 9744 9745 9746
    helper.append_op(
        type='swish',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'slope': beta})
    return out


J
jerrywgz 已提交
9747 9748 9749 9750
def prelu(x, mode, param_attr=None, name=None):
    """
    Equation:

H
haowang101779990 已提交
9751 9752
    .. math::
        y = \max(0, x) + \\alpha * \min(0, x)
J
jerrywgz 已提交
9753

J
jerrywgz 已提交
9754 9755 9756 9757 9758 9759 9760 9761
    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
9762 9763
    Args:
        x (Variable): The input tensor.
J
jerrywgz 已提交
9764
        mode (string): The mode for weight sharing. 
J
jerrywgz 已提交
9765
        param_attr(ParamAttr|None): The parameter attribute for the learnable
J
jerrywgz 已提交
9766
          weight (alpha), it can be create by ParamAttr.
J
jerrywgz 已提交
9767
        name(str|None): A name for this layer(optional). If set None, the layer
T
Tink_Y 已提交
9768
          will be named automatically.
J
jerrywgz 已提交
9769 9770 9771 9772 9773 9774 9775 9776

    Returns:
        Variable: The output tensor with the same shape as input.

    Examples:

        .. code-block:: python

J
jerrywgz 已提交
9777 9778 9779
            import paddle.fluid as fluid
            from paddle.fluid.param_attr import ParamAttr
            x = fluid.layers.data(name="x", shape=[5,10,10], dtype="float32")
J
jerrywgz 已提交
9780
            mode = 'channel'
J
jerrywgz 已提交
9781 9782 9783
            output = fluid.layers.prelu(
                     x,mode,param_attr=ParamAttr(name='alpha'))

J
jerrywgz 已提交
9784 9785 9786 9787 9788 9789 9790 9791 9792 9793 9794
    """
    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':
        alpha_shape = x.shape
    dtype = helper.input_dtype(input_param_name='x')
    alpha = helper.create_parameter(
Q
Qiao Longfei 已提交
9795
        attr=helper.param_attr,
J
jerrywgz 已提交
9796 9797 9798 9799
        shape=alpha_shape,
        dtype='float32',
        is_bias=False,
        default_initializer=Constant(1.0))
X
Xin Pan 已提交
9800
    out = helper.create_variable_for_type_inference(dtype)
J
jerrywgz 已提交
9801 9802 9803 9804 9805 9806 9807 9808 9809
    helper.append_op(
        type="prelu",
        inputs={"X": x,
                'Alpha': alpha},
        attrs={"mode": mode},
        outputs={"Out": out})
    return out


9810 9811 9812 9813 9814 9815 9816 9817 9818 9819
@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): ${t_max_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
9820
    Returns:
9821
        output(${out_type}): ${out_comment}
9822 9823 9824

    Examples:

9825
    .. code-block:: python
9826

9827
            import paddle.fluid as fluid
H
haowang101779990 已提交
9828 9829
            x = fluid.layers.data(name="x", shape=[2,3,16,16], dtype="float32")
            y = fluid.layers.brelu(x, t_min=1.0, t_max=20.0)
9830 9831
    """
    helper = LayerHelper('brelu', **locals())
X
Xin Pan 已提交
9832
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
9833 9834 9835 9836 9837 9838 9839 9840 9841 9842 9843 9844 9845 9846 9847 9848 9849 9850
    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): ${alpha_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
9851
    Returns:
9852
        output(${out_type}): ${out_comment}
9853 9854 9855 9856 9857

    Examples:

        .. code-block:: python

9858
            import paddle.fluid as fluid
H
haowang101779990 已提交
9859 9860
            x = fluid.layers.data(name="x", shape=[2,3,16,16], dtype="float32")
            y = fluid.layers.leaky_relu(x, alpha=0.01)
9861 9862
    """
    helper = LayerHelper('leaky_relu', **locals())
X
Xin Pan 已提交
9863
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
9864 9865 9866 9867 9868 9869 9870 9871 9872 9873 9874 9875 9876 9877 9878 9879 9880
    helper.append_op(
        type='leaky_relu',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'alpha': alpha})
    return out


@templatedoc()
def soft_relu(x, threshold=40.0, name=None):
    """
    ${comment}
    Args:
        x(${x_type}): ${x_comment}
        threshold(${threshold_type}|40.0): ${threshold_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
9881
    Returns:
9882
        output(${out_type}): ${out_comment}
9883 9884 9885

    Examples:

9886 9887 9888 9889 9890
        .. code-block:: python 
 
            import paddle.fluid as fluid
   
            x = fluid.layers.data(name="x", shape=[3,16,16], dtype="float32")
H
haowang101779990 已提交
9891
            y = fluid.layers.soft_relu(x, threshold=20.0)
9892 9893
    """
    helper = LayerHelper('soft_relu', **locals())
X
Xin Pan 已提交
9894
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
9895 9896 9897 9898 9899 9900 9901 9902
    helper.append_op(
        type='soft_relu',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'threshold': threshold})
    return out


9903 9904 9905 9906
def flatten(x, axis=1, name=None):
    """
    **Flatten layer**
    Flattens the input tensor into a 2D matrix.
M
minqiyang 已提交
9907

H
haowang101779990 已提交
9908
    For Example:
M
minqiyang 已提交
9909

H
haowang101779990 已提交
9910
    .. code-block:: text
9911

H
haowang101779990 已提交
9912 9913 9914 9915 9916 9917 9918 9919 9920 9921 9922 9923 9924 9925 9926 9927 9928 9929 9930 9931 9932
        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)
9933 9934 9935

    Args:
        x (Variable): A tensor of rank >= axis.
9936 9937
        axis (int): Indicate up to which input dimensions (exclusive) should
                    be flattened to the outer dimension of the output.
9938 9939 9940 9941 9942 9943 9944 9945
                    The value for axis must be in the range [0, R], where R
                    is the rank of the input tensor. When axis = 0, the shape
                    of the output tensor is (1, (d_0 X d_1 ... d_n), where the
                    shape of the input tensor is (d_0, d_1, ... d_n).
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
H
haowang101779990 已提交
9946 9947 9948
        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 \
9949 9950 9951 9952
                  inner dimension of the output.

    Raises:
        ValueError: If x is not a variable.
9953
        ValueError: If axis is not in range [0, rank(x)].
9954 9955 9956 9957 9958

    Examples:

        .. code-block:: python

9959
            import paddle.fluid as fluid
9960 9961 9962 9963 9964 9965 9966 9967 9968 9969 9970
            x = fluid.layers.data(name="x", shape=[4, 4, 3], dtype="float32")
            out = fluid.layers.flatten(x=x, axis=2)
    """
    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
9971 9972
    out = helper.create_variable_for_type_inference(x.dtype)
    x_shape = helper.create_variable_for_type_inference(x.dtype)
9973
    helper.append_op(
9974
        type='flatten2',
9975
        inputs={"X": x},
9976 9977
        outputs={'Out': out,
                 'XShape': x_shape},
9978 9979
        attrs={"axis": axis})
    return out
X
Xin Pan 已提交
9980 9981


C
chenweihang 已提交
9982
def sequence_enumerate(input, win_size, pad_value=0, name=None):
C
chenweihang 已提交
9983
    """
C
chenweihang 已提交
9984
    Generate a new sequence for the input index sequence, which enumerates all the
M
minqiyang 已提交
9985
    sub-sequences with length `win_size` of the input.
C
chenweihang 已提交
9986 9987
    The enumerated sequence has the same 1st dimension with variable `input`, and
    the 2nd dimension is `win_size`, padded by `pad_value` if necessary in generation.
M
minqiyang 已提交
9988

H
haowang101779990 已提交
9989 9990 9991 9992 9993 9994 9995 9996 9997 9998 9999 10000 10001 10002 10003 10004 10005
    .. code-block:: text

        Case 1:

          Input:
            X.lod = [[0, 3, 5]]
            X.data = [[1], [2], [3], [4], [5]]
            X.dims = [5, 1]

          Attrs:
            win_size = 2
            pad_value = 0

          Output:
            Out.lod = [[0, 3, 5]]
            Out.data = [[1, 2], [2, 3], [3, 0], [4, 5], [5, 0]]
            Out.dims = [5, 2]
C
chenweihang 已提交
10006 10007

    Args:
C
chenweihang 已提交
10008 10009 10010
        input (Variable): The input variable which is a index sequence.
        win_size (int): The window size for enumerating all sub-sequences.
        pad_value (int): The padding value, default 0.
C
chenweihang 已提交
10011 10012 10013 10014 10015 10016 10017

    Returns:
        Variable: The enumerate sequence variable which is a LoDTensor.

    Examples:
        .. code-block:: python

10018 10019 10020
            import paddle.fluid as fluid

            x = fluid.layers.data(name='x', shape=[-1, 1], dtype='int32', lod_level=1)
C
chenweihang 已提交
10021 10022
            out = fluid.layers.sequence_enumerate(input=x, win_size=3, pad_value=0)
    """
L
lujun 已提交
10023
    assert not in_dygraph_mode(), (
10024
        "sequence layer is not supported in dygraph mode yet.")
C
chenweihang 已提交
10025
    helper = LayerHelper('sequence_enumerate', **locals())
X
Xin Pan 已提交
10026 10027
    out = helper.create_variable_for_type_inference(
        helper.input_dtype(), stop_gradient=True)
C
chenweihang 已提交
10028 10029 10030 10031 10032 10033
    helper.append_op(
        type='sequence_enumerate',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={'win_size': win_size,
               'pad_value': pad_value})
M
minqiyang 已提交
10034
    return out
10035

10036

S
sneaxiy 已提交
10037 10038 10039 10040 10041 10042 10043 10044 10045
def sequence_mask(x, maxlen=None, dtype='int64', name=None):
    """
    **SequenceMask Layer**

    This layer outputs a mask according to the input :code:`x` and
    :code:`maxlen` with data type of :code:`dtype`.

    Supposing :code:`x` is a Tensor with shape [d_1, d_2, ..., d_n], the
    :code:`y` is a mask with shape [d_1, d_2, ..., d_n, maxlen], where:
10046

S
sneaxiy 已提交
10047
    .. math::
10048

S
sneaxiy 已提交
10049 10050 10051
        y(i_1, i_2,..., i_n, j) = (j < x(i_1, i_2,..., i_n))

    Args:
10052
        x (Variable): Input tensor of sequence_mask layer,
S
sneaxiy 已提交
10053 10054 10055 10056
                      whose elements are integers less than :code:`maxlen`.
        maxlen (int|None): Maximum length of the sequence. If :code:`maxlen`
                           is None, it would be replace with :math:`max(x)`.
        dtype (np.dtype|core.VarDesc.VarType|str): Data type of the output.
10057 10058 10059
        name (str|None): A name for this layer(optional). If set None, the
                         layer will be named automatically.

S
sneaxiy 已提交
10060 10061
    Returns:
        Variable: The output sequence mask.
10062

10063 10064 10065
    Examples:
        .. code-block:: python
	
10066
            import paddle.fluid as fluid
10067 10068 10069 10070 10071
            import paddle.fluid.layers as layers

            x = fluid.layers.data(name='x', shape=[10], dtype='float32', lod_level=1)
            mask = layers.sequence_mask(x=x)

S
sneaxiy 已提交
10072
    """
Q
qingqing01 已提交
10073
    helper = LayerHelper('sequence_mask', **locals())
S
sneaxiy 已提交
10074
    if name is None:
X
Xin Pan 已提交
10075
        out = helper.create_variable_for_type_inference(dtype=dtype)
S
sneaxiy 已提交
10076
    else:
X
Xin Pan 已提交
10077
        out = helper.create_variable_for_type_inference(dtype=dtype, name=name)
S
sneaxiy 已提交
10078

10079 10080 10081 10082 10083 10084 10085 10086
    inputs = {'X': [x]}
    attrs = {'out_dtype': out.dtype}
    if maxlen is not None:
        if isinstance(maxlen, Variable):
            inputs['MaxLenTensor'] = maxlen
        else:
            attrs['maxlen'] = maxlen

Q
qingqing01 已提交
10087
    helper.append_op(
10088 10089 10090
        type='sequence_mask', inputs=inputs, outputs={'Y': out}, attrs=attrs)

    out.stop_gradient = True
S
sneaxiy 已提交
10091
    return out
S
sneaxiy 已提交
10092 10093


X
Xin Pan 已提交
10094
def stack(x, axis=0):
S
sneaxiy 已提交
10095 10096 10097 10098
    """
    **Stack Layer**

    This layer stacks all of the input :code:`x` along axis.
10099 10100 10101 10102 10103 10104 10105

    Input :code:`x` can be a single variable, a :code:`list` of variables,
    or a :code:`tuple` of variables. If :code:`x` is a :code:`list` or
    :code:`tuple`, the shapes of all these variables must be the same.
    Supposing the shape of each input is :math:`[d_0, d_1, ..., d_{n-1}]`,
    the shape of the output variable would be
    :math:`[d_0, d_1, ..., d_{axis}=len(x), ..., d_{n-1}]`.
S
sneaxiy 已提交
10106
    If :code:`axis` < 0, it would be replaced with :code:`axis+rank(x[0])+1`.
10107
    If :code:`axis` is None, it would be replaced with 0.
S
sneaxiy 已提交
10108

C
chengduozh 已提交
10109 10110
    For Example:

C
chengduozh 已提交
10111 10112 10113 10114 10115 10116 10117 10118 10119 10120 10121 10122 10123 10124 10125 10126 10127 10128 10129 10130 10131 10132 10133 10134 10135 10136 10137 10138 10139 10140 10141 10142 10143 10144 10145 10146 10147 10148
    .. code-block:: text

        Case 1:
          Input:
            x[0].data = [ [1.0 , 2.0 ] ]
            x[0].dims = [1, 2]
            x[1].data = [ [3.0 , 4.0 ] ]
            x[1].dims = [1, 2]
            x[2].data = [ [5.0 , 6.0 ] ]
            x[2].dims = [1, 2]

          Attrs:
            axis = 0

          Output:
            Out.data =[ [ [1.0, 2.0] ],
                        [ [3.0, 4.0] ],
                        [ [5.0, 6.0] ] ]
            Out.dims = [3, 1, 2]

        Case 2:
          Given
            x[0].data = [ [1.0 , 2.0 ] ]
            x[0].dims = [1, 2]
            x[1].data = [ [3.0 , 4.0 ] ]
            x[1].dims = [1, 2]
            x[2].data = [ [5.0 , 6.0 ] ]
            x[2].dims = [1, 2]

          Attrs:
            axis = 1 or axis = -2

          Output:
            Out.data =[ [ [1.0, 2.0]
                          [3.0, 4.0]
                          [5.0, 6.0] ] ]
            Out.dims = [1, 3, 2]

S
sneaxiy 已提交
10149
    Args:
10150
        x (Variable|list(Variable)|tuple(Variable)): Input variables.
S
sneaxiy 已提交
10151
        axis (int|None): The axis along which all inputs are stacked.
10152

S
sneaxiy 已提交
10153 10154
    Returns:
        Variable: The stacked variable.
10155

10156 10157 10158
    Examples:
        .. code-block:: python

10159
            import paddle.fluid as fluid
10160
            import paddle.fluid.layers as layers
10161 10162
            x1 = layers.data(name='x1', shape=[1, 2], dtype='int32')
            x2 = layers.data(name='x2', shape=[1, 2], dtype='int32')
10163 10164
            data = layers.stack([x1,x2])

S
sneaxiy 已提交
10165 10166
    """

X
Xin Pan 已提交
10167 10168 10169 10170 10171 10172
    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
10173
    out = helper.create_variable_for_type_inference(x[0].dtype)
X
Xin Pan 已提交
10174
    helper.append_op(
S
sneaxiy 已提交
10175 10176
        type='stack', inputs={'X': x}, outputs={'Y': out},
        attrs={'axis': axis})
10177

X
Xin Pan 已提交
10178
    return out
D
dzhwinter 已提交
10179 10180


J
Jiawei Wang 已提交
10181 10182 10183 10184 10185 10186 10187 10188 10189 10190 10191 10192 10193 10194 10195 10196 10197 10198 10199 10200 10201 10202 10203 10204 10205 10206 10207 10208 10209 10210 10211 10212 10213 10214 10215 10216 10217 10218 10219 10220 10221 10222 10223 10224 10225 10226 10227 10228 10229 10230 10231 10232 10233 10234 10235 10236 10237 10238 10239 10240 10241 10242 10243 10244 10245 10246 10247 10248 10249 10250
@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
10251 10252 10253 10254 10255
def unstack(x, axis=0, num=None):
    """
    **UnStack Layer**

    This layer unstacks input :code:`x` into several tensors along axis.
M
minqiyang 已提交
10256

D
dzhwinter 已提交
10257 10258 10259
    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
10260
    raised.
D
dzhwinter 已提交
10261 10262

    Args:
M
minqiyang 已提交
10263
        x (Variable): Input variable.
D
dzhwinter 已提交
10264 10265
        axis (int): The axis along which the input is unstacked.
        num (int|None): The number of output variables.
M
minqiyang 已提交
10266

D
dzhwinter 已提交
10267 10268
    Returns:
        list(Variable): The unstacked variables.
M
minqiyang 已提交
10269

10270 10271 10272 10273 10274 10275
    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
            x = fluid.layers.data(name='x', shape=[5, 10], dtype='float32')
            y = fluid.layers.unstack(x, axis=1)
D
dzhwinter 已提交
10276 10277 10278 10279 10280 10281 10282 10283 10284 10285
    """

    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
10286
    for _ in range(num):
X
Xin Pan 已提交
10287
        outs.append(helper.create_variable_for_type_inference(x.dtype))
D
dzhwinter 已提交
10288 10289 10290 10291 10292 10293 10294 10295

    helper.append_op(
        type='unstack',
        inputs={'X': [x]},
        outputs={'Y': outs},
        attrs={'axis': axis,
               'num': num})
    return outs
W
whs 已提交
10296 10297 10298 10299 10300 10301 10302 10303 10304 10305 10306 10307


def expand(x, expand_times, name=None):
    """Expand operator tiles the input by given times number. You should set times
    number for each dimension by providing attribute 'expand_times'. The rank of X
    should be in [1, 6]. Please note that size of 'expand_times' 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]:
M
minqiyang 已提交
10308

W
whs 已提交
10309 10310 10311 10312
                [
                   [[1], [2], [3]],
                   [[4], [5], [6]]
                ]
M
minqiyang 已提交
10313

W
whs 已提交
10314
        Attr(expand_times):  [1, 2, 2]
M
minqiyang 已提交
10315

W
whs 已提交
10316
        Output(Out) is a 3-D tensor with shape [2, 6, 2]:
M
minqiyang 已提交
10317

W
whs 已提交
10318 10319 10320 10321
                [
                    [[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
10322

W
whs 已提交
10323 10324
    Args:
        x (Variable): A tensor with rank in [1, 6].
L
liym27 已提交
10325
        expand_times (list|tuple|Variable): Expand times number for each dimension.
W
whs 已提交
10326 10327 10328 10329 10330 10331 10332

    Returns:
        Variable: The expanded variable which is a LoDTensor. After expanding, size of each dimension of Output(Out) is equal to ithe size of the corresponding dimension of Input(X) multiplying the corresponding value given by expand_times.


    Examples:
        .. code-block:: python
L
liym27 已提交
10333

W
wangchaochaohu 已提交
10334
            import paddle.fluid as fluid
L
liym27 已提交
10335 10336 10337 10338 10339 10340 10341 10342 10343

            # 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])

            # 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)
W
whs 已提交
10344
    """
L
liym27 已提交
10345 10346 10347 10348 10349

    if not isinstance(expand_times, (list, tuple, Variable)):
        raise ValueError(
            "Input expand_times must be an Variable, python list or tuple.")

W
whs 已提交
10350
    helper = LayerHelper('expand', input=x, **locals())
L
liym27 已提交
10351 10352 10353 10354 10355 10356 10357 10358 10359 10360 10361 10362 10363 10364 10365 10366 10367 10368 10369 10370 10371 10372 10373 10374 10375 10376 10377 10378 10379 10380 10381 10382
    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
10383 10384 10385 10386 10387

    if in_dygraph_mode():
        inputs = {'X': x}
        attrs = {'expand_times': expand_times}
    else:
L
liym27 已提交
10388 10389 10390 10391 10392 10393 10394 10395
        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)
10396

L
liym27 已提交
10397 10398
    dtype = helper.input_dtype(input_param_name='x')
    out = helper.create_variable_for_type_inference(dtype)
W
whs 已提交
10399
    helper.append_op(
10400
        type='expand', inputs=inputs, outputs={'Out': out}, attrs=attrs)
W
whs 已提交
10401
    return out
S
sneaxiy 已提交
10402 10403


G
fix  
gongweibao 已提交
10404 10405 10406
from paddle.fluid.framework import convert_np_dtype_to_dtype_


G
gongweibao 已提交
10407
@templatedoc()
G
fix  
gongweibao 已提交
10408 10409 10410 10411 10412 10413 10414 10415 10416
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):
    """
G
gongweibao 已提交
10417
    ${comment}
G
fix  
gongweibao 已提交
10418 10419

    Args:
G
gongweibao 已提交
10420 10421 10422
        input (Variable): ${input_comment}
        shape (tuple|list): ${shape_comment}
        input_dim_idx (Int): ${input_dim_idx_comment}
G
gongweibao 已提交
10423
        output_dim_idx (Int): ${output_dim_idx_comment}
G
gongweibao 已提交
10424 10425 10426
        min (Float): ${min_comment}
        max (Float): ${max_comment}
        seed (Int): ${seed_comment}
G
fix  
gongweibao 已提交
10427 10428
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc
    Returns:
G
gongweibao 已提交
10429
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
10430

10431 10432 10433
    Examples:
        .. code-block:: python

10434
            import paddle.fluid as fluid
10435 10436
            import paddle.fluid.layers as layers 

10437 10438
            input = layers.data(name="input", shape=[13, 11], dtype='float32')
            out = layers.uniform_random_batch_size_like(input, [-1, 11])
G
fix  
gongweibao 已提交
10439 10440 10441
    """

    helper = LayerHelper('uniform_random_batch_size_like', **locals())
X
Xin Pan 已提交
10442
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
10443 10444 10445 10446 10447 10448 10449 10450 10451 10452 10453 10454 10455 10456 10457 10458
    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
10459 10460


G
gongweibao 已提交
10461
@templatedoc()
X
Xin Pan 已提交
10462
def gaussian_random(shape, mean=0.0, std=1.0, seed=0, dtype='float32'):
G
fix  
gongweibao 已提交
10463
    """
G
gongweibao 已提交
10464
    ${comment}
G
fix  
gongweibao 已提交
10465 10466

    Args:
G
gongweibao 已提交
10467 10468 10469 10470
        shape (tuple|list): ${shape_comment}
        mean (Float): ${mean_comment}
        std (Float): ${std_comment}
        seed (Int): ${seed_comment}
G
fix  
gongweibao 已提交
10471 10472 10473
        dtype(np.dtype|core.VarDesc.VarType|str): Output data type.

    Returns:
G
gongweibao 已提交
10474
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
10475

10476 10477 10478
    Examples:
        .. code-block:: python

10479
            import paddle.fluid as fluid
J
JesseyXujin 已提交
10480
            import paddle.fluid.layers as layers
10481
            out = layers.gaussian_random(shape=[20, 30])
G
fix  
gongweibao 已提交
10482 10483 10484
    """

    helper = LayerHelper('gaussian_random', **locals())
X
Xin Pan 已提交
10485
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
10486 10487 10488 10489 10490 10491 10492 10493 10494 10495
    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
10496
            'use_mkldnn': False
G
fix  
gongweibao 已提交
10497 10498 10499 10500 10501
        })

    return out


G
gongweibao 已提交
10502
@templatedoc()
G
fix  
gongweibao 已提交
10503
def sampling_id(x, min=0.0, max=1.0, seed=0, dtype='float32'):
G
fix  
gongweibao 已提交
10504
    """
G
gongweibao 已提交
10505
    ${comment}
G
fix  
gongweibao 已提交
10506 10507

    Args:
G
gongweibao 已提交
10508 10509 10510 10511
        x (Variable): ${x_comment}
        min (Float): ${min_comment}
        max (Float): ${max_comment}
        seed (Float): ${seed_comment}
G
fix  
gongweibao 已提交
10512
        dtype(np.dtype|core.VarDesc.VarType|str): The type of output data : float32, float_16, int etc
G
fix  
gongweibao 已提交
10513 10514

    Returns:
G
gongweibao 已提交
10515
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
10516

10517 10518 10519
    Examples:
        .. code-block:: python

10520
            import paddle.fluid as fluid
Y
Yibing Liu 已提交
10521
            x = fluid.layers.data(
10522 10523 10524 10525 10526
                name="X",
                shape=[13, 11],
                dtype='float32',
                append_batch_size=False)

Y
Yibing Liu 已提交
10527
            out = fluid.layers.sampling_id(x)
G
fix  
gongweibao 已提交
10528 10529 10530
    """

    helper = LayerHelper('sampling_id', **locals())
X
Xin Pan 已提交
10531
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
10532 10533 10534 10535 10536 10537 10538 10539 10540 10541 10542
    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
10543
@templatedoc()
G
fix  
gongweibao 已提交
10544 10545 10546 10547 10548 10549 10550 10551 10552
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
10553
    ${comment}
G
fix  
gongweibao 已提交
10554 10555

    Args:
G
gongweibao 已提交
10556 10557
        input (Variable): ${input_comment}
        shape (tuple|list): ${shape_comment}
G
gongweibao 已提交
10558
        input_dim_idx (Int): ${input_dim_idx_comment}
G
gongweibao 已提交
10559 10560 10561 10562
        output_dim_idx (Int): ${output_dim_idx_comment}
        mean (Float): ${mean_comment}
        std (Float): ${std_comment}
        seed (Int): ${seed_comment}
G
fix  
gongweibao 已提交
10563
        dtype(np.dtype|core.VarDesc.VarType|str): The type of output data : float32, float_16, int etc
G
fix  
gongweibao 已提交
10564 10565

    Returns:
G
gongweibao 已提交
10566
        out (Variable): ${out_comment}
10567 10568 10569 10570

    Examples:
        .. code-block:: python

10571
            import paddle.fluid as fluid
Y
Yibing Liu 已提交
10572
            input = fluid.layers.data(name="input", shape=[13, 11], dtype='float32')
10573

Y
Yibing Liu 已提交
10574
            out = fluid.layers.gaussian_random_batch_size_like(
10575
                input, shape=[-1, 11], mean=1.0, std=2.0)
G
fix  
gongweibao 已提交
10576 10577 10578
    """

    helper = LayerHelper('gaussian_random_batch_size_like', **locals())
X
Xin Pan 已提交
10579
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
10580 10581 10582 10583 10584 10585 10586 10587 10588 10589 10590 10591 10592 10593 10594 10595 10596 10597
    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
10598
@templatedoc()
X
Xin Pan 已提交
10599
def sum(x):
G
fix  
gongweibao 已提交
10600
    """
G
gongweibao 已提交
10601
    ${comment}
G
fix  
gongweibao 已提交
10602 10603

    Args:
G
gongweibao 已提交
10604
        x (Variable): ${x_comment}
G
fix  
gongweibao 已提交
10605 10606

    Returns:
G
gongweibao 已提交
10607
        out (Variable): ${out_comment}
10608 10609 10610 10611

    Examples:
        .. code-block:: python

10612
            import paddle.fluid as fluid
10613 10614 10615 10616
            import paddle.fluid.layers as layers
            input0 = layers.data(name="input0", shape=[13, 11], dtype='float32')
            input1 = layers.data(name="input1", shape=[13, 11], dtype='float32')
            out = layers.sum([input0,input1])
G
fix  
gongweibao 已提交
10617 10618 10619
    """

    helper = LayerHelper('sum', **locals())
X
Xin Pan 已提交
10620 10621
    out = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype('x'))
G
fix  
gongweibao 已提交
10622 10623 10624 10625
    helper.append_op(
        type='sum',
        inputs={'X': x},
        outputs={'Out': out},
X
Xin Pan 已提交
10626
        attrs={'use_mkldnn': False})
G
fix  
gongweibao 已提交
10627 10628 10629 10630

    return out


G
gongweibao 已提交
10631
@templatedoc()
G
fix  
gongweibao 已提交
10632 10633
def slice(input, axes, starts, ends):
    """
10634 10635 10636 10637 10638 10639 10640 10641 10642 10643 10644 10645 10646 10647 10648
    Slice Operator.

    Produces a slice of the input tensor 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, it uses this information
    to slice the input data tensor. If a negative value is passed for any of
    the start or end indices, it represents number of elements before the end
    of that dimension. If the value passed to start or end is larger than
    the 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\' and ends\'.
    Following examples will explain how slice works:

    .. code-block:: text
G
fix  
gongweibao 已提交
10649

10650 10651 10652 10653 10654 10655 10656 10657 10658 10659 10660 10661 10662 10663 10664 10665 10666
        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], ]
        
        Case2:
            Given:
                data = [ [1, 2, 3, 4], [5, 6, 7, 8], ]
                axes = [0, 1]
                starts = [0, 1]
                ends = [-1, 1000]
            Then:
                result = [ [2, 3, 4], ]
G
fix  
gongweibao 已提交
10667
    Args:
G
gongweibao 已提交
10668 10669
        input (Variable): ${input_comment}.
        axes (List): ${axes_comment}
10670 10671
        starts (List|Variable): ${starts_comment}
        ends (List|Variable): ${ends_comment}
G
fix  
gongweibao 已提交
10672 10673

    Returns:
G
gongweibao 已提交
10674
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
10675

10676 10677 10678
    Examples:
        .. code-block:: python

10679
            import paddle.fluid as fluid
10680

10681
            input = fluid.layers.data(
10682 10683
                name="input", shape=[3, 4, 5, 6], dtype='float32')

10684 10685 10686 10687 10688 10689 10690 10691 10692 10693 10694
            # 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)

            # 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)
G
fix  
gongweibao 已提交
10695 10696
    """

10697 10698 10699 10700 10701 10702 10703
    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
10704
    helper = LayerHelper('slice', **locals())
10705 10706 10707 10708 10709 10710 10711 10712 10713 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 10739 10740 10741 10742 10743 10744 10745 10746 10747 10748 10749 10750 10751 10752 10753 10754 10755 10756 10757 10758 10759 10760 10761 10762 10763 10764 10765 10766 10767 10768 10769 10770 10771 10772 10773 10774

    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
10775 10776
    out = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype('input'))
G
fix  
gongweibao 已提交
10777
    helper.append_op(
10778
        type='slice', inputs=inputs, attrs=attrs, outputs={'Out': out})
G
fix  
gongweibao 已提交
10779 10780 10781 10782 10783 10784

    return out


def shape(input):
    """
C
chengduozh 已提交
10785 10786
    **Shape Layer**

C
fix doc  
chengduozh 已提交
10787
    Get the shape of the input.
G
fix  
gongweibao 已提交
10788 10789

    Args:
C
chengduozh 已提交
10790
        input (Variable): The input variable.
G
fix  
gongweibao 已提交
10791 10792

    Returns:
C
fix doc  
chengduozh 已提交
10793
        Variable: The shape of the input variable.
G
fix  
gongweibao 已提交
10794

10795 10796 10797
    Examples:
        .. code-block:: python

10798 10799 10800
            import paddle.fluid as fluid

            input = fluid.layers.data(
10801
                name="input", shape=[3, 100, 100], dtype="float32")
10802
            out = fluid.layers.shape(input)
G
fix  
gongweibao 已提交
10803 10804 10805
    """

    helper = LayerHelper('shape', **locals())
10806
    out = helper.create_variable_for_type_inference(dtype='int32')
G
fix  
gongweibao 已提交
10807
    helper.append_op(
G
fix  
gongweibao 已提交
10808
        type='shape', inputs={'Input': input}, outputs={'Out': out})
G
fix  
gongweibao 已提交
10809 10810

    return out
G
merge  
gongweibao 已提交
10811 10812


Z
zhoukunsheng 已提交
10813 10814 10815 10816
def rank(input):
    """
    **Rank Layer**

Z
zhoukunsheng 已提交
10817
    Returns the number of dimensions for a tensor, which is a 0-D int32 Tensor.
Z
zhoukunsheng 已提交
10818 10819 10820 10821 10822 10823 10824 10825 10826 10827

    Args:
        input (Variable): The input variable.

    Returns:
        Variable: The rank of the input variable.

    Examples:
        .. code-block:: python

10828 10829 10830 10831
            import paddle.fluid as fluid

            input = fluid.layers.data(name="input", shape=[3, 100, 100], dtype="float32")
            rank = fluid.layers.rank(input) # 4
Z
zhoukunsheng 已提交
10832 10833 10834 10835 10836 10837 10838 10839
    """

    ndims = len(input.shape)
    out = assign(np.array(ndims, 'int32'))

    return out


Z
zhoukunsheng 已提交
10840 10841 10842 10843 10844 10845 10846 10847 10848 10849 10850 10851 10852 10853 10854 10855 10856 10857 10858 10859 10860 10861 10862 10863 10864 10865 10866 10867 10868
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
10869 10870 10871 10872
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
10873
    if in_dygraph_mode():
X
Xin Pan 已提交
10874 10875 10876
        x = base.to_variable(x)
        y = base.to_variable(y)

S
sneaxiy 已提交
10877 10878 10879 10880
    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)
    axis = helper.kwargs.get('axis', -1)
    use_mkldnn = helper.kwargs.get('use_mkldnn', False)
S
sneaxiy 已提交
10881 10882
    name = helper.kwargs.get('name', None)
    if name is None:
X
Xin Pan 已提交
10883
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
S
sneaxiy 已提交
10884 10885 10886
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)
S
sneaxiy 已提交
10887

S
sneaxiy 已提交
10888 10889 10890 10891 10892 10893 10894 10895 10896 10897 10898
    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)


@templatedoc()
S
sneaxiy 已提交
10899
def scale(x, scale=1.0, bias=0.0, bias_after_scale=True, act=None, name=None):
S
sneaxiy 已提交
10900 10901 10902 10903 10904 10905 10906 10907
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        scale(${scale_type}): ${scale_comment}
        bias(${bias_type}): ${bias_comment}
        bias_after_scale(${bias_after_scale_type}): ${bias_after_scale_comment}
S
sneaxiy 已提交
10908
        act(basestring|None): Activation applied to the output.
M
minqiyang 已提交
10909
        name(basestring|None): Name of the output.
S
sneaxiy 已提交
10910 10911 10912

    Returns:
        out(${out_type}): ${out_comment}
10913 10914 10915 10916 10917 10918 10919 10920

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid

            x = fluid.layers.data(name="X", shape=[1, 2, 5, 5], dtype='float32')
            y = fluid.layers.scale(x, scale = 2.0, bias = 1.0)
S
sneaxiy 已提交
10921 10922 10923
    """

    helper = LayerHelper('scale', **locals())
S
sneaxiy 已提交
10924
    if name is None:
X
Xin Pan 已提交
10925
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
S
sneaxiy 已提交
10926 10927 10928
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)
S
sneaxiy 已提交
10929 10930 10931 10932 10933 10934 10935 10936 10937 10938

    helper.append_op(
        type='scale',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={
            'scale': float(scale),
            'bias': float(bias),
            'bias_after_scale': bias_after_scale
        })
S
sneaxiy 已提交
10939
    return helper.append_activation(out)
S
sneaxiy 已提交
10940 10941


X
Xin Pan 已提交
10942
def elementwise_add(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
10943 10944 10945
    return _elementwise_op(LayerHelper('elementwise_add', **locals()))


X
Xin Pan 已提交
10946
def elementwise_div(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
10947 10948 10949
    return _elementwise_op(LayerHelper('elementwise_div', **locals()))


X
Xin Pan 已提交
10950
def elementwise_sub(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
10951 10952 10953
    return _elementwise_op(LayerHelper('elementwise_sub', **locals()))


X
Xin Pan 已提交
10954
def elementwise_mul(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
10955 10956 10957
    return _elementwise_op(LayerHelper('elementwise_mul', **locals()))


X
Xin Pan 已提交
10958
def elementwise_max(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
10959 10960 10961
    return _elementwise_op(LayerHelper('elementwise_max', **locals()))


X
Xin Pan 已提交
10962
def elementwise_min(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
10963 10964 10965
    return _elementwise_op(LayerHelper('elementwise_min', **locals()))


X
Xin Pan 已提交
10966
def elementwise_pow(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
10967 10968 10969
    return _elementwise_op(LayerHelper('elementwise_pow', **locals()))


10970 10971 10972 10973 10974 10975 10976 10977
def elementwise_mod(x, y, axis=-1, act=None, name=None):
    return _elementwise_op(LayerHelper('elementwise_mod', **locals()))


def elementwise_floordiv(x, y, axis=-1, act=None, name=None):
    return _elementwise_op(LayerHelper('elementwise_floordiv', **locals()))


S
sneaxiy 已提交
10978
for func in [
10979 10980 10981 10982 10983 10984 10985 10986 10987
        elementwise_add,
        elementwise_div,
        elementwise_sub,
        elementwise_mul,
        elementwise_max,
        elementwise_min,
        elementwise_pow,
        elementwise_mod,
        elementwise_floordiv,
S
sneaxiy 已提交
10988 10989 10990 10991 10992
]:
    op_proto = OpProtoHolder.instance().get_op_proto(func.__name__)
    func.__doc__ = _generate_doc_string_(
        op_proto,
        additional_args_lines=[
S
sneaxiy 已提交
10993 10994
            "act (basestring|None): Activation applied to the output.",
            "name (basestring|None): Name of the output."
S
sneaxiy 已提交
10995
        ])
10996 10997 10998 10999 11000 11001 11002 11003 11004 11005 11006 11007 11008 11009 11010 11011 11012 11013 11014 11015 11016 11017 11018 11019 11020 11021 11022 11023 11024 11025 11026 11027 11028 11029 11030 11031 11032
    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
11033 11034


11035
def _logical_op(op_name, x, y, out=None, name=None, binary_op=True):
M
minqiyang 已提交
11036 11037
    helper = LayerHelper(op_name, **locals())

M
minqiyang 已提交
11038 11039
    if binary_op:
        assert x.dtype == y.dtype
M
minqiyang 已提交
11040 11041 11042

    if out is None:
        if name is None:
X
Xin Pan 已提交
11043
            out = helper.create_variable_for_type_inference(dtype=x.dtype)
M
minqiyang 已提交
11044 11045 11046 11047 11048 11049 11050 11051 11052 11053 11054 11055 11056 11057 11058
        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()
11059
def logical_and(x, y, out=None, name=None):
M
minqiyang 已提交
11060 11061 11062 11063 11064 11065 11066 11067 11068 11069 11070
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        y(${y_type}): ${y_comment}
        out(Tensor): Output tensor of logical operation.
        name(basestring|None): Name of the output.

    Returns:
        out(${out_type}): ${out_comment}
11071 11072 11073 11074

    Examples:
        .. code-block:: python

11075
            import paddle.fluid as fluid
11076
            left = fluid.layers.data(
石晓伟 已提交
11077
                name='left', shape=[1], dtype='bool')
11078
            right = fluid.layers.data(
石晓伟 已提交
11079
                name='right', shape=[1], dtype='bool')
11080
            result = fluid.layers.logical_and(x=left, y=right)
M
minqiyang 已提交
11081 11082 11083 11084 11085 11086 11087
    """

    return _logical_op(
        op_name="logical_and", x=x, y=y, name=name, out=out, binary_op=True)


@templatedoc()
11088
def logical_or(x, y, out=None, name=None):
M
minqiyang 已提交
11089 11090 11091 11092 11093 11094 11095 11096 11097 11098 11099
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        y(${y_type}): ${y_comment}
        out(Tensor): Output tensor of logical operation.
        name(basestring|None): Name of the output.

    Returns:
        out(${out_type}): ${out_comment}
11100 11101 11102 11103

    Examples:
        .. code-block:: python

11104
            import paddle.fluid as fluid
11105
            left = fluid.layers.data(
石晓伟 已提交
11106
                name='left', shape=[1], dtype='bool')
11107
            right = fluid.layers.data(
石晓伟 已提交
11108
                name='right', shape=[1], dtype='bool')
11109
            result = fluid.layers.logical_or(x=left, y=right)
M
minqiyang 已提交
11110 11111 11112 11113 11114 11115 11116
    """

    return _logical_op(
        op_name="logical_or", x=x, y=y, name=name, out=out, binary_op=True)


@templatedoc()
11117
def logical_xor(x, y, out=None, name=None):
M
minqiyang 已提交
11118 11119 11120 11121 11122 11123 11124 11125 11126 11127 11128
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        y(${y_type}): ${y_comment}
        out(Tensor): Output tensor of logical operation.
        name(basestring|None): Name of the output.

    Returns:
        out(${out_type}): ${out_comment}
11129 11130 11131 11132

    Examples:
        .. code-block:: python

11133
            import paddle.fluid as fluid
11134
            left = fluid.layers.data(
石晓伟 已提交
11135
                name='left', shape=[1], dtype='bool')
11136
            right = fluid.layers.data(
石晓伟 已提交
11137
                name='right', shape=[1], dtype='bool')
11138
            result = fluid.layers.logical_xor(x=left, y=right)
M
minqiyang 已提交
11139 11140 11141 11142 11143 11144 11145
    """

    return _logical_op(
        op_name="logical_xor", x=x, y=y, name=name, out=out, binary_op=True)


@templatedoc()
11146
def logical_not(x, out=None, name=None):
M
minqiyang 已提交
11147 11148 11149 11150 11151 11152 11153 11154 11155 11156
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        out(Tensor): Output tensor of logical operation.
        name(basestring|None): Name of the output.

    Returns:
        out(${out_type}): ${out_comment}
11157 11158 11159 11160

    Examples:
        .. code-block:: python

11161
            import paddle.fluid as fluid
11162
            left = fluid.layers.data(
石晓伟 已提交
11163
                name='left', shape=[1], dtype='bool')
11164
            result = fluid.layers.logical_not(x=left)
M
minqiyang 已提交
11165 11166 11167 11168
    """

    return _logical_op(
        op_name="logical_not", x=x, y=None, name=name, out=out, binary_op=False)
11169 11170 11171 11172 11173 11174 11175 11176 11177 11178 11179 11180 11181 11182 11183


@templatedoc()
def clip(x, min, max, name=None):
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        min(${min_type}): ${min_comment}
        max(${max_type}): ${max_comment}
        name(basestring|None): Name of the output.

    Returns:
        out(${out_type}): ${out_comment}
11184 11185 11186 11187

    Examples:
        .. code-block:: python

S
SunGaofeng 已提交
11188
            import paddle.fluid as fluid
11189 11190 11191
            input = fluid.layers.data(
                name='data', shape=[1], dtype='float32')
            reward = fluid.layers.clip(x=input, min=-1.0, max=1.0)
11192 11193 11194 11195 11196
    """

    helper = LayerHelper("clip", **locals())

    if name is None:
11197 11198
        name = unique_name.generate_with_ignorable_key(".".join(
            [helper.name, 'tmp']))
S
sneaxiy 已提交
11199 11200 11201

    out = helper.create_variable(
        type=x.type, name=name, dtype=x.dtype, persistable=False)
11202 11203 11204 11205 11206 11207 11208 11209 11210 11211 11212 11213 11214 11215 11216 11217 11218 11219 11220 11221 11222 11223 11224

    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}
        name(basestring|None): Name of the output.

    Returns:
        out(${out_type}): ${out_comment}
11225 11226 11227 11228

    Examples:
        .. code-block:: python

11229
            import paddle.fluid as fluid
11230 11231 11232
            input = fluid.layers.data(
                name='data', shape=[1], dtype='float32')
            reward = fluid.layers.clip_by_norm(x=input, max_norm=1.0)
11233 11234 11235 11236 11237
    """

    helper = LayerHelper("clip_by_norm", **locals())

    if name is None:
11238 11239
        name = unique_name.generate_with_ignorable_key(".".join(
            [helper.name, 'tmp']))
S
sneaxiy 已提交
11240 11241 11242

    out = helper.create_variable(
        type=x.type, name=name, dtype=x.dtype, persistable=False)
11243 11244 11245 11246 11247 11248 11249 11250

    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
11251 11252 11253 11254 11255 11256 11257 11258 11259 11260 11261 11262 11263


@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}
11264 11265 11266 11267

    Examples:
        .. code-block:: python

11268
            import paddle.fluid as fluid
11269 11270 11271
            input = fluid.layers.data(
                name='data', shape=[2, 3], dtype='float32')
            mean = fluid.layers.mean(input)
X
Xin Pan 已提交
11272 11273 11274 11275 11276
    """

    helper = LayerHelper("mean", **locals())

    if name is None:
X
Xin Pan 已提交
11277
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
11278 11279 11280 11281 11282 11283 11284 11285 11286 11287
    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
11288 11289 11290 11291 11292 11293 11294 11295 11296 11297 11298
@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}
11299 11300 11301 11302

    Examples:
        .. code-block:: python

11303
            import paddle.fluid as fluid
11304 11305 11306 11307 11308
            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
11309 11310 11311 11312 11313 11314 11315 11316 11317 11318 11319 11320
    """

    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
11321 11322 11323 11324 11325 11326 11327 11328 11329 11330 11331 11332 11333 11334
@templatedoc()
def mul(x, y, x_num_col_dims=1, y_num_col_dims=1, name=None):
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        y(${y_type}): ${y_comment}
        x_num_col_dims(${x_num_col_dims_type}): ${x_num_col_dims_comment}
        y_num_col_dims(${y_num_col_dims_type}): ${y_num_col_dims_comment}
        name(basestring|None): Name of the output.

    Returns:
        out(${out_type}): ${out_comment}
11335 11336 11337 11338 11339 11340 11341 11342 11343 11344 11345 11346

    Examples:
        .. code-block:: python
            
            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
11347 11348 11349 11350 11351
    """

    helper = LayerHelper("mul", **locals())

    if name is None:
X
Xin Pan 已提交
11352
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
11353 11354 11355 11356 11357 11358 11359 11360 11361
    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
11362 11363
            "x_num_col_dims": x_num_col_dims,
            "y_num_col_dims": y_num_col_dims
X
Xin Pan 已提交
11364 11365 11366 11367 11368 11369
        },
        outputs={"Out": out})
    return out


@templatedoc()
J
jerrywgz 已提交
11370 11371 11372
def sigmoid_cross_entropy_with_logits(x,
                                      label,
                                      ignore_index=kIgnoreIndex,
11373 11374
                                      name=None,
                                      normalize=False):
X
Xin Pan 已提交
11375 11376 11377 11378 11379 11380
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        label(${label_type}): ${label_comment}
11381
        ignore_index(&{ignore_index}): ${ignore_index_comment}
X
Xin Pan 已提交
11382
        name(basestring|None): Name of the output.
11383 11384
        normalize(bool): If true, divide the output by the number of
            targets != ignore_index.
X
Xin Pan 已提交
11385 11386 11387

    Returns:
        out(${out_type}): ${out_comment}
11388 11389 11390 11391

    Examples:
        .. code-block:: python

11392
            import paddle.fluid as fluid
11393 11394 11395 11396 11397 11398 11399 11400 11401 11402
            input = fluid.layers.data(
                name='data', shape=[10], dtype='float32')
            label = fluid.layers.data(
                name='data', shape=[10], dtype='float32')
            loss = fluid.layers.sigmoid_cross_entropy_with_logits(
                x=input,
                label=label,
                ignore_index=-1,
                normalize=True) # or False
            # loss = fluid.layers.reduce_sum(loss) # summation of loss
X
Xin Pan 已提交
11403 11404 11405 11406 11407
    """

    helper = LayerHelper("sigmoid_cross_entropy_with_logits", **locals())

    if name is None:
X
Xin Pan 已提交
11408
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
11409 11410 11411 11412 11413 11414 11415 11416
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    helper.append_op(
        type="sigmoid_cross_entropy_with_logits",
        inputs={"X": x,
                "Label": label},
11417 11418
        attrs={"ignore_index": ignore_index,
               'normalize': normalize},
X
Xin Pan 已提交
11419 11420 11421 11422 11423 11424 11425 11426 11427 11428 11429 11430 11431 11432 11433 11434
        outputs={"Out": out})
    return out


@templatedoc()
def maxout(x, groups, name=None):
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        groups(${groups_type}): ${groups_comment}
        name(basestring|None): Name of the output.

    Returns:
        out(${out_type}): ${out_comment}
J
jerrywgz 已提交
11435 11436 11437 11438

    Examples:
        .. code-block:: python

11439
            import paddle.fluid as fluid
J
jerrywgz 已提交
11440 11441 11442 11443 11444
            input = fluid.layers.data(
                name='data', 
                shape=[256, 32, 32], 
                dtype='float32')
            out = fluid.layers.maxout(input, groups=2)
X
Xin Pan 已提交
11445 11446 11447 11448
    """
    helper = LayerHelper("maxout", **locals())

    if name is None:
X
Xin Pan 已提交
11449
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
11450 11451 11452 11453 11454 11455 11456 11457 11458 11459
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    helper.append_op(
        type="maxout",
        inputs={"X": x},
        attrs={"groups": groups},
        outputs={"Out": out})
    return out
11460 11461


J
JiabinYang 已提交
11462
def space_to_depth(x, blocksize, name=None):
J
JiabinYang 已提交
11463
    """
J
JiabinYang 已提交
11464
    Gives a blocksize to space_to_depth the input LoDtensor with Layout: [batch, channel, height, width]
11465 11466 11467

    This op rearranges blocks of spatial data, into depth. More specifically, this op outputs a copy of the
    input LoDtensor where values from the height and width dimensions are moved to the channel dimension.
J
JiabinYang 已提交
11468
    The attr blocksize indicates the input block size.
11469 11470

    space_to_depth will reorgnize the elements of input with shape[batch, channel, height, width] according
J
JiabinYang 已提交
11471
    to blocksize to construct output with shape [batch, channel * blocksize * blocksize, height/blocksize, width/blocksize]:
11472 11473

    space_to_depth is used to This operation is useful for resizing the activations between convolutions
J
JiabinYang 已提交
11474
    (but keeping all data)
J
JiabinYang 已提交
11475

J
JiabinYang 已提交
11476
    - Non-overlapping blocks of size block_size x block size are rearranged into depth at each location.
11477
    - The depth of the output tensor is block_size * block_size * input channel
J
JiabinYang 已提交
11478 11479 11480 11481 11482
    - 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


J
JiabinYang 已提交
11483
    Args:
J
JiabinYang 已提交
11484
        x(variable): The input LoDtensor.
J
JiabinYang 已提交
11485
        blocksize(variable): The blocksize to select the element on each feature map should be > 2
J
JiabinYang 已提交
11486 11487

    Returns:
J
JiabinYang 已提交
11488
        Variable: The output LoDtensor.
J
JiabinYang 已提交
11489 11490

    Raises:
J
JiabinYang 已提交
11491
        TypeError: blocksize type must be a long.
J
JiabinYang 已提交
11492 11493 11494

    Examples:
        .. code-block:: python
11495 11496 11497
	
            import paddle.fluid as fluid
            import numpy as np
J
JiabinYang 已提交
11498 11499

            data = fluid.layers.data(
11500
                name='data', shape=[1, 4, 2, 2], dtype='float32', append_batch_size=False)
J
JiabinYang 已提交
11501
            space_to_depthed = fluid.layers.space_to_depth(
J
JiabinYang 已提交
11502
                x=data, blocksize=2)
11503

11504
            exe = fluid.Executor(fluid.CPUPlace())
11505 11506 11507 11508
            data_np = np.arange(0,16).reshape((1,4,2,2)).astype('float32')
            out_main = exe.run(fluid.default_main_program(),
                          feed={'data': data_np},
                          fetch_list=[space_to_depthed])
11509

J
JiabinYang 已提交
11510 11511
    """

J
JiabinYang 已提交
11512
    helper = LayerHelper("space_to_depth", **locals())
J
JiabinYang 已提交
11513

J
JiabinYang 已提交
11514 11515
    if not (isinstance(blocksize, int)):
        raise ValueError("blocksize must be a python Int")
J
JiabinYang 已提交
11516 11517

    if name is None:
J
JiabinYang 已提交
11518 11519
        out = helper.create_variable_for_type_inference(
            dtype=x.dtype)  #fix create
J
JiabinYang 已提交
11520 11521 11522 11523 11524
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    helper.append_op(
J
JiabinYang 已提交
11525
        type="space_to_depth",
J
JiabinYang 已提交
11526
        inputs={"X": x},
J
JiabinYang 已提交
11527
        attrs={"blocksize": blocksize},
J
JiabinYang 已提交
11528
        outputs={"Out": out})
J
JiabinYang 已提交
11529 11530
    return out

J
JiabinYang 已提交
11531

S
sneaxiy 已提交
11532 11533
@templatedoc()
def sequence_reverse(x, name=None):
11534
    """
S
sneaxiy 已提交
11535 11536 11537 11538 11539 11540 11541 11542
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        name(basestring|None): Name of the output.

    Returns:
        out(${y_type}): ${y_comment}
B
bdzhuxiaoning 已提交
11543 11544 11545 11546 11547 11548 11549

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
            x = fluid.layers.data(name='x', shape=[2, 6], dtype='float32')
            x_reversed = fluid.layers.sequence_reverse(x)
S
sneaxiy 已提交
11550
    """
L
lujun 已提交
11551
    assert not in_dygraph_mode(), (
11552
        "sequence layer is not supported in dygraph mode yet.")
S
sneaxiy 已提交
11553 11554
    helper = LayerHelper("sequence_reverse", **locals())
    if name is None:
S
sneaxiy 已提交
11555
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
S
sneaxiy 已提交
11556 11557 11558 11559 11560 11561 11562 11563 11564 11565
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    helper.append_op(
        type="sequence_reverse",
        inputs={"X": x},
        outputs={"Y": out},
        attrs=dict())
    return out
S
sneaxiy 已提交
11566 11567


11568 11569 11570 11571 11572 11573 11574 11575 11576 11577 11578 11579 11580 11581 11582 11583 11584 11585 11586 11587 11588 11589 11590 11591 11592 11593 11594 11595 11596 11597 11598 11599 11600 11601 11602 11603 11604 11605 11606 11607 11608 11609 11610 11611 11612 11613 11614 11615 11616 11617 11618 11619 11620 11621 11622 11623 11624 11625 11626 11627 11628 11629 11630 11631 11632 11633 11634
def sequence_topk_avg_pooling(input, row, col, topks, channel_num):
    """
    The :attr:`topks` is a list with incremental values in this function. For each topk,
    it will average the topk features as an output feature for each channel of every 
    input sequence. Both :attr:`row` and :attr:`col` are LodTensor, which provide height 
    and width information for :attr:`input` tensor. If feature size of input sequence is less 
    than topk, it will padding 0 at the back.

    .. code-block:: text

            If channel_num is 2 and given row LoDTensor and col LoDTensor as follows:
                row.lod = [[5, 4]]
                col.lod = [[6, 7]]

            input is a LoDTensor with input.lod[0][i] = channel_num * row.lod[0][i] * col.lod[0][i] 
                input.lod = [[60, 56]]  # where 60 = channel_num * 5 * 6
                input.dims = [116, 1]   # where 116 = 60 + 56

            If topks is [1, 3, 5], then we get a 1-level LoDTensor:
                out.lod =  [[5, 4]] 	# share Lod info with row LodTensor
                out.dims = [9, 6]   	# where 6 = len(topks) * channel_num

    Args:
        input (Variable): The input should be 2D LodTensor with dims[1] equals 1.
        row (Variable): The row shoud be 1-level LodTensor to provide the height information
                        of the input tensor data.
        col (Variable): The col shoud be 1-level LodTensor to provide the width information
                        of the input tensor data.
        topks (list): A list of incremental value to average the topk feature.
        channel_num (int): The number of input channel.

    Returns:
        Variable: output LodTensor specified by this layer.

    Examples:

        .. code-block:: python

            import numpy as np
            from paddle.fluid import layers

            x_lod_tensor = layers.data(name='x', shape=[1], lod_level=1)
            row_lod_tensor = layers.data(name='row', shape=[6], lod_level=1)
            col_lod_tensor = layers.data(name='col', shape=[6], lod_level=1)
            out = layers.sequence_topk_avg_pooling(input=x_lod_tensor,
                                                   row=row_lod_tensor,
                                                   col=col_lod_tensor,
                                                   topks=[1, 3, 5],
                                                   channel_num=5)
    """
    helper = LayerHelper('sequence_topk_avg_pooling', **locals())
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
    pos = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype(), stop_gradient=True)
    helper.append_op(
        type='sequence_topk_avg_pooling',
        inputs={'X': input,
                'ROW': row,
                'COLUMN': col},
        outputs={'Out': out,
                 'pos': pos},
        attrs={'topks': topks,
               'channel_num': channel_num})

    return out


11635 11636 11637 11638 11639 11640
def affine_channel(x,
                   scale=None,
                   bias=None,
                   data_layout='NCHW',
                   name=None,
                   act=None):
11641 11642 11643 11644 11645
    """
    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.
11646

11647 11648 11649 11650 11651 11652 11653 11654 11655 11656 11657 11658
    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
            is applied in the second dimension.
        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
            the input.
        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.
        data_layout (string, default NCHW): NCHW or NHWC. If input is 2D
            tensor, you can ignore data_layout.
        name (str, default None): The name of this layer.
11659
        act (str, default None): Activation to be applied to the output of this layer.
11660 11661 11662

    Returns:
        out (Variable): A tensor of the same shape and data layout with x.
B
Bai Yifan 已提交
11663 11664 11665 11666 11667 11668 11669 11670 11671 11672 11673 11674 11675 11676

    Examples:
        .. code-block:: python
            
            import paddle.fluid as fluid
            data = fluid.layers.data(name='data', shape=[3, 32, 32],
                                     dtype='float32')
            input_scale = fluid.layers.create_parameter(shape=[3],
                                     dtype="float32")
            input_bias = fluid.layers.create_parameter(shape=[3],
                                     dtype="float32")
            out = fluid.layers.affine_channel(data,scale=input_scale,
                                     bias=input_bias)

11677 11678 11679 11680
    """
    helper = LayerHelper("affine_channel", **locals())

    if name is None:
X
Xin Pan 已提交
11681
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
11682 11683 11684 11685 11686 11687 11688 11689 11690 11691 11692
    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})
11693
    return helper.append_activation(out)
11694 11695


B
barrierye 已提交
11696
def similarity_focus(input, axis, indexes, name=None):
11697
    """
B
barrierye 已提交
11698
    SimilarityFocus Operator
B
barrierye 已提交
11699 11700

    Generate a similarity focus mask with the same shape of input using the following method:
M
minqiyang 已提交
11701

11702 11703 11704
    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
11705
       is (BatchSize, A, B, C), the shape of tensor T is (BatchSize, B, C).
11706 11707 11708 11709 11710 11711 11712
    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
11713
       each index.
B
barrierye 已提交
11714 11715 11716 11717
    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
11718 11719 11720 11721 11722 11723 11724 11725 11726 11727 11728 11729 11730 11731 11732 11733 11734 11735 11736 11737 11738 11739 11740 11741 11742 11743 11744 11745 11746 11747 11748 11749 11750 11751 11752 11753 11754 11755 11756 11757 11758 11759 11760 11761 11762 11763 11764 11765 11766
    .. 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
11767
    Args:
11768
        input(Variable): The input tensor variable(default float). It should
B
barrierye 已提交
11769
            be a 4-D tensor with shape [BatchSize, A, B, C].
B
barrierye 已提交
11770
        axis(int): Indicating the dimension to be selected. It can only be
B
barrierye 已提交
11771
            1, 2 or 3.
B
barrierye 已提交
11772
        indexes(list): Indicating the indexes of the selected dimension.
B
barrierye 已提交
11773 11774

    Returns:
H
haowang101779990 已提交
11775 11776
        Variable: A tensor variable with the same shape and same type \
                  as the input.
11777

B
barrierye 已提交
11778 11779
    Examples:
        .. code-block:: python
H
haowang101779990 已提交
11780

11781
            import paddle.fluid as fluid
B
barrierye 已提交
11782
            data = fluid.layers.data(
Y
Yibing Liu 已提交
11783 11784
                name='data', shape=[-1, 3, 2, 2], dtype='float32')
            fluid.layers.similarity_focus(input=data, axis=1, indexes=[0])
B
barrierye 已提交
11785 11786 11787 11788 11789 11790 11791 11792 11793 11794 11795 11796
    """
    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
11797 11798 11799 11800 11801
    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
11802 11803 11804 11805 11806 11807 11808
    helper.append_op(
        type='similarity_focus',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={"axis": axis,
               "indexes": indexes})
    return out
B
barrierye 已提交
11809 11810


M
minqiyang 已提交
11811 11812
def hash(input, hash_size, num_hash=1, name=None):
    """
M
minqiyang 已提交
11813 11814
    Hash the input to an integer whose value is less than the given hash size.

M
minqiyang 已提交
11815 11816
    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
11817 11818 11819 11820 11821 11822 11823 11824

    A simple example as below:

    .. code-block:: text

        Given:

        # shape [2, 2]
11825
        input.data = 
11826
            [[1, 2],
11827
             [3, 4]]
M
minqiyang 已提交
11828 11829 11830 11831 11832 11833 11834 11835 11836 11837 11838 11839 11840

        hash_size = 10000

        num_hash = 4

        Then:

        Hash op will take all number in input's 2nd dimension as hash algorithm's
        input for each time. Each input will be hashed for 4 times, and get an
        array whose length is 4. Each value in the array ranges from 0 to 9999.

        # shape [2, 4]
        output.data = [
11841 11842
            [[9662, 9217, 1129, 8487],
             [8310, 1327, 1654, 4567]],
M
minqiyang 已提交
11843 11844 11845 11846
        ]

    Args:
        input (Variable): The input variable which is a one-hot word. The
11847
            dimensions of the input variable must be 2. Both Tensor and LoDTensor are supported.
M
minqiyang 已提交
11848 11849
        hash_size (int): The space size for hash algorithm. The output value
            will keep in the range:math:`[0, hash_size - 1]`.
M
minqiyang 已提交
11850
        num_hash (int): The times of hash, default 1.
M
minqiyang 已提交
11851
        name (str, default None): The name of this layer.
M
minqiyang 已提交
11852 11853

    Returns:
11854
       Variable: The hash result variable, which the same variable type as `input`.
M
minqiyang 已提交
11855 11856 11857

    Examples:
       .. code-block:: python
H
haowang101779990 已提交
11858

11859 11860
            import paddle.fluid as fluid

11861 11862 11863 11864
            # titles has shape [batch, 1]
            titles = fluid.layers.data(name='titles', shape=[1], dtype='int32', lod_level=0)
            # hash_r has shape [batch, 2]
            hash_r = fluid.layers.hash(name='hash_x', input=titles, num_hash=2, hash_size=1000)
11865 11866


11867 11868 11869 11870
            # titles has shape [batch, 1] and lod information
            titles = fluid.layers.data(name='titles', shape=[1], dtype='int32', lod_level=1)
            # hash_r has shape [batch, 2] and inherits lod information from titles
            hash_r = fluid.layers.hash(name='hash_x', input=titles, num_hash=2, hash_size=1000)
M
minqiyang 已提交
11871 11872
    """
    helper = LayerHelper('hash', **locals())
M
minqiyang 已提交
11873 11874
    out = helper.create_variable_for_type_inference(
        helper.input_dtype(), stop_gradient=True)
M
minqiyang 已提交
11875 11876 11877 11878 11879 11880 11881
    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
11882 11883


D
dengkaipeng 已提交
11884
@templatedoc()
11885 11886
def grid_sampler(x, grid, name=None):
    """
11887
    This operation samples input X by using bilinear interpolation based on
H
haowang101779990 已提交
11888
    flow field grid, which is usually gennerated by :code:`affine_grid` . The grid of
11889 11890 11891 11892
    shape [N, H, W, 2] is the concatenation of (grid_x, grid_y) coordinates
    with shape [N, H, W] each, where grid_x is indexing the 4th dimension
    (in width dimension) of input data x and grid_y is indexng the 3rd
    dimention (in height dimension), finally results is the bilinear
11893
    interpolation value of 4 nearest corner points.
11894

H
haowang101779990 已提交
11895
    .. code-block:: text
11896

H
haowang101779990 已提交
11897 11898
        Step 1:
        Get (x, y) grid coordinates and scale to [0, H-1/W-1].
11899

H
haowang101779990 已提交
11900 11901
        grid_x = 0.5 * (grid[:, :, :, 0] + 1) * (W - 1)
        grid_y = 0.5 * (grid[:, :, :, 1] + 1) * (H - 1)
11902

H
haowang101779990 已提交
11903 11904 11905
        Step 2:
        Indices input data X with grid (x, y) in each [H, W] area, and bilinear
        interpolate point value by 4 nearest points.
11906

H
haowang101779990 已提交
11907 11908 11909 11910 11911 11912 11913 11914 11915
          wn ------- y_n ------- en
          |           |           |
          |          d_n          |
          |           |           |
         x_w --d_w-- grid--d_e-- x_e
          |           |           |
          |          d_s          |
          |           |           |
          ws ------- y_s ------- wn
11916

H
haowang101779990 已提交
11917 11918 11919 11920
        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
11921

H
haowang101779990 已提交
11922 11923 11924 11925
        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
11926

H
haowang101779990 已提交
11927 11928 11929 11930
        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
11931

H
haowang101779990 已提交
11932 11933
        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
11934 11935

    Args:
11936 11937 11938
        x(Variable): Input data of shape [N, C, H, W].
        grid(Variable): Input grid tensor of shape [N, H, W, 2].
        name (str, default None): The name of this layer.
D
dengkaipeng 已提交
11939 11940

    Returns:
H
haowang101779990 已提交
11941
        Variable: Output of shape [N, C, H, W] data samples input X
11942 11943
        using bilnear interpolation based on input grid.

H
haowang101779990 已提交
11944 11945 11946 11947
    Examples:

        .. code-block:: python

K
Kaipeng Deng 已提交
11948 11949 11950 11951 11952
            import paddle.fluid as fluid

            x = fluid.layers.data(name='x', shape=[10, 32, 32], dtype='float32')
            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
11953
            out = fluid.layers.grid_sampler(x=x, grid=grid)
11954

D
dengkaipeng 已提交
11955 11956 11957 11958 11959 11960 11961 11962 11963
    """
    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")

11964
    out = helper.create_variable_for_type_inference(x.dtype)
D
dengkaipeng 已提交
11965 11966
    ipts = {'X': x, 'Grid': grid}

11967
    helper.append_op(type='grid_sampler', inputs=ipts, outputs={'Output': out})
11968 11969 11970
    return out


G
gmcather 已提交
11971 11972 11973 11974 11975 11976 11977 11978 11979 11980 11981 11982 11983 11984 11985 11986 11987 11988 11989 11990 11991 11992 11993 11994 11995 11996 11997
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:
        input (Variable|list):  a 2-D tensor with shape [N x 1], where N is the
                                batch size. This input is a probability computed
                                by the previous operator.
        label (Variable|list):  the ground truth which is a 2-D tensor with
                                shape [N x 1], where N is the batch size.
        epsilon (float): epsilon
        name (string): the name of log_loss

    Returns:
        Variable: A 2-D tensor with shape [N x 1], the negative log loss.

    Examples:
        .. code-block:: python

11998
          import paddle.fluid as fluid
Y
Yibing Liu 已提交
11999 12000
          label = fluid.layers.data(name='label', shape=[1], dtype='int64')
          prob = fluid.layers.data(name='prob', shape=[10], dtype='float32')
G
gmcather 已提交
12001 12002 12003 12004 12005 12006 12007 12008 12009 12010 12011 12012 12013 12014 12015 12016 12017 12018 12019
          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


H
heqiaozhi 已提交
12020 12021 12022 12023 12024 12025 12026 12027 12028 12029 12030 12031 12032 12033 12034 12035 12036 12037 12038
def teacher_student_sigmoid_loss(input,
                                 label,
                                 soft_max_up_bound=15.0,
                                 soft_max_lower_bound=-15.0):
    """
    **Teacher Student Log Loss Layer**

    This layer accepts input predictions and target label and returns the
    teacher_student loss.

    .. math::
        loss = max(x, 0) - x * z + log(1 + exp(-abs(x))) + max(x, 0) - x * z' + log(1 + exp(-abs(x)))

    Args:
        input (Variable|list):  a 2-D tensor with shape [N x 1], where N is the
                                batch size. This input is a probability computed
                                by the previous operator.
        label (Variable|list):  the ground truth which is a 2-D tensor with
                                shape [N x 1], where N is the batch size.
M
minqiyang 已提交
12039
        soft_max_up_bound  (float):  if input > soft_max_up_bound, will be bound
H
heqiaozhi 已提交
12040 12041 12042 12043 12044 12045 12046
        soft_max_lower_bound (float): if input < soft_max_lower_bound, will be bound

    Returns:
        Variable: A 2-D tensor with shape [N x 1], the teacher_student_sigmoid_loss.

    Examples:
        .. code-block:: python
12047 12048
          
          import paddle.fluid as fluid
H
heqiaozhi 已提交
12049

12050 12051 12052 12053 12054
          batch_size = 64
          label = fluid.layers.data(
                    name="label", shape=[batch_size, 1], dtype="int64", append_batch_size=False)
          similarity = fluid.layers.data(
                    name="similarity", shape=[batch_size, 1], dtype="float32", append_batch_size=False)
H
heqiaozhi 已提交
12055
          cost = fluid.layers.teacher_student_sigmoid_loss(input=similarity, label=label)
12056

H
heqiaozhi 已提交
12057 12058 12059 12060 12061 12062 12063 12064 12065 12066 12067 12068 12069
    """
    helper = LayerHelper('teacher_student_sigmoid_loss', **locals())
    out = helper.create_variable(dtype=input.dtype)
    helper.append_op(
        type='teacher_student_sigmoid_loss',
        inputs={'X': [input],
                'Label': [label]},
        outputs={'Y': [out]},
        attrs={"soft_max_lower_bound": float(soft_max_lower_bound), \
                "soft_max_up_bound": float(soft_max_up_bound)})
    return out


G
gmcather 已提交
12070 12071 12072 12073
def add_position_encoding(input, alpha, beta, name=None):
    """
    **Add Position Encoding Layer**

H
haowang101779990 已提交
12074
    This layer accepts an input 3D-Tensor of shape [N x M x P], and returns an
G
gmcather 已提交
12075 12076
    output Tensor of shape [N x M x P] with positional encoding value.

H
haowang101779990 已提交
12077
    Refer to `Attention Is All You Need <http://arxiv.org/pdf/1706.03762.pdf>`_ .
G
gmcather 已提交
12078 12079

    .. math::
H
haowang101779990 已提交
12080 12081 12082
        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
12083 12084

    Where:
H
haowang101779990 已提交
12085 12086
      - :math:`PE(pos, 2i)` : the increment for the number at even position
      - :math:`PE(pos, 2i + 1)` : the increment for the number at odd position
G
gmcather 已提交
12087 12088 12089 12090 12091 12092 12093 12094 12095 12096 12097 12098 12099

    Args:
        input (Variable): 3-D input tensor with shape [N x M x P]
        alpha (float): multiple of Input Tensor
        beta (float): multiple of Positional Encoding Tensor
        name (string): the name of position encoding layer

    Returns:
        Variable: A 3-D Tensor of shape [N x M x P] with positional encoding.

    Examples:
        .. code-block:: python

12100 12101 12102 12103 12104 12105 12106 12107 12108
          import paddle.fluid as fluid

          tensor = fluid.layers.data(
              name='tensor',
              shape=[32, 64, 512],
              dtype='float32',
              append_batch_size=False)
          position_tensor = fluid.layers.add_position_encoding(
              input=tensor, alpha=1.0, beta=1.0)
H
haowang101779990 已提交
12109

G
gmcather 已提交
12110 12111 12112 12113 12114 12115 12116 12117 12118 12119 12120 12121 12122 12123 12124 12125
    """
    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
12126 12127 12128 12129 12130 12131 12132 12133 12134 12135


def bilinear_tensor_product(x,
                            y,
                            size,
                            act=None,
                            name=None,
                            param_attr=None,
                            bias_attr=None):
    """
Q
Qiao Longfei 已提交
12136
    **Add Bilinear Tensor Product Layer**
Q
Qiao Longfei 已提交
12137

Q
Qiao Longfei 已提交
12138
    This layer performs bilinear tensor product on two inputs.
Q
Qiao Longfei 已提交
12139 12140 12141
    For example:

    .. math::
H
haowang101779990 已提交
12142
       out_{i} = x * W_{i} * {y^\mathrm{T}}, i=0,1,...,size-1
Q
Qiao Longfei 已提交
12143

Q
Qiao Longfei 已提交
12144
    In this formula:
12145 12146
      - :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].
Q
Qiao Longfei 已提交
12147
      - :math:`W_{i}`: the i-th learned weight, shape is [M, N]
H
haowang101779990 已提交
12148
      - :math:`out_{i}`: the i-th element of out, shape is [batch_size, size].
Q
Qiao Longfei 已提交
12149 12150 12151
      - :math:`y^\mathrm{T}`: the transpose of :math:`y_{2}`.

    Args:
12152 12153
        x (Variable): 2-D input tensor with shape [batch_size, M]
        y (Variable): 2-D input tensor with shape [batch_size, N]
Q
Qiao Longfei 已提交
12154 12155 12156
        size (int): The dimension of this layer.
        act (str, default None): Activation to be applied to the output of this layer.
        name (str, default None): The name of this layer.
Q
Qiao Longfei 已提交
12157
        param_attr (ParamAttr, default None): The parameter attribute for the learnable w.
Q
Qiao Longfei 已提交
12158
            parameters/weights of this layer.
Q
Qiao Longfei 已提交
12159
        bias_attr (ParamAttr, default None): The parameter attribute for the bias
Q
Qiao Longfei 已提交
12160 12161 12162 12163
            of this layer. If it is set to False, no bias will be added to the output units.
            If it is set to None, the bias is initialized zero. Default: None.

    Returns:
Q
Qiao Longfei 已提交
12164
        Variable: A 2-D Tensor of shape [batch_size, size].
Q
Qiao Longfei 已提交
12165 12166 12167 12168

    Examples:
        .. code-block:: python

12169
          import paddle.fluid as fluid
Y
Yibing Liu 已提交
12170 12171 12172
          layer1 = fluid.layers.data("t1", shape=[-1, 5], dtype="float32")
          layer2 = fluid.layers.data("t2", shape=[-1, 4], dtype="float32")
          tensor = fluid.layers.bilinear_tensor_product(x=layer1, y=layer2, size=1000)
Q
Qiao Longfei 已提交
12173 12174
    """
    helper = LayerHelper('bilinear_tensor_product', **locals())
Q
Qiao Longfei 已提交
12175
    dtype = helper.input_dtype('x')
Q
Qiao Longfei 已提交
12176 12177 12178 12179

    param_shape = [size, x.shape[1], y.shape[1]]

    w = helper.create_parameter(
Q
Qiao Longfei 已提交
12180
        attr=helper.param_attr, shape=param_shape, dtype=dtype, is_bias=False)
Q
Qiao Longfei 已提交
12181 12182 12183 12184 12185 12186 12187 12188 12189 12190 12191 12192 12193 12194 12195 12196 12197

    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
12198 12199 12200 12201 12202 12203 12204 12205 12206 12207 12208 12209 12210


@templatedoc()
def get_tensor_from_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}
B
bdzhuxiaoning 已提交
12211 12212 12213 12214 12215 12216 12217 12218

    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
12219 12220 12221 12222 12223 12224 12225 12226 12227 12228
    """

    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
12229 12230


S
shippingwang 已提交
12231
def shuffle_channel(x, group, name=None):
S
shippingwang 已提交
12232 12233
    """
    **Shuffle Channel Operator**
12234

S
shippingwang 已提交
12235 12236 12237 12238 12239 12240
    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
12241
    
S
shippingwang 已提交
12242
    .. code-block:: text
12243

S
shippingwang 已提交
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
        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
12272
    Args: 
S
shippingwang 已提交
12273 12274
        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
12275 12276

    Returns:
S
shippingwang 已提交
12277 12278
        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
12279 12280

    Raises:
S
shippingwang 已提交
12281
        ValueError: If group is not an int type variable.
S
shippingwang 已提交
12282 12283 12284

    Examples:
        .. code-block:: python
12285

12286
            import paddle.fluid as fluid
12287
            input = fluid.layers.data(name='input', shape=[4,2,2], dtype='float32')
S
shippingwang 已提交
12288
            out = fluid.layers.shuffle_channel(x=input, group=2)
S
shippingwang 已提交
12289 12290 12291
    """
    helper = LayerHelper("shuffle_channel", **locals())

S
shippingwang 已提交
12292
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
S
shippingwang 已提交
12293 12294 12295 12296 12297 12298 12299 12300 12301

    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
12302
    return out
S
Add  
shippingwang 已提交
12303 12304


12305
@templatedoc()
D
dengkaipeng 已提交
12306
def temporal_shift(x, seg_num, shift_ratio=0.25, name=None):
12307 12308 12309 12310 12311 12312 12313 12314
    """
    **Temporal Shift Operator**
    
    ${comment}
                        
    Args: 
        x(Variable): ${x_comment}
        seg_num(int): ${seg_num_comment}
D
dengkaipeng 已提交
12315
        shift_ratio(float): ${shift_ratio_comment}
D
dengkaipeng 已提交
12316
        name (str, default None): The name of this layer.
12317 12318 12319 12320 12321 12322 12323 12324 12325 12326 12327

    Returns:
        out(Variable): The temporal shifting result is a tensor variable with the 
        same shape and same type as the input.

    Raises:
        TypeError: seg_num must be int type.

    Examples:
        .. code-block:: python

12328
            import paddle.fluid as fluid
12329
            input = fluid.layers.data(name='input', shape=[4,2,2], dtype='float32')
D
dengkaipeng 已提交
12330
            out = fluid.layers.temporal_shift(x=input, seg_num=2, shift_ratio=0.2)
12331 12332 12333 12334 12335 12336 12337 12338 12339 12340 12341 12342
    """
    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
12343 12344
        attrs={"seg_num": seg_num,
               "shift_ratio": shift_ratio})
12345 12346 12347
    return out


S
sneaxiy 已提交
12348
class PyFuncRegistry(object):
S
sneaxiy 已提交
12349 12350 12351
    _register_funcs = []

    def __init__(self, func):
S
sneaxiy 已提交
12352
        if func is None or not callable(func):
S
sneaxiy 已提交
12353 12354 12355
            raise TypeError('func must be a Python function')

        self._func = func
M
minqiyang 已提交
12356
        # find named args using reflection
S
sneaxiy 已提交
12357 12358 12359 12360 12361 12362 12363
        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
12364 12365 12366
        '''
        Why record self here?

M
minqiyang 已提交
12367 12368
        1. For debug usage. Users can call
           :code:`py_func.registered_func(idx)` method
S
sneaxiy 已提交
12369
           to find the registered function corresponding
M
minqiyang 已提交
12370
           to :code:`idx`.
S
sneaxiy 已提交
12371

M
minqiyang 已提交
12372 12373
        2. For increasing reference count of self.
           It seems that to release Python object
S
sneaxiy 已提交
12374
           whose reference count is 1 would cause
M
minqiyang 已提交
12375
           segmentation fault error in C++ side.
S
sneaxiy 已提交
12376 12377
           May be lack of Python GC in C++ side?
        '''
S
sneaxiy 已提交
12378
        PyFuncRegistry._register_funcs.append(self)
S
sneaxiy 已提交
12379 12380 12381 12382 12383 12384 12385 12386 12387 12388 12389 12390 12391 12392

    @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
12393 12394 12395 12396 12397 12398 12399 12400 12401
        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
12402

S
sneaxiy 已提交
12403 12404
        if not isinstance(func_ret, (list, tuple)):
            func_ret = (func_ret, )
S
sneaxiy 已提交
12405 12406

        ret = []
S
sneaxiy 已提交
12407 12408 12409
        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
12410 12411
                continue

S
sneaxiy 已提交
12412 12413
            if not isinstance(each_ret, np.ndarray):
                each_ret = np.array(each_ret)
S
sneaxiy 已提交
12414

S
sneaxiy 已提交
12415 12416 12417
            tensor = core.LoDTensor()
            tensor.set(each_ret, core.CPUPlace())
            ret.append(tensor)
S
sneaxiy 已提交
12418

S
sneaxiy 已提交
12419
        return tuple(ret)
S
sneaxiy 已提交
12420 12421


S
sneaxiy 已提交
12422 12423 12424 12425
@templatedoc()
def py_func(func, x, out, backward_func=None, skip_vars_in_backward_input=None):
    """
    PyFunc Operator.
M
minqiyang 已提交
12426

S
sneaxiy 已提交
12427 12428 12429 12430 12431 12432 12433 12434
    User can use :code:`py_func` to register operators in Python side.
    The inputs of :code:`func` is :code:`LoDTensor` and outputs can be
    numpy array or :code:`LoDTensor`. Paddle would call the registered
    :code:`func` in forward part, and call :code:`backward_func` in
    backward part (if :code:`backward_func` is not None).

    User should set the right data type and shape of :code:`out` before
    calling this function. However, data types and shapes of gradients of
S
sneaxiy 已提交
12435
    :code:`out` and :code:`x` would be inferred automatically.
S
sneaxiy 已提交
12436

S
sneaxiy 已提交
12437 12438
    Input orders of :code:`backward_func` would be: forward inputs
    :code:`x`, forward outputs :code:`out` and backward input gradients of
S
sneaxiy 已提交
12439 12440 12441 12442
    :code:`out`. If some variables of :code:`out` have no gradient, the input
    tensor would be None in Python side. If some variables of :code:`in` have
    no gradient, users should return None.

S
sneaxiy 已提交
12443
    This function can also be used to debug the running network. User can
M
minqiyang 已提交
12444
    add a :code:`py_func` operator without output, and print input
S
sneaxiy 已提交
12445 12446
    :code:`x` inside :code:`func`.

S
sneaxiy 已提交
12447 12448 12449 12450 12451
    Args:
        func (callable): forward Python function.
        x (Variable|list(Variable)|tuple(Variable)): inputs of :code:`func`.
        out (Variable|list(Variable)|tuple(Variable)): outputs of :code:`func`.
            Paddle cannot infer shapes and data types of :code:`out`. Users
M
minqiyang 已提交
12452
            should create :code:`out` beforehand.
S
sneaxiy 已提交
12453
        backward_func (callable|None): backward Python function.
M
minqiyang 已提交
12454
                                       None means no backward. Default None.
S
sneaxiy 已提交
12455
        skip_vars_in_backward_input (Variable|list(Variable)|tuple(Variable)):
M
minqiyang 已提交
12456
            Variables that are not needed in :code:`backward_func` inputs.
S
sneaxiy 已提交
12457 12458
            These variables must be any of :code:`x` and :code:`out`.
            If set, these vars would not be inputs of :code:`backward_func`,
M
minqiyang 已提交
12459
            Only useful when :code:`backward_func` is not None. Default None.
S
sneaxiy 已提交
12460 12461 12462

    Returns:
        out (Variable|list(Variable)|tuple(Variable)): input :code:`out`
S
sneaxiy 已提交
12463 12464

    Examples:
M
minqiyang 已提交
12465

S
sneaxiy 已提交
12466 12467 12468 12469 12470
        >>> import paddle.fluid as fluid
        >>> import six
        >>>
        >>> def create_tmp_var(name, dtype, shape):
        >>>     return fluid.default_main_program().current_block().create_var(
M
minqiyang 已提交
12471
        >>>         name=name, dtype=dtype, shape=shape)
S
sneaxiy 已提交
12472 12473
        >>>
        >>> # tanh activation has been provided by Paddle C++ op
M
minqiyang 已提交
12474
        >>> # Here, we only use tanh to be an example to show the usage
S
sneaxiy 已提交
12475 12476 12477
        >>> # of py_func
        >>> def tanh(x):
        >>>     return np.tanh(x)
M
minqiyang 已提交
12478
        >>>
S
sneaxiy 已提交
12479 12480 12481 12482 12483
        >>> # forward input x is skipped
        >>> def tanh_grad(y, dy):
        >>>     return np.array(dy) * (1 - np.square(np.array(y)))
        >>>
        >>> def debug_func(x):
M
minqiyang 已提交
12484
        >>>     print(x)
S
sneaxiy 已提交
12485 12486 12487 12488 12489 12490
        >>>
        >>> 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),
M
minqiyang 已提交
12491
        >>>             dtype=hidden.dtype, shape=hidden.shape)
S
sneaxiy 已提交
12492 12493
        >>>
        >>>         # user-defined layers with forward and backward
M
minqiyang 已提交
12494 12495
        >>>         hidden = fluid.layers.py_func(func=tanh, x=hidden,
        >>>             out=new_hidden, backward_func=tanh_grad,
S
sneaxiy 已提交
12496 12497 12498 12499 12500 12501 12502 12503
        >>>             skip_vars_in_backward_input=hidden)
        >>>
        >>>         # user-defined debug layers to print variables
        >>>         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
12504
    """
S
sneaxiy 已提交
12505
    helper = LayerHelper('py_func', **locals())
S
sneaxiy 已提交
12506 12507 12508
    if x is None:
        x = []
    elif isinstance(x, Variable):
S
sneaxiy 已提交
12509
        x = [x]
S
sneaxiy 已提交
12510 12511
    elif not isinstance(x, (list, tuple)):
        raise TypeError('Input must be Variable/list(Variable)/tuple(Variable)')
S
sneaxiy 已提交
12512

S
sneaxiy 已提交
12513 12514 12515
    if out is None:
        out_list = []
    elif isinstance(out, Variable):
S
sneaxiy 已提交
12516
        out_list = [out]
S
sneaxiy 已提交
12517
    elif isinstance(out, (list, tuple)):
S
sneaxiy 已提交
12518
        out_list = out
S
sneaxiy 已提交
12519 12520 12521
    else:
        raise TypeError(
            'Output must be Variable/list(Variable)/tuple(Variable)')
S
sneaxiy 已提交
12522

S
sneaxiy 已提交
12523 12524
    fwd_func_id = PyFuncRegistry(func).id
    bwd_func_id = PyFuncRegistry(
S
sneaxiy 已提交
12525
        backward_func).id if backward_func is not None else -1
S
sneaxiy 已提交
12526 12527

    for each_out in out_list:
S
sneaxiy 已提交
12528 12529
        if len(each_out.shape) == 0:
            raise ValueError(
S
sneaxiy 已提交
12530 12531
                'Output shapes of py_func op should be provided by users manually'
            )
S
sneaxiy 已提交
12532

S
sneaxiy 已提交
12533 12534 12535 12536 12537 12538 12539 12540 12541 12542 12543 12544 12545 12546 12547
    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
12548 12549 12550 12551

    helper.append_op(
        type='py_func',
        inputs={'X': x},
S
sneaxiy 已提交
12552 12553
        outputs={'Out': out_list},
        attrs={
S
sneaxiy 已提交
12554 12555 12556
            'forward_callable_id': fwd_func_id,
            'backward_callable_id': bwd_func_id,
            'backward_skip_vars': list(backward_skip_vars)
S
sneaxiy 已提交
12557
        })
S
sneaxiy 已提交
12558
    return out
S
sneaxiy 已提交
12559 12560 12561


# For debug usage
S
sneaxiy 已提交
12562 12563 12564 12565
py_func.registered_func = PyFuncRegistry.registered_func
py_func.registered_func_num = PyFuncRegistry.registered_func_num


12566 12567 12568 12569 12570 12571 12572 12573 12574 12575 12576 12577 12578
@templatedoc()
def psroi_pool(input,
               rois,
               output_channels,
               spatial_scale,
               pooled_height,
               pooled_width,
               name=None):
    """
    ${comment}

    Args:
        input (Variable): ${x_comment}
S
SunGaofeng 已提交
12579 12580 12581 12582 12583
        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.
12584 12585 12586 12587 12588 12589 12590 12591 12592 12593 12594 12595
        output_channels (integer): ${output_channels_comment}
        spatial_scale (float): ${spatial_scale_comment} Default: 1.0
        pooled_height (integer): ${pooled_height_comment} Default: 1
        pooled_width (integer): ${pooled_width_comment} Default: 1
        name (str, default None): The name of this layer.

    Returns:
        Variable: ${out_comment}.

    Examples:
        .. code-block:: python

S
SunGaofeng 已提交
12596 12597 12598 12599
            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')
            pool_out = fluid.layers.psroi_pool(x, rois, 10, 1.0, 7, 7)
12600 12601 12602 12603 12604 12605 12606 12607 12608 12609 12610 12611 12612 12613 12614 12615 12616 12617 12618 12619 12620 12621 12622 12623 12624
    """
    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
12625

M
minqiyang 已提交
12626

M
minqiyang 已提交
12627
def huber_loss(input, label, delta):
12628
    """
M
minqiyang 已提交
12629 12630 12631
    Huber loss is a loss function used in robust.
    Huber loss can evaluate the fitness of input to label.
    Different from MSE loss, Huber loss is more robust for outliers.
12632 12633 12634 12635

    When the difference between input and label is large than delta
    .. math::

M
minqiyang 已提交
12636
        huber\_loss = delta * (label - input) - 0.5 * delta * delta
12637 12638 12639 12640

    When the difference between input and label is less than delta
    .. math::

M
minqiyang 已提交
12641
        huber\_loss = 0.5 * (label - input) * (label - input)
12642 12643 12644 12645 12646 12647 12648


    Args:
        input (Variable): This input is a probability computed by the previous operator.
                          The first dimension is batch size, and the last dimension is 1.
        label (Variable): The groud truth whose first dimension is batch size
                          and last dimension is 1.
M
minqiyang 已提交
12649
        delta (float): The parameter of huber loss, which controls
12650 12651 12652
                       the range of outliers

    Returns:
M
minqiyang 已提交
12653
        huber\_loss (Variable): The huber loss with shape [batch_size, 1].
12654 12655 12656 12657

    Examples:
        .. code-block:: python

12658 12659 12660 12661 12662 12663 12664 12665 12666
            import paddle.fluid as fluid

            x = fluid.layers.data(name='x', shape=[13], dtype='float32')
            predict = fluid.layers.fc(input=x, size=1)
            label = fluid.layers.data(
                name='label', shape=[1], dtype='float32')
            loss = fluid.layers.huber_loss(
                input=predict, label=label, delta=1.0)

12667
    """
M
minqiyang 已提交
12668
    helper = LayerHelper('huber_loss', **locals())
12669 12670 12671 12672 12673 12674 12675 12676 12677 12678 12679
    residual = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype())
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
    helper.append_op(
        type='huber_loss',
        inputs={'X': input,
                'Y': label},
        outputs={'Out': out,
                 'Residual': residual},
        attrs={'delta': delta})
    return out
Z
zhaozhehao 已提交
12680 12681


D
dengkaipeng 已提交
12682 12683 12684 12685 12686 12687 12688 12689 12690 12691 12692 12693 12694 12695 12696 12697 12698
@templatedoc()
def kldiv_loss(x, target, reduction='mean', name=None):
    """
    ${comment}

    Args:
        x (Variable): ${x_comment}
        target (Variable): ${target_comment}
        reduction (Variable): ${reduction_comment}
        name (str, default None): The name of this layer.

    Returns:
        kldiv\_loss (Variable): The KL divergence loss.

    Examples:
        .. code-block:: python

12699
            import paddle.fluid as fluid
D
dengkaipeng 已提交
12700 12701 12702 12703 12704 12705 12706 12707 12708 12709 12710 12711 12712 12713 12714
            x = fluid.layers.data(name='x', shape=[4,2,2], dtype='float32')
            target = fluid.layers.data(name='target', shape=[4,2,2], dtype='float32')
            loss = fluid.layers.kldiv_loss(x=x, target=target, reduction='batchmean')
    """
    helper = LayerHelper('kldiv_loss', **locals())
    loss = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
        type='kldiv_loss',
        inputs={'X': x,
                'Target': target},
        outputs={'Loss': loss},
        attrs={'reduction': reduction})
    return loss


Z
zhaozhehao 已提交
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
@templatedoc()
def tree_conv(nodes_vector,
              edge_set,
              output_size,
              num_filters=1,
              max_depth=2,
              act='tanh',
              param_attr=None,
              bias_attr=None,
              name=None):
    """ 
    ${comment}
    		
    Args:
        nodes_vector(${nodes_vector_type}): ${nodes_vector_comment}
        edge_set(${edge_set_type}): ${edge_set_comment}
        output_size(int): output feature width
        num_filters(int): number of filters, Default 1
        max_depth(int): max depth of filters, Default 2
        act(str): activation function, Default tanh
        param_attr(ParamAttr): the parameter attribute for the filters, Default None
        bias_attr(ParamAttr): the parameter attribute for the bias of this layer, Default None
        name(str): a name of this layer(optional). If set None, the layer will be named automatically, Default None

    Returns:
        out(${out_type}): ${out_comment}

    Examples:
        .. code-block:: python

12745
          import paddle.fluid as fluid
T
Tao Luo 已提交
12746 12747 12748
          # 10 for max_node_size of dataset, 5 for vector width
          nodes_vector = fluid.layers.data(name='vectors', shape=[10, 5], dtype='float32')
          # 10 for max_node_size of dataset, 2 for every edge has two nodes
Z
zhaozhehao 已提交
12749
          # edges must be directional
T
Tao Luo 已提交
12750 12751 12752 12753
          edge_set = fluid.layers.data(name='edge_set', shape=[10, 2], dtype='float32')
          # the shape of output will be [10, 6, 1],
          # 10 for max_node_size of dataset, 6 for output size, 1 for 1 filter
          out_vector = fluid.layers.tree_conv(nodes_vector, edge_set, 6, 1, 2)
Z
zhaozhehao 已提交
12754
          # After reshape, output tensor could be nodes_vector for next tree convolution
T
Tao Luo 已提交
12755 12756
          out_vector = fluid.layers.reshape(out_vector, shape=[-1, 10, 6])
          out_vector_2 = fluid.layers.tree_conv(out_vector, edge_set, 3, 4, 2)
Z
zhaozhehao 已提交
12757
          # also output tensor could be pooling(the pooling in paper called global pooling)
T
Tao Luo 已提交
12758
          pooled = fluid.layers.reduce_max(out_vector, dim=2) # global pooling
Z
zhaozhehao 已提交
12759 12760 12761 12762 12763 12764 12765 12766 12767 12768 12769 12770 12771 12772 12773 12774 12775 12776 12777 12778 12779 12780 12781
    """
    helper = LayerHelper("tree_conv", **locals())
    dtype = helper.input_dtype('nodes_vector')
    feature_size = nodes_vector.shape[2]
    W_shape = [feature_size, 3, output_size, num_filters]
    W = helper.create_parameter(
        attr=param_attr, shape=W_shape, dtype=dtype, is_bias=False)
    if name == 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='tree_conv',
        inputs={'NodesVector': nodes_vector,
                'EdgeSet': edge_set,
                'Filter': W},
        outputs={'Out': out, },
        attrs={'max_depth': max_depth})
    if helper.bias_attr:
        pre_activation = helper.append_bias_op(out)
    else:
        pre_activation = out
    return helper.append_activation(pre_activation)
C
ceci3 已提交
12782 12783


C
ceci3 已提交
12784
from .ops import square
C
ceci3 已提交
12785
from .control_flow import equal
C
ceci3 已提交
12786 12787


C
ceci3 已提交
12788 12789 12790
def npair_loss(anchor, positive, labels, l2_reg=0.002):
    '''
  **Npair Loss Layer**
C
ceci3 已提交
12791

C
ceci3 已提交
12792
  Read `Improved Deep Metric Learning with Multi class N pair Loss Objective <http://www.nec-labs.com/uploads/images/Department-Images/MediaAnalytics/papers/nips16_npairmetriclearning.pdf>`_ .
C
ceci3 已提交
12793 12794

  Npair loss requires paired data. Npair loss has two parts: the first part is L2
C
ceci3 已提交
12795
  regularizer on the embedding vector; the second part is cross entropy loss which
C
ceci3 已提交
12796 12797 12798 12799 12800
  takes the similarity matrix of anchor and positive as logits.

  Args:
    anchor(Variable): embedding vector for the anchor image. shape=[batch_size, embedding_dims]
    positive(Variable): embedding vector for the positive image. shape=[batch_size, embedding_dims]
C
ceci3 已提交
12801 12802
    labels(Variable): 1-D tensor. shape=[batch_size]
    l2_reg(float32): L2 regularization term on embedding vector, default: 0.002
C
ceci3 已提交
12803 12804 12805 12806 12807 12808 12809

  Returns:
    npair loss(Variable): return npair loss, shape=[1]

  Examples:
    .. code-block:: python

12810
       import paddle.fluid as fluid
C
ceci3 已提交
12811 12812 12813 12814 12815 12816 12817 12818
       anchor = fluid.layers.data(
                     name = 'anchor', shape = [18, 6], dtype = 'float32', append_batch_size=False)
       positive = fluid.layers.data(
                     name = 'positive', shape = [18, 6], dtype = 'float32', append_batch_size=False)
       labels = fluid.layers.data(
                     name = 'labels', shape = [18], dtype = 'float32', append_batch_size=False)

       npair_loss = fluid.layers.npair_loss(anchor, positive, labels, l2_reg = 0.002)
C
ceci3 已提交
12819 12820 12821 12822 12823 12824 12825
  '''
    Beta = 0.25
    batch_size = labels.shape[0]

    labels = reshape(labels, shape=[batch_size, 1], inplace=True)
    labels = expand(labels, expand_times=[1, batch_size])

C
ceci3 已提交
12826
    labels = equal(labels, transpose(labels, perm=[1, 0])).astype('float32')
C
ceci3 已提交
12827 12828
    labels = labels / reduce_sum(labels, dim=1, keep_dim=True)

C
ceci3 已提交
12829 12830
    l2loss = reduce_mean(reduce_sum(square(anchor), 1)) \
             + reduce_mean(reduce_sum(square(positive), 1))
C
ceci3 已提交
12831 12832 12833 12834
    l2loss = l2loss * Beta * l2_reg

    similarity_matrix = matmul(
        anchor, positive, transpose_x=False, transpose_y=True)
C
ceci3 已提交
12835 12836 12837
    softmax_ce = softmax_with_cross_entropy(
        logits=similarity_matrix, label=labels, soft_label=True)
    cross_entropy = reduce_sum(labels * softmax_ce, 0)
C
ceci3 已提交
12838 12839 12840
    celoss = reduce_mean(cross_entropy)

    return l2loss + celoss
12841 12842


R
ruri 已提交
12843 12844 12845 12846 12847 12848 12849 12850 12851 12852 12853 12854 12855 12856 12857 12858 12859 12860 12861 12862 12863 12864 12865 12866 12867 12868 12869 12870 12871
def pixel_shuffle(x, upscale_factor):
    """

    **Pixel Shuffle Layer**

    This layer rearranges elements in a tensor of shape [N, C, H, W]
    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.

        .. code-block:: text
        
            Given a 4-D tensor with the shape:
                x.shape = [1, 9, 4, 4]
            Given upscale_factor:
                upscale_factor= 3
            output shape is:
                [1, 1, 12, 12]
    
    Args:

        x(Variable): The input tensor variable.
        upscale_factor(int): factor to increase spatial resolution

    Returns:

12872
        Out(Variable): Reshaped tensor according to the new dimension.
R
ruri 已提交
12873 12874 12875 12876 12877 12878 12879 12880 12881

    Raises:

        ValueError: If the square of upscale_factor cannot divide the channels of input.

    Examples:

        .. code-block:: python

12882
            import paddle.fluid as fluid
R
ruri 已提交
12883
            input = fluid.layers.data(name="input", shape=[9,4,4])
R
ruri 已提交
12884 12885 12886 12887 12888 12889 12890 12891 12892 12893 12894 12895 12896 12897 12898 12899 12900 12901 12902
            output = fluid.layers.pixel_shuffle(x=input, upscale_factor=3)

    """

    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


12903 12904 12905 12906 12907 12908 12909 12910 12911 12912 12913 12914 12915 12916 12917 12918 12919 12920 12921 12922 12923 12924 12925 12926 12927 12928 12929 12930 12931 12932 12933
def fsp_matrix(x, y):
    """

    **FSP matrix op**

    This op is used to calculate the flow of solution procedure (FSP) matrix of two feature maps.
    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:

        x (Variable): A feature map with shape [batch_size, x_channel, height, width].
        y (Variable): A feature map with shape [batch_size, y_channel, height, width].
                      The y_channel can be different with the x_channel of Input(X)
                      while the other dimensions must be the same with Input(X)'s.

    Returns:

        fsp matrix (Variable): The output of FSP op with shape [batch_size, x_channel, y_channel].
        The x_channel is the channel of x and the y_channel is the channel of y.

    Examples:

        .. code-block:: python

B
Bai Yifan 已提交
12934 12935 12936 12937 12938 12939
            import paddle.fluid as fluid
            data = fluid.layers.data(name='data', shape=[3, 32, 32])
            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)
12940 12941 12942 12943 12944 12945 12946 12947
            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
12948 12949 12950 12951


def continuous_value_model(input, cvm, use_cvm=True):
    """
H
fix doc  
heqiaozhi 已提交
12952

H
heqiaozhi 已提交
12953
    **continuous_value_model layers**
H
fix doc  
heqiaozhi 已提交
12954

H
fix doc  
heqiaozhi 已提交
12955
    continuous value model(cvm). Now, it only considers show and click value in CTR project.
H
fix doc  
heqiaozhi 已提交
12956 12957 12958
    We assume that input is an embedding vector with cvm_feature, whose shape is [N * D] (D is 2 + embedding dim).
    If use_cvm is True, it will log(cvm_feature), and output shape is [N * D].
    If use_cvm is False, it will remove cvm_feature from input, and output shape is [N * (D - 2)].
H
heqiaozhi 已提交
12959
    
H
fix doc  
heqiaozhi 已提交
12960
    This layer accepts a tensor named input which is ID after embedded(lod level is 1), cvm is a show_click info.
H
fix doc  
heqiaozhi 已提交
12961

H
heqiaozhi 已提交
12962
    Args:
H
fix doc  
heqiaozhi 已提交
12963 12964

        input (Variable): a 2-D LodTensor with shape [N x D], where N is the batch size, D is 2 + the embedding dim. lod level = 1.
H
heqiaozhi 已提交
12965 12966
        cvm (Variable):   a 2-D Tensor with shape [N x 2], where N is the batch size, 2 is show and click.
        use_cvm  (bool):  use cvm or not. if use cvm, the output dim is the same as input
H
fix doc  
heqiaozhi 已提交
12967
                          if don't use cvm, the output dim is input dim - 2(remove show and click)
12968
                          (cvm op is a customized op, which input is a sequence has embed_with_cvm default, so we need an op named cvm to decided whever use it or not.)
H
fix doc  
heqiaozhi 已提交
12969

H
heqiaozhi 已提交
12970
    Returns:
H
fix doc  
heqiaozhi 已提交
12971 12972 12973

        Variable: A 2-D LodTensor with shape [N x D], if use cvm, D is equal to input dim, if don't use cvm, D is equal to input dim - 2. 

H
heqiaozhi 已提交
12974
    Examples:
H
fix doc  
heqiaozhi 已提交
12975

H
heqiaozhi 已提交
12976
        .. code-block:: python
H
fix doc  
heqiaozhi 已提交
12977

12978
          import paddle.fluid as fluid
H
heqiaozhi 已提交
12979 12980 12981 12982 12983 12984 12985 12986 12987 12988
          input = fluid.layers.data(name="input", shape=[-1, 1], lod_level=1, append_batch_size=False, dtype="int64")#, stop_gradient=False)
          label = fluid.layers.data(name="label", shape=[-1, 1], append_batch_size=False, dtype="int64")
          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
12989

H
heqiaozhi 已提交
12990 12991 12992 12993 12994 12995 12996 12997 12998
    """
    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
12999
    return out
Z
zhoukunsheng 已提交
13000 13001 13002 13003 13004 13005 13006 13007 13008 13009 13010 13011 13012 13013 13014 13015 13016 13017


def where(condition):
    """
    Return an int64 tensor with rank 2, specifying the coordinate of true element in `condition`.

    Output's first dimension is the number of true element, second dimension is rank(number of dimension) of `condition`.
    If there is zero true element, then an empty tensor will be generated.  

    Args:
        condition(Variable): A bool tensor with rank at least 1.

    Returns:
        Variable: The tensor variable storing a 2-D tensor. 

    Examples:
        .. code-block:: python

13018
             import paddle.fluid as fluid
13019 13020 13021
             import paddle.fluid.layers as layers
             import numpy as np

Z
zhoukunsheng 已提交
13022
             # condition is a tensor [True, False, True]
13023 13024 13025
             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
13026 13027

             # condition is a tensor [[True, False], [False, True]]
13028 13029 13030
             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
13031 13032

             # condition is a tensor [False, False, False]
13033 13034 13035 13036
             condition = layers.assign(np.array([0, 0, 0], dtype='int32'))
             condition = layers.cast(condition, 'bool')
             out = layers.where(condition) # [[]]

Z
zhoukunsheng 已提交
13037 13038 13039 13040 13041 13042 13043 13044 13045
    """
    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
13046 13047 13048 13049 13050 13051 13052 13053 13054 13055 13056 13057 13058 13059 13060 13061 13062


def sign(x):
    """
    **sign**

    This function returns sign of every element in `x`: 1 for positive, -1 for negative and 0 for zero.

    Args:
        x(Variable|numpy.ndarray): The input tensor.

    Returns:
        Variable: The output sign tensor with identical shape and dtype to `x`.

    Examples:
        .. code-block:: python

13063 13064 13065
          import paddle.fluid as fluid
          import numpy as np

Z
zhoukunsheng 已提交
13066
          # [1, 0, -1]
13067 13068
          data = fluid.layers.sign(np.array([3, 0, -2], dtype='int32')) 

Z
zhoukunsheng 已提交
13069 13070 13071 13072 13073 13074 13075 13076 13077 13078 13079 13080
    """

    helper = LayerHelper("sign", **locals())

    if not isinstance(x, Variable):
        x = assign(x)

    out = helper.create_variable_for_type_inference(dtype=x.dtype)

    helper.append_op(type='sign', inputs={'X': [x]}, outputs={'Out': [out]})

    return out
13081 13082


Z
zhoukunsheng 已提交
13083 13084 13085 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
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


13122 13123 13124 13125 13126 13127 13128 13129 13130 13131 13132 13133 13134 13135 13136 13137 13138 13139 13140 13141 13142 13143 13144 13145 13146 13147 13148 13149 13150 13151 13152 13153 13154 13155 13156 13157 13158 13159 13160 13161 13162 13163 13164 13165 13166 13167 13168 13169 13170 13171 13172 13173
def unique_with_counts(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, count). `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, \
            `count` is count of unqiue element in the `x`.

    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]
    """
    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


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 13203 13204 13205 13206 13207 13208 13209 13210 13211 13212 13213 13214 13215 13216 13217 13218 13219 13220 13221 13222 13223 13224 13225 13226 13227 13228 13229 13230 13231 13232 13233 13234 13235 13236 13237 13238 13239 13240 13241 13242 13243 13244 13245 13246 13247 13248 13249 13250 13251 13252 13253 13254 13255 13256 13257 13258 13259 13260 13261 13262 13263 13264 13265 13266 13267 13268 13269 13270 13271 13272 13273 13274 13275
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,
                    name=None):
    """
    **Deformable Convolution Layer**

    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:
    
    .. math::

        y(p) = \sum_{k=1}^{K}{w_k * x(p + p_k + \Delta p_k) * \Delta m_k}
    
    Where :math:`\Delta p_k` and :math:`\Delta m_k` are the learnable offset and modulation scalar for the k-th location, respectively.
    Refer to `Deformable ConvNets v2: More Deformable, Better Results
    <https://arxiv.org/abs/1811.11168v2>`_ .
    
    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:
        input (Variable): The input image with [N, C, H, W] format.
        offset (Variable): The input coord offset of deformable convolution layer.
        Mask (Variable): The input mask of deformable covolution layer.
        num_filters(int): The number of filter. It is as same as the output
            image channel.
        filter_size (int|tuple|None): The filter size. If filter_size is a tuple,
            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.
        param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
            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.
        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of
            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.
        name (str|None): A name for this layer(optional). If set None, the layer
            will be named automatically. Default: None
    Returns:
        Variable: The tensor variable storing the deformable convolution \
                  result.
    Raises:
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
    Examples:
        .. code-block:: python

13276
          import paddle.fluid as fluid
13277 13278 13279 13280 13281 13282 13283 13284 13285 13286 13287 13288 13289 13290 13291 13292 13293 13294 13295 13296 13297 13298 13299 13300 13301 13302 13303 13304 13305 13306 13307 13308 13309 13310 13311 13312 13313 13314 13315 13316 13317 13318 13319 13320 13321 13322 13323 13324 13325 13326 13327 13328 13329 13330 13331 13332 13333 13334 13335 13336 13337 13338 13339 13340 13341 13342 13343 13344
          data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32')
          offset = fluid.layers.data(name='offset', shape=[18, 32, 32], dtype='float32')
          mask = fluid.layers.data(name='mask', shape=[9, 32, 32], dtype='float32')
          out = fluid.layers.deformable_conv(input=data, offset=offset, mask=mask,
                                             num_filters=2, filter_size=3, padding=1)
    """

    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 not isinstance(mask, Variable):
        raise TypeError("Input Mask 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)

    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,
        })

    output = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
    return output
13345 13346 13347 13348 13349 13350 13351 13352 13353 13354 13355 13356 13357 13358 13359 13360 13361 13362 13363 13364 13365 13366 13367 13368 13369 13370 13371 13372 13373 13374 13375 13376 13377 13378 13379 13380 13381 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 13417 13418 13419 13420 13421 13422 13423 13424 13425 13426 13427 13428 13429 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


def unfold(x, kernel_sizes, strides=1, paddings=0, dilations=1, name=None):
    """

    This function returns a col buffer of sliding local blocks of input x, also known
    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.

    For each input :math:`X` with shape [N, C, H, W], the output shape [N, Cout, Lout]
    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


    Args:
        x(Varaible):              The input tensor of format [N, C, H, W].
        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].

    
    Returns:
        Variable: 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.

    Examples:

        .. code-block:: python

            import paddle.fluid as fluid
            x = fluid.layers.data(name = 'data', shape = [3, 224, 224], dtype = 'float32')
            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
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 13505 13506 13507


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):
    """
    Deformable PSROI Pooling Layer
    
    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.
       trans (Variable): Offset of features on ROIs while pooling.The format is NCHW, 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 is True or False. Default: False.
       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.
       group_size (list|tuple): The number of groups which input channels are divided.(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 (integer): The pooled output height. Default: 1.
       pooled_width (integer): The pooled output width. Default: 1.
       part_size (list|tuple): The height and width of offset, eg.(4, 6), which height is 4 and width is 6, Default: 
                        if None, default value is [pooled_height, pooled_width].
       sample_per_part (integer): The number of samples in each bin. Default: 1.
       trans_std (float): Coefficient of offset. Default: 0.1.
       position_sensitive (bool): Whether to choose deformable psroi pooling mode or not. Default: False.
       name (str): Name of layer. Default: None.
    Returns:
        Variable: The tensor variable storing the deformable psroi pooling \
                  result.


    Examples:
      .. code-block:: python

13508
        import paddle.fluid as fluid
C
cjt222 已提交
13509 13510 13511 13512 13513 13514 13515 13516 13517 13518 13519 13520 13521 13522 13523 13524 13525 13526 13527 13528 13529 13530 13531 13532 13533 13534 13535 13536 13537 13538 13539 13540 13541 13542 13543 13544 13545 13546 13547 13548 13549 13550 13551 13552 13553 13554 13555 13556 13557 13558 13559 13560 13561 13562 13563 13564 13565 13566 13567 13568 13569
        input = fluid.layers.data(name="input",
                                  shape=[2, 192, 64, 64], 
                                  dtype='float32', 
                                  append_batch_size=False)                   
        rois = fluid.layers.data(name="rois",
                                 shape=[4],
                                 dtype='float32', 
                                 lod_level=1)
        trans = fluid.layers.data(name="trans",
                                  shape=[2, 384, 64, 64], 
                                  dtype='float32', 
                                  append_batch_size=False) 
        x = fluid.layers.nn.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)
    """

    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
13570 13571


K
Kevin 已提交
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 13604 13605 13606 13607 13608 13609 13610 13611 13612 13613 13614 13615 13616 13617 13618 13619 13620 13621 13622 13623 13624 13625 13626 13627 13628 13629 13630 13631 13632 13633 13634 13635 13636 13637 13638 13639 13640 13641 13642 13643 13644 13645 13646 13647 13648 13649 13650 13651 13652 13653 13654 13655 13656 13657 13658 13659 13660 13661 13662 13663 13664 13665 13666 13667 13668 13669 13670 13671 13672 13673 13674 13675 13676 13677 13678 13679 13680 13681 13682 13683 13684 13685 13686
def var_conv_2d(input,
                row,
                col,
                input_channel,
                output_channel,
                filter_size,
                stride=1,
                param_attr=None,
                act=None,
                dtype='float32',
                name=None):
    """
    The var_conv_2d layer calculates the output base on the :attr:`input` with variable length,
    row, col, input channel, filter size and strides. Both :attr:`input`, :attr:`row`,
    and :attr:`col` are 1-level LodTensor. The covolution operation is same as conv2d layer with 
    padding. Besides, input.dims[1] should be 1. 

    .. code-block:: text
            
            If input_channel is 2 and given row lodTensor and col lodTensor as follows:
                row.lod = [[5, 4]]
                col.lod = [[6, 7]]
            input is a lodTensor: 
                input.lod = [[60, 56]]	# where 60 = input_channel * 5 * 6
                input.dims = [116, 1]	# where 116 = 60 + 56
            
            If set output_channel is 3, filter_size is [3, 3], stride is [1, 1]:
                output.lod = [[90, 84]] # where 90 = output_channel * [(5-1)/stride + 1] * [(6-1)/stride + 1]
                output.dims = [174, 1]  # where 174 = 90 + 84

    Args:
        input (Variable): The input shoud be 1-level LodTensor with dims[1] equals 1.
        row (Variable): The row shoud be 1-level LodTensor to provide height information.
        col (Variable): The col shoud be 1-level LodTensor to provide width information.
        input_channel (int): The number of input channel.
        output_channel (int): The number of output channel.
        filter_size (int|tuple|None): The filter size. If filter_size is a tuple,
            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.
        param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
            of var_conv2d. If it is set to None or one attribute of ParamAttr, var_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)`,
            and the :math:`std` is :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None.
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None
        dtype ('float32'): The data type of parameter and output.
        name (str|None): A name for this layer(optional). If set None, the layer
            will be named automatically. Default: None

    Returns:
        Variable: Output variable with LoD specified by this layer.

    Examples:
        .. code-block:: python

            import numpy as np
            from paddle.fluid import layers

            x_lod_tensor = layers.data(name='x', shape=[1], lod_level=1)
            row_lod_tensor = layers.data(name='row', shape=[6], lod_level=1)
            col_lod_tensor = layers.data(name='col', shape=[6], lod_level=1)
            out = layers.var_conv_2d(input=x_lod_tensor, 
                                     row=row_lod_tensor,
                                     col=col_lod_tensor,
                                     input_channel=3,
                                     output_channel=5,
                                     filter_size=[3, 3],
                                     stride=1)
    """
    helper = LayerHelper('var_conv_2d', **locals())
    x_shape = list(input.shape)
    assert len(x_shape) == 2

    filter_size = utils.convert_to_list(filter_size, 2, 'filter_size')
    stride = utils.convert_to_list(stride, 2, 'stride')

    filter_shape = [
        int(output_channel),
        int(input_channel) * filter_size[0] * filter_size[1]
    ]
    filter_param = helper.create_parameter(
        attr=helper.param_attr,
        shape=filter_shape,
        dtype=dtype, )

    conv_res = helper.create_variable_for_type_inference(dtype)
    tmp_res = helper.create_variable_for_type_inference(
        dtype, stop_gradient=True)

    helper.append_op(
        type='var_conv_2d',
        inputs={
            'X': input,
            'ROW': row,
            'COLUMN': col,
            'W': filter_param,
        },
        outputs={"Out": conv_res,
                 "Col": tmp_res},
        attrs={
            'InputChannel': input_channel,
            'OutputChannel': output_channel,
            'StrideH': stride[0],
            'StrideW': stride[1],
            'KernelH': filter_size[0],
            'KernelW': filter_size[1],
        })

    return helper.append_activation(conv_res)


A
Aurelius84 已提交
13687 13688 13689 13690 13691 13692 13693 13694 13695 13696 13697 13698 13699 13700 13701 13702 13703 13704 13705 13706 13707 13708 13709 13710 13711 13712 13713 13714 13715 13716 13717 13718 13719 13720 13721 13722 13723 13724 13725 13726 13727 13728 13729 13730 13731 13732 13733 13734 13735 13736 13737 13738 13739 13740 13741 13742 13743 13744 13745 13746 13747 13748 13749 13750 13751 13752 13753 13754 13755 13756 13757 13758 13759 13760 13761 13762 13763 13764 13765 13766 13767 13768
def match_matrix_tensor(x,
                        y,
                        channel_num,
                        act=None,
                        param_attr=None,
                        dtype='float32',
                        name=None):
    """
    Calculate the semantic matching matrix of two word sequences with variable length.
    Given a query A of length `n` and a title B of length `m`, the input shape are respectively
    [n, h] and [m, h], which h is hidden_size. If :attr:`channel_num` is set to 3,
    it will generate a learnable parameter matrix W with shape [h, 3, h].
    Then the semantic matching matrix of query A and title B is calculated by 
    A * W * B.T = [n, h]*[h, 3, h]*[h, m] = [n, 3, m]. The learnable parameter matrix `W` 
    is equivalent to a fully connected layer in the calculation process. If :attr:`act` is provided, 
    the corresponding activation function will be applied to output matrix.
    The :attr:`x` and :attr:`y` should be LodTensor and only one level LoD is supported.

    .. code-block:: text

            Given a 1-level LoDTensor x:
                x.lod =  [[2,                     3,                               ]]
                x.data = [[0.3, 0.1], [0.2, 0.3], [0.5, 0.6], [0.7, 0.1], [0.3, 0.4]]
                x.dims = [5, 2]
            y is a Tensor:
                y.lod =  [[3,                                 1,       ]]
                y.data = [[0.1, 0.2], [0.3, 0.7], [0.9, 0.2], [0.4, 0.1]]
                y.dims = [4, 2]
            set channel_num 2, then we get a 1-level LoDTensor:
                out.lod =  [[12, 6]]   # where 12 = channel_num * x.lod[0][0] * y.lod[0][0]
                out.dims = [18, 1]     # where 18 = 12 + 6

    Args:
        x (Variable): Input variable x which should be 1-level LodTensor.
        y (Variable): Input variable y which should be 1-level LodTensor.
        channel_num (int): The channel number of learnable parameter W.
        act (str, default None): Activation to be applied to the output of this layer.
        param_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for learnable
            parameters/weights of this layer.
        dtype ('float32'): The data type of w data.
        name (str|None): A name for this layer(optional). If set None, the layer will be named automatically. Default: None

    Returns:
        Variable: output with LoD specified by this layer.

    Examples:
        .. code-block:: python

            import numpy as np
            from paddle.fluid import layers

            x_lod_tensor = layers.data(name='x', shape=[10], lod_level=1)
            y_lod_tensor = layers.data(name='y', shape=[10], lod_level=1)
            out, out_tmp = layers.match_matrix_tensor(x=x_lod_tensor, y=y_lod_tensor, channel_num=3)
    """
    helper = LayerHelper('match_matrix_tensor', **locals())

    x_shape = list(x.shape)
    y_shape = list(y.shape)
    assert len(x_shape) == 2 and len(y_shape) == 2 and x_shape[-1] == y_shape[
        -1]

    weight_shape = [x_shape[-1], channel_num, y_shape[-1]]
    w = helper.create_parameter(
        attr=helper.param_attr, shape=weight_shape, dtype=dtype, is_bias=False)
    mm_res = helper.create_variable_for_type_inference(dtype)
    tmp_res = helper.create_variable_for_type_inference(
        dtype, stop_gradient=True)
    helper.append_op(
        type='match_matrix_tensor',
        inputs={
            'X': x,
            'Y': y,
            'W': w,
        },
        outputs={"Out": mm_res,
                 "Tmp": tmp_res},
        attrs={'dim_t': channel_num})

    return helper.append_activation(mm_res), tmp_res


13769 13770 13771 13772 13773 13774 13775 13776 13777 13778 13779 13780 13781 13782 13783 13784 13785 13786 13787 13788 13789 13790 13791 13792 13793 13794 13795 13796 13797 13798 13799 13800 13801 13802 13803 13804 13805 13806 13807 13808 13809 13810 13811 13812 13813 13814 13815 13816 13817 13818 13819 13820 13821 13822 13823 13824 13825 13826 13827 13828 13829 13830 13831 13832 13833 13834 13835 13836 13837 13838 13839 13840 13841 13842 13843 13844 13845 13846 13847 13848 13849 13850
def shard_index(input, index_num, nshards, shard_id, ignore_value=-1):
    """
    This layer creates the sharded index for input. This layers is used in
    model- and data- parallel mixed training generally, in which the index
    data (usually the label) should be recaculated in each trainer according
    to 

    .. math::
        
        assert index_num % nshards == 0

        shard_size = index_num / nshards

        y = x % shard_size if x / shard_size == shard_id else ignore_value

    We take the distributed one-hot representation to show what this layer is
    used for. The distributed one-hot representation is seperated into multiple
    shards, and each shard is filling zeros except the one with the index
    inside. In order to create these sharded representation in each trainer,
    the original index should be recalculated (i.e. sharded) before.

    Examples:
    
        X is a Tensor of integer values:
          X.shape = [4, 1]
          X.data = [[1], [6], [12], [19]]
        
        suppose index_num = 20 and nshards = 2, then we get shard_size = 10
        
        if shard_id == 0, we get the Out:
          Out.shape = [4, 1]
          Out.data = [[1], [6], [-1], [-1]]
        
        if shard_id == 1, we get the Out:
          Out.shape = [4, 1]
          Out.data = [[-1], [-1], [2], [9]]
    
        the default `ignore_value` -1 is used in this example.
    
    Args:
        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

    Returns:
        Variable: The shard index of input.

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
            label = fluid.layers.data(name="label", shape=[1], dtype="int64")
            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 index_num % nshards != 0:
        raise ValueError(
            'The index_num(%d) cannot be evenly divided by nshards(%d)' %
            (index_num, nshards))
    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
13851 13852 13853 13854 13855 13856 13857 13858 13859 13860 13861 13862 13863 13864 13865 13866 13867 13868 13869 13870 13871 13872 13873 13874 13875 13876 13877 13878 13879 13880 13881 13882 13883 13884 13885


@templatedoc()
def hard_swish(x, threshold=6.0, scale=6.0, offset=3.0, name=None):
    """
    ${comment}
    Args:
        x(Varaible): Input of HardSwish operator.
        threshold(float): The threshold parameter of HardSwish operator. Default:threshold=6.0
        scale(float): The scale parameter of HardSwish operator. Default:scale=6.0
        offset(float): The offset parameter of HardSwish operator. Default:offset=3.0
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        Variable: The output tensor with the same shape as input.

    Examples:

        .. code-block:: python

            import paddle.fluid as fluid
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
            y = fluid.layers.hard_swish(x)
    """
    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