nn.py 308.9 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
P
peizhilin 已提交
21
import os
Y
Yu Yang 已提交
22 23
from ..layer_helper import LayerHelper
from ..initializer import Normal, Constant
S
sneaxiy 已提交
24
from ..framework import Variable, OpProtoHolder
Y
yangyaming 已提交
25
from ..param_attr import ParamAttr
S
sneaxiy 已提交
26
from .layer_function_generator import autodoc, templatedoc, _generate_doc_string_
27 28
from .tensor import concat
from . import utils
F
fengjiayi 已提交
29
from .. import unique_name
30
from functools import reduce
31
from .. import core
Y
Yu Yang 已提交
32 33

__all__ = [
X
Xin Pan 已提交
34 35
    'fc',
    'embedding',
P
peizhilin 已提交
36
    'dynamic_lstm',
X
Xin Pan 已提交
37 38 39 40
    'dynamic_lstmp',
    'dynamic_gru',
    'gru_unit',
    'linear_chain_crf',
P
peizhilin 已提交
41
    'crf_decoding',
X
Xin Pan 已提交
42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
    'cos_sim',
    'cross_entropy',
    'square_error_cost',
    'chunk_eval',
    'sequence_conv',
    'conv2d',
    'conv3d',
    'sequence_pool',
    'sequence_softmax',
    'softmax',
    'pool2d',
    'pool3d',
    'batch_norm',
    'beam_search_decode',
    'conv2d_transpose',
    'conv3d_transpose',
    'sequence_expand',
    'sequence_expand_as',
    'sequence_pad',
Y
Yibing Liu 已提交
61
    'sequence_unpad',
X
Xin Pan 已提交
62 63 64 65 66 67 68 69
    'lstm_unit',
    'reduce_sum',
    'reduce_mean',
    'reduce_max',
    'reduce_min',
    'reduce_prod',
    'sequence_first_step',
    'sequence_last_step',
Y
Yibing Liu 已提交
70
    'sequence_slice',
X
Xin Pan 已提交
71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87
    'dropout',
    'split',
    'ctc_greedy_decoder',
    'edit_distance',
    'l2_normalize',
    'matmul',
    'topk',
    'warpctc',
    'sequence_reshape',
    'transpose',
    'im2sequence',
    'nce',
    'hsigmoid',
    'beam_search',
    'row_conv',
    'multiplex',
    'layer_norm',
D
Dun 已提交
88
    'group_norm',
X
Xin Pan 已提交
89 90 91 92 93 94 95 96 97 98 99 100
    'softmax_with_cross_entropy',
    'smooth_l1',
    'one_hot',
    'autoincreased_step_counter',
    'reshape',
    'squeeze',
    'unsqueeze',
    'lod_reset',
    'lrn',
    'pad',
    'pad_constant_like',
    'label_smooth',
P
peizhilin 已提交
101
    'roi_pool',
J
jerrywgz 已提交
102
    'roi_align',
X
Xin Pan 已提交
103 104 105 106
    'dice_loss',
    'image_resize',
    'image_resize_short',
    'resize_bilinear',
107
    'resize_nearest',
X
Xin Pan 已提交
108 109 110 111 112 113
    'gather',
    'scatter',
    'sequence_scatter',
    'random_crop',
    'mean_iou',
    'relu',
C
chengduo 已提交
114
    'selu',
X
Xin Pan 已提交
115 116 117
    'log',
    'crop',
    'rank_loss',
M
minqiyang 已提交
118
    'margin_rank_loss',
X
Xin Pan 已提交
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161
    'elu',
    'relu6',
    'pow',
    'stanh',
    'hard_sigmoid',
    'swish',
    'prelu',
    'brelu',
    'leaky_relu',
    'soft_relu',
    'flatten',
    'sequence_mask',
    'stack',
    'pad2d',
    'unstack',
    'sequence_enumerate',
    'expand',
    'sequence_concat',
    'scale',
    'elementwise_add',
    'elementwise_div',
    'elementwise_sub',
    'elementwise_mul',
    'elementwise_max',
    'elementwise_min',
    'elementwise_pow',
    'uniform_random_batch_size_like',
    'gaussian_random',
    'sampling_id',
    'gaussian_random_batch_size_like',
    'sum',
    'slice',
    'shape',
    '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
162
    'space_to_depth',
W
whs 已提交
163
    'affine_grid',
S
sneaxiy 已提交
164
    'sequence_reverse',
165
    'affine_channel',
B
barrierye 已提交
166
    'similarity_focus',
M
minqiyang 已提交
167
    'hash',
D
dengkaipeng 已提交
168
    'grid_sampler',
G
gmcather 已提交
169 170
    'log_loss',
    'add_position_encoding',
Q
Qiao Longfei 已提交
171
    'bilinear_tensor_product',
Y
Yu Yang 已提交
172 173 174 175 176 177 178 179 180
]


def fc(input,
       size,
       num_flatten_dims=1,
       param_attr=None,
       bias_attr=None,
       act=None,
J
Jacek Czaja 已提交
181
       is_test=False,
182
       name=None):
Y
Yu Yang 已提交
183
    """
184
    **Fully Connected Layer**
Y
Yu Yang 已提交
185

186 187 188 189 190 191 192 193
    This function creates a fully connected layer in the network. It can take
    multiple tensors as its inputs. It creates a variable called weights for
    each input tensor, 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 coresponding weight to produce an output Tensor.
    If multiple input tensors are given, the results of multiple multiplications
    will be sumed up. If bias_attr is not None, a bias variable will be created
    and added to the output. Finally, if activation is not None, it will be applied
F
fengjiayi 已提交
194
    to the output as well.
C
caoying03 已提交
195

C
caoying03 已提交
196
    This process can be formulated as follows:
197 198 199

    .. math::

200
        Out = Act({\sum_{i=0}^{N-1}X_iW_i + b})
201 202 203

    In the above equation:

C
caoying03 已提交
204 205 206 207
    * :math:`N`: Number of the input.
    * :math:`X_i`: The input tensor.
    * :math:`W`: The weights created by this layer.
    * :math:`b`: The bias parameter created by this layer (if needed).
208
    * :math:`Act`: The activation function.
C
caoying03 已提交
209
    * :math:`Out`: The output tensor.
Y
Yu Yang 已提交
210 211

    Args:
R
ranqiu 已提交
212 213 214 215 216 217 218 219 220 221 222 223 224 225 226
        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
            `X` is a 6-dimensional tensor with a shape [2, 3, 4, 5, 6], and `num_flatten_dims` = 3.
            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
227 228
            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
229
        act (str, default None): Activation to be applied to the output of this layer.
J
Jacek Czaja 已提交
230
        is_test(bool): A flag indicating whether execution is in test phase.
R
ranqiu 已提交
231
        name (str, default None): The name of this layer.
Y
Yu Yang 已提交
232

233
    Returns:
F
fengjiayi 已提交
234
        Variable: The transformation result.
235 236

    Raises:
C
caoying03 已提交
237
        ValueError: If rank of the input tensor is less than 2.
238 239 240 241

    Examples:
        .. code-block:: python

F
fengjiayi 已提交
242
          data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
243
          fc = fluid.layers.fc(input=data, size=1000, act="tanh")
Y
Yu Yang 已提交
244
    """
C
caoying03 已提交
245

C
caoying03 已提交
246
    helper = LayerHelper("fc", **locals())
Y
Yu Yang 已提交
247 248 249 250

    dtype = helper.input_dtype()

    mul_results = []
251 252
    for input_var, param_attr in helper.iter_inputs_and_params():
        input_shape = input_var.shape
Y
Yu Yang 已提交
253 254 255
        param_shape = [
            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
        ] + [size]
Y
ying 已提交
256

Y
Yu Yang 已提交
257
        w = helper.create_parameter(
258
            attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False)
X
Xin Pan 已提交
259
        tmp = helper.create_variable_for_type_inference(dtype)
260
        helper.append_op(
261 262 263
            type="mul",
            inputs={"X": input_var,
                    "Y": w},
264
            outputs={"Out": tmp},
M
mozga-intel 已提交
265 266
            attrs={"x_num_col_dims": num_flatten_dims,
                   "y_num_col_dims": 1})
267 268 269 270
        mul_results.append(tmp)

    if len(mul_results) == 1:
        pre_bias = mul_results[0]
271
    else:
X
Xin Pan 已提交
272
        pre_bias = helper.create_variable_for_type_inference(dtype)
273
        helper.append_op(
274 275 276
            type="sum",
            inputs={"X": mul_results},
            outputs={"Out": pre_bias},
X
Xin Pan 已提交
277
            attrs={"use_mkldnn": False})
278 279 280 281
    # 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
282 283


284 285 286
def embedding(input,
              size,
              is_sparse=False,
287
              is_distributed=False,
288 289 290
              padding_idx=None,
              param_attr=None,
              dtype='float32'):
Y
Yu Yang 已提交
291
    """
292 293
    **Embedding Layer**

294
    This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in
295 296
    a lookup table. The result of this lookup is the embedding of each ID in the
    :attr:`input`.
297 298 299

    All the input variables are passed in as local variables to the LayerHelper
    constructor.
Y
Yu Yang 已提交
300 301

    Args:
302 303 304 305 306
        input(Variable): The tensor variable containing the IDs.
        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.
307
        is_distributed(bool): Whether to run lookup table from remote parameter server.
308 309
        padding_idx(int|long|None): If :attr:`None`, it makes no effect to lookup.
            Otherwise the given :attr:`padding_idx` indicates padding the output
310
            with zeros whenever lookup encounters it in :attr:`input`. If
311
            :math:`padding_idx < 0`, the :attr:`padding_idx` to use in lookup is
312 313
            :math:`size[0] + dim`.
        param_attr(ParamAttr): Parameters for this layer
314
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc
Y
Yu Yang 已提交
315

316 317 318
    Returns:
        Variable: The tensor variable storing the embeddings of the \
                  supplied inputs.
Y
Yu Yang 已提交
319

320 321
    Examples:
        .. code-block:: python
Y
Yu Yang 已提交
322

C
chengduoZH 已提交
323
          dict_size = len(dataset.ids)
324
          data = fluid.layers.data(name='ids', shape=[32, 32], dtype='float32')
C
chengduoZH 已提交
325
          fc = fluid.layers.embedding(input=data, size=[dict_size, 16])
Y
Yu Yang 已提交
326 327 328 329 330
    """

    helper = LayerHelper('embedding', **locals())
    w = helper.create_parameter(
        attr=helper.param_attr, shape=size, dtype=dtype, is_bias=False)
X
Xin Pan 已提交
331
    tmp = helper.create_variable_for_type_inference(dtype)
332 333
    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
334 335 336 337 338
    helper.append_op(
        type='lookup_table',
        inputs={'Ids': input,
                'W': w},
        outputs={'Out': tmp},
339 340 341 342 343
        attrs={
            'is_sparse': is_sparse,
            'is_distributed': is_distributed,
            'padding_idx': padding_idx
        })
Y
Yu Yang 已提交
344 345 346
    return tmp


W
wopeizl 已提交
347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362
@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}
P
peizhilin 已提交
363

W
wopeizl 已提交
364 365 366 367 368 369 370 371 372 373 374
    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
375

W
wopeizl 已提交
376 377 378 379
                               - 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
380

W
wopeizl 已提交
381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466
                               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

            hidden_dim = 512
            forward_proj = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
                                           bias_attr=False)
            forward, _ = fluid.layers.dynamic_lstm(
                input=forward_proj, size=hidden_dim * 4, use_peepholes=False)
    """
    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
467 468


Y
Yibing Liu 已提交
469 470 471 472 473 474 475 476 477 478 479
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',
                  proj_activation='tanh',
480 481
                  dtype='float32',
                  name=None):
Y
Yibing Liu 已提交
482 483 484
    """
    **Dynamic LSTMP Layer**

485 486 487 488 489 490
    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
491 492 493 494 495

    The formula is as follows:

    .. math::

496
        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
497

498
        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
499

500
        \\tilde{c_t} & = act_g(W_{cx}x_t + W_{cr}r_{t-1} + b_c)
Y
Yibing Liu 已提交
501

502
        o_t & = \sigma(W_{ox}x_{t} + W_{or}r_{t-1} + W_{oc}c_t + b_o)
Y
Yibing Liu 已提交
503

504
        c_t & = f_t \odot c_{t-1} + i_t \odot \\tilde{c_t}
Y
Yibing Liu 已提交
505

506
        h_t & = o_t \odot act_h(c_t)
Y
Yibing Liu 已提交
507

508
        r_t & = \overline{act_h}(W_{rh}h_t)
Y
Yibing Liu 已提交
509

Y
Yibing Liu 已提交
510 511 512 513 514 515
    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, \
516
          we use vectors to reprenset these diagonal weight matrices.
Y
Yibing Liu 已提交
517
    * :math:`b`: Denotes bias vectors (e.g. :math:`b_i` is the input gate \
518
          bias vector).
Y
Yibing Liu 已提交
519 520 521
    * :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 \
522
          the same size as the cell output activation vector :math:`h`.
Y
Yibing Liu 已提交
523
    * :math:`h`: The hidden state.
524
    * :math:`r`: The recurrent projection of the hidden state.
Y
Yibing Liu 已提交
525 526
    * :math:`\\tilde{c_t}`: The candidate hidden state, whose \
          computation is based on the current input and previous hidden state.
527
    * :math:`\odot`: The element-wise product of the vectors.
Y
Yibing Liu 已提交
528
    * :math:`act_g` and :math:`act_h`: The cell input and cell output \
529
          activation functions and `tanh` is usually used for them.
Y
Yibing Liu 已提交
530 531
    * :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
532 533 534 535

    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.
536

Y
Yibing Liu 已提交
537 538 539 540 541 542 543 544 545 546 547 548
    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.
549
        param_attr(ParamAttr|None): The parameter attribute for the learnable
Y
Yibing Liu 已提交
550 551
                               hidden-hidden weight and projection weight.

552 553
                               - Hidden-hidden weight = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}.
554 555
                               - 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
556 557
                                 size.
                               - Projection weight = {:math:`W_{rh}`}.
558
                               - The shape of projection weight is (D x P).
C
chengduo 已提交
559 560 561 562 563

                               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.
564
        bias_attr(ParamAttr|None): The bias attribute for the learnable bias
Y
Yibing Liu 已提交
565 566 567 568 569 570
                              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`}.
571
                                - The shape is (1 x 4D).
Y
Yibing Liu 已提交
572 573 574
                              2. `use_peepholes = True`
                                - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
                                                 W_{fc}, W_{oc}`}.
575
                                - The shape is (1 x 7D).
C
chengduo 已提交
576 577 578 579 580

                              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
581 582 583 584 585 586 587 588 589
        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.
590
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
Y
Yibing Liu 已提交
591 592
                              default "tanh".
        proj_activation(str): The activation for projection output.
593
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
Y
Yibing Liu 已提交
594 595
                              default "tanh".
        dtype(str): Data type. Choices = ["float32", "float64"], default "float32".
596 597
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
Y
Yibing Liu 已提交
598 599

    Returns:
600 601 602 603
        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
604 605

    Examples:
606

Y
Yibing Liu 已提交
607 608
        .. code-block:: python

609 610 611 612
            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
613
            hidden_dim, proj_dim = 512, 256
614
            fc_out = fluid.layers.fc(input=emb, size=hidden_dim * 4,
Y
Yibing Liu 已提交
615
                                     act=None, bias_attr=None)
616 617 618
            proj_out, _ = fluid.layers.dynamic_lstmp(input=fc_out,
                                                     size=hidden_dim * 4,
                                                     proj_size=proj_dim,
Y
Yibing Liu 已提交
619 620 621 622
                                                     use_peepholes=False,
                                                     is_reverse=True,
                                                     cell_activation="tanh",
                                                     proj_activation="tanh")
Y
Yibing Liu 已提交
623
    """
624

C
chengduo 已提交
625
    assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp."
Y
Yibing Liu 已提交
626
    helper = LayerHelper('lstmp', **locals())
M
minqiyang 已提交
627
    size = size // 4
Y
Yibing Liu 已提交
628 629 630 631 632 633 634 635 636 637
    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
638 639 640 641 642 643
    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)
Y
Yibing Liu 已提交
644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671

    helper.append_op(
        type='lstmp',
        inputs={
            'Input': input,
            'Weight': weight,
            'ProjWeight': proj_weight,
            'Bias': bias
        },
        outputs={
            'Projection': projection,
            'Cell': cell,
            'OrderedP0': ordered_proj0,
            'BatchHidden': batch_hidden,
            '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,
            'proj_activation': proj_activation
        })
    return projection, cell


G
guosheng 已提交
672 673 674 675 676 677 678 679 680
def dynamic_gru(input,
                size,
                param_attr=None,
                bias_attr=None,
                is_reverse=False,
                gate_activation='sigmoid',
                candidate_activation='tanh',
                h_0=None):
    """
681
    **Gated Recurrent Unit (GRU) Layer**
G
guosheng 已提交
682

683
    Refer to `Empirical Evaluation of Gated Recurrent Neural Networks on
684
    Sequence Modeling <https://arxiv.org/abs/1412.3555>`_ .
685

G
guosheng 已提交
686 687 688 689 690 691 692 693 694
    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)
695

G
guosheng 已提交
696
        h_t & = (1-u_t) \odot h_{t-1} + u_t \odot \\tilde{h_t}
697

G
guosheng 已提交
698
    The :math:`\odot` is the element-wise product of the vectors. :math:`act_g`
699 700
    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
701 702 703 704
    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
705
    to use fully-connect layer before GRU layer.
G
guosheng 已提交
706 707

    Args:
708 709
        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
710
            Variable is a matrix with shape :math:`(T \\times 3D)`, where
711
            :math:`T` is the total time steps in this mini-batch, :math:`D`
G
guosheng 已提交
712 713
            is the hidden size.
        size(int): The dimension of the gru cell.
714
        param_attr(ParamAttr|None): The parameter attribute for the learnable
G
guosheng 已提交
715 716
            hidden-hidden weight matrix. Note:

717
            - The shape of the weight matrix is :math:`(T \\times 3D)`, where
G
guosheng 已提交
718
              :math:`D` is the hidden size.
719
            - All elements in the weight matrix can be divided into two parts.
G
guosheng 已提交
720
              The first part are weights of the update gate and reset gate with
721
              shape :math:`(D \\times 2D)`, and the second part are weights for
G
guosheng 已提交
722
              candidate hidden state with shape :math:`(D \\times D)`.
723 724 725 726 727

            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
728
            of GRU. Note that the bias with :math:`(1 \\times 3D)` concatenates
729
            the bias in the update gate, reset gate and candidate calculations.
730 731 732
            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
733 734
            bias_attr. If the Initializer of the bias_attr is not set, the bias
            is initialized zero. Default: None.
735
        is_reverse(bool): Whether to compute reversed GRU, default
G
guosheng 已提交
736 737 738
            :attr:`False`.
        gate_activation(str): The activation for update gate and reset gate.
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "sigmoid".
739
        candidate_activation(str): The activation for candidate hidden state.
G
guosheng 已提交
740
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh".
741 742 743 744
        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
745 746

    Returns:
G
guosheng 已提交
747
        Variable: The hidden state of GRU. The shape is :math:`(T \\times D)`, \
748
            and sequence length is the same with the input.
749

G
guosheng 已提交
750
    Examples:
751

G
guosheng 已提交
752 753
        .. code-block:: python

754 755 756 757
            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
758
            hidden_dim = 512
759
            x = fluid.layers.fc(input=emb, size=hidden_dim * 3)
G
guosheng 已提交
760 761 762 763 764 765 766 767 768 769
            hidden = fluid.layers.dynamic_gru(input=x, dim=hidden_dim)
    """

    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
770
    batch_size = input.shape[0]
G
guosheng 已提交
771
    inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
S
sneaxiy 已提交
772
    if h_0:
G
guosheng 已提交
773
        assert h_0.shape == (
Y
Yancey 已提交
774 775 776
            batch_size, size
        ), 'The shape of h0 should be(batch_size, %d)' % size
        inputs['H0'] = h_0
G
guosheng 已提交
777

X
Xin Pan 已提交
778 779 780 781
    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
782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799

    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,
            'activation': candidate_activation
        })
    return hidden


Y
Yu Yang 已提交
800 801 802
def gru_unit(input,
             hidden,
             size,
803 804
             param_attr=None,
             bias_attr=None,
Y
Yu Yang 已提交
805
             activation='tanh',
806
             gate_activation='sigmoid'):
Y
Yu Yang 已提交
807
    """
808
    GRU unit layer. The equation of a gru step is:
Y
Yu Yang 已提交
809

810 811
        .. math::
            u_t & = actGate(xu_{t} + W_u h_{t-1} + b_u)
Y
Yu Yang 已提交
812

813
            r_t & = actGate(xr_{t} + W_r h_{t-1} + b_r)
Y
Yu Yang 已提交
814

815
            m_t & = actNode(xm_t + W_c dot(r_t, h_{t-1}) + b_m)
816

817
            h_t & = dot((1-u_t), m_t) + dot(u_t, h_{t-1})
818 819

    The inputs of gru unit includes :math:`z_t`, :math:`h_{t-1}`. In terms
820 821 822
    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
823 824
    connected layer has to be applied, such that :math:`z_t = W_{fc}x_t`.

825 826
    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
827 828 829
    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`.
830 831 832

    Args:
        input (Variable): The fc transformed input value of current step.
833
        hidden (Variable): The hidden value of gru unit from previous step.
834
        size (integer): The input dimension value.
835 836 837 838 839 840 841 842 843 844 845 846 847 848
        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
849
            of GRU. Note that the bias with :math:`(1 \\times 3D)` concatenates
850
            the bias in the update gate, reset gate and candidate calculations.
851 852 853
            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
854 855
            bias_attr. If the Initializer of the bias_attr is not set, the bias
            is initialized zero. Default: None.
856 857 858 859
        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
860

861 862 863 864 865 866
    Returns:
        tuple: The hidden value, reset-hidden value and gate values.

    Examples:

        .. code-block:: python
Y
Yu Yang 已提交
867

868
             # assuming we have x_t_data and prev_hidden of size=10
869
             x_t = fluid.layers.fc(input=x_t_data, size=30)
870 871
             hidden_val, r_h_val, gate_val = fluid.layers.gru_unit(input=x_t,
                                                    hidden = prev_hidden)
Y
Yu Yang 已提交
872 873 874 875 876 877 878 879 880 881 882 883

    """
    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
884
    size = size // 3
Y
Yu Yang 已提交
885 886

    # create weight
887 888
    weight = helper.create_parameter(
        attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype)
Y
Yu Yang 已提交
889

X
Xin Pan 已提交
890 891 892
    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)
893
    inputs = {'Input': input, 'HiddenPrev': hidden, 'Weight': weight}
Y
Yu Yang 已提交
894
    # create bias
895
    if helper.bias_attr:
Y
Yu Yang 已提交
896 897 898
        bias_size = [1, 3 * size]
        bias = helper.create_parameter(
            attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
899
        inputs['Bias'] = bias
Y
Yu Yang 已提交
900 901 902

    helper.append_op(
        type='gru_unit',
903
        inputs=inputs,
Y
Yu Yang 已提交
904 905 906 907 908 909
        outputs={
            'Gate': gate,
            'ResetHiddenPrev': reset_hidden_pre,
            'Hidden': updated_hidden,
        },
        attrs={
910 911
            'activation': 2,  # tanh
            'gate_activation': 1,  # sigmoid
Y
Yu Yang 已提交
912 913 914 915 916
        })

    return updated_hidden, reset_hidden_pre, gate


Y
yuyang18 已提交
917
@templatedoc()
918
def linear_chain_crf(input, label, param_attr=None):
Y
yuyang18 已提交
919 920 921 922 923 924 925
    """
    Linear Chain CRF.

    ${comment}

    Args:
        input(${emission_type}): ${emission_comment}
D
dzhwinter 已提交
926
        input(${transition_type}): ${transition_comment}
Y
yuyang18 已提交
927 928 929 930
        label(${label_type}): ${label_comment}
        param_attr(ParamAttr): The attribute of the learnable parameter.

    Returns:
D
dzhwinter 已提交
931 932 933
        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
934 935

    """
Y
Yu Yang 已提交
936 937 938 939 940 941
    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
942 943 944 945 946 947 948 949
    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())
Y
Yu Yang 已提交
950 951 952 953 954 955 956 957 958 959 960 961 962 963 964
    helper.append_op(
        type='linear_chain_crf',
        inputs={"Emission": [input],
                "Transition": transition,
                "Label": label},
        outputs={
            "Alpha": [alpha],
            "EmissionExps": [emission_exps],
            "TransitionExps": transition_exps,
            "LogLikelihood": log_likelihood
        })

    return log_likelihood


W
wopeizl 已提交
965 966 967 968
@templatedoc()
def crf_decoding(input, param_attr, label=None):
    """
    ${comment}
Y
yuyang18 已提交
969

W
wopeizl 已提交
970 971
    Args:
        input(${emission_type}): ${emission_comment}
Y
yi.wu 已提交
972

W
wopeizl 已提交
973
        param_attr(ParamAttr): The parameter attribute for training.
Y
yi.wu 已提交
974

W
wopeizl 已提交
975
        label(${label_type}): ${label_comment}
Y
yuyang18 已提交
976

W
wopeizl 已提交
977 978
    Returns:
        Variable: ${viterbi_path_comment}
979

W
wopeizl 已提交
980 981
    Examples:
        .. code-block:: python
Y
yi.wu 已提交
982

W
wopeizl 已提交
983 984 985 986 987 988 989 990 991 992
           crf_decode = layers.crf_decoding(
                input=hidden, param_attr=ParamAttr(name="crfw"))
    """
    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],
P
peizhilin 已提交
993
                "Transition": transition,
W
wopeizl 已提交
994 995
                "Label": label},
        outputs={"ViterbiPath": [viterbi_path]})
Y
Yu Yang 已提交
996

W
wopeizl 已提交
997
    return viterbi_path
Y
Yu Yang 已提交
998 999


Y
yi.wu 已提交
1000
@templatedoc()
F
fengjiayi 已提交
1001
def cos_sim(X, Y):
Y
Yu Yang 已提交
1002
    """
Y
yi.wu 已提交
1003 1004 1005
    ${comment}

    Args:
1006 1007
        X (Variable): ${x_comment}.
        Y (Variable): ${y_comment}.
F
fengjiayi 已提交
1008

Y
yi.wu 已提交
1009
    Returns:
1010
        Variable: the output of cosine(X, Y).
Y
Yu Yang 已提交
1011
    """
F
fengjiayi 已提交
1012
    helper = LayerHelper('cos_sim', **locals())
X
Xin Pan 已提交
1013 1014 1015
    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
1016 1017 1018 1019 1020 1021 1022 1023 1024 1025
    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
1026 1027 1028 1029 1030
def dropout(x,
            dropout_prob,
            is_test=False,
            seed=None,
            name=None,
P
phlrain 已提交
1031
            dropout_implementation="downgrade_in_infer"):
1032 1033 1034 1035 1036
    """
    Computes dropout.

    Drop or keep each element of `x` independently. Dropout is a regularization
    technique for reducing overfitting by preventing neuron co-adaption during
1037
    training. The dropout operator randomly sets (according to the given dropout
1038 1039 1040 1041
    probability) the outputs of some units to zero, while others are remain
    unchanged.

    Args:
1042 1043
        x (Variable): The input tensor variable.
        dropout_prob (float): Probability of setting units to zero.
1044 1045 1046 1047 1048 1049 1050
        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.
P
phlrain 已提交
1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061
        dropout_implementation(string): ['downgrade_in_infer'(defauld)|'upscale_in_train']
                                        1. downgrade_in_infer(default), downgrade the outcome at inference
                                           train: out = input * mask
                                           inference: out = input * dropout_prob
                                           (make is a tensor same shape with input, value is 0 or 1
                                            ratio of 0 is dropout_prob)
                                        2. upscale_in_train, upscale the outcome at training time
                                           train: out = input * mask / ( 1.0 - dropout_prob )
                                           inference: out = input
                                           (make is a tensor same shape with input, value is 0 or 1
                                            ratio of 0 is dropout_prob)
1062
                                           dropout op can be removed from the program.
P
phlrain 已提交
1063
                                           the program will be efficient
1064

P
phlrain 已提交
1065

1066 1067

    Returns:
1068
        Variable: A tensor variable is the shape with `x`.
1069 1070

    Examples:
1071

1072 1073
        .. code-block:: python

1074 1075
            x = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
            droped = fluid.layers.dropout(x, dropout_prob=0.5)
1076 1077
    """

F
fengjiayi 已提交
1078
    helper = LayerHelper('dropout', **locals())
X
Xin Pan 已提交
1079 1080 1081
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    mask = helper.create_variable_for_type_inference(
        dtype=x.dtype, stop_gradient=True)
C
chengduo 已提交
1082 1083 1084 1085

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

1086 1087 1088 1089 1090
    helper.append_op(
        type='dropout',
        inputs={'X': [x]},
        outputs={'Out': [out],
                 'Mask': [mask]},
1091 1092 1093 1094
        attrs={
            'dropout_prob': dropout_prob,
            'is_test': is_test,
            'fix_seed': seed is not None,
P
phlrain 已提交
1095 1096
            'seed': seed if seed is not None else 0,
            'dropout_implementation': dropout_implementation,
1097
        })
1098 1099 1100
    return out


1101
def cross_entropy(input, label, soft_label=False, ignore_index=-100):
Y
Yu Yang 已提交
1102
    """
Y
Yibing Liu 已提交
1103 1104
    **Cross Entropy Layer**

1105 1106 1107
    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
1108 1109

    1) One-hot cross-entropy:
F
fengjiayi 已提交
1110
        `soft_label = False`, `Label[i, 0]` indicates the class index for sample i:
Y
yangyaming 已提交
1111

Y
Yibing Liu 已提交
1112
        .. math::
Y
yangyaming 已提交
1113

Y
Yibing Liu 已提交
1114 1115 1116
            Y[i] = -\log(X[i, Label[i]])

    2) Soft-label cross-entropy:
F
fengjiayi 已提交
1117 1118
        `soft_label = True`, `Label[i, j]` indicates the soft label of class j
        for sample i:
Y
Yibing Liu 已提交
1119 1120 1121 1122 1123

        .. math::

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

Y
Yibing Liu 已提交
1124
       Please make sure that in this case the summation of each row of `label`
Y
Yibing Liu 已提交
1125 1126 1127
       equals one.

    3) One-hot cross-entropy with vecterized `label`:
F
fengjiayi 已提交
1128 1129
         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
1130
         to a one-hot cross-entropy with one-hot label representation.
Y
yangyaming 已提交
1131

Y
Yibing Liu 已提交
1132
    Args:
Y
yangyaming 已提交
1133
        input (Variable|list):  a 2-D tensor with shape [N x D], where N is the
1134 1135 1136 1137
                                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
1138
        label (Variable|list): the ground truth which is a 2-D tensor. When
1139 1140 1141 1142
                               `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
1143
        soft_label (bool): a flag indicating whether to
1144
                                           interpretate the given labels as soft
1145
                                           labels. Default: `False`.
M
minqiyang 已提交
1146 1147
        ignore_index (int): Specifies a target value that is ignored and does
                            not contribute to the input gradient. Only valid
1148
                            if soft_label is set to False. Default: -100
Y
Yibing Liu 已提交
1149 1150 1151 1152 1153

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

    Raises:
1154 1155 1156 1157 1158
        `ValueError`: 1) the 1st dimension of `input` and `label` are not equal.
                      2) when `soft_label == True`, and the 2nd dimension of
                         `input` and `label` are not equal.
                      3) when `soft_label == False`, and the 2nd dimension of
                         `label` is not 1.
Y
Yibing Liu 已提交
1159 1160 1161 1162 1163 1164

    Examples:
        .. code-block:: python

          predict = fluid.layers.fc(input=net, size=classdim, act='softmax')
          cost = fluid.layers.cross_entropy(input=predict, label=label)
Y
Yu Yang 已提交
1165
    """
F
fengjiayi 已提交
1166
    helper = LayerHelper('cross_entropy', **locals())
X
Xin Pan 已提交
1167
    out = helper.create_variable_for_type_inference(dtype=input.dtype)
Y
Yu Yang 已提交
1168 1169 1170 1171 1172
    helper.append_op(
        type='cross_entropy',
        inputs={'X': [input],
                'Label': [label]},
        outputs={'Y': [out]},
1173 1174
        attrs={"soft_label": soft_label,
               "ignore_index": ignore_index})
Y
Yu Yang 已提交
1175 1176 1177
    return out


F
fengjiayi 已提交
1178
def square_error_cost(input, label):
Y
Yu Yang 已提交
1179
    """
1180 1181
    **Square error cost layer**

1182 1183
    This layer accepts input predictions and target label and returns the
    squared error cost.
Y
ying 已提交
1184

1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197
    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:
1198 1199
        input (Variable): Input tensor, has predictions.
        label (Variable): Label tensor, has target labels.
1200 1201

    Returns:
G
guosheng 已提交
1202
        Variable: The tensor variable storing the element-wise squared error \
1203
                  difference of input and label.
1204 1205 1206 1207 1208 1209 1210 1211

    Examples:
        .. code-block:: python

          y = layers.data(name='y', shape=[1], dtype='float32')
          y_predict = layers.data(name='y_predict', shape=[1], dtype='float32')
          cost = layers.square_error_cost(input=y_predict, label=y)

Y
Yu Yang 已提交
1212
    """
F
fengjiayi 已提交
1213
    helper = LayerHelper('square_error_cost', **locals())
X
Xin Pan 已提交
1214
    minus_out = helper.create_variable_for_type_inference(dtype=input.dtype)
Y
Yu Yang 已提交
1215 1216 1217 1218 1219 1220
    helper.append_op(
        type='elementwise_sub',
        inputs={'X': [input],
                'Y': [label]},
        outputs={'Out': [minus_out]})

X
Xin Pan 已提交
1221
    square_out = helper.create_variable_for_type_inference(dtype=input.dtype)
Y
Yu Yang 已提交
1222
    helper.append_op(
F
fengjiayi 已提交
1223 1224
        type='square', inputs={'X': [minus_out]},
        outputs={'Out': [square_out]})
Y
Yu Yang 已提交
1225 1226 1227
    return square_out


Y
yi.wu 已提交
1228
@templatedoc()
Y
Yu Yang 已提交
1229 1230 1231 1232
def chunk_eval(input,
               label,
               chunk_scheme,
               num_chunk_types,
F
fengjiayi 已提交
1233
               excluded_chunk_types=None):
Y
Yu Yang 已提交
1234
    """
Y
yi.wu 已提交
1235
    **Chunk Evaluator**
Y
yi.wu 已提交
1236

Y
yangyaming 已提交
1237
    This function computes and outputs the precision, recall and
1238
    F1-score of chunk detection.
Y
yi.wu 已提交
1239

Y
yi.wu 已提交
1240 1241 1242 1243 1244 1245 1246 1247
    For some basics of chunking, please refer to
    'Chunking with Support Vector Machines <https://aclanthology.info/pdf/N/N01/N01-1025.pdf>'.

    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
1248

Y
yi.wu 已提交
1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273
       ====== ====== ======  =====  ==  ============   =====  ===== =====  ==  =========
              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
1274

Y
yi.wu 已提交
1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298
       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
1299
    Args:
1300 1301 1302 1303 1304
        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}
F
fengjiayi 已提交
1305

Y
yi.wu 已提交
1306
    Returns:
Y
update  
yi.wu 已提交
1307 1308 1309
        tuple: tuple containing: precision, recall, f1_score,
        num_infer_chunks, num_label_chunks,
        num_correct_chunks
1310

Y
yi.wu 已提交
1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322
    Examples:
        .. code-block:: python

            crf = fluid.layers.linear_chain_crf(
                input=hidden, label=label, param_attr=ParamAttr(name="crfw"))
            crf_decode = fluid.layers.crf_decoding(
                input=hidden, param_attr=ParamAttr(name="crfw"))
            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
1323
    """
F
fengjiayi 已提交
1324
    helper = LayerHelper("chunk_eval", **locals())
Y
Yu Yang 已提交
1325 1326

    # prepare output
X
Xin Pan 已提交
1327 1328 1329 1330 1331 1332 1333
    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
1334 1335 1336 1337 1338 1339 1340 1341

    helper.append_op(
        type="chunk_eval",
        inputs={"Inference": [input],
                "Label": [label]},
        outputs={
            "Precision": [precision],
            "Recall": [recall],
1342 1343 1344 1345
            "F1-Score": [f1_score],
            "NumInferChunks": [num_infer_chunks],
            "NumLabelChunks": [num_label_chunks],
            "NumCorrectChunks": [num_correct_chunks]
Y
Yu Yang 已提交
1346 1347 1348
        },
        attrs={
            "num_chunk_types": num_chunk_types,
G
guosheng 已提交
1349 1350
            "chunk_scheme": chunk_scheme,
            "excluded_chunk_types": excluded_chunk_types or []
Y
Yu Yang 已提交
1351
        })
1352 1353
    return (precision, recall, f1_score, num_infer_chunks, num_label_chunks,
            num_correct_chunks)
Y
Yu Yang 已提交
1354 1355


1356
@templatedoc()
Y
Yu Yang 已提交
1357 1358 1359 1360 1361 1362 1363
def sequence_conv(input,
                  num_filters,
                  filter_size=3,
                  filter_stride=1,
                  padding=None,
                  bias_attr=None,
                  param_attr=None,
C
chengduo 已提交
1364 1365
                  act=None,
                  name=None):
Y
Yu Yang 已提交
1366 1367 1368 1369
    """
    This function creates the op for sequence_conv, using the inputs and
    other convolutional configurations for the filters and stride as given
    in the input parameters to the function.
1370 1371 1372 1373 1374 1375 1376

    Args:
        input (Variable): ${x_comment}
        num_filters (int): number of filters.
        filter_size (int): the filter size (H and W).
        filter_stride (int): stride of the filter.
        padding (bool): if True, add paddings.
C
chengduo 已提交
1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389
        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
1390

1391 1392
    Returns:
        Variable: output of sequence_conv
Y
Yu Yang 已提交
1393 1394 1395 1396 1397 1398 1399
    """

    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
1400
    pre_bias = helper.create_variable_for_type_inference(dtype)
Y
Yu Yang 已提交
1401 1402 1403 1404 1405 1406 1407 1408 1409 1410

    helper.append_op(
        type='sequence_conv',
        inputs={
            'X': [input],
            'Filter': [filter_param],
        },
        outputs={"Out": pre_bias},
        attrs={
            'contextStride': filter_stride,
M
minqiyang 已提交
1411
            'contextStart': -int(filter_size // 2),
Y
Yu Yang 已提交
1412 1413 1414 1415 1416 1417
            'contextLength': filter_size
        })
    pre_act = helper.append_bias_op(pre_bias)
    return helper.append_activation(pre_act)


C
chengduo 已提交
1418
def sequence_softmax(input, use_cudnn=False, name=None):
1419 1420 1421
    """
    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
1422
    input Tensor can be either :math:`[N, 1]` or :math:`[N]`, where :math:`N`
1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438
    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
1439 1440 1441
            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.
1442

1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453
    Returns:
        Variable: output of sequence_softmax

    Examples:

        .. code-block:: python

             x = fluid.layers.data(name='x', shape=[7, 1],
                              dtype='float32', lod_level=1)
             x_sequence_softmax = fluid.layers.sequence_softmax(input=x)
    """
1454 1455
    helper = LayerHelper('sequence_softmax', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
1456
    softmax_out = helper.create_variable_for_type_inference(dtype)
1457 1458 1459 1460 1461 1462 1463 1464
    helper.append_op(
        type="sequence_softmax",
        inputs={"X": input},
        outputs={"Out": softmax_out},
        attrs={"use_cudnn": use_cudnn})
    return softmax_out


C
chengduo 已提交
1465
def softmax(input, use_cudnn=True, name=None):
Q
qiaolongfei 已提交
1466
    """
1467
    The input of the softmax operator is a tensor of any rank. The output tensor
F
fengjiayi 已提交
1468
    has the same shape as the input.
Q
qiaolongfei 已提交
1469

1470 1471 1472 1473 1474 1475
    The input tensor will first be logically flattened to a 2-D matrix. The matrix's
    second dimension(row length) is as same as the last dimension of the input
    tensor, and the first dimension(column length) is the product of all other
    dimensions of the input tensor. For each row of the matrix, the softmax operator
    squashes the K-dimensional(K is the width of the matrix, which is also the size
    of the input tensor's last dimension) vector of arbitrary real values to a
F
fengjiayi 已提交
1476
    K-dimensional vector of real values in the range [0, 1] that add up to 1.
Q
qiaolongfei 已提交
1477 1478 1479 1480 1481 1482 1483

    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
1484
    For each row :math:`i` and each column :math:`j` in the matrix, we have:
Q
qiaolongfei 已提交
1485 1486 1487 1488 1489 1490 1491 1492

    .. 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 \
C
chengduo 已提交
1493 1494 1495
            library is installed.
        name (str|None): A name for this layer(optional). If set None, the layer
            will be named automatically. Default: None.
Q
qiaolongfei 已提交
1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507

    Returns:
        Variable: output of softmax

    Examples:

        .. code-block:: python

             fc = fluid.layers.fc(input=x, size=10)
             softmax = fluid.layers.softmax(input=fc)

    """
1508 1509
    helper = LayerHelper('softmax', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
1510
    softmax_out = helper.create_variable_for_type_inference(dtype)
1511 1512 1513 1514 1515 1516 1517 1518
    helper.append_op(
        type="softmax",
        inputs={"X": input},
        outputs={"Out": softmax_out},
        attrs={"use_cudnn": use_cudnn})
    return softmax_out


Y
Yu Yang 已提交
1519 1520 1521
def conv2d(input,
           num_filters,
           filter_size,
C
chengduoZH 已提交
1522 1523
           stride=1,
           padding=0,
1524
           dilation=1,
Y
Yu Yang 已提交
1525 1526 1527
           groups=None,
           param_attr=None,
           bias_attr=None,
C
chengduoZH 已提交
1528
           use_cudnn=True,
1529 1530
           act=None,
           name=None):
Y
Yu Yang 已提交
1531
    """
C
chengduoZH 已提交
1532
    The convolution2D layer calculates the output based on the input, filter
T
tensor-tang 已提交
1533 1534
    and strides, paddings, dilations, groups parameters. Input and
    Output are in NCHW format, where N is batch size, C is the number of
1535
    channels, H is the height of the feature, and W is the width of the feature.
T
tensor-tang 已提交
1536 1537 1538 1539 1540 1541 1542
    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/>`_
    for more detials.
1543 1544 1545
    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
1546

1547
    For each input :math:`X`, the equation is:
C
refine  
chengduoZH 已提交
1548

C
chengduoZH 已提交
1549 1550
    .. math::

C
refine  
chengduoZH 已提交
1551
        Out = \sigma (W \\ast X + b)
C
chengduoZH 已提交
1552

T
tensor-tang 已提交
1553
    Where:
C
chengduoZH 已提交
1554

1555 1556 1557 1558 1559
    * :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
1560
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
1561 1562 1563

    Example:

1564 1565
        - Input:

W
weixing02 已提交
1566
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
C
refine  
chengduoZH 已提交
1567

W
weixing02 已提交
1568
          Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)`
C
refine  
chengduoZH 已提交
1569

1570
        - Output:
T
tensor-tang 已提交
1571

W
weixing02 已提交
1572
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
C
refine  
chengduoZH 已提交
1573

C
chengduoZH 已提交
1574
        Where
1575 1576

        .. math::
C
chengduoZH 已提交
1577

W
weixing02 已提交
1578 1579
            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
1580 1581

    Args:
1582
        input (Variable): The input image with [N, C, H, W] format.
T
tensor-tang 已提交
1583
        num_filters(int): The number of filter. It is as same as the output
1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600
            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 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
1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611
            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)`,
             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 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.
1612 1613
        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
1614 1615
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None
1616
        name (str|None): A name for this layer(optional). If set None, the layer
C
chengduo 已提交
1617
            will be named automatically. Default: None
C
chengduoZH 已提交
1618 1619

    Returns:
G
guosheng 已提交
1620
        Variable: The tensor variable storing the convolution and \
C
chengduoZH 已提交
1621 1622
                  non-linearity activation result.

C
refine  
chengduoZH 已提交
1623
    Raises:
1624 1625
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
C
refine  
chengduoZH 已提交
1626

C
chengduoZH 已提交
1627 1628 1629
    Examples:
        .. code-block:: python

1630 1631
          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
1632 1633 1634
    """

    num_channels = input.shape[1]
C
chengduo 已提交
1635
    assert param_attr is not False, "param_attr should not be False here."
1636
    l_type = 'conv2d'
X
xzl 已提交
1637 1638
    if (num_channels == groups and num_filters % num_channels == 0 and
            not use_cudnn):
1639
        l_type = 'depthwise_conv2d'
1640 1641 1642 1643

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

Y
Yu Yang 已提交
1644 1645 1646 1647 1648
    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
1649
        num_filter_channels = num_channels // groups
Y
Yu Yang 已提交
1650

C
chengduoZH 已提交
1651 1652 1653
    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')
1654
    dilation = utils.convert_to_list(dilation, 2, 'dilation')
C
chengduoZH 已提交
1655

C
chengduoZH 已提交
1656 1657
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1658 1659

    input_shape = input.shape
M
minqiyang 已提交
1660
    filter_shape = [num_filters, int(num_filter_channels)] + filter_size
Y
Yu Yang 已提交
1661 1662

    def _get_default_param_initializer():
C
chengduo 已提交
1663 1664
        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
1665 1666 1667 1668 1669 1670 1671 1672
        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
1673
    pre_bias = helper.create_variable_for_type_inference(dtype)
Y
Yu Yang 已提交
1674

1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688
    if use_cudnn:
        helper.create_variable(
            name="kCUDNNFwdAlgoCache",
            persistable=True,
            type=core.VarDesc.VarType.RAW)
        helper.create_variable(
            name="kCUDNNBwdDataAlgoCache",
            persistable=True,
            type=core.VarDesc.VarType.RAW)
        helper.create_variable(
            name="kCUDNNBwdFilterAlgoCache",
            persistable=True,
            type=core.VarDesc.VarType.RAW)

Y
Yu Yang 已提交
1689
    helper.append_op(
1690
        type=l_type,
Y
Yu Yang 已提交
1691 1692 1693 1694 1695
        inputs={
            'Input': input,
            'Filter': filter_param,
        },
        outputs={"Output": pre_bias},
C
chengduoZH 已提交
1696 1697 1698
        attrs={
            'strides': stride,
            'paddings': padding,
1699
            'dilations': dilation,
C
chengduoZH 已提交
1700
            'groups': groups,
1701
            'use_cudnn': use_cudnn,
1702
            'use_mkldnn': False,
C
chengduoZH 已提交
1703
        })
Y
Yu Yang 已提交
1704 1705 1706 1707 1708 1709

    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
1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726
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
1727 1728 1729 1730 1731 1732
    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
1733 1734 1735 1736 1737 1738 1739 1740 1741

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

    .. math::

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

    In the above equation:

1742 1743
    * :math:`X`: Input value, a tensor with NCDHW format.
    * :math:`W`: Filter value, a tensor with MCDHW format.
C
chengduoZH 已提交
1744 1745 1746
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
1747
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772

    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.
            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,
1773
            it must contain three integers, (filter_size_D, filter_size_H, filter_size_W).
C
chengduoZH 已提交
1774 1775
            Otherwise, the filter will be a square.
        stride (int|tuple): The stride size. If stride is a tuple, it must
1776
            contain three integers, (stride_D, stride_H, stride_W). Otherwise, the
C
chengduoZH 已提交
1777 1778
            stride_D = stride_H = stride_W = stride. Default: stride = 1.
        padding (int|tuple): The padding size. If padding is a tuple, it must
1779
            contain three integers, (padding_D, padding_H, padding_W). Otherwise, the
C
chengduoZH 已提交
1780 1781
            padding_D = padding_H = padding_W = padding. Default: padding = 0.
        dilation (int|tuple): The dilation size. If dilation is a tuple, it must
1782
            contain three integers, (dilation_D, dilation_H, dilation_W). Otherwise, the
C
chengduoZH 已提交
1783 1784 1785 1786 1787 1788
            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
1789 1790 1791 1792 1793 1794 1795 1796 1797 1798
        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
1799 1800
        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
1801 1802
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None.
C
chengduoZH 已提交
1803
        name (str|None): A name for this layer(optional). If set None, the layer
C
chengduo 已提交
1804
            will be named automatically. Default: None.
C
chengduoZH 已提交
1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816

    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

1817 1818
          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
1819 1820 1821
    """

    l_type = 'conv3d'
C
chengduo 已提交
1822
    assert param_attr is not False, "param_attr should not be False here."
C
chengduoZH 已提交
1823 1824 1825 1826 1827 1828 1829 1830 1831 1832
    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
1833
        num_filter_channels = num_channels // groups
C
chengduoZH 已提交
1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846

    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
1847 1848 1849
        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
1850 1851 1852 1853 1854 1855 1856 1857
        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
1858
    pre_bias = helper.create_variable_for_type_inference(dtype)
C
chengduoZH 已提交
1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872

    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
1873
            'use_mkldnn': False
C
chengduoZH 已提交
1874 1875
        })

1876
    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
C
chengduoZH 已提交
1877 1878 1879 1880

    return helper.append_activation(pre_act)


J
Jacek Czaja 已提交
1881
def sequence_pool(input, pool_type, is_test=False):
Y
Yu Yang 已提交
1882
    """
Y
yangyaming 已提交
1883 1884 1885
    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
1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896

    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

       x is a 1-level LoDTensor:
1897
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
1898 1899 1900 1901 1902
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
1903
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
1904 1905 1906 1907 1908 1909 1910

       for different pool_type:
         average: out.data = [2, 4, 3], where 2=(1+3)/2, 4=(2+4+6)/3, 3=(5+1)/2
         sum    : out.data = [4, 12, 6], where 4=1+3, 12=2+4+6, 6=5+1
         sqrt   : out.data = [2.82, 6.93, 4.24], where 2.82=(1+3)/sqrt(2),
                    6.93=(2+4+6)/sqrt(3), 4.24=(5+1)/sqrt(2)
         max    : out.data = [3, 6, 5], where 3=max(1,3), 6=max(2,4,6), 5=max(5,1)
1911 1912
         last   : out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1)
         first  : out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1)
F
fengjiayi 已提交
1913

L
Luo Tao 已提交
1914 1915
    Args:
        input(variable): The input variable which is a LoDTensor.
Y
yangyaming 已提交
1916
        pool_type (string): The pooling type of sequence_pool.
L
Luo Tao 已提交
1917
            It supports average, sum, sqrt and max.
J
Jacek Czaja 已提交
1918
        is_test(bool, Default False): Used distinguish training from scoring mode.
L
Luo Tao 已提交
1919 1920 1921 1922 1923 1924 1925

    Returns:
        The sequence pooling variable which is a Tensor.

    Examples:

        .. code-block:: python
F
fengjiayi 已提交
1926

Y
yangyaming 已提交
1927
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1928 1929 1930 1931 1932
                              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')
1933 1934
             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
1935
    """
F
fengjiayi 已提交
1936
    helper = LayerHelper('sequence_pool', **locals())
Y
Yu Yang 已提交
1937
    dtype = helper.input_dtype()
X
Xin Pan 已提交
1938 1939
    pool_out = helper.create_variable_for_type_inference(dtype)
    max_index = helper.create_variable_for_type_inference(dtype)
Y
Yu Yang 已提交
1940 1941 1942 1943 1944 1945

    helper.append_op(
        type="sequence_pool",
        inputs={"X": input},
        outputs={"Out": pool_out,
                 "MaxIndex": max_index},
J
Jacek Czaja 已提交
1946 1947
        attrs={"pooltype": pool_type.upper(),
               "is_test": is_test})
Y
Yu Yang 已提交
1948

Y
yangyaming 已提交
1949 1950 1951 1952 1953
    # 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
1954 1955 1956
    return pool_out


C
add doc  
chengduoZH 已提交
1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975
@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

           out = fluid.layers.sequence_concat(input=[seq1, seq2, seq3])
    """
    helper = LayerHelper('sequence_concat', **locals())
X
Xin Pan 已提交
1976
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
C
add doc  
chengduoZH 已提交
1977 1978 1979 1980 1981
    helper.append_op(
        type='sequence_concat', inputs={'X': input}, outputs={'Out': [out]})
    return out


F
fengjiayi 已提交
1982
def sequence_first_step(input):
L
Luo Tao 已提交
1983
    """
L
Luo Tao 已提交
1984
    This function gets the first step of sequence.
L
Luo Tao 已提交
1985 1986 1987 1988

    .. code-block:: text

       x is a 1-level LoDTensor:
1989
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
1990 1991 1992 1993 1994
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
1995
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
1996
         out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1)
F
fengjiayi 已提交
1997

L
Luo Tao 已提交
1998 1999 2000 2001 2002 2003 2004 2005 2006
    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
2007

Y
yangyaming 已提交
2008
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
2009 2010 2011
                              dtype='float32', lod_level=1)
             x_first_step = fluid.layers.sequence_first_step(input=x)
    """
2012 2013 2014
    return sequence_pool(input=input, pool_type="first")


F
fengjiayi 已提交
2015
def sequence_last_step(input):
L
Luo Tao 已提交
2016
    """
L
Luo Tao 已提交
2017
    This function gets the last step of sequence.
L
Luo Tao 已提交
2018 2019 2020 2021

    .. code-block:: text

       x is a 1-level LoDTensor:
2022
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
2023 2024 2025 2026 2027
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
2028
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
2029
         out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1)
F
fengjiayi 已提交
2030

L
Luo Tao 已提交
2031 2032 2033 2034 2035 2036 2037 2038 2039
    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
2040

Y
yangyaming 已提交
2041
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
2042 2043 2044
                              dtype='float32', lod_level=1)
             x_last_step = fluid.layers.sequence_last_step(input=x)
    """
2045 2046 2047
    return sequence_pool(input=input, pool_type="last")


Y
Yibing Liu 已提交
2048 2049 2050 2051
def sequence_slice(input, offset, length, name=None):
    """
    **Sequence Slice Layer**

2052
    The layer crops a subsequence from given sequence with given start
Y
Yibing Liu 已提交
2053 2054 2055 2056 2057
    offset and subsequence length.

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

    .. code-block:: text
2058

Y
Yibing Liu 已提交
2059 2060
	- Case:

2061
            Given the input Variable **input**:
2062

2063 2064 2065
                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
2066

2067
            with offset.data = [[0], [1]] and length.data = [[2], [1]],
Y
Yibing Liu 已提交
2068

2069
            the output Variable will be
2070

2071 2072 2073
                out.data = [[a1, a2], [b1, b2], [e1, e2]],
                out.lod = [[2, 1]],
                out.dims = (3, 2).
2074 2075

    NOTE: The first dimension size of **input**, **offset** and **length**
2076
          should be equal. The **offset** should start from 0.
2077

Y
Yibing Liu 已提交
2078
    Args:
2079
        input(Variable): The input Variable which consists of the complete
Y
Yibing Liu 已提交
2080
                         sequences.
Y
Yibing Liu 已提交
2081 2082 2083 2084 2085 2086
        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
2087
        Variable: The output subsequences.
Y
Yibing Liu 已提交
2088 2089 2090 2091 2092 2093 2094 2095 2096 2097

    Examples:

        .. code-block:: python

             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"))
2098
             subseqs = fluid.layers.sequence_slice(input=seqs, offset=offset,
Y
Yibing Liu 已提交
2099 2100 2101 2102
                                                   length=length)
    """
    helper = LayerHelper("sequence_slice", **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
2103
    out = helper.create_variable_for_type_inference(dtype)
Y
Yibing Liu 已提交
2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117

    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
2118
@templatedoc()
Y
Yu Yang 已提交
2119
def pool2d(input,
C
chengduoZH 已提交
2120 2121
           pool_size=-1,
           pool_type="max",
C
chengduoZH 已提交
2122 2123
           pool_stride=1,
           pool_padding=0,
C
caoying03 已提交
2124
           global_pooling=False,
C
chengduoZH 已提交
2125
           use_cudnn=True,
2126
           ceil_mode=False,
2127 2128
           name=None,
           exclusive=True):
Y
Yu Yang 已提交
2129
    """
F
fengjiayi 已提交
2130
    ${comment}
2131 2132

    Args:
2133 2134 2135
        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
2136
                          feature, and W is the width of the feature.
J
JiabinYang 已提交
2137
        pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list,
J
JiabinYang 已提交
2138 2139
            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
2140
        pool_type: ${pooling_type_comment}
J
JiabinYang 已提交
2141 2142 2143 2144 2145 2146
        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.
2147 2148 2149
        global_pooling (bool): ${global_pooling_comment}
        use_cudnn (bool): ${use_cudnn_comment}
        ceil_mode (bool): ${ceil_mode_comment}
2150
        name (str|None): A name for this layer(optional). If set None, the
F
fengjiayi 已提交
2151
                        layer will be named automatically.
2152
        exclusive (bool): Whether to exclude padding points in average pooling
2153
                          mode, default is true
F
fengjiayi 已提交
2154

2155
    Returns:
F
fengjiayi 已提交
2156
        Variable: The pooling result.
F
fengjiayi 已提交
2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169

    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

          data = fluid.layers.data(
              name='data', shape=[3, 32, 32], dtype='float32')
          conv2d = fluid.layers.pool2d(
2170 2171 2172 2173
                            input=data,
                            pool_size=2,
                            pool_type='max',
                            pool_stride=1,
F
fengjiayi 已提交
2174
                            global_pooling=False)
Y
Yu Yang 已提交
2175 2176 2177 2178 2179
    """
    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
2180

C
chengduoZH 已提交
2181 2182 2183 2184 2185
    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
2186 2187 2188 2189
    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
2190 2191
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
2192

C
Add doc  
chengduoZH 已提交
2193
    l_type = 'pool2d'
2194 2195

    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
2196
    dtype = helper.input_dtype()
X
Xin Pan 已提交
2197
    pool_out = helper.create_variable_for_type_inference(dtype)
Y
Yu Yang 已提交
2198 2199

    helper.append_op(
2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210
        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,
2211 2212
            "use_mkldnn": False,
            "exclusive": exclusive,
2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225
        })

    return pool_out


def pool3d(input,
           pool_size=-1,
           pool_type="max",
           pool_stride=1,
           pool_padding=0,
           global_pooling=False,
           use_cudnn=True,
           ceil_mode=False,
2226 2227
           name=None,
           exclusive=True):
2228 2229
    """
    This function adds the operator for pooling in 3-dimensions, using the
Y
Yu Yang 已提交
2230
    pooling configurations mentioned in input parameters.
2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242

    Args:
        input (Variable): ${input_comment}
        pool_size (int): ${ksize_comment}
        pool_type (str): ${pooling_type_comment}
        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.
2243
        exclusive (bool): Whether to exclude padding points in average pooling
2244
                          mode, default is true
2245

2246
    Returns:
2247
        Variable: output of pool3d layer.
Y
Yu Yang 已提交
2248 2249 2250 2251 2252
    """
    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
2253

C
chengduoZH 已提交
2254 2255 2256 2257 2258
    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))

2259 2260 2261
    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
2262

C
chengduoZH 已提交
2263 2264
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
2265

2266 2267
    l_type = "pool3d"
    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
2268
    dtype = helper.input_dtype()
X
Xin Pan 已提交
2269
    pool_out = helper.create_variable_for_type_inference(dtype)
Y
Yu Yang 已提交
2270 2271

    helper.append_op(
2272
        type=l_type,
Y
Yu Yang 已提交
2273 2274 2275 2276 2277 2278 2279
        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
2280
            "paddings": pool_padding,
2281
            "use_cudnn": use_cudnn,
2282
            "ceil_mode": ceil_mode,
2283 2284
            "use_mkldnn": False,
            "exclusive": exclusive,
Y
Yu Yang 已提交
2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296
        })

    return pool_out


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
2297
               data_layout='NCHW',
Y
Yang Yang 已提交
2298
               in_place=False,
2299 2300
               name=None,
               moving_mean_name=None,
W
wanghaoshuang 已提交
2301
               moving_variance_name=None,
2302 2303
               do_model_average_for_mean_and_var=False,
               fuse_with_relu=False):
Y
Yu Yang 已提交
2304
    """
Q
qiaolongfei 已提交
2305 2306 2307 2308
    **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
2309

Q
qiaolongfei 已提交
2310
    1. NHWC `[batch, in_height, in_width, in_channels]`
Q
qiaolongfei 已提交
2311

Q
qiaolongfei 已提交
2312 2313
    2. NCHW `[batch, in_channels, in_height, in_width]`

Q
qiaolongfei 已提交
2314 2315 2316
    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
2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328

    :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
2329 2330

    Args:
Q
qiaolongfei 已提交
2331
        input(variable): The input variable which is a LoDTensor.
Q
qiaolongfei 已提交
2332 2333 2334 2335
        act(string, Default None): Activation type, linear|relu|prelu|...
        is_test(bool, Default False): Used for training or training.
        momentum(float, Default 0.9):
        epsilon(float, Default 1e-05):
C
chengduo 已提交
2336 2337 2338 2339 2340 2341 2342 2343
        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
             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 parameter attribute for the bias of batch_norm.
             If it is set to None or one attribute of ParamAttr, batch_norm
             will create ParamAttr as bias_attr. If the Initializer of the bias_attr
             is not set, the bias is initialized zero. Default: None.
Q
qiaolongfei 已提交
2344
        data_layout(string, default NCHW): NCHW|NHWC
2345
        in_place(bool, Default False): Make the input and output of batch norm reuse memory.
Q
qiaolongfei 已提交
2346 2347 2348 2349
        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.
Q
qiaolongfei 已提交
2350
        do_model_average_for_mean_and_var(bool, Default False): Do model average for mean and variance or not.
2351
        fuse_with_relu (bool): if True, this OP performs relu after batch norm.
2352 2353

    Returns:
Q
qiaolongfei 已提交
2354
        Variable: A tensor variable which is the result after applying batch normalization on the input.
Q
qiaolongfei 已提交
2355 2356 2357 2358 2359 2360 2361

    Examples:

        .. code-block:: python

            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
2362
    """
C
chengduo 已提交
2363
    assert bias_attr is not False, "bias_attr should not be False in batch_norm."
Y
Yu Yang 已提交
2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385
    helper = LayerHelper('batch_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]

    # create parameter
    scale = helper.create_parameter(
        attr=helper.param_attr,
        shape=param_shape,
        dtype=dtype,
        default_initializer=Constant(1.0))

    bias = helper.create_parameter(
2386
        attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
Y
Yu Yang 已提交
2387

2388 2389
    mean = helper.create_parameter(
        attr=ParamAttr(
W
wanghaoshuang 已提交
2390 2391 2392
            name=moving_mean_name,
            initializer=Constant(0.0),
            trainable=False,
W
wanghaoshuang 已提交
2393
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
2394
        shape=param_shape,
2395 2396 2397 2398 2399 2400 2401
        dtype=input.dtype)
    mean.stop_gradient = True

    variance = helper.create_parameter(
        attr=ParamAttr(
            name=moving_variance_name,
            initializer=Constant(1.0),
W
wanghaoshuang 已提交
2402
            trainable=False,
W
wanghaoshuang 已提交
2403
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
2404
        shape=param_shape,
2405 2406
        dtype=input.dtype)
    variance.stop_gradient = True
Y
Yu Yang 已提交
2407 2408 2409 2410 2411 2412

    # 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
2413 2414 2415 2416
    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
2417

X
Xin Pan 已提交
2418 2419
    batch_norm_out = input if in_place else helper.create_variable_for_type_inference(
        dtype)
Y
Yu Yang 已提交
2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436

    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
        },
2437 2438 2439 2440
        attrs={
            "momentum": momentum,
            "epsilon": epsilon,
            "is_test": is_test,
X
Xin Pan 已提交
2441
            "use_mkldnn": False,
2442
            "fuse_with_relu": fuse_with_relu
2443
        })
Y
Yu Yang 已提交
2444 2445 2446 2447

    return helper.append_activation(batch_norm_out)


Y
yuyang18 已提交
2448
@templatedoc()
G
guosheng 已提交
2449 2450 2451 2452 2453 2454 2455 2456 2457 2458
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
2459
    ${comment}
G
guosheng 已提交
2460 2461 2462

    The formula is as follows:

Y
yuyang18 已提交
2463
    ..  math::
G
guosheng 已提交
2464 2465 2466 2467 2468 2469 2470

        \\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
2471 2472 2473 2474 2475 2476 2477 2478
    * :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
2479

G
guosheng 已提交
2480 2481
    Args:
        input(Variable): The input tensor variable.
2482
        scale(bool): Whether to learn the adaptive gain :math:`g` after
S
sneaxiy 已提交
2483
            normalization. Default True.
2484
        shift(bool): Whether to learn the adaptive bias :math:`b` after
S
sneaxiy 已提交
2485 2486
            normalization. Default True.
        begin_norm_axis(int): The normalization will be performed along
G
guosheng 已提交
2487
            dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`.
S
sneaxiy 已提交
2488
            Default 1.
2489
        epsilon(float): The small value added to the variance to prevent
S
sneaxiy 已提交
2490
            division by zero. Default 1e-05.
G
guosheng 已提交
2491
        param_attr(ParamAttr|None): The parameter attribute for the learnable
S
sneaxiy 已提交
2492 2493
            gain :math:`g`. If :attr:`scale` is False, :attr:`param_attr` is
            omitted. If :attr:`scale` is True and :attr:`param_attr` is None,
2494 2495
            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
2496
        bias_attr(ParamAttr|None): The parameter attribute for the learnable
S
sneaxiy 已提交
2497 2498
            bias :math:`b`. If :attr:`shift` is False, :attr:`bias_attr` is
            omitted. If :attr:`shift` is True and :attr:`param_attr` is None,
2499
            a default :code:`ParamAttr` would be added as bias. The
S
sneaxiy 已提交
2500
            :attr:`bias_attr` is initialized as 0 if it is added. Default None.
G
guosheng 已提交
2501
        act(str): Activation to be applied to the output of layer normalizaiton.
S
sneaxiy 已提交
2502 2503 2504
                  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
2505 2506

    Returns:
Y
yuyang18 已提交
2507
        ${y_comment}
G
guosheng 已提交
2508 2509 2510

    Examples:

Y
yuyang18 已提交
2511 2512 2513
        >>> 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
2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528
    """
    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
2529
    if shift:
G
guosheng 已提交
2530 2531 2532 2533 2534 2535
        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
2536 2537 2538 2539 2540
    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
2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555

    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
2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633
@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**

    Refer to `Group Normalization <https://arxiv.org/abs/1803.08494>`

    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:

        >>> 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
    mean_out = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
    variance_out = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
    group_norm_out = helper.create_tmp_variable(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)


Y
Yu Yang 已提交
2634 2635 2636 2637
def conv2d_transpose(input,
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
2638 2639 2640
                     padding=0,
                     stride=1,
                     dilation=1,
2641
                     groups=None,
C
caoying03 已提交
2642
                     param_attr=None,
2643
                     bias_attr=None,
C
chengduoZH 已提交
2644
                     use_cudnn=True,
2645
                     act=None,
C
caoying03 已提交
2646
                     name=None):
Y
Yu Yang 已提交
2647
    """
2648 2649 2650 2651 2652 2653 2654 2655
    **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
2656 2657
    layer, please refer to the following explanation and references
    `therein <http://www.matthewzeiler.com/wp-content/uploads/2017/07/cvpr2010.pdf>`_.
2658 2659 2660
    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.
2661 2662 2663 2664 2665

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

    .. math::

2666
        Out = \sigma (W \\ast X + b)
2667

2668
    Where:
2669 2670 2671

    * :math:`X`: Input value, a tensor with NCHW format.
    * :math:`W`: Filter value, a tensor with MCHW format.
2672 2673 2674 2675
    * :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
2676

2677 2678 2679 2680
    Example:

        - Input:

2681
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
2682

2683
          Filter shape: :math:`(C_{in}, C_{out}, H_f, W_f)`
2684 2685 2686

        - Output:

2687
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
2688 2689

        Where
Y
Yu Yang 已提交
2690

2691 2692
        .. math::

2693 2694 2695 2696
           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 \\\\
           H_{out} \in [ H^\prime_{out}, H^\prime_{out} + strides[0] ) \\\\
           W_{out} \in [ W^\prime_{out}, W^\prime_{out} + strides[1] )
Y
Yu Yang 已提交
2697 2698

    Args:
2699 2700 2701 2702
        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
2703 2704 2705 2706
            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.
2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724
        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
2725 2726 2727 2728 2729 2730 2731 2732 2733 2734
            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.
2735
        use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
C
chengduo 已提交
2736 2737 2738
            library is installed. Default: True.
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None.
2739
        name(str|None): A name for this layer(optional). If set None, the layer
C
chengduo 已提交
2740
            will be named automatically. Default: True.
Y
Yu Yang 已提交
2741 2742

    Returns:
2743
        Variable: The tensor variable storing the convolution transpose result.
2744 2745

    Raises:
2746 2747
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
2748 2749 2750 2751

    Examples:
       .. code-block:: python

2752 2753
          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
2754
    """
C
chengduo 已提交
2755
    assert param_attr is not False, "param_attr should not be False in conv2d_transpose."
2756 2757 2758 2759 2760 2761 2762 2763
    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
2764 2765 2766
    if not isinstance(input, Variable):
        raise TypeError("Input of conv2d_transpose must be Variable")

C
chengduoZH 已提交
2767 2768 2769
    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
2770

C
chengduoZH 已提交
2771 2772
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
G
guosheng 已提交
2773

Y
Yu Yang 已提交
2774 2775 2776 2777 2778
    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
2779

Y
Yu Yang 已提交
2780 2781
        h_in = input.shape[2]
        w_in = input.shape[3]
G
guosheng 已提交
2782

C
chengduoZH 已提交
2783
        filter_size_h = (output_size[0] - (h_in - 1) * stride[0] + 2 *
M
minqiyang 已提交
2784
                         padding[0] - 1) // dilation[0] + 1
C
chengduoZH 已提交
2785
        filter_size_w = (output_size[1] - (w_in - 1) * stride[1] + 2 *
M
minqiyang 已提交
2786
                         padding[1] - 1) // dilation[1] + 1
Y
Yu Yang 已提交
2787
        filter_size = [filter_size_h, filter_size_w]
C
chengduoZH 已提交
2788 2789 2790
    else:
        filter_size = utils.convert_to_list(filter_size, 2,
                                            'conv2d_transpose.filter_size')
C
chengduo 已提交
2791

2792 2793 2794 2795 2796 2797 2798
    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')
2799
    groups = 1 if groups is None else groups
M
minqiyang 已提交
2800
    filter_shape = [input_channel, num_filters // groups] + filter_size
C
chengduo 已提交
2801

Y
Yu Yang 已提交
2802 2803 2804
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

X
Xin Pan 已提交
2805
    pre_bias = helper.create_variable_for_type_inference(dtype=input.dtype)
Y
Yu Yang 已提交
2806
    helper.append_op(
2807
        type=op_type,
Y
Yu Yang 已提交
2808 2809
        inputs={'Input': [input],
                'Filter': [img_filter]},
2810
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
2811
        attrs={
2812
            'output_size': output_size,
2813 2814 2815 2816 2817
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn
Y
Yu Yang 已提交
2818 2819
        })

2820 2821 2822
    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
2823 2824


2825
def conv3d_transpose(input,
Y
Yu Yang 已提交
2826 2827 2828
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
2829 2830 2831
                     padding=0,
                     stride=1,
                     dilation=1,
2832
                     groups=None,
C
caoying03 已提交
2833
                     param_attr=None,
2834
                     bias_attr=None,
C
chengduoZH 已提交
2835
                     use_cudnn=True,
2836
                     act=None,
C
caoying03 已提交
2837
                     name=None):
Y
Yu Yang 已提交
2838
    """
2839
    **Convlution3D transpose layer**
2840

2841
    The convolution3D transpose layer calculates the output based on the input,
2842
    filter, and dilations, strides, paddings. Input(Input) and output(Output)
2843 2844 2845 2846 2847 2848
    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
    explanation and references `therein <http://www.matthewzeiler.com/wp-content/uploads/2017/07/cvpr2010.pdf>`_.
2849 2850 2851
    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.
2852 2853 2854 2855 2856

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

    .. math::

2857
        Out = \sigma (W \\ast X + b)
2858 2859 2860

    In the above equation:

2861 2862
    * :math:`X`: Input value, a tensor with NCDHW format.
    * :math:`W`: Filter value, a tensor with MCDHW format.
2863 2864 2865 2866
    * :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
2867

2868 2869 2870 2871
    Example:

        - Input:

2872
          Input shape: :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})`
2873

2874
          Filter shape: :math:`(C_{in}, C_{out}, D_f, H_f, W_f)`
2875 2876 2877

        - Output:

2878
          Output shape: :math:`(N, C_{out}, D_{out}, H_{out}, W_{out})`
2879 2880

        Where
Y
Yu Yang 已提交
2881

2882 2883
        .. math::

2884 2885 2886
           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
2887 2888

    Args:
2889
        input(Variable): The input image with [N, C, D, H, W] format.
2890 2891 2892
        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
2893
            tuple, it must contain three integers, (image_D, image_H, image_W). This
2894 2895
            parameter only works when filter_size is None.
        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
2896
            it must contain three integers, (filter_size_D, filter_size_H, filter_size_W).
2897 2898 2899
            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
2900 2901
            contain three integers, (padding_D, padding_H, padding_W). Otherwise, the
            padding_D = padding_H = padding_W = padding. Default: padding = 0.
2902
        stride(int|tuple): The stride size. If stride is a tuple, it must
2903 2904
            contain three integers, (stride_D, stride_H, stride_W). Otherwise, the
            stride_D = stride_H = stride_W = stride. Default: stride = 1.
2905
        dilation(int|tuple): The dilation size. If dilation is a tuple, it must
2906 2907 2908
            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
2909 2910 2911 2912 2913
            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
2914 2915 2916 2917 2918 2919 2920 2921 2922
        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.
2923 2924
        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
2925 2926
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None.
2927 2928
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.
Y
Yu Yang 已提交
2929 2930

    Returns:
2931
        Variable: The tensor variable storing the convolution transpose result.
2932 2933

    Raises:
2934 2935
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
2936 2937 2938 2939

    Examples:
       .. code-block:: python

2940 2941
          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
2942
    """
C
chengduo 已提交
2943
    assert param_attr is not False, "param_attr should not be False in conv3d_transpose."
2944 2945
    l_type = "conv3d_transpose"
    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
2946
    if not isinstance(input, Variable):
2947
        raise TypeError("Input of conv3d_transpose must be Variable")
Y
Yu Yang 已提交
2948 2949
    input_channel = input.shape[1]

2950 2951 2952
    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
2953

C
chengduoZH 已提交
2954 2955 2956
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")

Y
Yu Yang 已提交
2957 2958 2959 2960 2961 2962
    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]

2963 2964 2965
        d_in = input.shape[2]
        h_in = input.shape[3]
        w_in = input.shape[4]
C
chengduoZH 已提交
2966

2967
        filter_size_d = (output_size[0] - (d_in - 1) * stride[0] + 2 *
M
minqiyang 已提交
2968
                         padding[0] - 1) // dilation[0] + 1
2969
        filter_size_h = (output_size[1] - (h_in - 1) * stride[1] + 2 *
M
minqiyang 已提交
2970
                         padding[1] - 1) // dilation[1] + 1
2971
        filter_size_w = (output_size[2] - (w_in - 1) * stride[2] + 2 *
M
minqiyang 已提交
2972
                         padding[2] - 1) // dilation[2] + 1
2973
        filter_size = [filter_size_d, filter_size_h, filter_size_w]
C
chengduoZH 已提交
2974
    else:
2975 2976
        filter_size = utils.convert_to_list(filter_size, 3,
                                            'conv3d_transpose.filter_size')
Y
Yu Yang 已提交
2977

2978
    groups = 1 if groups is None else groups
M
minqiyang 已提交
2979
    filter_shape = [input_channel, num_filters // groups] + filter_size
Y
Yu Yang 已提交
2980 2981 2982
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

X
Xin Pan 已提交
2983
    pre_bias = helper.create_variable_for_type_inference(dtype=input.dtype)
Y
Yu Yang 已提交
2984
    helper.append_op(
2985
        type=l_type,
Y
Yu Yang 已提交
2986 2987
        inputs={'Input': [input],
                'Filter': [img_filter]},
2988
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
2989 2990 2991 2992
        attrs={
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
2993
            'groups': groups,
C
chengduoZH 已提交
2994 2995
            'use_cudnn': use_cudnn
        })
Y
Yu Yang 已提交
2996

2997 2998
    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
2999
    return out
Y
yangyaming 已提交
3000 3001


Y
yangyaming 已提交
3002
def sequence_expand(x, y, ref_level=-1, name=None):
3003
    """Sequence Expand Layer. This layer will expand the input variable **x**
Y
yangyaming 已提交
3004 3005 3006 3007
    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:
3008 3009 3010 3011 3012

    .. code-block:: text

        * Case 1
            x is a LoDTensor:
3013
                x.lod  = [[2,        2]]
Y
yangyaming 已提交
3014
                x.data = [[a], [b], [c], [d]]
3015 3016 3017
                x.dims = [4, 1]

            y is a LoDTensor:
3018 3019
                y.lod = [[2,    2],
                         [3, 3, 1, 1]]
3020

Y
yangyaming 已提交
3021
            ref_level: 0
3022

Y
yangyaming 已提交
3023
            then output is a 1-level LoDTensor:
3024
                out.lod =  [[2,        2,        2,        2]]
Y
yangyaming 已提交
3025
                out.data = [[a], [b], [a], [b], [c], [d], [c], [d]]
3026 3027 3028 3029
                out.dims = [8, 1]

        * Case 2
            x is a Tensor:
Y
yangyaming 已提交
3030
                x.data = [[a], [b], [c]]
3031 3032 3033
                x.dims = [3, 1]

            y is a LoDTensor:
3034
                y.lod = [[2, 0, 3]]
3035

Y
yangyaming 已提交
3036
            ref_level: -1
3037

Y
yangyaming 已提交
3038 3039 3040
            then output is a Tensor:
                out.data = [[a], [a], [c], [c], [c]]
                out.dims = [5, 1]
3041 3042 3043
    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
3044 3045
        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
3046
        name(str|None): A name for this layer(optional). If set None, the layer
Y
yangyaming 已提交
3047
                        will be named automatically.
3048 3049 3050 3051 3052 3053 3054 3055 3056 3057

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

    Examples:
        .. code-block:: python

            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
3058
            out = layers.sequence_expand(x=x, y=y, ref_level=0)
3059
    """
Y
yangyaming 已提交
3060
    helper = LayerHelper('sequence_expand', input=x, **locals())
3061
    dtype = helper.input_dtype()
X
Xin Pan 已提交
3062
    tmp = helper.create_variable_for_type_inference(dtype)
3063
    helper.append_op(
Y
yangyaming 已提交
3064 3065 3066 3067 3068
        type='sequence_expand',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': tmp},
        attrs={'ref_level': ref_level})
3069
    return tmp
3070 3071


C
chengduo 已提交
3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127
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

            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)
    """
    helper = LayerHelper('sequence_expand_as', input=x, **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
3128
    tmp = helper.create_variable_for_type_inference(dtype)
C
chengduo 已提交
3129 3130 3131 3132 3133 3134 3135 3136
    helper.append_op(
        type='sequence_expand_as',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': tmp})
    return tmp


F
fengjiayi 已提交
3137
@templatedoc()
3138
def sequence_pad(x, pad_value, maxlen=None, name=None):
F
fengjiayi 已提交
3139 3140 3141 3142 3143
    """
    ${comment}

    Args:
        x(Variable): Input variable which should contain lod information.
M
minqiyang 已提交
3144 3145 3146
        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
3147
            automatically broadcasted to the shape of time step.
M
minqiyang 已提交
3148 3149 3150 3151
        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
3152 3153 3154
            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
3155

F
fengjiayi 已提交
3156
    Returns:
M
minqiyang 已提交
3157
        Variable: The padded sequence batch and the original lengths before
3158
                  padding. All sequences has the same length.
M
minqiyang 已提交
3159

F
fengjiayi 已提交
3160 3161 3162 3163 3164 3165 3166
    Examples:
        .. code-block:: python

            import numpy

            x = fluid.layers.data(name='y', shape=[10, 5],
                             dtype='float32', lod_level=1)
G
gmcather 已提交
3167
            pad_value = fluid.layers.assign(
D
dongzhihong 已提交
3168
                input=numpy.array([0.0], dtype=numpy.float32))
F
fengjiayi 已提交
3169 3170 3171 3172 3173
            out = fluid.layers.sequence_pad(x=x, pad_value=pad_value)
    """

    helper = LayerHelper('sequence_pad', input=x, **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
3174 3175
    out = helper.create_variable_for_type_inference(dtype)
    length = helper.create_variable_for_type_inference(dtype)
3176 3177 3178 3179

    pad_value.stop_gradient = True
    length.stop_gradient = True

F
fengjiayi 已提交
3180 3181 3182 3183 3184 3185
    if maxlen is None:
        maxlen = -1
    helper.append_op(
        type='sequence_pad',
        inputs={'X': x,
                'PadValue': pad_value},
3186 3187
        outputs={'Out': out,
                 'Length': length},
F
fengjiayi 已提交
3188
        attrs={'padded_length': maxlen})
3189
    return out, length
F
fengjiayi 已提交
3190 3191


3192
def sequence_unpad(x, length, name=None):
Y
Yibing Liu 已提交
3193
    """
3194
    **Sequence Unpad Layer**
Y
Yibing Liu 已提交
3195

3196 3197
    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
3198 3199 3200 3201 3202 3203 3204 3205 3206
    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],
3207 3208 3209
		      [11.0, 12.0, 13.0, 14.0, 15.0]],

	in which there are 3 sequences padded to length 5, and the acutal length
3210
	specified by input Variable **length**:
Y
Yibing Liu 已提交
3211 3212 3213 3214 3215 3216

	    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]]
3217
	    out.lod = [[2, 3, 4]]
Y
Yibing Liu 已提交
3218 3219 3220 3221 3222 3223

    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.
3224 3225
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.
Y
Yibing Liu 已提交
3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239

    Returns:
        Variable: The Variable contains the unpadded sequences.

    Examples:
        .. code-block:: python

            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)
    """

    helper = LayerHelper('sequence_unpad', input=x, **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
3240
    out = helper.create_variable_for_type_inference(dtype)
Y
Yibing Liu 已提交
3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251

    length.stop_gradient = True

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


3252 3253 3254 3255 3256 3257 3258 3259 3260
def beam_search(pre_ids,
                pre_scores,
                ids,
                scores,
                beam_size,
                end_id,
                level=0,
                name=None):
    """
3261 3262
    Beam search is a classical algorithm for selecting candidate words in a
    machine translation task.
Y
Yan Chunwei 已提交
3263 3264 3265

    Refer to `Beam search <https://en.wikipedia.org/wiki/Beam_search>`_
    for more details.
M
minqiyang 已提交
3266 3267

    This layer does the search in beams for one time step. Specifically, it
3268 3269 3270 3271 3272 3273
    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
    computation cell. 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.
M
minqiyang 已提交
3274

3275 3276 3277 3278 3279 3280 3281 3282
    Note that the :attr:`scores` passed in should be accumulated scores, and
    length penalty should be done with extra operators before calculating the
    accumulated scores if needed, also suggest finding top-K before it and
    using the top-K candidates following.

    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
3283

3284
    Args:
3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309
        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.
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
F
fengjiayi 已提交
3310

3311
    Returns:
3312 3313
        Variable: The LodTensor pair containing the selected ids and the \
            corresponding scores.
Y
Yan Chunwei 已提交
3314 3315 3316 3317

    Examples:
        .. code-block:: python

3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334
            # Suppose `probs` contains predicted results from the computation
            # cell and `pre_ids` and `pre_scores` is the output of beam_search
            # at previous step.
            topk_scores, topk_indices = layers.topk(probs, k=beam_size)
            accu_scores = layers.elementwise_add(
                x=layers.log(x=topk_scores)),
                y=layers.reshape(
                    pre_scores, shape=[-1]),
                axis=0)
            selected_ids, selected_scores = layers.beam_search(
                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
3335 3336 3337 3338
    helper = LayerHelper('beam_search', **locals())
    score_type = scores.dtype
    id_type = ids.dtype

X
Xin Pan 已提交
3339 3340 3341
    selected_scores = helper.create_variable_for_type_inference(
        dtype=score_type)
    selected_ids = helper.create_variable_for_type_inference(dtype=id_type)
Q
Qiao Longfei 已提交
3342 3343 3344 3345 3346

    helper.append_op(
        type='beam_search',
        inputs={
            'pre_ids': pre_ids,
3347
            'pre_scores': pre_scores,
Q
Qiao Longfei 已提交
3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364
            'ids': ids,
            'scores': scores,
        },
        outputs={
            'selected_ids': selected_ids,
            'selected_scores': selected_scores,
        },
        attrs={
            # TODO(ChunweiYan) to assure other value support
            'level': level,
            'beam_size': beam_size,
            'end_id': end_id,
        })

    return selected_ids, selected_scores


3365 3366 3367 3368 3369 3370 3371
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
3372

3373 3374 3375 3376 3377 3378 3379 3380 3381
    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
3382

3383 3384 3385 3386 3387 3388
    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
3389

3390 3391 3392 3393 3394 3395 3396 3397
    Examples:
        .. code-block:: python
            # Suppose `ids` and `scores` are LodTensorArray variables reserving
            # the selected ids and scores of all steps
            finished_ids, finished_scores = layers.beam_search_decode(
                ids, scores, beam_size=5, end_id=0)
    """
    helper = LayerHelper('beam_search_decode', **locals())
X
Xin Pan 已提交
3398 3399
    sentence_ids = helper.create_variable_for_type_inference(dtype=ids.dtype)
    sentence_scores = helper.create_variable_for_type_inference(dtype=ids.dtype)
3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414

    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
3415 3416 3417 3418
def lstm_unit(x_t,
              hidden_t_prev,
              cell_t_prev,
              forget_bias=0.0,
Y
yangyaming 已提交
3419
              param_attr=None,
C
caoying03 已提交
3420 3421
              bias_attr=None,
              name=None):
Y
yangyaming 已提交
3422 3423 3424 3425
    """Lstm unit layer. The equation of a lstm step is:

        .. math::

3426
            i_t & = \sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i)
Y
yangyaming 已提交
3427

3428
            f_t & = \sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + b_f)
Y
yangyaming 已提交
3429

3430
            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
3431

3432
            o_t & = \sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + b_o)
Y
yangyaming 已提交
3433 3434 3435

            h_t & = o_t tanh(c_t)

3436 3437 3438 3439 3440 3441
    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
3442 3443 3444

        .. math::

3445
            L_{i_t} = W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i
Y
yangyaming 已提交
3446 3447 3448 3449 3450 3451 3452 3453

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

        .. math::

            i_t = \sigma(L_{i_t})

Y
yangyaming 已提交
3454
    This layer has two outputs including :math:`h_t` and :math:`o_t`.
Y
yangyaming 已提交
3455 3456

    Args:
Y
yangyaming 已提交
3457 3458 3459 3460 3461 3462
        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
3463
        forget_bias (float): The forget bias of lstm unit.
C
chengduo 已提交
3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475
        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
3476 3477
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
Y
yangyaming 已提交
3478 3479

    Returns:
Y
yangyaming 已提交
3480
        tuple: The hidden value and cell value of lstm unit.
Y
yangyaming 已提交
3481 3482

    Raises:
3483 3484 3485 3486
        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
3487 3488 3489 3490 3491 3492

    Examples:

        .. code-block:: python

             x_t = fluid.layers.fc(input=x_t_data, size=10)
3493
             prev_hidden = fluid.layers.fc(input=prev_hidden_data, size=30)
Y
yangyaming 已提交
3494
             prev_cell = fluid.layers.fc(input=prev_cell_data, size=30)
Y
yangyaming 已提交
3495
             hidden_value, cell_value = fluid.layers.lstm_unit(x_t=x_t,
Y
yangyaming 已提交
3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511
                                                    hidden_t_prev=prev_hidden,
                                                    cell_t_prev=prev_cell)
    """
    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
3512
        raise ValueError("The 1st dimensions of x_t, hidden_t_prev and "
3513 3514 3515 3516
                         "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
3517 3518
                         "cell_t_prev must be the same.")

Y
yangyaming 已提交
3519 3520 3521
    if bias_attr is None:
        bias_attr = ParamAttr()

Y
yangyaming 已提交
3522
    size = cell_t_prev.shape[1]
3523
    concat_out = concat(input=[x_t, hidden_t_prev], axis=1)
Y
yangyaming 已提交
3524 3525
    fc_out = fc(input=concat_out,
                size=4 * size,
Y
yangyaming 已提交
3526
                param_attr=param_attr,
3527
                bias_attr=bias_attr)
Y
yangyaming 已提交
3528
    dtype = x_t.dtype
X
Xin Pan 已提交
3529 3530
    c = helper.create_variable_for_type_inference(dtype)
    h = helper.create_variable_for_type_inference(dtype)
Y
yangyaming 已提交
3531 3532 3533 3534 3535 3536 3537 3538 3539

    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
3540
    return h, c
G
guosheng 已提交
3541 3542


C
caoying03 已提交
3543
def reduce_sum(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
3544
    """
Y
yangyaming 已提交
3545
    Computes the sum of tensor elements over the given dimension.
G
guosheng 已提交
3546 3547 3548

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
3549
        dim (list|int|None): The dimensions along which the sum is performed. If
Y
yangyaming 已提交
3550 3551
            :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
3552 3553
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
3554
        keep_dim (bool|False): Whether to reserve the reduced dimension in the
Y
yangyaming 已提交
3555
            output Tensor. The result tensor will have one fewer dimension
G
guosheng 已提交
3556
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
3557 3558
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
3559 3560 3561

    Returns:
        Variable: The reduced Tensor variable.
F
fengjiayi 已提交
3562

G
guosheng 已提交
3563 3564 3565 3566 3567 3568
    Examples:
        .. code-block:: python

            # 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
3569
            # Each example is followed by the corresponding output tensor.
G
guosheng 已提交
3570 3571 3572 3573
            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
3574 3575 3576 3577

            # x is a Tensor variable with shape [2, 2, 2] and elements as below:
            #      [[[1, 2], [3, 4]],
            #      [[5, 6], [7, 8]]]
Q
qiaolongfei 已提交
3578
            # Each example is followed by the corresponding output tensor.
W
whs 已提交
3579 3580 3581
            fluid.layers.reduce_sum(x, dim=[1, 2]) # [10, 26]
            fluid.layers.reduce_sum(x, dim=[0, 1]) # [16, 20]

G
guosheng 已提交
3582 3583
    """
    helper = LayerHelper('reduce_sum', **locals())
X
Xin Pan 已提交
3584
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
3585 3586
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
G
guosheng 已提交
3587 3588 3589 3590 3591
    helper.append_op(
        type='reduce_sum',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3592
            'dim': dim if dim != None else [0],
G
guosheng 已提交
3593 3594 3595 3596
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
3597 3598


C
caoying03 已提交
3599
def reduce_mean(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
3600
    """
Y
Yibing Liu 已提交
3601
    Computes the mean of the input tensor's elements along the given dimension.
G
guosheng 已提交
3602 3603 3604

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
Yibing Liu 已提交
3605 3606 3607
        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
3608
            must be in the range :math:`[-rank(input), rank(input))`. If
3609
            :math:`dim[i] < 0`, the dimension to reduce is
Y
Yibing Liu 已提交
3610
            :math:`rank(input) + dim[i]`.
Y
yangyaming 已提交
3611 3612
        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
3613
            than the :attr:`input` unless :attr:`keep_dim` is true.
Y
Yibing Liu 已提交
3614
        name(str|None): A name for this layer(optional). If set `None`, the layer
C
caoying03 已提交
3615
                       will be named automatically.
G
guosheng 已提交
3616 3617

    Returns:
Y
Yibing Liu 已提交
3618
        Variable: The reduced mean Variable.
F
fengjiayi 已提交
3619

G
guosheng 已提交
3620 3621 3622 3623 3624 3625 3626 3627 3628 3629
    Examples:
        .. code-block:: python

            # 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.
            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]
F
stash  
fengjiayi 已提交
3630 3631
            fluid.layers.reduce_mean(
                x, dim=1, keep_dim=True)  # [[0.475], [0.4]]
W
whs 已提交
3632 3633 3634 3635 3636 3637 3638

            # x is a Tensor variable with shape [2, 2, 2] and elements as below:
            #      [[[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.
            fluid.layers.reduce_mean(x, dim=[1, 2]) # [2.5, 6.5]
            fluid.layers.reduce_mean(x, dim=[0, 1]) # [4.0, 5.0]
G
guosheng 已提交
3639 3640
    """
    helper = LayerHelper('reduce_mean', **locals())
X
Xin Pan 已提交
3641
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
3642 3643
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
G
guosheng 已提交
3644 3645 3646 3647 3648
    helper.append_op(
        type='reduce_mean',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3649
            'dim': dim if dim != None else [0],
G
guosheng 已提交
3650 3651 3652 3653
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
3654 3655


C
caoying03 已提交
3656
def reduce_max(input, dim=None, keep_dim=False, name=None):
3657
    """
Y
yangyaming 已提交
3658
    Computes the maximum of tensor elements over the given dimension.
3659 3660 3661

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
3662
        dim (list|int|None): The dimension along which the maximum is computed.
Y
yangyaming 已提交
3663 3664 3665
            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
3666
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
Y
yangyaming 已提交
3667 3668
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
3669
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
3670 3671
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
3672 3673 3674

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
3675

3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686
    Examples:
        .. code-block:: python

            # 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.
            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
3687 3688 3689 3690 3691 3692 3693

            # x is a Tensor variable with shape [2, 2, 2] and elements as below:
            #      [[[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.
            fluid.layers.reduce_max(x, dim=[1, 2]) # [4.0, 8.0]
            fluid.layers.reduce_max(x, dim=[0, 1]) # [7.0, 8.0]
3694 3695
    """
    helper = LayerHelper('reduce_max', **locals())
X
Xin Pan 已提交
3696
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
3697 3698
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3699 3700 3701 3702 3703
    helper.append_op(
        type='reduce_max',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3704
            'dim': dim if dim != None else [0],
3705 3706 3707 3708 3709 3710
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


C
caoying03 已提交
3711
def reduce_min(input, dim=None, keep_dim=False, name=None):
3712
    """
Y
yangyaming 已提交
3713
    Computes the minimum of tensor elements over the given dimension.
3714 3715 3716

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
3717
        dim (list|int|None): The dimensions along which the minimum is computed.
Y
yangyaming 已提交
3718 3719 3720
            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
3721
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
Y
yangyaming 已提交
3722 3723
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
3724
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
3725 3726
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
3727 3728 3729

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
3730

3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741
    Examples:
        .. code-block:: python

            # 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.
            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
3742 3743 3744 3745 3746 3747 3748

            # x is a Tensor variable with shape [2, 2, 2] and elements as below:
            #      [[[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.
            fluid.layers.reduce_min(x, dim=[1, 2]) # [1.0, 5.0]
            fluid.layers.reduce_min(x, dim=[0, 1]) # [1.0, 2.0]
3749 3750
    """
    helper = LayerHelper('reduce_min', **locals())
X
Xin Pan 已提交
3751
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
3752 3753
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3754 3755 3756 3757 3758
    helper.append_op(
        type='reduce_min',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3759
            'dim': dim if dim != None else [0],
3760 3761 3762 3763
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
3764 3765


3766 3767 3768 3769 3770 3771
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
3772
        dim (list|int|None): The dimensions along which the product is performed. If
3773 3774
            :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
3775 3776
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
3777 3778 3779
        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
3780
        name(str|None): A name for this layer(optional). If set None, the
Z
zhouhanqing 已提交
3781
            layer will be named automatically.
3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795

    Returns:
        Variable: The reduced Tensor variable.

    Examples:
        .. code-block:: python

            # 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.
            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
3796
            fluid.layers.reduce_prod(x, dim=1,
Z
zhouhanqing 已提交
3797
                                     keep_dim=True)  # [[0.027], [0.0084]]
W
whs 已提交
3798 3799 3800 3801 3802 3803 3804

            # x is a Tensor variable with shape [2, 2, 2] and elements as below:
            #      [[[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.
            fluid.layers.reduce_prod(x, dim=[1, 2]) # [24.0, 1680.0]
            fluid.layers.reduce_prod(x, dim=[0, 1]) # [105.0, 384.0]
3805 3806
    """
    helper = LayerHelper('reduce_prod', **locals())
X
Xin Pan 已提交
3807
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
3808 3809
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3810 3811 3812 3813 3814
    helper.append_op(
        type='reduce_prod',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3815
            'dim': dim if dim != None else [0],
3816 3817 3818 3819 3820 3821
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


C
caoying03 已提交
3822
def split(input, num_or_sections, dim=-1, name=None):
G
guosheng 已提交
3823
    """
C
caoying03 已提交
3824
    Split the input tensor into multiple sub-tensors.
G
guosheng 已提交
3825 3826 3827

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
C
caoying03 已提交
3828 3829 3830 3831 3832
        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
3833
            :attr:`dim` dimension orderly.
C
caoying03 已提交
3834
        dim (int): The dimension along which to split. If :math:`dim < 0`, the
G
guosheng 已提交
3835
            dimension to split along is :math:`rank(input) + dim`.
C
caoying03 已提交
3836 3837
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
3838 3839

    Returns:
D
dzhwinter 已提交
3840
        list(Variable): The list of segmented tensor variables.
G
guosheng 已提交
3841 3842 3843 3844 3845 3846 3847 3848 3849

    Examples:
        .. code-block:: python

            # x is a Tensor variable with shape [3, 9, 5]:
            x0, x1, x2 = fluid.layers.split(x, num_or_sections=3, dim=1)
            x0.shape  # [3, 3, 5]
            x1.shape  # [3, 3, 5]
            x2.shape  # [3, 3, 5]
F
stash  
fengjiayi 已提交
3850 3851
            x0, x1, x2 = fluid.layers.split(
                x, num_or_sections=[2, 3, 4], dim=1)
G
guosheng 已提交
3852 3853 3854 3855 3856 3857 3858 3859 3860 3861 3862 3863 3864 3865 3866
            x0.shape  # [3, 2, 5]
            x1.shape  # [3, 3, 5]
            x2.shape  # [3, 4, 5]
    """
    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:
        assert len(num_or_sections) < input_shape[
            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
3867
        helper.create_variable_for_type_inference(dtype=helper.input_dtype())
G
guosheng 已提交
3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880
        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
3881 3882 3883 3884 3885 3886 3887 3888 3889


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

3890
    .. math::
3891 3892

        y = \\frac{x}{ \sqrt{\sum {x^2} + epsion }}
C
caoying03 已提交
3893 3894 3895 3896 3897

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

    Args:
3898
        x(Variable|list): The input tensor to l2_normalize layer.
3899
        axis(int): The axis on which to apply normalization. If `axis < 0`, \
3900 3901
            the dimension to normalization is rank(X) + axis. -1 is the
            last dimension.
3902
        epsilon(float): The epsilon value is used to avoid division by zero, \
3903
            the defalut value is 1e-10.
3904
        name(str|None): A name for this layer(optional). If set None, the layer \
3905
            will be named automatically.
C
caoying03 已提交
3906 3907

    Returns:
3908
        Variable: The output tensor variable is the same shape with `x`.
C
caoying03 已提交
3909 3910

    Examples:
3911

C
caoying03 已提交
3912 3913
        .. code-block:: python

3914 3915 3916 3917
            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
3918 3919
    """

F
fengjiayi 已提交
3920 3921
    if len(x.shape) == 1:
        axis = 0
C
caoying03 已提交
3922 3923
    helper = LayerHelper("l2_normalize", **locals())

X
Xin Pan 已提交
3924 3925
    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
3926
    helper.append_op(
3927 3928 3929 3930
        type="norm",
        inputs={"X": x},
        outputs={"Out": out,
                 "Norm": norm},
C
caoying03 已提交
3931
        attrs={
3932 3933
            "axis": 1 if axis is None else axis,
            "epsilon": epsilon,
C
caoying03 已提交
3934 3935
        })
    return out
3936 3937


S
sneaxiy 已提交
3938
def matmul(x, y, transpose_x=False, transpose_y=False, alpha=1.0, name=None):
G
guosheng 已提交
3939
    """
Y
ying 已提交
3940 3941 3942 3943
    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
3944

C
chengduoZH 已提交
3945
    The actual behavior depends on the shapes of :math:`x`, :math:`y` and the
3946
    flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically:
G
guosheng 已提交
3947

3948 3949 3950 3951 3952
    - 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
3953
      :math:`[1, D]` in transposed form.
G
guosheng 已提交
3954

C
chengduoZH 已提交
3955
    - After transpose, the two tensors are 2-D or n-D and matrix multiplication
3956
      performs in the following way.
G
guosheng 已提交
3957

3958
      - If both are 2-D, they are multiplied like conventional matrices.
C
chengduoZH 已提交
3959
      - If either is n-D, it is treated as a stack of matrices residing in the
Y
ying 已提交
3960
        last two dimensions and a batched matrix multiply supporting broadcast
3961
        applies on the two tensors.
G
guosheng 已提交
3962

Y
ying 已提交
3963 3964
    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
3965
    removed after matrix multiplication.
G
guosheng 已提交
3966 3967 3968

    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
3969 3970 3971
        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
3972
        alpha (float): The scale of output. Default 1.0.
3973
        name(str|None): A name for this layer(optional). If set None, the layer
3974
            will be named automatically.
G
guosheng 已提交
3975 3976

    Returns:
3977
        Variable: The product Tensor variable.
G
guosheng 已提交
3978

G
guosheng 已提交
3979 3980 3981
    Examples:
        .. code-block:: python

3982
            # Examples to clarify shapes of the inputs and output
C
chengduoZH 已提交
3983 3984
            # x: [B, ..., M, K], y: [B, ..., K, N]
            fluid.layers.matmul(x, y)  # out: [B, ..., M, N]
Y
ying 已提交
3985

3986 3987
            # x: [B, M, K], y: [B, K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
3988

3989 3990
            # x: [B, M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
3991

3992 3993
            # x: [M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [M, N]
Y
ying 已提交
3994 3995 3996 3997

            # x: [B, M, K], y: [K]
            fluid.layers.matmul(x, y)  # out: [B, M]

3998 3999
            # x: [K], y: [K]
            fluid.layers.matmul(x, y)  # out: [1]
4000

Y
ying 已提交
4001
            # x: [M], y: [N]
4002
            fluid.layers.matmul(x, y, True, True)  # out: [M, N]
G
guosheng 已提交
4003
    """
Y
ying 已提交
4004 4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015

    def __check_input(x, y):
        if len(y.shape) > len(x.shape):
            raise ValueError(
                "Invalid inputs for matmul. "
                "x's rank should be always greater than or equal to y'rank.")

        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
4016
            y_shape = y_shape + [1]
Y
ying 已提交
4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031 4032

        # 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]:
            raise ValueError("Invalid inputs for matmul.")

        if len(y_shape) > 2:
            for i, dim_x in enumerate(x_shape[:-2]):
                if dim_x != y_shape[i]:
                    raise ValueError("Invalid inputs for matmul.")

    __check_input(x, y)

4033
    helper = LayerHelper('matmul', **locals())
X
Xin Pan 已提交
4034
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
G
guosheng 已提交
4035
    helper.append_op(
4036 4037 4038 4039
        type='matmul',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
S
sneaxiy 已提交
4040 4041 4042
        attrs={
            'transpose_X': transpose_x,
            'transpose_Y': transpose_y,
S
sneaxiy 已提交
4043
            'alpha': float(alpha),
S
sneaxiy 已提交
4044
        })
4045
    return out
4046 4047


4048
def topk(input, k, name=None):
Q
qingqing01 已提交
4049 4050 4051 4052
    """
    This operator is used to find values and indices of the k largest entries
    for the last dimension.

F
fengjiayi 已提交
4053
    If the input is a vector (1-D Tensor), finds the k largest entries in the vector
Q
qingqing01 已提交
4054 4055 4056 4057 4058 4059
    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
4060 4061 4062 4063 4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077 4078 4079 4080
    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
4081 4082 4083
    Args:
        input(Variable): The input variable which can be a vector or Tensor with
            higher rank.
4084
        k(int):  The number of top elements to look for along the last dimension
F
fengjiayi 已提交
4085
                 of input.
4086
        name(str|None): A name for this layer(optional). If set None, the layer
4087
                       will be named automatically.
F
fengjiayi 已提交
4088
                       Default: None
Q
qingqing01 已提交
4089 4090

    Returns:
4091 4092 4093
        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
4094
        within the last dimension of input.
Q
qingqing01 已提交
4095

F
fengjiayi 已提交
4096 4097
    Raises:
        ValueError: If k < 1 or k is not less than the last dimension of input
Q
qingqing01 已提交
4098 4099 4100 4101 4102 4103 4104

    Examples:
        .. code-block:: python

            top5_values, top5_indices = layers.topk(input, k=5)
    """
    helper = LayerHelper("top_k", **locals())
X
Xin Pan 已提交
4105 4106
    values = helper.create_variable_for_type_inference(dtype=input.dtype)
    indices = helper.create_variable_for_type_inference(dtype="int64")
Q
qingqing01 已提交
4107 4108 4109 4110 4111 4112 4113 4114 4115 4116 4117
    helper.append_op(
        type="top_k",
        inputs={"X": [input]},
        outputs={"Out": [values],
                 "Indices": [indices]},
        attrs={"k": k})
    values.stop_gradient = True
    indices.stop_gradient = True
    return values, indices


4118
def edit_distance(input, label, normalized=True, ignored_tokens=None):
4119
    """
Y
ying 已提交
4120 4121 4122 4123 4124 4125 4126 4127 4128
    EditDistance operator computes the edit distances between a batch of
    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
4129

Y
ying 已提交
4130
    "kitten" -> "sitten" -> "sittin" -> "sitting"
W
wanghaoshuang 已提交
4131

4132
    The input is a LoDTensor consisting of all the hypothesis strings with
Y
ying 已提交
4133 4134
    the total number denoted by `batch_size`, and the separation is specified
    by the LoD information. And the `batch_size` reference strings are arranged
4135
    in order in the same way in the input LoDTensor.
W
wanghaoshuang 已提交
4136

4137
    The output contains the `batch_size` results and each stands for the edit
Y
ying 已提交
4138 4139
    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
4140

4141 4142 4143
    Args:
        input(Variable): The indices for hypothesis strings.
        label(Variable): The indices for reference strings.
4144
        normalized(bool, default True): Indicated whether to normalize the edit distance by
Y
ying 已提交
4145
                          the length of reference string.
4146
        ignored_tokens(list<int>, default None): Tokens that should be removed before
Y
ying 已提交
4147
                                     calculating edit distance.
4148
        name (str): The name of this layer. It is optional.
4149

W
wanghaoshuang 已提交
4150
    Returns:
W
wanghaoshuang 已提交
4151
        Variable: sequence-to-sequence edit distance in shape [batch_size, 1].
W
wanghaoshuang 已提交
4152 4153 4154 4155

    Examples:
        .. code-block:: python

T
tink2123 已提交
4156 4157
            x = fluid.layers.data(name='x', shape=[1], dtype='float32')
            y = fluid.layers.data(name='y', shape=[1], dtype='float32')
4158
            cost = fluid.layers.edit_distance(input=x,label=y)
4159
    """
4160
    helper = LayerHelper("edit_distance", **locals())
4161

4162
    # remove some tokens from input and labels
W
wanghaoshuang 已提交
4163
    if ignored_tokens is not None and len(ignored_tokens) > 0:
X
Xin Pan 已提交
4164 4165
        erased_input = helper.create_variable_for_type_inference(dtype="int64")
        erased_label = helper.create_variable_for_type_inference(dtype="int64")
4166 4167 4168 4169 4170

        helper.append_op(
            type="sequence_erase",
            inputs={"X": [input]},
            outputs={"Out": [erased_input]},
W
wanghaoshuang 已提交
4171
            attrs={"tokens": ignored_tokens})
4172 4173 4174 4175 4176
        input = erased_input

        helper.append_op(
            type="sequence_erase",
            inputs={"X": [label]},
W
whs 已提交
4177
            outputs={"Out": [erased_label]},
W
wanghaoshuang 已提交
4178
            attrs={"tokens": ignored_tokens})
4179 4180
        label = erased_label

4181
    # edit distance op
X
Xin Pan 已提交
4182 4183
    edit_distance_out = helper.create_variable_for_type_inference(dtype="int64")
    sequence_num = helper.create_variable_for_type_inference(dtype="int64")
4184 4185 4186 4187
    helper.append_op(
        type="edit_distance",
        inputs={"Hyps": [input],
                "Refs": [label]},
4188 4189
        outputs={"Out": [edit_distance_out],
                 "SequenceNum": [sequence_num]},
4190 4191
        attrs={"normalized": normalized})

4192
    return edit_distance_out, sequence_num
4193 4194 4195 4196 4197


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
4198

Y
ying 已提交
4199 4200 4201 4202
    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.
4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219

    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]]

4220
        input.lod = [[4, 4]]
4221 4222 4223 4224 4225 4226 4227

        Then:

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

4228
        output.lod = [[2, 1]]
4229 4230 4231

    Args:

Y
ying 已提交
4232 4233 4234 4235 4236 4237 4238 4239 4240
        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).
4241
        name (str): The name of this layer. It is optional.
4242 4243

    Returns:
4244
        Variable: CTC greedy decode result. If all the sequences in result were
4245
        empty, the result LoDTensor will be [-1] with LoD [[]] and dims [1, 1].
4246 4247 4248 4249 4250

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
W
wanghaoshuang 已提交
4251

4252
            cost = fluid.layers.ctc_greedy_decoder(input=x, blank=0)
W
wanghaoshuang 已提交
4253
    """
4254
    helper = LayerHelper("ctc_greedy_decoder", **locals())
Q
qingqing01 已提交
4255
    _, topk_indices = topk(input, k=1)
4256 4257

    # ctc align op
X
Xin Pan 已提交
4258
    ctc_out = helper.create_variable_for_type_inference(dtype="int64")
4259 4260 4261
    helper.append_op(
        type="ctc_align",
        inputs={"Input": [topk_indices]},
W
wanghaoshuang 已提交
4262
        outputs={"Output": [ctc_out]},
4263 4264
        attrs={"merge_repeated": True,
               "blank": blank})
4265
    return ctc_out
4266 4267


W
Wu Yi 已提交
4268
def warpctc(input, label, blank=0, norm_by_times=False, use_cudnn=False):
W
wanghaoshuang 已提交
4269
    """
4270 4271
    An operator integrating the open source Warp-CTC library
    (https://github.com/baidu-research/warp-ctc)
W
wanghaoshuang 已提交
4272
    to compute Connectionist Temporal Classification (CTC) loss.
4273 4274
    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
4275 4276 4277
    input tensor.

    Args:
4278
       input (Variable): The unscaled probabilities of variable-length sequences,
W
wanghaoshuang 已提交
4279 4280 4281 4282
         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).
4283
       label (Variable): The ground truth of variable-length sequence,
4284 4285 4286
         which is a 2-D Tensor with LoD information. It is of the shape [Lg, 1],
         where Lg is th sum of all labels' length.
       blank (int, default 0): The blank label index of Connectionist
W
wanghaoshuang 已提交
4287 4288
         Temporal Classification (CTC) loss, which is in the
         half-opened interval [0, num_classes + 1).
4289 4290 4291
       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
4292
         follewed by a mean_op.
W
Wu Yi 已提交
4293
       use_cudnn (bool, default false): Whether to use cudnn.
W
wanghaoshuang 已提交
4294 4295

    Returns:
4296 4297
        Variable: The Connectionist Temporal Classification (CTC) loss,
        which is a 2-D Tensor of the shape [batch_size, 1].
W
wanghaoshuang 已提交
4298 4299

    Examples:
4300

W
wanghaoshuang 已提交
4301
        .. code-block:: python
4302

4303 4304 4305
            label = fluid.layers.data(shape=[11, 8], dtype='float32', lod_level=1)
            predict = fluid.layers.data(shape=[11, 1], dtype='float32')
            cost = fluid.layers.warpctc(input=predict, label=label)
W
wanghaoshuang 已提交
4306 4307

    """
F
fengjiayi 已提交
4308
    helper = LayerHelper('warpctc', **locals())
X
Xin Pan 已提交
4309 4310
    loss_out = helper.create_variable_for_type_inference(dtype=input.dtype)
    grad_out = helper.create_variable_for_type_inference(dtype=input.dtype)
W
wanghaoshuang 已提交
4311 4312 4313 4314 4315 4316
    helper.append_op(
        type='warpctc',
        inputs={'Logits': [input],
                'Label': [label]},
        outputs={'WarpCTCGrad': [grad_out],
                 'Loss': [loss_out]},
W
Wu Yi 已提交
4317 4318 4319 4320 4321
        attrs={
            'blank': blank,
            'norm_by_times': norm_by_times,
            'use_cudnn': use_cudnn
        })
W
wanghaoshuang 已提交
4322
    return loss_out
4323 4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335 4336 4337


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]]
4338 4339 4340
            x.data = [[1,  2], [3,  4],
                      [5,  6], [7,  8],
                      [9, 10], [11, 12]]
4341 4342 4343 4344 4345
            x.dims = [6, 2]

        set new_dim = 4

        then out is a LoDTensor:
4346

4347
            out.lod  = [[0, 1, 3]]
4348 4349 4350 4351

            out.data = [[1,  2,  3,  4],
                        [5,  6,  7,  8],
                        [9, 10, 11, 12]]
4352 4353 4354 4355 4356 4357 4358
            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:
4359 4360 4361

       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.
4362 4363

    Returns:
4364

4365 4366 4367 4368 4369
        Variable: Reshaped LoDTensor according to new dimension.

    Examples:
        .. code-block:: python

4370
            x = fluid.layers.data(shape=[5, 20], dtype='float32', lod_level=1)
4371
            x_reshaped = fluid.layers.sequence_reshape(input=x, new_dim=10)
4372 4373
    """
    helper = LayerHelper('sequence_reshape', **locals())
X
Xin Pan 已提交
4374
    out = helper.create_variable_for_type_inference(helper.input_dtype())
4375 4376 4377 4378 4379 4380
    helper.append_op(
        type='sequence_reshape',
        inputs={'X': [input]},
        outputs={'Out': [out]},
        attrs={'new_dim': new_dim})
    return out
Y
ying 已提交
4381 4382


4383 4384 4385 4386
# 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
4387 4388 4389 4390 4391 4392
def nce(input,
        label,
        num_total_classes,
        sample_weight=None,
        param_attr=None,
        bias_attr=None,
C
chengduo 已提交
4393
        num_neg_samples=None,
4394 4395 4396 4397
        name=None,
        sampler="uniform",
        custom_dist=None,
        seed=0):
4398 4399 4400 4401 4402 4403 4404
    """
    ${comment}

    Args:
        input (Variable): input variable.
        label (Variable): label.
        num_total_classes (int):${num_total_classes_comment}
4405 4406
        sample_weight (Variable|None): A Variable of shape [batch_size, 1]
            storing a weight for each sample. The default weight for each
4407
            sample is 1.0.
C
chengduo 已提交
4408 4409 4410 4411 4412 4413 4414 4415 4416
        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.
4417
        num_neg_samples (int): ${num_neg_samples_comment}
C
chengduo 已提交
4418 4419
        name (str|None): A name for this layer(optional). If set None, the layer
             will be named automatically. Default: None.
4420 4421 4422
        sampler (str): The sampler used to sample class from negtive classes.
                       It can be 'uniform', 'log_uniform' or 'custom_dist'.
                       default: 'uniform'.
4423
        custom_dist (Variable): A tensor with shape [num_total_classes].
4424 4425 4426 4427
                       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.
F
fengjiayi 已提交
4428

4429
    Returns:
Y
Yibing Liu 已提交
4430 4431 4432 4433 4434 4435 4436 4437 4438 4439 4440 4441 4442 4443 4444 4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456
        Variable: The output nce loss.

    Examples:
        .. code-block:: python

            window_size = 5
            words = []
            for i in xrange(window_size):
                words.append(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 = layers.embedding(input=words[i], size=[dict_size, 32],
                                       param_attr='emb.w', is_sparse=True)
                embs.append(emb)

            embs = layers.concat(input=embs, axis=1)
            loss = layers.nce(input=embs, label=words[label_word],
                          num_total_classes=dict_size, param_attr='nce.w',
                          bias_attr='nce.b')
4457 4458 4459 4460 4461 4462 4463 4464 4465

            #or use custom distribution
            dist = fluid.layers.assign(input=np.array([0.05,0.5,0.1,0.3,0.05]).astype("float32"))
            loss = layers.nce(input=embs, label=words[label_word],
                          num_total_classes=5, param_attr='nce.w',
                          bias_attr='nce.b',
                          num_neg_samples=3,
                          sampler="custom_dist",
                          custom_dist=dist)
4466

4467
    """
Y
Yang Yu 已提交
4468 4469 4470
    helper = LayerHelper('nce', **locals())
    assert isinstance(input, Variable)
    assert isinstance(label, Variable)
C
chengduo 已提交
4471 4472

    dim = input.shape[1]
Y
Yang Yu 已提交
4473 4474 4475 4476 4477 4478
    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)
C
chengduo 已提交
4479 4480 4481 4482 4483 4484 4485 4486 4487 4488 4489 4490 4491
    inputs = {
        'Input': input,
        'Label': label,
        'Weight': w,
        'SampleWeight': sample_weight if sample_weight is not None else []
    }
    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
4492 4493 4494
    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
4495

Y
Yang Yu 已提交
4496 4497 4498 4499 4500
    if num_neg_samples is None:
        num_neg_samples = 10
    else:
        num_neg_samples = int(num_neg_samples)

4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520
    inputs = {
        'Input': input,
        'Label': label,
        'Weight': w,
        'Bias': b,
        'SampleWeight': sample_weight if sample_weight is not None else []
    }

    if sampler == "uniform":
        sampler = 0
    elif sampler == "log_uniform":
        sampler = 1
    elif sampler == "custom_dist":
        assert custom_dist is not None
        assert isinstance(custom_dist, Variable)
        inputs['CustomDistribution'] = custom_dist
        sampler = 2
    else:
        raise Exception("Unsupported sampler type.")

Y
Yang Yu 已提交
4521 4522
    attrs = {
        'num_total_classes': int(num_total_classes),
4523 4524 4525
        'num_neg_samples': num_neg_samples,
        'seed': seed,
        'sampler': sampler
Y
Yang Yu 已提交
4526
    }
Y
Yang Yu 已提交
4527 4528 4529

    helper.append_op(
        type='nce',
C
chengduo 已提交
4530
        inputs=inputs,
Y
Yang Yu 已提交
4531 4532 4533 4534 4535 4536
        outputs={
            'Cost': cost,
            'SampleLogits': sample_logits,
            'SampleLabels': sample_labels
        },
        attrs=attrs)
Y
Yang Yu 已提交
4537
    return cost / (num_neg_samples + 1)
4538 4539


C
chengduo 已提交
4540 4541 4542 4543 4544 4545
def hsigmoid(input,
             label,
             num_classes,
             param_attr=None,
             bias_attr=None,
             name=None):
W
weixing02 已提交
4546 4547
    """
    The hierarchical sigmoid operator is used to accelerate the training
M
minqiyang 已提交
4548
    process of language model. This operator organizes the classes into a
G
guosheng 已提交
4549 4550 4551 4552 4553 4554 4555 4556 4557
    complete binary tree, each leaf node represents a class(a word) and each
    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.

    Refer to `Hierarchical Probabilistic Neural Network Language Model
    <http://www.iro.umontreal.ca/~lisa/pointeurs/hierarchical-nnlm-aistats05.pdf>`_
M
minqiyang 已提交
4558

W
weixing02 已提交
4559
    Args:
M
minqiyang 已提交
4560
        input (Variable): The input tensor variable with shape
G
guosheng 已提交
4561 4562 4563 4564 4565
            :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]`.
        num_classes: (int), The number of classes, must not be less than 2.
C
chengduo 已提交
4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576
        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.
W
weixing02 已提交
4577 4578 4579 4580 4581 4582 4583 4584

    Returns:
        Out: (Tensor) The cost of hierarchical sigmoid operator. the shape is [N, 1]

    Examples:

        .. code-block:: python

G
guosheng 已提交
4585 4586 4587
            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
4588 4589 4590 4591
    """

    helper = LayerHelper('hierarchical_sigmoid', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
4592 4593
    out = helper.create_variable_for_type_inference(dtype)
    pre_out = helper.create_variable_for_type_inference(dtype)
W
weixing02 已提交
4594 4595
    dim = input.shape[1]
    if num_classes < 2:
G
guosheng 已提交
4596
        raise ValueError("num_classes must not be less than 2.")
W
weixing02 已提交
4597 4598 4599 4600 4601
    weights = helper.create_parameter(
        attr=helper.param_attr,
        shape=[num_classes - 1, dim],
        is_bias=False,
        dtype=input.dtype)
W
weixing02 已提交
4602 4603 4604 4605 4606 4607 4608 4609
    inputs = {"X": input, "W": weights, "Label": label}
    if helper.bias_attr:
        bias = helper.create_parameter(
            attr=helper.bias_attr,
            shape=[1, num_classes - 1],
            is_bias=True,
            dtype=input.dtype)
        inputs['Bias'] = bias
W
weixing02 已提交
4610 4611
    helper.append_op(
        type="hierarchical_sigmoid",
W
weixing02 已提交
4612
        inputs=inputs,
W
weixing02 已提交
4613 4614 4615 4616 4617 4618
        outputs={"Out": out,
                 "PreOut": pre_out},
        attrs={"num_classes": num_classes})
    return out


Y
fix ci.  
ying 已提交
4619
def transpose(x, perm, name=None):
Y
ying 已提交
4620 4621 4622 4623 4624 4625 4626
    """
    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:
4627 4628 4629
        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
4630 4631 4632 4633 4634 4635 4636

    Returns:
        Variable: A transposed Tensor.

    Examples:
        .. code-block:: python

4637
            # use append_batch_size=False to avoid prepending extra
4638
            # batch size in shape
4639
            x = fluid.layers.data(name='x', shape=[5, 10, 15],
4640
                            dtype='float32', append_batch_size=False)
Y
fix ci.  
ying 已提交
4641
            x_transposed = layers.transpose(x, perm=[1, 0, 2])
Y
ying 已提交
4642 4643
    """

Y
fix ci.  
ying 已提交
4644
    if len(perm) != len(x.shape):
Y
ying 已提交
4645 4646 4647
        raise ValueError(
            "Input(perm) is the permutation of dimensions of Input(input). "
            "It's length shoud be equal to Input(input)'s rank.")
Y
ying 已提交
4648 4649 4650 4651 4652 4653
    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
4654 4655

    helper = LayerHelper('transpose', **locals())
X
Xin Pan 已提交
4656 4657
    out = helper.create_variable_for_type_inference(x.dtype)
    x_shape = helper.create_variable_for_type_inference(x.dtype)
Y
ying 已提交
4658
    helper.append_op(
4659
        type='transpose2',
Y
fix ci.  
ying 已提交
4660
        inputs={'X': [x]},
4661 4662
        outputs={'Out': [out],
                 'XShape': [x_shape]},
Y
ying 已提交
4663 4664
        attrs={'axis': perm})
    return out
4665 4666


4667 4668 4669 4670 4671 4672 4673
def im2sequence(input,
                filter_size=1,
                stride=1,
                padding=0,
                input_image_size=None,
                out_stride=1,
                name=None):
4674
    """
4675 4676 4677 4678 4679 4680 4681
    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:
4682 4683 4684 4685 4686 4687 4688 4689 4690 4691

    .. 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
4692 4693 4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706 4707 4708 4709

        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.

4710 4711 4712 4713 4714 4715 4716 4717 4718
        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.

4719 4720 4721
        name (int): The name of this layer. It is optional.

    Returns:
W
wanghaoshuang 已提交
4722 4723 4724 4725 4726
        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.
4727 4728 4729 4730 4731 4732 4733 4734 4735 4736 4737 4738 4739 4740 4741 4742 4743 4744 4745 4746 4747 4748 4749 4750 4751 4752 4753

    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
4754 4755 4756
            filter = [2, 2]
            stride = [1, 1]
            padding = [0, 0]
4757 4758 4759 4760 4761 4762 4763 4764 4765 4766 4767 4768

            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.]]

4769
            output.dims = {8, 8}
4770

4771
            output.lod = [[4, 4]]
4772

D
dzhwinter 已提交
4773
     Examples:
4774 4775 4776

        .. code-block:: python

4777 4778
            output = fluid.layers.im2sequence(
                input=layer, stride=[1, 1], filter_size=[2, 2])
4779 4780

    """
W
wanghaoshuang 已提交
4781 4782 4783 4784 4785 4786 4787 4788 4789 4790

    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])
4791 4792 4793 4794 4795 4796 4797
    inputs = {"X": input}
    attrs = {"kernels": filter_size, "strides": stride, "padding": padding}
    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
4798
    helper = LayerHelper('im2sequence', **locals())
X
Xin Pan 已提交
4799
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
4800
    helper.append_op(
4801
        type='im2sequence', inputs=inputs, outputs={'Out': out}, attrs=attrs)
4802
    return out
4803 4804


Y
yuyang18 已提交
4805
@templatedoc()
4806
def row_conv(input, future_context_size, param_attr=None, act=None):
Y
yuyang18 已提交
4807 4808
    """
    ${comment}
4809 4810

    Args:
Y
yuyang18 已提交
4811
        input (${x_type}): ${x_comment}.
Y
yangyaming 已提交
4812 4813
        future_context_size (int): Future context size. Please note, the shape
            of convolution kernel is [future_context_size + 1, D].
4814 4815 4816 4817 4818
        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
4819
        ${out_comment}.
4820 4821

    Examples:
Y
yuyang18 已提交
4822 4823 4824 4825
        >>> 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)
4826 4827 4828 4829 4830 4831
    """
    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
4832
    out = helper.create_variable_for_type_inference(dtype)
4833 4834 4835 4836 4837
    helper.append_op(
        type='row_conv',
        inputs={'X': [input],
                'Filter': [filter_param]},
        outputs={'Out': [out]})
Y
yangyaming 已提交
4838
    return helper.append_activation(out)
4839 4840


Y
yuyang18 已提交
4841
@templatedoc()
4842 4843
def multiplex(inputs, index):
    """
Y
yuyang18 已提交
4844 4845 4846 4847 4848 4849 4850
    ${comment}

    >>> 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)
4851 4852

    Args:
Y
yuyang18 已提交
4853 4854
       inputs (list): ${x_comment}.
       index (${ids_type}): ${ids_comment}.
4855 4856

    Returns:
Y
yuyang18 已提交
4857
        ${out_comment}.
4858 4859
    """
    helper = LayerHelper('multiplex', **locals())
Y
yangyaming 已提交
4860 4861 4862 4863 4864

    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
4865
    out = helper.create_variable_for_type_inference(inputs[0].dtype)
4866 4867 4868 4869 4870 4871
    helper.append_op(
        type='multiplex',
        inputs={'X': inputs,
                'Ids': index},
        outputs={'Out': [out]})
    return out
4872 4873


4874 4875 4876
def softmax_with_cross_entropy(logits,
                               label,
                               soft_label=False,
S
sneaxiy 已提交
4877
                               ignore_index=-100,
4878 4879
                               numeric_stable_mode=False,
                               return_softmax=False):
4880 4881
    """
    **Softmax With Cross Entropy Operator.**
4882

4883 4884 4885 4886
    Cross entropy loss with softmax is used as the output layer extensively. This
    operator computes the softmax normalized values for each row of the input
    tensor, after which cross-entropy loss is computed. This provides a more
    numerically stable gradient.
4887

4888 4889 4890
    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.
4891

4892 4893 4894
    When the attribute soft_label is set 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.
4895

4896
    The equation is as follows:
4897

4898
    1) Hard label (one-hot label, so every sample has exactly one class)
4899

4900 4901 4902 4903
    .. math::

        loss_j =  -\\text{logit}_{label_j} +
        \\log\\left(\\sum_{i=0}^{K}\\exp(\\text{logit}_i)\\right), j = 1,..., K
4904

4905 4906 4907
    2) Soft label (each sample can have a distribution over all classes)

    .. math::
4908

4909 4910 4911 4912
        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

S
sneaxiy 已提交
4913 4914 4915
    3) If numeric_stable_mode is True, softmax is calculated first by:

    .. math::
4916

S
sneaxiy 已提交
4917 4918 4919 4920 4921 4922 4923 4924
        max_j = \\max_{i=0}^{K}{\\text{logit}_i}

        log\\_max\\_sum_j = \\log\\sum_{i=0}^{K}\\exp(logit_i - max_j)

        softmax_j = \\exp(logit_j - max_j - {log\\_max\\_sum}_j)

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

4925 4926 4927 4928 4929 4930 4931 4932
    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. If soft_label
            is set to false, Label is a Tensor<int64> with shape [N x 1]. If
            soft_label is set to true, Label is a Tensor<float/double> with
        soft_label (bool): A flag to indicate whether to interpretate the given
            labels as soft labels. By default, `soft_label` is set to False.
M
minqiyang 已提交
4933 4934
        ignore_index (int): Specifies a target value that is ignored and does
                            not contribute to the input gradient. Only valid
4935
                            if soft_label is set to False. Default: -100
S
sneaxiy 已提交
4936 4937 4938
        numeric_stable_mode (bool): A flag to indicate whether to use a more
                                    numerically stable algorithm. Only valid
                                    when soft_label is False and GPU is used.
4939 4940 4941
                                    When soft_label is True or CPU is used,
                                    the algorithm is always numerically stable.
                                    Note that the speed may be slower when use
S
sneaxiy 已提交
4942
                                    stable algorithm. Default: False
4943
        return_softmax (bool): A flag indicating whether to return the softmax
4944
                               along with the cross entropy loss. Default: False
4945

4946
    Returns:
4947 4948 4949 4950
        Variable or Tuple of two Variables: Return the cross entropy loss if
                              `return_softmax` is False, otherwise the tuple
                              (loss, softmax), where the cross entropy loss is
                              a 2-D tensor with shape [N x 1], and softmax is a
4951
                              2-D tensor with shape [N x K].
4952 4953 4954 4955 4956 4957 4958

    Examples:
        .. code-block:: python

            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
4959 4960
            out = fluid.layers.softmax_with_cross_entropy(
                logits=fc, label=label)
4961 4962
    """
    helper = LayerHelper('softmax_with_cross_entropy', **locals())
X
Xin Pan 已提交
4963 4964
    softmax = helper.create_variable_for_type_inference(dtype=logits.dtype)
    loss = helper.create_variable_for_type_inference(dtype=logits.dtype)
4965 4966 4967 4968 4969 4970
    helper.append_op(
        type='softmax_with_cross_entropy',
        inputs={'Logits': logits,
                'Label': label},
        outputs={'Softmax': softmax,
                 'Loss': loss},
S
sneaxiy 已提交
4971 4972 4973 4974 4975
        attrs={
            'soft_label': soft_label,
            'ignore_index': ignore_index,
            'numeric_stable_mode': numeric_stable_mode
        })
4976 4977 4978 4979

    if return_softmax:
        return loss, softmax

4980 4981 4982 4983 4984
    return loss


def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None):
    """
Y
Yibing Liu 已提交
4985 4986
    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
4987
    For each instance, it computes the smooth L1 loss element by element first
4988
    and then sums all the losses. So the shape of ouput Variable is
4989
    [batch_size, 1].
4990

4991 4992
    Args:
        x (Variable): A tensor with rank at least 2. The input value of smooth
Q
qingqing01 已提交
4993
            L1 loss op with shape [batch_size, dim1, ..., dimN].
4994
        y (Variable): A tensor with rank at least 2. The target value of smooth
Y
Yibing Liu 已提交
4995
            L1 loss op with same shape as :attr:`x`.
4996
        inside_weight (Variable|None):  A tensor with rank at least 2. This
4997 4998
            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
4999
            by this tensor element by element.
5000
        outside_weight (Variable|None): A tensor with rank at least 2. This
5001 5002
            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
5003
            element by element.
5004
        sigma (float|None): Hyper parameter of smooth L1 loss layer. A float
5005 5006
           scalar with default value 1.0.

5007
    Returns:
5008
        Variable: The output smooth L1 loss with shape [batch_size, 1].
5009 5010 5011 5012 5013

    Examples:
        .. code-block:: python

            data = fluid.layers.data(name='data', shape=[128], dtype='float32')
F
stash  
fengjiayi 已提交
5014 5015
            label = fluid.layers.data(
                name='label', shape=[100], dtype='float32')
5016
            fc = fluid.layers.fc(input=data, size=100)
F
fengjiayi 已提交
5017
            out = fluid.layers.smooth_l1(x=fc, y=label)
5018
    """
5019

5020
    helper = LayerHelper('smooth_l1_loss', **locals())
X
Xin Pan 已提交
5021 5022
    diff = helper.create_variable_for_type_inference(dtype=x.dtype)
    loss = helper.create_variable_for_type_inference(dtype=x.dtype)
5023 5024 5025 5026 5027 5028 5029 5030 5031 5032 5033 5034
    helper.append_op(
        type='smooth_l1_loss',
        inputs={
            'X': x,
            'Y': y,
            'InsideWeight': inside_weight,
            'OutsideWeight': outside_weight
        },
        outputs={'Diff': diff,
                 'Out': loss},
        attrs={'sigma': sigma})
    return loss
5035 5036 5037 5038


def one_hot(input, depth):
    """
Y
Yibing Liu 已提交
5039
    This layer creates the one-hot representations for input indices.
5040 5041

    Args:
Y
Yibing Liu 已提交
5042 5043
        input(Variable): Input indices, last dimension must be 1.
        depth(scalar): An interger defining the depth of the one-hot dimension.
5044 5045

    Returns:
Y
Yibing Liu 已提交
5046
        Variable: The one-hot representations of input.
5047 5048

    Examples:
C
caoying03 已提交
5049
        .. code-block:: python
5050

Y
Yibing Liu 已提交
5051 5052
            label = layers.data(name="label", shape=[1], dtype="float32")
            one_hot_label = layers.one_hot(input=label, depth=10)
5053 5054
    """
    helper = LayerHelper("one_hot", **locals())
X
Xin Pan 已提交
5055
    one_hot_out = helper.create_variable_for_type_inference(dtype='float32')
5056 5057 5058 5059 5060 5061
    helper.append_op(
        type="one_hot",
        inputs={'X': input},
        attrs={'depth': depth},
        outputs={'Out': one_hot_out})
    return one_hot_out
Y
Yu Yang 已提交
5062 5063


Y
Yu Yang 已提交
5064
def autoincreased_step_counter(counter_name=None, begin=1, step=1):
Y
Yu Yang 已提交
5065
    """
Y
yi.wu 已提交
5066 5067 5068
    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
5069 5070 5071 5072 5073 5074

    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.

5075 5076
    Returns:
        Variable: The global run counter.
Y
yi.wu 已提交
5077 5078 5079 5080 5081 5082

    Examples:
        .. code-block:: python

           global_step = fluid.layers.autoincreased_step_counter(
               counter_name='@LR_DECAY_COUNTER@', begin=begin, step=1)
Y
Yu Yang 已提交
5083 5084
    """
    helper = LayerHelper('global_step_counter')
Y
Yu Yang 已提交
5085 5086
    if counter_name is None:
        counter_name = '@STEP_COUNTER@'
Y
Yu Yang 已提交
5087 5088 5089 5090 5091
    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
5092
                value=begin - 1, force_cpu=True))
W
Wu Yi 已提交
5093
        helper.main_program.global_block()._prepend_op(
Y
Yu Yang 已提交
5094 5095
            type='increment',
            inputs={'X': [counter]},
Y
Yu Yang 已提交
5096 5097
            outputs={'Out': [counter]},
            attrs={'step': float(step)})
Y
Yu Yang 已提交
5098 5099 5100
        counter.stop_gradient = True

    return counter
Y
yangyaming 已提交
5101 5102


5103
def reshape(x, shape, actual_shape=None, act=None, inplace=False, name=None):
C
caoying03 已提交
5104
    """
C
caoying03 已提交
5105 5106
    Gives a new shape to the input Tensor without changing its data.

5107 5108 5109 5110 5111
    The target shape can be given by :attr:`shape` or :attr:`actual_shape`.
    :attr:`shape` is a list of integer while :attr:`actual_shape` is a tensor
    variable. :attr:`actual_shape` has a higher priority than :attr:`shape`
    if it is provided, while :attr:`shape` still should be set correctly to
    gurantee shape inference in compile-time.
C
caoying03 已提交
5112

5113
    Some tricks exist when specifying the target shape.
C
caoying03 已提交
5114

5115 5116 5117 5118
    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.

5119
    2. 0 means the actual dimension value is going to be copied from the
5120 5121 5122 5123
    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
5124 5125

    1. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
W
wanghaoshuang 已提交
5126
    is [6, 8], the reshape operator will transform x into a 2-D tensor with
5127
    shape [6, 8] and leaving x's data unchanged.
C
caoying03 已提交
5128

5129
    2. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
5130 5131
    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
5132 5133
    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
5134
    dimensions.
C
caoying03 已提交
5135

5136
    3. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
5137 5138 5139 5140
    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
5141 5142

    Args:
5143
        x(variable): The input tensor.
C
caoying03 已提交
5144 5145
        shape(list): The new shape. At most one dimension of the new shape can
                     be -1.
5146 5147 5148 5149 5150
        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
                                than :attr:`shape`.
5151 5152
        act (str): The non-linear activation to be applied to the reshaped tensor
                   variable.
5153 5154 5155 5156 5157 5158 5159
        inplace(bool): Must use :attr:`False` if :attr:`x` is used in multiple
                       operators. If this flag is set :attr:`True`, reuse input
                       :attr:`x` to reshape, which will change the shape of
                       tensor variable :attr:`x` and might cause errors when
                       :attr:`x` is used in multiple operators. If :attr:`False`,
                       preserve the shape :attr:`x` and create a new output tensor
                       variable whose data is copied from input x but reshaped.
5160
        name (str): The name of this layer. It is optional.
C
caoying03 已提交
5161

5162
    Returns:
G
guosheng 已提交
5163 5164 5165 5166
        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
5167

X
Xin Pan 已提交
5168 5169 5170
    Raises:
        TypeError: if actual_shape is neither Variable nor None.

C
caoying03 已提交
5171 5172
    Examples:
        .. code-block:: python
G
guosheng 已提交
5173

5174
            data = fluid.layers.data(
5175
                name='data', shape=[2, 4, 6], dtype='float32')
C
caoying03 已提交
5176
            reshaped = fluid.layers.reshape(
G
guosheng 已提交
5177
                x=data, shape=[-1, 0, 3, 2], inplace=True)
C
caoying03 已提交
5178 5179 5180
    """

    if not (isinstance(shape, list) or isinstance(shape, tuple)):
L
luotao1 已提交
5181
        raise ValueError("Input shape must be a python list or tuple.")
X
Xin Pan 已提交
5182 5183 5184 5185 5186
    inputs = {"X": x}
    if isinstance(actual_shape, Variable):
        inputs["Shape"] = actual_shape
    elif actual_shape is not None:
        raise TypeError("actual_shape should either be Variable or None")
C
caoying03 已提交
5187

5188 5189 5190 5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201 5202
    # Validate the shape
    unk_dim_idx = -1
    for dim_idx, dim_size in enumerate(shape):
        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.")

5203
    helper = LayerHelper("reshape2", **locals())
5204 5205
    out = x if inplace else helper.create_variable_for_type_inference(
        dtype=x.dtype)
X
Xin Pan 已提交
5206
    x_shape = helper.create_variable_for_type_inference(dtype=x.dtype)
C
caoying03 已提交
5207
    helper.append_op(
5208
        type="reshape2",
X
Xin Pan 已提交
5209
        inputs=inputs,
D
dzhwinter 已提交
5210
        attrs={"shape": shape},
5211 5212
        outputs={"Out": out,
                 "XShape": x_shape})
C
caoying03 已提交
5213

D
dzhwinter 已提交
5214
    return helper.append_activation(out)
5215

5216

5217
def squeeze(input, axes, name=None):
Y
Yibing Liu 已提交
5218
    """
M
minqiyang 已提交
5219 5220 5221
    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
5222
    selected with shape entry not equal to one, an error is raised.
M
minqiyang 已提交
5223

Y
Yibing Liu 已提交
5224 5225
    Examples:
    Case 1:
M
minqiyang 已提交
5226
      Given
Y
Yibing Liu 已提交
5227 5228 5229 5230 5231 5232 5233 5234
        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)
M
minqiyang 已提交
5235
        and
Y
Yibing Liu 已提交
5236 5237 5238
          axes = []
        we get:
          Out.shape = (3, 5)
M
minqiyang 已提交
5239

Y
Yibing Liu 已提交
5240
    Args:
5241
        input (Variable): The input variable to be squeezed.
Y
Yibing Liu 已提交
5242
        axes (list): List of integers, indicating the dimensions to be squeezed.
5243
        name (str|None): Name for this layer.
Y
Yibing Liu 已提交
5244 5245 5246 5247 5248 5249 5250 5251

    Returns:
        Variable: Output squeezed variable.

    Examples:
        .. code-block:: python

            x = layers.data(name='x', shape=[5, 1, 10])
5252
            y = layers.sequeeze(input=x, axes=[1])
Y
Yibing Liu 已提交
5253 5254
    """
    helper = LayerHelper("squeeze", **locals())
X
Xin Pan 已提交
5255 5256
    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
5257
    helper.append_op(
5258
        type="squeeze2",
5259
        inputs={"X": input},
Y
Yibing Liu 已提交
5260
        attrs={"axes": axes},
5261 5262
        outputs={"Out": out,
                 "XShape": x_shape})
Y
Yibing Liu 已提交
5263

5264 5265 5266
    return out


5267
def unsqueeze(input, axes, name=None):
Y
Yibing Liu 已提交
5268
    """
M
minqiyang 已提交
5269 5270 5271
    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
5272

M
minqiyang 已提交
5273 5274
    For example:
      Given a tensor such that tensor with shape [3, 4, 5],
Y
Yibing Liu 已提交
5275
      then Unsqueezed tensor with axes=[0, 4] has shape [1, 3, 4, 5, 1].
M
minqiyang 已提交
5276

Y
Yibing Liu 已提交
5277
    Args:
5278
        input (Variable): The input variable to be unsqueezed.
Y
Yibing Liu 已提交
5279
        axes (list): List of integers, indicating the dimensions to be inserted.
5280
        name (str|None): Name for this layer.
Y
Yibing Liu 已提交
5281 5282 5283 5284 5285 5286 5287 5288

    Returns:
        Variable: Output unsqueezed variable.

    Examples:
        .. code-block:: python

            x = layers.data(name='x', shape=[5, 10])
5289
            y = layers.unsequeeze(input=x, axes=[1])
Y
Yibing Liu 已提交
5290 5291
    """
    helper = LayerHelper("unsqueeze", **locals())
X
Xin Pan 已提交
5292 5293
    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
5294
    helper.append_op(
5295
        type="unsqueeze2",
5296
        inputs={"X": input},
Y
Yibing Liu 已提交
5297
        attrs={"axes": axes},
5298 5299
        outputs={"Out": out,
                 "XShape": x_shape})
Y
Yibing Liu 已提交
5300

5301 5302
    return out

5303

Y
yangyaming 已提交
5304
def lod_reset(x, y=None, target_lod=None):
Y
yangyaming 已提交
5305
    """
Y
Yibing Liu 已提交
5306
    Set LoD of :attr:`x` to a new one specified by :attr:`y` or
5307 5308 5309 5310
    :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
Y
Yibing Liu 已提交
5311
    :attr:`Y.data` or :attr:`target_lod`, only one level LoD is supported.
Y
yangyaming 已提交
5312 5313 5314 5315 5316 5317

    .. code-block:: text

        * Example 1:

            Given a 1-level LoDTensor x:
5318
                x.lod =  [[ 2,           3,                   1 ]]
Y
yangyaming 已提交
5319 5320 5321
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

5322
            target_lod: [4, 2]
Y
yangyaming 已提交
5323 5324

            then we get a 1-level LoDTensor:
5325
                out.lod =  [[4,                          2]]
Y
yangyaming 已提交
5326 5327 5328 5329 5330 5331
                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:
5332
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
5333 5334 5335 5336
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a Tensor:
5337
                y.data = [[2, 4]]
Y
yangyaming 已提交
5338 5339 5340
                y.dims = [1, 3]

            then we get a 1-level LoDTensor:
5341
                out.lod =  [[2,            4]]
Y
yangyaming 已提交
5342 5343 5344 5345 5346 5347
                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:
5348
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
5349 5350 5351 5352
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a 2-level LoDTensor:
5353
                y.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
5354 5355 5356 5357
                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:
5358
                out.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
5359 5360 5361 5362 5363
                out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                out.dims = [6, 1]

    Args:
        x (Variable): Input variable which could be a Tensor or LodTensor.
5364
        y (Variable|None): If provided, output's LoD would be derived
Y
Yibing Liu 已提交
5365
                           from :attr:`y`.
Y
yangyaming 已提交
5366
        target_lod (list|tuple|None): One level LoD which should be considered
Y
Yibing Liu 已提交
5367
                                      as target LoD when :attr:`y` not provided.
Y
yangyaming 已提交
5368 5369

    Returns:
Y
Yibing Liu 已提交
5370
        Variable: Output variable with LoD specified by this layer.
Y
yangyaming 已提交
5371 5372

    Raises:
Y
Yibing Liu 已提交
5373
        ValueError: If :attr:`y` and :attr:`target_lod` are both None.
Y
yangyaming 已提交
5374 5375 5376 5377 5378 5379 5380 5381 5382

    Examples:
        .. code-block:: python

            x = layers.data(name='x', shape=[10])
            y = layers.data(name='y', shape=[10, 20], lod_level=2)
            out = layers.lod_reset(x=x, y=y)
    """
    helper = LayerHelper("lod_reset", **locals())
X
Xin Pan 已提交
5383
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
Y
yangyaming 已提交
5384 5385 5386 5387 5388 5389 5390 5391 5392 5393 5394 5395 5396 5397
    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:
        raise ValueError("y and target_lod should not be both None.")

    return out
D
dragonwarrior 已提交
5398 5399 5400 5401 5402 5403 5404 5405 5406 5407 5408


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::

D
dzhwinter 已提交
5409
      Output(i, x, y) = Input(i, x, y) / \\left(k + \\alpha \\sum\\limits^{\\min(C, c + n/2)}_{j = \\max(0, c - n/2)}(Input(j, x, y))^2\\right)^{\\beta}
D
dragonwarrior 已提交
5410 5411 5412 5413 5414 5415 5416 5417 5418 5419 5420 5421 5422 5423 5424 5425 5426 5427 5428 5429 5430 5431 5432 5433 5434 5435 5436 5437

    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

F
stash  
fengjiayi 已提交
5438 5439
          data = fluid.layers.data(
              name="data", shape=[3, 112, 112], dtype="float32")
D
dragonwarrior 已提交
5440 5441 5442 5443 5444 5445 5446 5447 5448 5449 5450 5451
          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
5452 5453 5454
    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
5455 5456 5457 5458 5459 5460 5461 5462 5463 5464 5465 5466 5467
    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
5468 5469 5470 5471


def pad(x, paddings, pad_value=0., name=None):
    """
G
guosheng 已提交
5472
    Pads a tensor with a constant value given by :attr:`pad_value`, and the
W
wanghaoshuang 已提交
5473
    padded width is specified by :attr:`paddings`.
G
guosheng 已提交
5474

G
guosheng 已提交
5475 5476 5477 5478
    Specifically, the number of values padded before the contents of :attr:`x`
    in dimension :attr:`i` is indicated by :attr:`paddings[i]`, and the number
    of values padded after the contents of :attr:`x` in dimension :attr:`i` is
    indicated by :attr:`paddings[i+1]`.
G
guosheng 已提交
5479 5480 5481 5482 5483 5484 5485 5486 5487 5488 5489 5490 5491 5492 5493 5494 5495 5496 5497 5498 5499 5500

    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
5501
                         The length of :attr:paddings must be
G
guosheng 已提交
5502 5503 5504 5505 5506 5507 5508 5509 5510 5511
                         :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
5512

G
guosheng 已提交
5513 5514 5515 5516 5517 5518
            # x is a rank 2 tensor variable.
            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
5519
    out = helper.create_variable_for_type_inference(dtype)
G
guosheng 已提交
5520 5521 5522 5523 5524 5525 5526
    helper.append_op(
        type='pad',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'paddings': paddings,
               'pad_value': float(pad_value)})
    return out
5527 5528


C
chengduo 已提交
5529 5530 5531 5532 5533 5534 5535 5536 5537 5538 5539 5540 5541 5542 5543 5544 5545 5546 5547 5548 5549 5550 5551 5552 5553 5554 5555 5556 5557 5558 5559 5560 5561 5562 5563 5564 5565 5566 5567 5568 5569 5570 5571 5572 5573 5574 5575 5576 5577 5578 5579 5580 5581 5582 5583 5584 5585 5586 5587 5588 5589 5590 5591 5592 5593 5594 5595 5596 5597 5598
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)

    And
        pad_value = -1,

    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)

    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)
            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
5599
    out = helper.create_variable_for_type_inference(dtype)
C
chengduo 已提交
5600 5601 5602 5603 5604 5605 5606 5607 5608
    helper.append_op(
        type='pad_constant_like',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
        attrs={'pad_value': float(pad_value)})
    return out


5609 5610 5611 5612 5613 5614 5615
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
5616 5617
    called label-smoothing regularization (LSR).

5618 5619 5620 5621 5622 5623 5624 5625 5626 5627 5628 5629 5630 5631 5632 5633 5634 5635 5636 5637 5638 5639 5640
    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
5641
                              be :math:`(1, class\_num)`.
5642 5643
        epsilon(float): The weight used to mix up the original ground-truth
                        distribution and the fixed distribution.
5644
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32,
5645 5646 5647 5648 5649 5650 5651 5652 5653 5654 5655 5656 5657 5658 5659 5660 5661 5662 5663
                                                  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

            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
5664
    smooth_label = helper.create_variable_for_type_inference(dtype)
5665 5666 5667 5668 5669 5670 5671
    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
5672 5673


W
wopeizl 已提交
5674 5675 5676 5677 5678 5679 5680 5681 5682 5683 5684 5685 5686 5687 5688 5689 5690 5691 5692 5693 5694 5695 5696 5697 5698 5699 5700 5701 5702 5703 5704 5705 5706 5707 5708 5709
@templatedoc()
def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0):
    """
    ${comment}

    Args:
        input (Variable): ${x_comment}
        rois (Variable): ROIs (Regions of Interest) to pool over.
        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

            pool_out = fluid.layers.roi_pool(input=x, rois=rois, 7, 7, 1.0)
    """
    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
5710 5711


J
jerrywgz 已提交
5712 5713 5714 5715 5716 5717
@templatedoc()
def roi_align(input,
              rois,
              pooled_height=1,
              pooled_width=1,
              spatial_scale=1.0,
J
jerrywgz 已提交
5718 5719
              sampling_ratio=-1,
              name=None):
J
jerrywgz 已提交
5720 5721 5722 5723 5724 5725 5726 5727 5728 5729 5730 5731 5732 5733 5734 5735
    """
    ${comment}

    Args:
        input (Variable): ${x_comment}
        rois (Variable): ROIs (Regions of Interest) to pool over.
        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

5736 5737 5738
            align_out = fluid.layers.roi_align(input=x,
                                               rois=rois,
                                               pooled_height=7,
J
jerrywgz 已提交
5739 5740 5741 5742 5743 5744
                                               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
5745
    align_out = helper.create_variable_for_type_inference(dtype)
J
jerrywgz 已提交
5746 5747 5748 5749 5750 5751 5752 5753 5754 5755 5756 5757 5758 5759
    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
5760 5761 5762 5763 5764 5765 5766 5767 5768 5769 5770 5771 5772 5773 5774 5775 5776 5777 5778 5779 5780 5781 5782 5783 5784 5785
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:
5786 5787
        .. code-block:: python

W
whs 已提交
5788 5789 5790 5791
            predictions = fluid.layers.softmax(x)
            loss = fluid.layers.dice_loss(input=predictions, label=label, 2)
    """
    label = one_hot(label, depth=input.shape[-1])
5792
    reduce_dim = list(range(1, len(input.shape)))
W
whs 已提交
5793 5794 5795 5796 5797 5798
    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)
5799 5800


5801 5802 5803 5804
def image_resize(input,
                 out_shape=None,
                 scale=None,
                 name=None,
5805 5806
                 resample='BILINEAR',
                 actual_shape=None):
5807
    """
Q
qiaolongfei 已提交
5808
    **Resize a Batch of Images**
F
stash  
fengjiayi 已提交
5809

5810
    The input must be a tensor of the shape (num_batches, channels, in_h, in_w),
5811 5812 5813
    and the resizing only applies on the last two dimensions(hight and width).

    Supporting resample methods:
Q
update  
qiaolongfei 已提交
5814

5815
        'BILINEAR' : Bilinear interpolation
5816
        'NEAREST' : Nearest neighbor interpolation
F
stash  
fengjiayi 已提交
5817

5818
    Args:
5819
        input (Variable): The input tensor of image resize layer,
5820 5821
                          This is a 4-D tensor of the shape
                          (num_batches, channels, in_h, in_w).
5822
        out_shape(list|tuple|Variable|None): Output shape of image resize
5823 5824
                                    layer, the shape is (out_h, out_w).
                                    Default: None
B
baiyf 已提交
5825
        scale(float|None): The multiplier for the input height or width.
5826 5827 5828
                         At least one of out_shape or scale must be set.
                         And out_shape has a higher priority than scale.
                         Default: None
5829 5830
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
5831
        resample(str): The resample method. It supports 'BILINEAR' and 'NEAREST'
5832
                       currently.
5833
                       Default: 'BILINEAR'
5834 5835 5836
        actual_shape(Variable): An optional input to specify output shape
                                dynamically. If provided, image resize
                                according to this given shape rather than
5837
                                :attr:`out_shape` and :attr:`scale` specifying
5838 5839 5840 5841 5842 5843 5844
                                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
5845 5846
                                constructing stage.
                                Default: None
5847 5848

    Returns:
Q
update  
qiaolongfei 已提交
5849 5850
        Variable: The output is a 4-D tensor of the shape
        (num_batches, channls, out_h, out_w).
F
stash  
fengjiayi 已提交
5851

5852 5853 5854
    Raises:
        TypeError: out_shape should be a list or tuple or Variable.
        TypeError: actual_shape should either be Variable or None.
5855
        ValueError: The 'resample' of image_resize can only be 'BILINEAR'
5856 5857 5858 5859
                    or 'NEAREST' currently.
        ValueError: One of out_shape and scale must not be None.
        ValueError: out_shape length should be 2.

5860 5861 5862
    Examples:
        .. code-block:: python

5863
            out = fluid.layers.image_resize(input, out_shape=[12, 12], resample="NEAREST")
5864
    """
5865 5866 5867 5868
    resample_methods = {
        'BILINEAR': 'bilinear',
        'NEAREST': 'nearest',
    }
5869 5870
    if resample not in resample_methods:
        raise ValueError(
5871
            "The 'resample' of image_resize can only be 'BILINEAR' or 'NEAREST' currently."
5872
        )
5873
    if out_shape is None and scale is None:
5874
        raise ValueError("One of out_shape and scale must not be None.")
5875
    helper = LayerHelper('interpolate', **locals())
5876
    dtype = helper.input_dtype()
5877 5878 5879 5880

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

5881 5882 5883
    out_h = 0
    out_w = 0
    inputs = {"X": input}
5884
    if out_shape is not None:
5885 5886 5887 5888
        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.")
5889
            inputs['OutSize'] = out_shape
5890 5891 5892 5893 5894 5895 5896 5897
        elif not (_is_list_or_turple_(out_shape)):
            raise TypeError("out_shape should be a list or tuple or Variable.")
        elif len(out_shape) != 2:
            raise ValueError("out_shape length should be 2.")

        out_shape = list(map(int, out_shape))
        out_h = out_shape[0]
        out_w = out_shape[1]
5898 5899 5900 5901
    else:
        out_h = int(input.shape[2] * scale)
        out_w = int(input.shape[3] * scale)

5902 5903 5904 5905 5906
    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
5907
    out = helper.create_variable_for_type_inference(dtype)
5908
    helper.append_op(
5909
        type='interpolate',
5910
        inputs=inputs,
5911
        outputs={"Out": out},
5912 5913 5914 5915 5916
        attrs={
            "out_h": out_h,
            "out_w": out_w,
            "interp_method": resample_methods[resample]
        })
5917
    return out
F
stash  
fengjiayi 已提交
5918 5919


5920
@templatedoc(op_type="interpolate")
5921 5922 5923 5924 5925
def resize_bilinear(input,
                    out_shape=None,
                    scale=None,
                    name=None,
                    actual_shape=None):
5926
    """
5927 5928
    Resize input by performing bilinear interpolation based on given
    output shape which specified by actual_shape, out_shape and scale
5929 5930
    in priority order.

5931 5932 5933 5934
    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
5935 5936
    again in the other direction.

5937
    For details of bilinear interpolation, please refer to Wikipedia:
5938
    https://en.wikipedia.org/wiki/Bilinear_interpolation
Y
yuyang18 已提交
5939 5940 5941 5942 5943

    Args:
        input(${x_type}): ${x_comment}.

        out_shape(${out_size_type}): ${out_size_comment}.
5944

Y
yuyang18 已提交
5945 5946 5947 5948 5949
        scale(float|None): The multiplier for the input height or width. At
             least one of out_shape or scale must be set. And out_shape has
             a higher priority than scale. Default: None.

        name(str|None): The output variable name.
5950 5951 5952
        actual_shape(Variable): An optional input to specify output shape
                                dynamically. If provided, image resize
                                according to this given shape rather than
5953
                                :attr:`out_shape` and :attr:`scale` specifying
5954 5955 5956 5957 5958 5959 5960
                                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
5961 5962
                                constructing stage.
                                Default: None
Y
yuyang18 已提交
5963 5964 5965

    Returns:
        ${out_comment}.
5966 5967 5968 5969 5970

    Examples:
        .. code-block:: python

            out = fluid.layers.resize_bilinear(input, out_shape=[12, 12])
5971 5972
    """

5973
    return image_resize(input, out_shape, scale, name, 'BILINEAR', actual_shape)
5974 5975


5976
@templatedoc(op_type="interpolate")
5977 5978 5979 5980 5981
def resize_nearest(input,
                   out_shape=None,
                   scale=None,
                   name=None,
                   actual_shape=None):
5982
    """
5983
    Resize input by performing nearest neighbor interpolation in both the
5984 5985
    3rd dimention(in height direction) and the 4th dimention(in width
    direction) based on given output shape which specified by actual_shape,
5986 5987
    out_shape and scale in priority order.

5988
    For details of nearest neighbor interpolation, please refer to Wikipedia:
5989
    https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation
Y
yuyang18 已提交
5990 5991 5992 5993 5994

    Args:
        input(${x_type}): ${x_comment}.

        out_shape(${out_size_type}): ${out_size_comment}.
5995

Y
yuyang18 已提交
5996 5997 5998 5999 6000
        scale(float|None): The multiplier for the input height or width. At
             least one of out_shape or scale must be set. And out_shape has
             a higher priority than scale. Default: None.

        name(str|None): The output variable name.
6001 6002 6003
        actual_shape(Variable): An optional input to specify output shape
                                dynamically. If provided, image resize
                                according to this given shape rather than
6004
                                :attr:`out_shape` and :attr:`scale` specifying
6005 6006 6007 6008 6009 6010 6011
                                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
6012 6013
                                constructing stage.
                                Default: None
Y
yuyang18 已提交
6014 6015 6016

    Returns:
        ${out_comment}.
6017 6018 6019 6020 6021

    Examples:
        .. code-block:: python

            out = fluid.layers.resize_nearest(input, out_shape=[12, 12])
6022 6023
    """

6024
    return image_resize(input, out_shape, scale, name, 'NEAREST', actual_shape)
6025 6026 6027 6028


def image_resize_short(input, out_short_len, resample='BILINEAR'):
    """
6029 6030 6031
    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
6032 6033 6034 6035 6036 6037 6038
    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.
6039
        resample (str): resample method, default: BILINEAR.
F
fengjiayi 已提交
6040

6041
    Returns:
Q
update  
qiaolongfei 已提交
6042
        Variable: The output is a 4-D tensor of the shape
6043
        (num_batches, channls, out_h, out_w).
6044 6045 6046 6047 6048 6049 6050 6051 6052 6053
    """
    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
6054 6055 6056
    out_shape[long_idx] = int(
        float(out_shape[long_idx]) * (float(out_short_len) / float(hw[
            short_idx])) + 0.5)
6057 6058 6059
    return image_resize(input=input, out_shape=out_shape, resample=resample)


W
whs 已提交
6060 6061
def gather(input, index):
    """
Q
qiaolongfei 已提交
6062 6063
    **Gather Layer**

6064
    Output is obtained by gathering entries of the outer-most dimension
W
whs 已提交
6065 6066 6067 6068
    of X indexed by `index` and concatenate them together.

    .. math::

6069
        Out = X[Index]
W
whs 已提交
6070 6071 6072 6073 6074 6075 6076


    .. code-block:: text


                Given:

6077 6078
                X = [[1, 2],
                     [3, 4],
W
whs 已提交
6079 6080 6081 6082 6083 6084 6085 6086 6087 6088
                     [5, 6]]

                Index = [1, 2]

                Then:

                Out = [[3, 4],
                       [5, 6]]

    Args:
6089
        input (Variable): The source input with rank>=1.
W
whs 已提交
6090 6091 6092 6093 6094 6095
        index (Variable): The index input with rank=1.

    Returns:
        output (Variable): The output is a tensor with the same rank as input.

    Examples:
W
whs 已提交
6096

W
whs 已提交
6097 6098 6099 6100 6101 6102
        .. code-block:: python

            output = fluid.layers.gather(x, index)
    """
    helper = LayerHelper('gather', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
6103
    out = helper.create_variable_for_type_inference(dtype)
W
whs 已提交
6104 6105 6106 6107 6108 6109 6110 6111
    helper.append_op(
        type="gather",
        inputs={"X": input,
                "Index": index},
        outputs={"Out": out})
    return out


6112 6113 6114 6115 6116 6117 6118 6119 6120 6121 6122 6123 6124 6125 6126 6127 6128 6129 6130 6131 6132 6133 6134 6135 6136 6137 6138 6139 6140 6141 6142
def scatter(input, index, updates, name=None):
    """
    **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.

    Returns:
        output (Variable): The output is a tensor with the same shape as input.

    Examples:

        .. code-block:: python

            output = fluid.layers.scatter(input, index, updates)

    """
    helper = LayerHelper('scatter', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
6143
    out = helper.create_variable_for_type_inference(dtype)
6144 6145 6146 6147 6148 6149 6150 6151 6152
    helper.append_op(
        type="scatter",
        inputs={"X": input,
                "Ids": index,
                "Updates": updates},
        outputs={"Out": out})
    return out


Q
Qingsheng Li 已提交
6153 6154 6155 6156 6157 6158 6159 6160 6161 6162 6163 6164 6165 6166 6167 6168 6169 6170 6171 6172 6173 6174 6175 6176 6177 6178 6179 6180 6181 6182 6183 6184 6185 6186 6187 6188 6189 6190 6191 6192 6193 6194 6195 6196 6197 6198 6199 6200 6201 6202
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:
    Given the following input:
    .. code-block:: text
        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:
    .. code-block:: text
        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:
        output (Variable): The output is a tensor with the same shape as input.

    Examples:

        .. code-block:: python

            output = fluid.layers.sequence_scatter(input, index, updates)

    """
    helper = LayerHelper('sequence_scatter', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
6203
    out = helper.create_variable_for_type_inference(dtype)
Q
Qingsheng Li 已提交
6204 6205 6206 6207 6208 6209 6210 6211 6212
    helper.append_op(
        type="sequence_scatter",
        inputs={"X": input,
                "Ids": index,
                "Updates": updates},
        outputs={"Out": out})
    return out


Y
yuyang18 已提交
6213 6214 6215 6216 6217 6218 6219 6220 6221 6222 6223 6224 6225
@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}
6226

6227 6228 6229
    Examples:
        >>> img = fluid.layers.data("img", [3, 256, 256])
        >>> cropped_img = fluid.layers.random_crop(img, shape=[3, 224, 224])
Y
yuyang18 已提交
6230
    """
F
stash  
fengjiayi 已提交
6231
    helper = LayerHelper("random_crop", **locals())
F
fengjiayi 已提交
6232
    dtype = x.dtype
X
Xin Pan 已提交
6233
    out = helper.create_variable_for_type_inference(dtype)
Y
yuyang18 已提交
6234
    if seed is None:
6235
        seed = np.random.randint(-65536, 65536)
F
fengjiayi 已提交
6236
    op_attrs = {"shape": shape}
F
stash  
fengjiayi 已提交
6237
    if isinstance(seed, int):
F
fengjiayi 已提交
6238 6239 6240 6241 6242
        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
6243 6244 6245 6246
    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
6247
        inputs={"X": x,
F
stash  
fengjiayi 已提交
6248 6249
                "Seed": seed},
        outputs={"Out": out,
F
fengjiayi 已提交
6250 6251
                 "SeedOut": seed},
        attrs=op_attrs)
F
stash  
fengjiayi 已提交
6252
    return out
W
whs 已提交
6253 6254


6255
def log(x, name=None):
W
wanghaoshuang 已提交
6256 6257 6258 6259 6260
    """
    Calculates the natural log of the given input tensor, element-wise.

    .. math::

6261
        Out = \\ln(x)
W
wanghaoshuang 已提交
6262 6263

    Args:
6264
        x (Variable): Input tensor.
6265 6266
        name (str|None, default None): A name for this layer If set None,
            the layer will be named automatically.
W
wanghaoshuang 已提交
6267 6268 6269 6270 6271 6272 6273 6274

    Returns:
        Variable: The natural log of the input tensor computed element-wise.

    Examples:

        .. code-block:: python

6275
            output = fluid.layers.log(x)
W
wanghaoshuang 已提交
6276 6277
    """
    helper = LayerHelper('log', **locals())
W
wanghaoshuang 已提交
6278
    dtype = helper.input_dtype(input_param_name='x')
X
Xin Pan 已提交
6279
    out = helper.create_variable_for_type_inference(dtype)
W
wanghaoshuang 已提交
6280
    helper.append_op(type="log", inputs={"X": x}, outputs={"Out": out})
W
wanghaoshuang 已提交
6281 6282 6283
    return out


6284
def relu(x, name=None):
W
wanghaoshuang 已提交
6285 6286
    """
    Relu takes one input data (Tensor) and produces one output data (Tensor)
6287
    where the rectified linear function, y = max(0, x), is applied to
W
wanghaoshuang 已提交
6288 6289 6290 6291
    the tensor elementwise.

    .. math::

6292
        Out = \\max(0, x)
W
wanghaoshuang 已提交
6293 6294

    Args:
6295
        x (Variable): The input tensor.
6296 6297
        name (str|None, default None): A name for this layer If set None,
            the layer will be named automatically.
W
wanghaoshuang 已提交
6298 6299 6300 6301 6302 6303 6304 6305

    Returns:
        Variable: The output tensor with the same shape as input.

    Examples:

        .. code-block:: python

6306
            output = fluid.layers.relu(x)
W
wanghaoshuang 已提交
6307 6308
    """
    helper = LayerHelper('relu', **locals())
W
wanghaoshuang 已提交
6309
    dtype = helper.input_dtype(input_param_name='x')
X
Xin Pan 已提交
6310
    out = helper.create_variable_for_type_inference(dtype)
W
wanghaoshuang 已提交
6311
    helper.append_op(type="relu", inputs={"X": x}, outputs={"Out": out})
W
wanghaoshuang 已提交
6312
    return out
6313 6314


C
chengduo 已提交
6315 6316 6317 6318 6319 6320 6321 6322 6323 6324 6325 6326 6327 6328 6329 6330 6331 6332 6333 6334 6335 6336 6337 6338 6339 6340 6341 6342 6343 6344 6345 6346 6347 6348 6349 6350 6351 6352 6353 6354 6355
@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

            output = fluid.layers.selu(x)
    """
    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
6356 6357 6358
def mean_iou(input, label, num_classes):
    """
    Mean Intersection-Over-Union is a common evaluation metric for
6359 6360 6361 6362
    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
6363
    .. math::
6364 6365

        IOU = \\frac{true\_positiv}{(true\_positive + false\_positive + false\_negative)}.
W
whs 已提交
6366

6367
    The predictions are accumulated in a confusion matrix and mean-IOU
W
whs 已提交
6368 6369 6370 6371 6372
    is then calculated from it.


    Args:
        input (Variable): A Tensor of prediction results for semantic labels with type int32 or int64.
6373
        label (Variable): A Tensor of ground truth labels with type int32 or int64.
W
whs 已提交
6374
                           Its shape should be the same as input.
6375
        num_classes (int): The possible number of labels.
W
whs 已提交
6376 6377 6378 6379

    Returns:
        mean_iou (Variable): A Tensor representing the mean intersection-over-union with shape [1].
        out_wrong(Variable): A Tensor with shape [num_classes]. The wrong numbers of each class.
6380
        out_correct(Variable): A Tensor with shape [num_classes]. The correct numbers of each class.
W
whs 已提交
6381 6382 6383 6384

    Examples:

        .. code-block:: python
6385

W
whs 已提交
6386 6387 6388 6389
            iou, wrongs, corrects = fluid.layers.mean_iou(predict, label, num_classes)
    """
    helper = LayerHelper('mean_iou', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
6390 6391 6392
    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
6393 6394
    helper.append_op(
        type="mean_iou",
W
whs 已提交
6395 6396
        inputs={"Predictions": input,
                "Labels": label},
W
whs 已提交
6397
        outputs={
W
whs 已提交
6398 6399 6400
            "OutMeanIou": out_mean_iou,
            "OutWrong": out_wrong,
            "OutCorrect": out_correct
W
whs 已提交
6401 6402 6403
        },
        attrs={"num_classes": num_classes})
    return out_mean_iou, out_wrong, out_correct
6404 6405 6406 6407 6408 6409 6410 6411 6412 6413 6414 6415 6416 6417 6418 6419 6420 6421 6422 6423 6424 6425 6426 6427 6428 6429 6430 6431 6432 6433 6434 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 6462 6463 6464 6465 6466 6467 6468 6469 6470 6471 6472 6473 6474 6475 6476 6477


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`
        offsets (Variable|list/tuple of integer|None): Specifies the copping
            offsets at each dimension. It can be a Variable or 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 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

            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")
            crop = fluid.layers.crop(z, shape=[2, 3])

    """
    helper = LayerHelper('crop', **locals())

    if not (isinstance(shape, list) or isinstance(shape, tuple) or \
C
chengduo 已提交
6478
                    isinstance(shape, Variable)):
6479 6480 6481 6482 6483
        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
6484
    out = helper.create_variable_for_type_inference(x.dtype)
6485 6486 6487 6488 6489 6490 6491 6492 6493 6494 6495 6496 6497 6498 6499 6500 6501
    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
6502 6503


W
whs 已提交
6504 6505 6506 6507 6508 6509 6510 6511 6512 6513 6514 6515 6516 6517 6518 6519 6520
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]]]
6521

W
whs 已提交
6522
              out_shape = [2, 3, 5, 5]
6523

W
whs 已提交
6524
          Step 1:
6525

W
whs 已提交
6526 6527 6528
              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:
6529

W
whs 已提交
6530 6531 6532 6533 6534 6535 6536 6537 6538 6539 6540 6541 6542 6543 6544 6545 6546 6547 6548 6549 6550 6551 6552 6553 6554 6555 6556 6557 6558 6559 6560 6561 6562 6563 6564 6565 6566 6567 6568 6569 6570 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 6603 6604 6605 6606 6607 6608 6609 6610 6611 6612 6613 6614 6615 6616 6617 6618 6619 6620 6621
              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].
        out_shape (Variable | list | tuple): The shape of target output with format [N, C, H, W].
        out_shape can be a Variable or a list or tuple.
        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
            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 \
        isinstance(out_shape, Variable)):
        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


6622 6623 6624 6625 6626 6627 6628 6629
def rank_loss(label, left, right, name=None):
    """
    **Rank loss layer for RankNet**

    RankNet(http://icml.cc/2015/wp-content/uploads/2015/06/icml_ranking.pdf)
    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
6630

6631 6632
    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
6633

6634 6635 6636 6637
    Rank loss layer takes three inputs: left (o_i), right (o_j) and
    label (P_{i,j}). The inputs respectively represent RankNet's output scores
    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
6638

6639 6640 6641 6642 6643
    $$
      C_{i,j} = -\tilde{P_{ij}} * o_{i,j} + \log(1 + e^{o_{i,j}}) \\
      o_{i,j} =  o_i - o_j  \\
      \tilde{P_{i,j}} = \left \{0, 0.5, 1 \right \} \ or \ \left \{0, 1 \right \}
    $$
M
minqiyang 已提交
6644 6645 6646

    Rank loss layer takes batch inputs with size batch_size (batch_size >= 1).

6647 6648 6649 6650 6651 6652 6653 6654 6655 6656 6657 6658 6659 6660 6661 6662 6663 6664 6665 6666 6667 6668 6669 6670 6671 6672 6673 6674 6675 6676 6677 6678 6679 6680 6681
    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

            label = fluid.layers.data(name="label", shape=[4, 1], dtype="float32")
            left = fluid.layers.data(name="left", shape=[4, 1], dtype="float32")
            right = fluid.layers.data(name="right", shape=[4, 1], dtype="float32")
            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
6682
    out = helper.create_variable_for_type_inference("float32")
6683 6684 6685 6686 6687 6688 6689 6690

    helper.append_op(
        type='rank_loss',
        inputs={"Label": label,
                "Left": left,
                "Right": right},
        outputs={'Out': out})
    return out
6691 6692


M
minqiyang 已提交
6693 6694
def margin_rank_loss(label, left, right, margin=0.1, name=None):
    """
M
minqiyang 已提交
6695
    Margin Ranking Loss Layer for ranking problem,
M
minqiyang 已提交
6696
    which compares left score and right score passed in.
M
minqiyang 已提交
6697
    The ranking loss can be defined as following equation:
M
minqiyang 已提交
6698 6699 6700 6701 6702 6703

    .. math::

        rank\_loss &= max(0, -label * (left - right) + margin)

    Args:
M
minqiyang 已提交
6704
       label (Variable): Indicates whether the left is ranked higher than the right or not.
M
minqiyang 已提交
6705 6706
       left (Variable): Ranking score for left.
       right (Variable): Ranking score for right.
M
minqiyang 已提交
6707
       margin (float): Indicates the given margin.
M
minqiyang 已提交
6708 6709 6710
       name (str|None): A name for this layer (optional). If set None, the layer
                       will be named automatically.
    Returns:
M
minqiyang 已提交
6711
       Variable: The ranking loss.
M
minqiyang 已提交
6712
    Raises:
M
minqiyang 已提交
6713
       ValueError: Any of label, left, and right is not a Variable.
M
minqiyang 已提交
6714 6715 6716 6717 6718 6719 6720
    Examples:
        .. code-block:: python
           label = fluid.layers.data(name="label", shape=[4, 1], dtype="float32")
           left = fluid.layers.data(name="left", shape=[4, 1], dtype="float32")
           right = fluid.layers.data(name="right", shape=[4, 1], dtype="float32")
           out = fluid.layers.margin_rank_loss(label, left, right)
    """
M
minqiyang 已提交
6721
    helper = LayerHelper('margin_rank_loss', **locals())
M
minqiyang 已提交
6722 6723 6724 6725 6726 6727
    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
6728 6729
    out = helper.create_variable_for_type_inference(left.dtype)
    act = helper.create_variable_for_type_inference(left.dtype)
M
minqiyang 已提交
6730 6731 6732 6733 6734 6735 6736 6737 6738 6739 6740
    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
6741 6742 6743 6744 6745 6746 6747 6748 6749 6750 6751 6752 6753 6754
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:

      Given that X is a channel of image from input:
M
minqiyang 已提交
6755

W
whs 已提交
6756 6757
      X = [[1, 2, 3],
           [4, 5, 6]]
M
minqiyang 已提交
6758

W
whs 已提交
6759
      Case 0:
M
minqiyang 已提交
6760

W
whs 已提交
6761 6762 6763
        paddings = [0, 1, 2, 3],
        mode = 'constant'
        pad_value = 0
M
minqiyang 已提交
6764

W
whs 已提交
6765 6766 6767
        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
6768

W
whs 已提交
6769
      Case 1:
M
minqiyang 已提交
6770

W
whs 已提交
6771 6772
        paddings = [0, 1, 2, 1],
        mode = 'reflect'
M
minqiyang 已提交
6773

W
whs 已提交
6774 6775 6776
        Out = [[3, 2, 1, 2, 3, 2]
               [6, 5, 4, 5, 6, 5]
               [3, 2, 1, 2, 3, 2]]
M
minqiyang 已提交
6777

W
whs 已提交
6778
      Case 2:
M
minqiyang 已提交
6779

W
whs 已提交
6780 6781
        paddings = [0, 1, 2, 1],
        mode = 'edge'
M
minqiyang 已提交
6782

W
whs 已提交
6783 6784 6785
        Out = [[1, 1, 1, 2, 3, 3]
               [4, 4, 4, 5, 6, 6]
               [4, 4, 4, 5, 6, 6]]
M
minqiyang 已提交
6786 6787


W
whs 已提交
6788 6789 6790 6791 6792 6793 6794 6795 6796 6797 6798 6799 6800 6801 6802 6803 6804 6805 6806 6807 6808 6809 6810 6811 6812 6813
    Args:
        input (Variable): The input image with [N, C, H, W] format or [N, H, W, C] format.
        paddings (tuple|list): The padding size. If padding is a tuple, it must
            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

          data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32')
          result = fluid.layers.pad2d(input=data, padding=[1,2,3,4], mode='reflect')
    """

    helper = LayerHelper('pad2d', **locals())
    dtype = helper.input_dtype(input_param_name='input')
X
Xin Pan 已提交
6814
    out = helper.create_variable_for_type_inference(dtype)
W
whs 已提交
6815 6816 6817 6818 6819 6820 6821 6822 6823 6824 6825 6826 6827 6828
    helper.append_op(
        type='pad2d',
        inputs={'X': input},
        outputs={"Out": out},
        attrs={
            'paddings': paddings,
            'mode': mode,
            'pad_value': pad_value,
            'data_frmat': data_format
        })

    return out


6829 6830 6831 6832 6833 6834 6835 6836 6837 6838 6839 6840
@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
6841 6842 6843 6844 6845

    Examples:

        .. code-block:: python

Z
ZhenWang 已提交
6846 6847
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
            y = fluid.layers.elu(x, alpha=0.2)
6848 6849
    """
    helper = LayerHelper('elu', **locals())
X
Xin Pan 已提交
6850
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
6851 6852 6853 6854 6855 6856 6857 6858 6859 6860 6861 6862 6863 6864 6865 6866 6867 6868 6869 6870
    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
6871 6872 6873 6874 6875

    Examples:

        .. code-block:: python

Z
ZhenWang 已提交
6876 6877
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
            y = fluid.layers.relu6(x, threshold=6.0)
6878 6879
    """
    helper = LayerHelper('relu6', **locals())
X
Xin Pan 已提交
6880
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
6881 6882 6883 6884 6885 6886 6887 6888 6889 6890 6891 6892 6893 6894 6895 6896 6897 6898 6899 6900
    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
6901 6902 6903 6904 6905

    Examples:

        .. code-block:: python

Z
ZhenWang 已提交
6906 6907
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
            y = fluid.layers.pow(x, factor=2.0)
6908 6909
    """
    helper = LayerHelper('pow', **locals())
X
Xin Pan 已提交
6910
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
6911 6912 6913 6914 6915 6916 6917 6918 6919 6920 6921 6922 6923 6924 6925 6926 6927 6928 6929 6930 6931
    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
6932 6933 6934 6935 6936

    Examples:

        .. code-block:: python

Z
ZhenWang 已提交
6937 6938
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
            y = fluid.layers.stanh(x, scale_a=0.6667, scale_b=1.7159)
6939 6940
    """
    helper = LayerHelper('stanh', **locals())
X
Xin Pan 已提交
6941
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
6942 6943 6944 6945 6946 6947 6948 6949 6950 6951 6952 6953 6954 6955 6956 6957 6958 6959 6960 6961 6962 6963
    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
6964 6965 6966 6967 6968

    Examples:

        .. code-block:: python

Z
ZhenWang 已提交
6969 6970
            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)
6971 6972
    """
    helper = LayerHelper('hard_sigmoid', **locals())
X
Xin Pan 已提交
6973
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
6974 6975 6976 6977 6978 6979 6980 6981 6982 6983 6984 6985 6986 6987 6988 6989 6990 6991 6992 6993 6994
    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
6995 6996 6997 6998 6999

    Examples:

        .. code-block:: python

Z
ZhenWang 已提交
7000 7001
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
            y = fluid.layers.swish(x, beta=2.0)
7002 7003
    """
    helper = LayerHelper('swish', **locals())
X
Xin Pan 已提交
7004
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
7005 7006 7007 7008 7009 7010 7011 7012
    helper.append_op(
        type='swish',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'slope': beta})
    return out


J
jerrywgz 已提交
7013 7014 7015 7016
def prelu(x, mode, param_attr=None, name=None):
    """
    Equation:

J
jerrywgz 已提交
7017
        y = \max(0, x) + alpha * \min(0, x)
J
jerrywgz 已提交
7018 7019 7020

    Args:
        x (Variable): The input tensor.
J
jerrywgz 已提交
7021 7022
        param_attr(ParamAttr|None): The parameter attribute for the learnable
                       weight (alpha).
J
jerrywgz 已提交
7023 7024 7025 7026
        mode (string): The mode for weight sharing. It supports all, channel
                       and element. all: all elements share same weight
                       channel:elements in a channel share same weight
                       element:each element has a weight
J
jerrywgz 已提交
7027
        name(str|None): A name for this layer(optional). If set None, the layer
J
jerrywgz 已提交
7028
                       will be named automatically.
J
jerrywgz 已提交
7029 7030 7031 7032 7033 7034 7035 7036

    Returns:
        Variable: The output tensor with the same shape as input.

    Examples:

        .. code-block:: python

Z
ZhenWang 已提交
7037 7038 7039
            x = fluid.layers.data(name="x", shape=[10,10], dtype="float32")
            mode = 'channel'
            output = fluid.layers.prelu(x,mode)
J
jerrywgz 已提交
7040 7041 7042 7043 7044 7045 7046 7047 7048 7049 7050
    """
    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
7051
        attr=helper.param_attr,
J
jerrywgz 已提交
7052 7053 7054 7055
        shape=alpha_shape,
        dtype='float32',
        is_bias=False,
        default_initializer=Constant(1.0))
X
Xin Pan 已提交
7056
    out = helper.create_variable_for_type_inference(dtype)
J
jerrywgz 已提交
7057 7058 7059 7060 7061 7062 7063 7064 7065
    helper.append_op(
        type="prelu",
        inputs={"X": x,
                'Alpha': alpha},
        attrs={"mode": mode},
        outputs={"Out": out})
    return out


7066 7067 7068 7069 7070 7071 7072 7073 7074 7075
@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): ${t_max_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
7076
    Returns:
7077
        output(${out_type}): ${out_comment}
7078 7079 7080 7081 7082 7083 7084

    Examples:

        .. code-block:: python

        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)
7085 7086
    """
    helper = LayerHelper('brelu', **locals())
X
Xin Pan 已提交
7087
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
7088 7089 7090 7091 7092 7093 7094 7095 7096 7097 7098 7099 7100 7101 7102 7103 7104 7105
    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): ${alpha_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
7106
    Returns:
7107
        output(${out_type}): ${out_comment}
7108 7109 7110 7111 7112 7113 7114

    Examples:

        .. code-block:: python

        x = fluid.layers.data(name="x", shape=[2,3,16,16], dtype="float32")
        y = fluid.layers.leaky_relu(x, alpha=0.01)
7115 7116
    """
    helper = LayerHelper('leaky_relu', **locals())
X
Xin Pan 已提交
7117
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
7118 7119 7120 7121 7122 7123 7124 7125 7126 7127 7128 7129 7130 7131 7132 7133 7134
    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): ${threshold_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
7135
    Returns:
7136
        output(${out_type}): ${out_comment}
7137 7138 7139 7140 7141 7142 7143

    Examples:

        .. code-block:: python

        x = fluid.layers.data(name="x", shape=[2,3,16,16], dtype="float32")
        y = fluid.layers.soft_relu(x, threshold=20.0)
7144 7145
    """
    helper = LayerHelper('soft_relu', **locals())
X
Xin Pan 已提交
7146
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
7147 7148 7149 7150 7151 7152 7153 7154
    helper.append_op(
        type='soft_relu',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'threshold': threshold})
    return out


7155 7156 7157 7158 7159 7160 7161 7162 7163 7164 7165 7166 7167
def flatten(x, axis=1, name=None):
    """
    **Flatten layer**
    Flattens the input tensor into a 2D matrix.

    Examples:
    Case 1:
      Given
        X.shape = (3, 100, 100, 4)
      and
        axis = 2
      We get:
        Out.shape = (3 * 100, 4 * 100)
7168

7169 7170 7171 7172 7173 7174 7175 7176 7177 7178
    Case 2:
      Given
        X.shape = (3, 100, 100, 4)
      and
        axis = 0
      We get:
        Out.shape = (1, 3 * 100 * 100 * 4)

    Args:
        x (Variable): A tensor of rank >= axis.
7179 7180
        axis (int): Indicate up to which input dimensions (exclusive) should
                    be flattened to the outer dimension of the output.
7181 7182 7183 7184 7185 7186 7187 7188 7189 7190 7191 7192 7193 7194 7195
                    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:
        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
                  inner dimension of the output.

    Raises:
        ValueError: If x is not a variable.
7196
        ValueError: If axis is not in range [0, rank(x)].
7197 7198 7199 7200 7201 7202 7203 7204 7205 7206 7207 7208 7209 7210 7211 7212

    Examples:

        .. code-block:: python

            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
7213 7214
    out = helper.create_variable_for_type_inference(x.dtype)
    x_shape = helper.create_variable_for_type_inference(x.dtype)
7215
    helper.append_op(
7216
        type='flatten2',
7217
        inputs={"X": x},
7218 7219
        outputs={'Out': out,
                 'XShape': x_shape},
7220 7221
        attrs={"axis": axis})
    return out
X
Xin Pan 已提交
7222 7223


C
chenweihang 已提交
7224
def sequence_enumerate(input, win_size, pad_value=0, name=None):
C
chenweihang 已提交
7225
    """
C
chenweihang 已提交
7226
    Generate a new sequence for the input index sequence, which enumerates all the
M
minqiyang 已提交
7227
    sub-sequences with length `win_size` of the input.
C
chenweihang 已提交
7228 7229
    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
7230

C
chenweihang 已提交
7231 7232 7233 7234
    Examples:
    Case 1:
      Input:
        X.lod = [[0, 3, 5]]
7235
        X.data = [[1], [2], [3], [4], [5]]
C
chenweihang 已提交
7236 7237 7238 7239 7240 7241
        X.dims = [5, 1]
      Attrs:
        win_size = 2
        pad_value = 0
      Output:
        Out.lod = [[0, 3, 5]]
7242
        Out.data = [[1, 2], [2, 3], [3, 0], [4, 5], [5, 0]]
C
chenweihang 已提交
7243 7244 7245
        Out.dims = [5, 2]

    Args:
C
chenweihang 已提交
7246 7247 7248
        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
7249 7250 7251 7252 7253 7254 7255 7256 7257 7258 7259

    Returns:
        Variable: The enumerate sequence variable which is a LoDTensor.

    Examples:
        .. code-block:: python

            x = fluid.layers.data(shape[30, 1], dtype='int32', lod_level=1)
            out = fluid.layers.sequence_enumerate(input=x, win_size=3, pad_value=0)
    """
    helper = LayerHelper('sequence_enumerate', **locals())
X
Xin Pan 已提交
7260 7261
    out = helper.create_variable_for_type_inference(
        helper.input_dtype(), stop_gradient=True)
C
chenweihang 已提交
7262 7263 7264 7265 7266 7267
    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
7268
    return out
7269

7270

S
sneaxiy 已提交
7271 7272 7273 7274 7275 7276 7277 7278 7279
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:
7280

S
sneaxiy 已提交
7281
    .. math::
7282

S
sneaxiy 已提交
7283 7284 7285
        y(i_1, i_2,..., i_n, j) = (j < x(i_1, i_2,..., i_n))

    Args:
7286
        x (Variable): Input tensor of sequence_mask layer,
S
sneaxiy 已提交
7287 7288 7289 7290
                      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.
7291 7292 7293
        name (str|None): A name for this layer(optional). If set None, the
                         layer will be named automatically.

S
sneaxiy 已提交
7294 7295
    Returns:
        Variable: The output sequence mask.
7296

S
sneaxiy 已提交
7297 7298
    """

Q
qingqing01 已提交
7299
    helper = LayerHelper('sequence_mask', **locals())
S
sneaxiy 已提交
7300
    if name is None:
X
Xin Pan 已提交
7301
        out = helper.create_variable_for_type_inference(dtype=dtype)
S
sneaxiy 已提交
7302
    else:
X
Xin Pan 已提交
7303
        out = helper.create_variable_for_type_inference(dtype=dtype, name=name)
S
sneaxiy 已提交
7304

Q
qingqing01 已提交
7305 7306 7307
    helper.append_op(
        type='sequence_mask',
        inputs={'X': [x]},
S
sneaxiy 已提交
7308 7309
        outputs={'Y': out},
        attrs={
7310
            'maxlen': maxlen if maxlen is not None else -1,
S
sneaxiy 已提交
7311 7312 7313
            'out_dtype': out.dtype
        })
    return out
S
sneaxiy 已提交
7314 7315


X
Xin Pan 已提交
7316
def stack(x, axis=0):
S
sneaxiy 已提交
7317 7318 7319 7320
    """
    **Stack Layer**

    This layer stacks all of the input :code:`x` along axis.
7321 7322 7323 7324 7325 7326 7327

    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
7328
    If :code:`axis` < 0, it would be replaced with :code:`axis+rank(x[0])+1`.
7329
    If :code:`axis` is None, it would be replaced with 0.
S
sneaxiy 已提交
7330 7331

    Args:
7332
        x (Variable|list(Variable)|tuple(Variable)): Input variables.
S
sneaxiy 已提交
7333
        axis (int|None): The axis along which all inputs are stacked.
7334

S
sneaxiy 已提交
7335 7336
    Returns:
        Variable: The stacked variable.
7337

S
sneaxiy 已提交
7338 7339
    """

X
Xin Pan 已提交
7340 7341 7342 7343 7344 7345
    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
7346
    out = helper.create_variable_for_type_inference(x[0].dtype)
X
Xin Pan 已提交
7347
    helper.append_op(
S
sneaxiy 已提交
7348 7349
        type='stack', inputs={'X': x}, outputs={'Y': out},
        attrs={'axis': axis})
7350

X
Xin Pan 已提交
7351
    return out
D
dzhwinter 已提交
7352 7353 7354 7355 7356 7357 7358


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
7359

D
dzhwinter 已提交
7360 7361 7362
    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
7363
    raised.
D
dzhwinter 已提交
7364 7365

    Args:
M
minqiyang 已提交
7366
        x (Variable): Input variable.
D
dzhwinter 已提交
7367 7368
        axis (int): The axis along which the input is unstacked.
        num (int|None): The number of output variables.
M
minqiyang 已提交
7369

D
dzhwinter 已提交
7370 7371
    Returns:
        list(Variable): The unstacked variables.
M
minqiyang 已提交
7372

D
dzhwinter 已提交
7373 7374 7375 7376 7377 7378 7379 7380 7381 7382 7383
    """

    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 = []
    for _ in num:
X
Xin Pan 已提交
7384
        outs.append(helper.create_variable_for_type_inference(x.dtype))
D
dzhwinter 已提交
7385 7386 7387 7388 7389 7390 7391 7392

    helper.append_op(
        type='unstack',
        inputs={'X': [x]},
        outputs={'Y': outs},
        attrs={'axis': axis,
               'num': num})
    return outs
W
whs 已提交
7393 7394 7395 7396 7397 7398 7399 7400 7401 7402 7403 7404


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
7405

W
whs 已提交
7406 7407 7408 7409
                [
                   [[1], [2], [3]],
                   [[4], [5], [6]]
                ]
M
minqiyang 已提交
7410

W
whs 已提交
7411
        Attr(expand_times):  [1, 2, 2]
M
minqiyang 已提交
7412

W
whs 已提交
7413
        Output(Out) is a 3-D tensor with shape [2, 6, 2]:
M
minqiyang 已提交
7414

W
whs 已提交
7415 7416 7417 7418
                [
                    [[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
7419

W
whs 已提交
7420 7421 7422 7423 7424 7425 7426 7427 7428 7429 7430 7431 7432 7433 7434 7435
    Args:
        x (Variable): A tensor with rank in [1, 6].
        expand_times (list|tuple): Expand times number for each dimension.

    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

            x = fluid.layers.data(name='x', shape=[10], dtype='float32')
            out = fluid.layers.expand(x=x, expand_times=[1, 2, 2])
    """
    helper = LayerHelper('expand', input=x, **locals())
    dtype = helper.input_dtype(input_param_name='x')
X
Xin Pan 已提交
7436
    out = helper.create_variable_for_type_inference(dtype)
W
whs 已提交
7437 7438 7439 7440 7441 7442
    helper.append_op(
        type='expand',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'expand_times': expand_times})
    return out
S
sneaxiy 已提交
7443 7444


G
fix  
gongweibao 已提交
7445 7446 7447
from paddle.fluid.framework import convert_np_dtype_to_dtype_


G
gongweibao 已提交
7448
@templatedoc()
G
fix  
gongweibao 已提交
7449 7450 7451 7452 7453 7454 7455 7456 7457
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
7458
    ${comment}
G
fix  
gongweibao 已提交
7459 7460

    Args:
G
gongweibao 已提交
7461 7462 7463
        input (Variable): ${input_comment}
        shape (tuple|list): ${shape_comment}
        input_dim_idx (Int): ${input_dim_idx_comment}
G
gongweibao 已提交
7464
        output_dim_idx (Int): ${output_dim_idx_comment}
G
gongweibao 已提交
7465 7466 7467
        min (Float): ${min_comment}
        max (Float): ${max_comment}
        seed (Int): ${seed_comment}
G
fix  
gongweibao 已提交
7468 7469
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc
    Returns:
G
gongweibao 已提交
7470
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
7471 7472 7473 7474

    """

    helper = LayerHelper('uniform_random_batch_size_like', **locals())
X
Xin Pan 已提交
7475
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
7476 7477 7478 7479 7480 7481 7482 7483 7484 7485 7486 7487 7488 7489 7490 7491
    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
7492 7493


G
gongweibao 已提交
7494
@templatedoc()
X
Xin Pan 已提交
7495
def gaussian_random(shape, mean=0.0, std=1.0, seed=0, dtype='float32'):
G
fix  
gongweibao 已提交
7496
    """
G
gongweibao 已提交
7497
    ${comment}
G
fix  
gongweibao 已提交
7498 7499

    Args:
G
gongweibao 已提交
7500 7501 7502 7503
        shape (tuple|list): ${shape_comment}
        mean (Float): ${mean_comment}
        std (Float): ${std_comment}
        seed (Int): ${seed_comment}
G
fix  
gongweibao 已提交
7504 7505 7506
        dtype(np.dtype|core.VarDesc.VarType|str): Output data type.

    Returns:
G
gongweibao 已提交
7507
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
7508 7509 7510 7511

    """

    helper = LayerHelper('gaussian_random', **locals())
X
Xin Pan 已提交
7512
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
7513 7514 7515 7516 7517 7518 7519 7520 7521 7522
    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
7523
            'use_mkldnn': False
G
fix  
gongweibao 已提交
7524 7525 7526 7527 7528
        })

    return out


G
gongweibao 已提交
7529
@templatedoc()
G
fix  
gongweibao 已提交
7530
def sampling_id(x, min=0.0, max=1.0, seed=0, dtype='float32'):
G
fix  
gongweibao 已提交
7531
    """
G
gongweibao 已提交
7532
    ${comment}
G
fix  
gongweibao 已提交
7533 7534

    Args:
G
gongweibao 已提交
7535 7536 7537 7538
        x (Variable): ${x_comment}
        min (Float): ${min_comment}
        max (Float): ${max_comment}
        seed (Float): ${seed_comment}
G
fix  
gongweibao 已提交
7539
        dtype(np.dtype|core.VarDesc.VarType|str): The type of output data : float32, float_16, int etc
G
fix  
gongweibao 已提交
7540 7541

    Returns:
G
gongweibao 已提交
7542
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
7543 7544 7545 7546

    """

    helper = LayerHelper('sampling_id', **locals())
X
Xin Pan 已提交
7547
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
7548 7549 7550 7551 7552 7553 7554 7555 7556 7557 7558
    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
7559
@templatedoc()
G
fix  
gongweibao 已提交
7560 7561 7562 7563 7564 7565 7566 7567 7568
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
7569
    ${comment}
G
fix  
gongweibao 已提交
7570 7571

    Args:
G
gongweibao 已提交
7572 7573
        input (Variable): ${input_comment}
        shape (tuple|list): ${shape_comment}
G
gongweibao 已提交
7574
        input_dim_idx (Int): ${input_dim_idx_comment}
G
gongweibao 已提交
7575 7576 7577 7578
        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
7579
        dtype(np.dtype|core.VarDesc.VarType|str): The type of output data : float32, float_16, int etc
G
fix  
gongweibao 已提交
7580 7581

    Returns:
G
gongweibao 已提交
7582
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
7583 7584 7585
    """

    helper = LayerHelper('gaussian_random_batch_size_like', **locals())
X
Xin Pan 已提交
7586
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
7587 7588 7589 7590 7591 7592 7593 7594 7595 7596 7597 7598 7599 7600 7601 7602 7603 7604
    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
7605
@templatedoc()
X
Xin Pan 已提交
7606
def sum(x):
G
fix  
gongweibao 已提交
7607
    """
G
gongweibao 已提交
7608
    ${comment}
G
fix  
gongweibao 已提交
7609 7610

    Args:
G
gongweibao 已提交
7611
        x (Variable): ${x_comment}
G
fix  
gongweibao 已提交
7612 7613

    Returns:
G
gongweibao 已提交
7614
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
7615 7616 7617
    """

    helper = LayerHelper('sum', **locals())
X
Xin Pan 已提交
7618 7619
    out = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype('x'))
G
fix  
gongweibao 已提交
7620 7621 7622 7623
    helper.append_op(
        type='sum',
        inputs={'X': x},
        outputs={'Out': out},
X
Xin Pan 已提交
7624
        attrs={'use_mkldnn': False})
G
fix  
gongweibao 已提交
7625 7626 7627 7628

    return out


G
gongweibao 已提交
7629
@templatedoc()
G
fix  
gongweibao 已提交
7630 7631
def slice(input, axes, starts, ends):
    """
G
gongweibao 已提交
7632
    ${comment}
G
fix  
gongweibao 已提交
7633 7634

    Args:
G
gongweibao 已提交
7635 7636 7637 7638
        input (Variable): ${input_comment}.
        axes (List): ${axes_comment}
        starts (List): ${starts_comment}
        ends (List): ${ends_comment}
G
fix  
gongweibao 已提交
7639 7640

    Returns:
G
gongweibao 已提交
7641
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
7642 7643 7644 7645

    """

    helper = LayerHelper('slice', **locals())
X
Xin Pan 已提交
7646 7647
    out = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype('input'))
G
fix  
gongweibao 已提交
7648 7649 7650 7651 7652 7653 7654 7655 7656 7657 7658
    helper.append_op(
        type='slice',
        inputs={'Input': input},
        outputs={'Out': out},
        attrs={'axes': axes,
               'starts': starts,
               'ends': ends})

    return out


G
gongweibao 已提交
7659
@templatedoc()
G
fix  
gongweibao 已提交
7660 7661
def shape(input):
    """
G
gongweibao 已提交
7662
    ${comment}
G
fix  
gongweibao 已提交
7663 7664

    Args:
G
gongweibao 已提交
7665
        input (Variable): ${input_comment}
G
fix  
gongweibao 已提交
7666 7667

    Returns:
G
gongweibao 已提交
7668
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
7669 7670 7671 7672

    """

    helper = LayerHelper('shape', **locals())
X
Xin Pan 已提交
7673 7674
    out = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype('input'))
G
fix  
gongweibao 已提交
7675
    helper.append_op(
G
fix  
gongweibao 已提交
7676
        type='shape', inputs={'Input': input}, outputs={'Out': out})
G
fix  
gongweibao 已提交
7677 7678

    return out
G
merge  
gongweibao 已提交
7679 7680


S
sneaxiy 已提交
7681 7682 7683 7684 7685 7686 7687 7688
def _elementwise_op(helper):
    op_type = helper.layer_type
    x = helper.kwargs.get('x', None)
    y = helper.kwargs.get('y', None)
    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
7689 7690
    name = helper.kwargs.get('name', None)
    if name is None:
X
Xin Pan 已提交
7691
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
S
sneaxiy 已提交
7692 7693 7694
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)
S
sneaxiy 已提交
7695

S
sneaxiy 已提交
7696 7697 7698 7699 7700 7701 7702 7703 7704 7705 7706
    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
7707
def scale(x, scale=1.0, bias=0.0, bias_after_scale=True, act=None, name=None):
S
sneaxiy 已提交
7708 7709 7710 7711 7712 7713 7714 7715
    """
    ${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
7716
        act(basestring|None): Activation applied to the output.
M
minqiyang 已提交
7717
        name(basestring|None): Name of the output.
S
sneaxiy 已提交
7718 7719 7720 7721 7722 7723

    Returns:
        out(${out_type}): ${out_comment}
    """

    helper = LayerHelper('scale', **locals())
S
sneaxiy 已提交
7724
    if name is None:
X
Xin Pan 已提交
7725
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
S
sneaxiy 已提交
7726 7727 7728
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)
S
sneaxiy 已提交
7729 7730 7731 7732 7733 7734 7735 7736 7737 7738

    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
7739
    return helper.append_activation(out)
S
sneaxiy 已提交
7740 7741


X
Xin Pan 已提交
7742
def elementwise_add(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
7743 7744 7745
    return _elementwise_op(LayerHelper('elementwise_add', **locals()))


X
Xin Pan 已提交
7746
def elementwise_div(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
7747 7748 7749
    return _elementwise_op(LayerHelper('elementwise_div', **locals()))


X
Xin Pan 已提交
7750
def elementwise_sub(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
7751 7752 7753
    return _elementwise_op(LayerHelper('elementwise_sub', **locals()))


X
Xin Pan 已提交
7754
def elementwise_mul(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
7755 7756 7757
    return _elementwise_op(LayerHelper('elementwise_mul', **locals()))


X
Xin Pan 已提交
7758
def elementwise_max(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
7759 7760 7761
    return _elementwise_op(LayerHelper('elementwise_max', **locals()))


X
Xin Pan 已提交
7762
def elementwise_min(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
7763 7764 7765
    return _elementwise_op(LayerHelper('elementwise_min', **locals()))


X
Xin Pan 已提交
7766
def elementwise_pow(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
7767 7768 7769 7770 7771 7772 7773 7774 7775 7776 7777
    return _elementwise_op(LayerHelper('elementwise_pow', **locals()))


for func in [
        elementwise_add, elementwise_div, elementwise_sub, elementwise_mul,
        elementwise_max, elementwise_min, elementwise_pow
]:
    op_proto = OpProtoHolder.instance().get_op_proto(func.__name__)
    func.__doc__ = _generate_doc_string_(
        op_proto,
        additional_args_lines=[
S
sneaxiy 已提交
7778 7779
            "act (basestring|None): Activation applied to the output.",
            "name (basestring|None): Name of the output."
S
sneaxiy 已提交
7780
        ])
M
minqiyang 已提交
7781 7782


7783
def _logical_op(op_name, x, y, out=None, name=None, binary_op=True):
M
minqiyang 已提交
7784 7785
    helper = LayerHelper(op_name, **locals())

M
minqiyang 已提交
7786 7787
    if binary_op:
        assert x.dtype == y.dtype
M
minqiyang 已提交
7788 7789 7790

    if out is None:
        if name is None:
X
Xin Pan 已提交
7791
            out = helper.create_variable_for_type_inference(dtype=x.dtype)
M
minqiyang 已提交
7792 7793 7794 7795 7796 7797 7798 7799 7800 7801 7802 7803 7804 7805 7806
        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()
7807
def logical_and(x, y, out=None, name=None):
M
minqiyang 已提交
7808 7809 7810 7811 7812 7813 7814 7815 7816 7817 7818
    """
    ${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}
7819 7820 7821 7822 7823 7824 7825 7826 7827

    Examples:
        .. code-block:: python

            left = fluid.layers.data(
                name='left', shape=[1], dtype='int32')
            right = fluid.layers.data(
                name='right', shape=[1], dtype='int32')
            result = fluid.layers.logical_and(x=left, y=right)
M
minqiyang 已提交
7828 7829 7830 7831 7832 7833 7834
    """

    return _logical_op(
        op_name="logical_and", x=x, y=y, name=name, out=out, binary_op=True)


@templatedoc()
7835
def logical_or(x, y, out=None, name=None):
M
minqiyang 已提交
7836 7837 7838 7839 7840 7841 7842 7843 7844 7845 7846
    """
    ${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}
7847 7848 7849 7850 7851 7852 7853 7854 7855

    Examples:
        .. code-block:: python

            left = fluid.layers.data(
                name='left', shape=[1], dtype='int32')
            right = fluid.layers.data(
                name='right', shape=[1], dtype='int32')
            result = fluid.layers.logical_or(x=left, y=right)
M
minqiyang 已提交
7856 7857 7858 7859 7860 7861 7862
    """

    return _logical_op(
        op_name="logical_or", x=x, y=y, name=name, out=out, binary_op=True)


@templatedoc()
7863
def logical_xor(x, y, out=None, name=None):
M
minqiyang 已提交
7864 7865 7866 7867 7868 7869 7870 7871 7872 7873 7874
    """
    ${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}
7875 7876 7877 7878 7879 7880 7881 7882 7883

    Examples:
        .. code-block:: python

            left = fluid.layers.data(
                name='left', shape=[1], dtype='int32')
            right = fluid.layers.data(
                name='right', shape=[1], dtype='int32')
            result = fluid.layers.logical_xor(x=left, y=right)
M
minqiyang 已提交
7884 7885 7886 7887 7888 7889 7890
    """

    return _logical_op(
        op_name="logical_xor", x=x, y=y, name=name, out=out, binary_op=True)


@templatedoc()
7891
def logical_not(x, out=None, name=None):
M
minqiyang 已提交
7892 7893 7894 7895 7896 7897 7898 7899 7900 7901
    """
    ${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}
7902 7903 7904 7905 7906 7907 7908

    Examples:
        .. code-block:: python

            left = fluid.layers.data(
                name='left', shape=[1], dtype='int32')
            result = fluid.layers.logical_not(x=left)
M
minqiyang 已提交
7909 7910 7911 7912
    """

    return _logical_op(
        op_name="logical_not", x=x, y=None, name=name, out=out, binary_op=False)
7913 7914 7915 7916 7917 7918 7919 7920 7921 7922 7923 7924 7925 7926 7927


@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}
7928 7929 7930 7931 7932 7933 7934

    Examples:
        .. code-block:: python

            input = fluid.layers.data(
                name='data', shape=[1], dtype='float32')
            reward = fluid.layers.clip(x=input, min=-1.0, max=1.0)
7935 7936 7937 7938 7939
    """

    helper = LayerHelper("clip", **locals())

    if name is None:
S
sneaxiy 已提交
7940 7941 7942 7943
        name = unique_name.generate(".".join([helper.name, 'tmp']))

    out = helper.create_variable(
        type=x.type, name=name, dtype=x.dtype, persistable=False)
7944 7945 7946 7947 7948 7949 7950 7951 7952 7953 7954 7955 7956 7957 7958 7959 7960 7961 7962 7963 7964 7965 7966

    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}
7967 7968 7969 7970 7971 7972 7973

    Examples:
        .. code-block:: python

            input = fluid.layers.data(
                name='data', shape=[1], dtype='float32')
            reward = fluid.layers.clip_by_norm(x=input, max_norm=1.0)
7974 7975 7976 7977 7978
    """

    helper = LayerHelper("clip_by_norm", **locals())

    if name is None:
S
sneaxiy 已提交
7979 7980 7981 7982
        name = unique_name.generate(".".join([helper.name, 'tmp']))

    out = helper.create_variable(
        type=x.type, name=name, dtype=x.dtype, persistable=False)
7983 7984 7985 7986 7987 7988 7989 7990

    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
7991 7992 7993 7994 7995 7996 7997 7998 7999 8000 8001 8002 8003 8004 8005 8006 8007 8008


@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}
    """

    helper = LayerHelper("mean", **locals())

    if name is None:
X
Xin Pan 已提交
8009
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
8010 8011 8012 8013 8014 8015 8016 8017 8018 8019 8020 8021 8022 8023 8024 8025 8026 8027 8028 8029 8030 8031 8032 8033 8034 8035 8036 8037 8038
    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


@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}
    """

    helper = LayerHelper("mul", **locals())

    if name is None:
X
Xin Pan 已提交
8039
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
8040 8041 8042 8043 8044 8045 8046 8047 8048
    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
8049 8050
            "x_num_col_dims": x_num_col_dims,
            "y_num_col_dims": y_num_col_dims
X
Xin Pan 已提交
8051 8052 8053 8054 8055 8056 8057 8058 8059 8060 8061 8062 8063 8064 8065 8066 8067 8068 8069 8070 8071 8072
        },
        outputs={"Out": out})
    return out


@templatedoc()
def sigmoid_cross_entropy_with_logits(x, label, name=None):
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        label(${label_type}): ${label_comment}
        name(basestring|None): Name of the output.

    Returns:
        out(${out_type}): ${out_comment}
    """

    helper = LayerHelper("sigmoid_cross_entropy_with_logits", **locals())

    if name is None:
X
Xin Pan 已提交
8073
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
8074 8075 8076 8077 8078 8079 8080 8081 8082 8083 8084 8085 8086 8087 8088 8089 8090 8091 8092 8093 8094 8095 8096 8097 8098 8099 8100 8101 8102
    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},
        attrs={},
        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}
    """
    helper = LayerHelper("maxout", **locals())

    if name is None:
X
Xin Pan 已提交
8103
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
8104 8105 8106 8107 8108 8109 8110 8111 8112 8113
    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
8114 8115


J
JiabinYang 已提交
8116
def space_to_depth(x, blocksize, name=None):
J
JiabinYang 已提交
8117
    """
J
JiabinYang 已提交
8118
    Gives a blocksize to space_to_depth the input LoDtensor with Layout: [batch, channel, height, width]
8119 8120 8121

    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
8122
    The attr blocksize indicates the input block size.
8123 8124

    space_to_depth will reorgnize the elements of input with shape[batch, channel, height, width] according
J
JiabinYang 已提交
8125
    to blocksize to construct output with shape [batch, channel * blocksize * blocksize, height/blocksize, width/blocksize]:
8126 8127

    space_to_depth is used to This operation is useful for resizing the activations between convolutions
J
JiabinYang 已提交
8128
    (but keeping all data)
J
JiabinYang 已提交
8129

J
JiabinYang 已提交
8130
    - Non-overlapping blocks of size block_size x block size are rearranged into depth at each location.
8131
    - The depth of the output tensor is block_size * block_size * input channel
J
JiabinYang 已提交
8132 8133 8134 8135 8136
    - 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
8137
    Args:
J
JiabinYang 已提交
8138
        x(variable): The input LoDtensor.
J
JiabinYang 已提交
8139
        blocksize(variable): The blocksize to select the element on each feature map should be > 2
J
JiabinYang 已提交
8140 8141

    Returns:
J
JiabinYang 已提交
8142
        Variable: The output LoDtensor.
J
JiabinYang 已提交
8143 8144

    Raises:
J
JiabinYang 已提交
8145
        TypeError: blocksize type must be a long.
J
JiabinYang 已提交
8146 8147 8148 8149 8150 8151

    Examples:
        .. code-block:: python

            data = fluid.layers.data(
                name='data', shape=[1, 4, 2, 2], dtype='float32')
J
JiabinYang 已提交
8152
            space_to_depthed = fluid.layers.space_to_depth(
J
JiabinYang 已提交
8153
                x=data, blocksize=2)
J
JiabinYang 已提交
8154 8155
    """

J
JiabinYang 已提交
8156
    helper = LayerHelper("space_to_depth", **locals())
J
JiabinYang 已提交
8157

J
JiabinYang 已提交
8158 8159
    if not (isinstance(blocksize, int)):
        raise ValueError("blocksize must be a python Int")
J
JiabinYang 已提交
8160 8161

    if name is None:
J
JiabinYang 已提交
8162 8163
        out = helper.create_variable_for_type_inference(
            dtype=x.dtype)  #fix create
J
JiabinYang 已提交
8164 8165 8166 8167 8168
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    helper.append_op(
J
JiabinYang 已提交
8169
        type="space_to_depth",
J
JiabinYang 已提交
8170
        inputs={"X": x},
J
JiabinYang 已提交
8171
        attrs={"blocksize": blocksize},
J
JiabinYang 已提交
8172
        outputs={"Out": out})
J
JiabinYang 已提交
8173 8174
    return out

J
JiabinYang 已提交
8175

S
sneaxiy 已提交
8176 8177
@templatedoc()
def sequence_reverse(x, name=None):
8178
    """
S
sneaxiy 已提交
8179 8180 8181 8182 8183 8184 8185 8186 8187 8188 8189
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        name(basestring|None): Name of the output.

    Returns:
        out(${y_type}): ${y_comment}
    """
    helper = LayerHelper("sequence_reverse", **locals())
    if name is None:
S
sneaxiy 已提交
8190
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
S
sneaxiy 已提交
8191 8192 8193 8194 8195 8196 8197 8198 8199 8200
    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
8201 8202


8203 8204 8205 8206 8207 8208
def affine_channel(x, scale=None, bias=None, data_layout='NCHW', name=None):
    """
    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.
8209

8210 8211 8212 8213 8214 8215 8216 8217 8218 8219 8220 8221 8222 8223 8224 8225 8226 8227 8228
    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.

    Returns:
        out (Variable): A tensor of the same shape and data layout with x.
    """
    helper = LayerHelper("affine_channel", **locals())

    if name is None:
X
Xin Pan 已提交
8229
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
8230 8231 8232 8233 8234 8235 8236 8237 8238 8239 8240 8241
    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})
    return out
8242 8243


B
barrierye 已提交
8244
def similarity_focus(input, axis, indexes, name=None):
8245
    """
B
barrierye 已提交
8246
    SimilarityFocus Operator
B
barrierye 已提交
8247 8248

    Generate a similarity focus mask with the same shape of input using the following method:
8249 8250 8251
    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
8252
       is (BatchSize, A, B, C), the shape of tensor T is (BatchSize, B, C).
8253 8254 8255 8256 8257 8258 8259
    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
8260
       each index.
B
barrierye 已提交
8261 8262 8263 8264
    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
8265 8266 8267 8268 8269 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
    .. 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
8314
    Args:
8315
        input(Variable): The input tensor variable(default float). It should
B
barrierye 已提交
8316
            be a 4-D tensor with shape [BatchSize, A, B, C].
B
barrierye 已提交
8317
        axis(int): Indicating the dimension to be selected. It can only be
B
barrierye 已提交
8318
            1, 2 or 3.
B
barrierye 已提交
8319
        indexes(list): Indicating the indexes of the selected dimension.
B
barrierye 已提交
8320 8321

    Returns:
8322
        Variable: A tensor variable with the same shape and same type
B
barrierye 已提交
8323
            as the input.
8324

B
barrierye 已提交
8325 8326 8327
    Examples:
        .. code-block:: python
            data = fluid.layers.data(
B
barrierye 已提交
8328 8329
              name='data', shape=[2, 3, 2, 2], dtype='float32')
            x = fluid.layers.layer_norm(input=data, axis=1, indexes=[0])
B
barrierye 已提交
8330 8331 8332 8333 8334 8335 8336 8337 8338 8339 8340 8341
    """
    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
8342 8343 8344 8345 8346
    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
8347 8348 8349 8350 8351 8352 8353
    helper.append_op(
        type='similarity_focus',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={"axis": axis,
               "indexes": indexes})
    return out
B
barrierye 已提交
8354 8355


M
minqiyang 已提交
8356 8357
def hash(input, hash_size, num_hash=1, name=None):
    """
M
minqiyang 已提交
8358 8359
    Hash the input to an integer whose value is less than the given hash size.

M
minqiyang 已提交
8360 8361
    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
8362 8363 8364 8365 8366 8367 8368 8369 8370 8371 8372 8373 8374 8375 8376 8377 8378 8379 8380 8381 8382 8383 8384 8385 8386 8387 8388 8389 8390 8391 8392 8393 8394 8395 8396 8397 8398 8399

    A simple example as below:

    .. code-block:: text

        Given:

        # shape [2, 2]
        input.data = [
            [[1], [2]],
            [[3], [4]],
        ]

        input.lod = [[0, 2]]

        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 = [
            [[9662], [9217], [1129], [8487]],
            [[8310], [1327], [1654], [4567]],
        ]

        output.lod = [[0, 2]]

    Args:
        input (Variable): The input variable which is a one-hot word. The
            dimensions of the input variable must be 2.
        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
8400
        num_hash (int): The times of hash, default 1.
M
minqiyang 已提交
8401
        name (str, default None): The name of this layer.
M
minqiyang 已提交
8402 8403 8404 8405 8406 8407 8408 8409 8410

    Returns:
       Variable: The hash result variable which is a LoDTensor.

    Examples:
       .. code-block:: python
           word_dict = paddle.dataset.imdb.word_dict()
           x = fluid.layers.data(shape[1], dtype='int32', lod_level=1)
           out = fluid.layers.hash(input=x, num_hash=4, hash_size=1000)
M
minqiyang 已提交
8411 8412
    """
    helper = LayerHelper('hash', **locals())
M
minqiyang 已提交
8413 8414
    out = helper.create_variable_for_type_inference(
        helper.input_dtype(), stop_gradient=True)
M
minqiyang 已提交
8415 8416 8417 8418 8419 8420 8421
    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
8422 8423


D
dengkaipeng 已提交
8424
@templatedoc()
8425 8426
def grid_sampler(x, grid, name=None):
    """
8427
    This operation samples input X by using bilinear interpolation based on
8428
    flow field grid, which is usually gennerated by affine_grid. The grid of
8429 8430 8431 8432
    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
8433
    interpolation value of 4 nearest corner points.
8434 8435 8436 8437 8438 8439 8440 8441

    Step 1:
    Get (x, y) grid coordinates and scale to [0, H-1/W-1].

    grid_x = 0.5 * (grid[:, :, :, 0] + 1) * (W - 1)
    grid_y = 0.5 * (grid[:, :, :, 1] + 1) * (H - 1)

    Step 2:
8442
    Indices input data X with grid (x, y) in each [H, W] area, and bilinear
8443 8444 8445 8446 8447 8448 8449 8450 8451 8452 8453 8454 8455 8456 8457 8458 8459 8460 8461 8462 8463 8464 8465 8466 8467 8468 8469 8470 8471
    interpolate point value by 4 nearest points.

      wn ------- y_n ------- en
      |           |           |
      |          d_n          |
      |           |           |
     x_w --d_w-- grid--d_e-- x_e
      |           |           |
      |          d_s          |
      |           |           |
      ws ------- y_s ------- wn

    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

    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

    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

    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
8472 8473

    Args:
8474 8475 8476
        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
8477 8478

    Returns:
8479
        out(Variable): Output of shape [N, C, H, W] data samples input X
8480 8481 8482 8483 8484 8485 8486 8487 8488
        using bilnear interpolation based on input grid.

    Exmples:
    .. code-block:: python

        x = fluid.layers.data(name='x', shape=[3, 10, 32, 32], dtype='float32')
        theta = fluid.layers.data(name='theta', shape=[3, 2, 3], dtype='float32')
        grid = fluid.layers.affine_grid(input=theta, size=[3, 10, 32, 32]})
        out = fluid.layers.grid_sampler(x=x, grid=grid)
D
dengkaipeng 已提交
8489 8490 8491 8492 8493 8494 8495 8496 8497
    """
    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")

8498
    out = helper.create_variable_for_type_inference(x.dtype)
D
dengkaipeng 已提交
8499 8500
    ipts = {'X': x, 'Grid': grid}

8501
    helper.append_op(type='grid_sampler', inputs=ipts, outputs={'Output': out})
8502 8503 8504
    return out


G
gmcather 已提交
8505 8506 8507 8508 8509 8510 8511 8512 8513 8514 8515 8516 8517 8518 8519 8520 8521 8522 8523 8524 8525 8526 8527 8528 8529 8530 8531 8532 8533 8534 8535 8536 8537 8538 8539 8540 8541 8542 8543 8544 8545 8546 8547 8548 8549 8550 8551 8552 8553 8554 8555 8556 8557 8558 8559 8560 8561 8562 8563 8564 8565 8566 8567 8568 8569 8570 8571 8572 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
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

          prob = fluid.layers.sigmoid(net)
          cost = fluid.layers.log_loss(input=prob, label=label)
    """
    helper = LayerHelper('log_loss', **locals())

    if name is None:
        loss = helper.create_variable_for_type_inference(dtype=input.dtype)
    else:
        loss = helper.create_variable(
            name=name, dtype=input.dtype, persistable=False)

    helper.append_op(
        type='log_loss',
        inputs={'Predicted': [input],
                'Labels': [label]},
        outputs={'Loss': [loss]},
        attrs={'epsilon': epsilon})
    return loss


def add_position_encoding(input, alpha, beta, name=None):
    """
    **Add Position Encoding Layer**

    This layer accepts an input 3D-Tensor of shape [N x M x P], and return an
    output Tensor of shape [N x M x P] with positional encoding value.

    Refer to `Attention Is All You Need<http://arxiv.org/pdf/1706.03762.pdf>`_ .

    .. math::
        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)

    Where:
    * PE(pos, 2i): the increment for the number at even position
    * PE(pos, 2i + 1): the increment for the number at odd position

    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

          position_tensor = fluid.layers.add_position_encoding(input=tensor)
    """
    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
8599 8600 8601 8602 8603 8604 8605 8606 8607 8608


def bilinear_tensor_product(x,
                            y,
                            size,
                            act=None,
                            name=None,
                            param_attr=None,
                            bias_attr=None):
    """
Q
Qiao Longfei 已提交
8609
    **Add Bilinear Tensor Product Layer**
Q
Qiao Longfei 已提交
8610

Q
Qiao Longfei 已提交
8611
    This layer performs bilinear tensor product on two inputs.
Q
Qiao Longfei 已提交
8612 8613 8614
    For example:

    .. math::
8615
       out{i} = x * W_{i} * {y^\mathrm{T}}, i=0,1,...,size-1
Q
Qiao Longfei 已提交
8616

Q
Qiao Longfei 已提交
8617
    In this formula:
8618 8619
      - :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
8620
      - :math:`W_{i}`: the i-th learned weight, shape is [M, N]
8621
      - :math:`out{i}`: the i-th element of out, shape is [batch_size, size].
Q
Qiao Longfei 已提交
8622 8623 8624
      - :math:`y^\mathrm{T}`: the transpose of :math:`y_{2}`.

    Args:
8625 8626
        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
8627 8628 8629
        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
8630
        param_attr (ParamAttr, default None): The parameter attribute for the learnable w.
Q
Qiao Longfei 已提交
8631
            parameters/weights of this layer.
Q
Qiao Longfei 已提交
8632
        bias_attr (ParamAttr, default None): The parameter attribute for the bias
Q
Qiao Longfei 已提交
8633 8634 8635 8636
            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
8637
        Variable: A 2-D Tensor of shape [batch_size, size].
Q
Qiao Longfei 已提交
8638 8639 8640 8641

    Examples:
        .. code-block:: python

Q
Qiao Longfei 已提交
8642
          tensor = bilinear_tensor_product(x=layer1, y=layer2, size=1000)
Q
Qiao Longfei 已提交
8643 8644
    """
    helper = LayerHelper('bilinear_tensor_product', **locals())
Q
Qiao Longfei 已提交
8645
    dtype = helper.input_dtype('x')
Q
Qiao Longfei 已提交
8646 8647 8648 8649

    param_shape = [size, x.shape[1], y.shape[1]]

    w = helper.create_parameter(
Q
Qiao Longfei 已提交
8650
        attr=helper.param_attr, shape=param_shape, dtype=dtype, is_bias=False)
Q
Qiao Longfei 已提交
8651 8652 8653 8654 8655 8656 8657 8658 8659 8660 8661 8662 8663 8664 8665 8666 8667

    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)