nn.py 367.8 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.
14
"""
15
All layers just related to the neural network.
16 17
"""

18 19
from __future__ import print_function

20
import numpy as np
S
sneaxiy 已提交
21
import six
22
import os
S
sneaxiy 已提交
23
import inspect
24
from ..layer_helper import LayerHelper
25
from ..initializer import Normal, Constant, NumpyArrayInitializer
S
sneaxiy 已提交
26
from ..framework import Variable, OpProtoHolder
27
from ..param_attr import ParamAttr
S
sneaxiy 已提交
28
from .layer_function_generator import autodoc, templatedoc, _generate_doc_string_
29
from .tensor import concat, assign
30
from . import utils
F
fengjiayi 已提交
31
from .. import unique_name
32
from functools import reduce
33
from .. import core
X
Xin Pan 已提交
34
from ..imperative import layers
35 36

__all__ = [
X
Xin Pan 已提交
37 38 39 40 41 42 43 44 45 46
    'fc',
    'embedding',
    'dynamic_lstm',
    'dynamic_lstmp',
    'dynamic_gru',
    'gru_unit',
    'linear_chain_crf',
    'crf_decoding',
    'cos_sim',
    'cross_entropy',
47
    'bpr_loss',
X
Xin Pan 已提交
48 49 50 51 52 53 54 55 56 57
    'square_error_cost',
    'chunk_eval',
    'sequence_conv',
    'conv2d',
    'conv3d',
    'sequence_pool',
    'sequence_softmax',
    'softmax',
    'pool2d',
    'pool3d',
58 59
    'adaptive_pool2d',
    'adaptive_pool3d',
X
Xin Pan 已提交
60
    'batch_norm',
H
heqiaozhi 已提交
61
    'data_norm',
X
Xin Pan 已提交
62 63 64 65 66 67
    'beam_search_decode',
    'conv2d_transpose',
    'conv3d_transpose',
    'sequence_expand',
    'sequence_expand_as',
    'sequence_pad',
68
    'sequence_unpad',
X
Xin Pan 已提交
69 70 71 72 73 74 75 76
    'lstm_unit',
    'reduce_sum',
    'reduce_mean',
    'reduce_max',
    'reduce_min',
    'reduce_prod',
    'sequence_first_step',
    'sequence_last_step',
77
    'sequence_slice',
X
Xin Pan 已提交
78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94
    'dropout',
    'split',
    'ctc_greedy_decoder',
    'edit_distance',
    'l2_normalize',
    'matmul',
    'topk',
    'warpctc',
    'sequence_reshape',
    'transpose',
    'im2sequence',
    'nce',
    'hsigmoid',
    'beam_search',
    'row_conv',
    'multiplex',
    'layer_norm',
D
Dun 已提交
95
    'group_norm',
X
Xin Pan 已提交
96 97 98 99 100 101 102 103 104 105 106 107 108
    'softmax_with_cross_entropy',
    'smooth_l1',
    'one_hot',
    'autoincreased_step_counter',
    'reshape',
    'squeeze',
    'unsqueeze',
    'lod_reset',
    'lrn',
    'pad',
    'pad_constant_like',
    'label_smooth',
    'roi_pool',
J
jerrywgz 已提交
109
    'roi_align',
X
Xin Pan 已提交
110 111 112 113
    'dice_loss',
    'image_resize',
    'image_resize_short',
    'resize_bilinear',
114
    'resize_nearest',
X
Xin Pan 已提交
115 116 117 118 119 120
    'gather',
    'scatter',
    'sequence_scatter',
    'random_crop',
    'mean_iou',
    'relu',
C
chengduo 已提交
121
    'selu',
X
Xin Pan 已提交
122 123 124
    'log',
    'crop',
    'rank_loss',
125
    'margin_rank_loss',
X
Xin Pan 已提交
126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168
    '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
169
    'space_to_depth',
170
    'affine_grid',
S
sneaxiy 已提交
171
    'sequence_reverse',
172
    'affine_channel',
B
barrierye 已提交
173
    'similarity_focus',
M
minqiyang 已提交
174
    'hash',
175
    'grid_sampler',
176 177
    'log_loss',
    'add_position_encoding',
178
    'bilinear_tensor_product',
C
chengduo 已提交
179 180
    'merge_selected_rows',
    'get_tensor_from_selected_rows',
P
phlrain 已提交
181
    'lstm',
S
shippingwang 已提交
182
    'shuffle_channel',
S
sneaxiy 已提交
183
    'py_func',
184
    'psroi_pool',
185
    'teacher_student_sigmoid_loss',
M
minqiyang 已提交
186
    'huber_loss',
Z
zhaozhehao 已提交
187
    'tree_conv',
188 189
]

J
jerrywgz 已提交
190 191
kIgnoreIndex = -100

192 193 194 195 196 197 198

def fc(input,
       size,
       num_flatten_dims=1,
       param_attr=None,
       bias_attr=None,
       act=None,
199
       is_test=False,
200
       name=None):
201
    """
202
    **Fully Connected Layer**
203

204 205 206 207 208 209 210 211
    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
212
    to the output as well.
C
caoying03 已提交
213

C
caoying03 已提交
214
    This process can be formulated as follows:
215 216 217

    .. math::

218
        Out = Act({\sum_{i=0}^{N-1}X_iW_i + b})
219 220 221

    In the above equation:

222 223 224 225
    * :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).
226
    * :math:`Act`: The activation function.
C
caoying03 已提交
227
    * :math:`Out`: The output tensor.
228 229

    Args:
R
ranqiu 已提交
230 231 232 233 234 235 236 237 238 239
        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
240
            `X` is a 5-dimensional tensor with a shape [2, 3, 4, 5, 6], and `num_flatten_dims` = 3.
R
ranqiu 已提交
241 242 243 244
            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
245 246
            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
247
        act (str, default None): Activation to be applied to the output of this layer.
248
        is_test(bool): A flag indicating whether execution is in test phase.
R
ranqiu 已提交
249
        name (str, default None): The name of this layer.
250

251
    Returns:
F
fengjiayi 已提交
252
        Variable: The transformation result.
253 254

    Raises:
255
        ValueError: If rank of the input tensor is less than 2.
256 257 258 259

    Examples:
        .. code-block:: python

F
fengjiayi 已提交
260
          data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
261
          fc = fluid.layers.fc(input=data, size=1000, act="tanh")
262
    """
C
caoying03 已提交
263

C
caoying03 已提交
264
    helper = LayerHelper("fc", **locals())
265 266 267 268

    dtype = helper.input_dtype()

    mul_results = []
269 270
    for input_var, param_attr in helper.iter_inputs_and_params():
        input_shape = input_var.shape
271 272 273
        param_shape = [
            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
        ] + [size]
274

275
        w = helper.create_parameter(
276
            attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False)
277
        tmp = helper.create_variable_for_type_inference(dtype)
278
        helper.append_op(
279 280 281
            type="mul",
            inputs={"X": input_var,
                    "Y": w},
282
            outputs={"Out": tmp},
M
mozga-intel 已提交
283 284
            attrs={"x_num_col_dims": num_flatten_dims,
                   "y_num_col_dims": 1})
285 286 287 288
        mul_results.append(tmp)

    if len(mul_results) == 1:
        pre_bias = mul_results[0]
289
    else:
290
        pre_bias = helper.create_variable_for_type_inference(dtype)
291
        helper.append_op(
292 293 294
            type="sum",
            inputs={"X": mul_results},
            outputs={"Out": pre_bias},
X
Xin Pan 已提交
295
            attrs={"use_mkldnn": False})
296 297 298 299
    # add bias
    pre_activation = helper.append_bias_op(pre_bias, dim_start=num_flatten_dims)
    # add activation
    return helper.append_activation(pre_activation)
300 301


302 303 304
def embedding(input,
              size,
              is_sparse=False,
305
              is_distributed=False,
306 307 308
              padding_idx=None,
              param_attr=None,
              dtype='float32'):
309
    """
310 311
    **Embedding Layer**

312
    This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in
313 314
    a lookup table. The result of this lookup is the embedding of each ID in the
    :attr:`input`.
315 316 317

    All the input variables are passed in as local variables to the LayerHelper
    constructor.
318 319

    Args:
320 321 322 323 324
        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.
325
        is_distributed(bool): Whether to run lookup table from remote parameter server.
326 327
        padding_idx(int|long|None): If :attr:`None`, it makes no effect to lookup.
            Otherwise the given :attr:`padding_idx` indicates padding the output
328
            with zeros whenever lookup encounters it in :attr:`input`. If
329
            :math:`padding_idx < 0`, the :attr:`padding_idx` to use in lookup is
330 331
            :math:`size[0] + dim`.
        param_attr(ParamAttr): Parameters for this layer
332
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc
333

334 335 336
    Returns:
        Variable: The tensor variable storing the embeddings of the \
                  supplied inputs.
337

338 339
    Examples:
        .. code-block:: python
340

C
chengduoZH 已提交
341
          dict_size = len(dataset.ids)
342
          data = fluid.layers.data(name='ids', shape=[32, 32], dtype='float32')
C
chengduoZH 已提交
343
          fc = fluid.layers.embedding(input=data, size=[dict_size, 16])
344 345 346
    """

    helper = LayerHelper('embedding', **locals())
347
    remote_prefetch = is_sparse and (not is_distributed)
Q
Qiao Longfei 已提交
348 349
    if remote_prefetch:
        assert is_sparse is True and is_distributed is False
350 351
    w = helper.create_parameter(
        attr=helper.param_attr, shape=size, dtype=dtype, is_bias=False)
352
    tmp = helper.create_variable_for_type_inference(dtype)
353 354
    padding_idx = -1 if padding_idx is None else padding_idx if padding_idx >= 0 else (
        size[0] + padding_idx)
355 356 357 358 359
    helper.append_op(
        type='lookup_table',
        inputs={'Ids': input,
                'W': w},
        outputs={'Out': tmp},
360 361 362
        attrs={
            'is_sparse': is_sparse,
            'is_distributed': is_distributed,
Q
Qiao Longfei 已提交
363
            'remote_prefetch': remote_prefetch,
364 365
            'padding_idx': padding_idx
        })
366 367 368
    return tmp


W
wopeizl 已提交
369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384
@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}
385

W
wopeizl 已提交
386 387 388 389 390 391 392 393 394 395 396
    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.
397

W
wopeizl 已提交
398 399 400 401
                               - Weights = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}
                               - The shape is (D x 4D), where D is the hidden
                                 size.
402

W
wopeizl 已提交
403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488
                               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
489 490


P
phlrain 已提交
491 492 493 494 495 496
def lstm(input,
         init_h,
         init_c,
         max_len,
         hidden_size,
         num_layers,
P
phlrain 已提交
497
         dropout_prob=0.0,
P
phlrain 已提交
498 499 500 501 502
         is_bidirec=False,
         is_test=False,
         name=None,
         default_initializer=None,
         seed=-1):
503
    """
P
phlrain 已提交
504
    If Device is GPU, This op will use cudnn LSTM implementation
505 506

    A four-gate Long Short-Term Memory network with no peephole connections.
M
minqiyang 已提交
507
    In the forward pass the output ht and cell output ct for a given iteration can be computed from the recurrent input ht-1,
508 509
    the cell input ct-1 and the previous layer input xt given matrices W, R and biases bW, bR from the following equations:

510
    .. math::
M
minqiyang 已提交
511 512 513 514 515 516 517

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

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

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

518
       \\tilde{c_t} &= tanh(W_{cx}x_t + W_{ch}h_{t-1} + bx_c + bh_c)
M
minqiyang 已提交
519 520 521 522

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

       h_t &= o_t \odot tanh(c_t)
523 524

    - $W$ terms denote weight matrices (e.g. $W_{ix}$ is the matrix
P
phlrain 已提交
525 526 527 528 529 530
      of weights from the input gate to the input)
    - The b terms denote bias vectors ($bx_i$ and $bh_i$ are the input gate bias vector).
    - sigmoid is the logistic sigmoid function.
    - $i, f, o$ and $c$ are the input gate, forget gate, output gate,
      and cell activation vectors, respectively, all of which have the same size as
      the cell output activation vector $h$.
531 532 533
    - The :math:`\odot` is the element-wise product of the vectors.
    - :math:`tanh` is the activation functions.
    - :math:`\\tilde{c_t}` is also called candidate hidden state,
P
phlrain 已提交
534
      which is computed based on the current input and the previous hidden state.
535

M
minqiyang 已提交
536
    Where sigmoid is the sigmoid operator: :math:`sigmoid(x) = 1 / (1 + e^{-x})` , * represents a point-wise multiplication,
537 538 539 540 541
    X represensts a matrix multiplication


    Args:
        input (Variable): LSTM input tensor, shape MUST be ( seq_len x batch_size x input_size )
M
minqiyang 已提交
542
        init_h(Variable): The initial hidden state of the LSTM
543 544 545 546 547
                       This is a tensor with shape ( num_layers x batch_size x hidden_size)
                       if is_bidirec = True, shape should be ( num_layers*2 x batch_size x hidden_size)
        init_c(Variable): The initial cell state of the LSTM.
                       This is a tensor with shape ( num_layers x batch_size x hidden_size )
                       if is_bidirec = True, shape should be ( num_layers*2 x batch_size x hidden_size)
M
minqiyang 已提交
548
        max_len (int): max length of LSTM. the first dim of input tensor CAN NOT greater than max_len
549 550
        hidden_size (int): hidden size of the LSTM
        num_layers (int): total layers number of the LSTM
P
phlrain 已提交
551 552
        dropout_prob(float|0.0): dropout prob, dropout ONLY work between rnn layers, NOT between time steps
                             There is NO dropout work on rnn output of the last RNN layers
553 554 555 556 557 558
        is_bidirec (bool): If it is bidirectional
        is_test (bool): If it is in test phrase
        name (str|None): A name for this layer(optional). If set None, the layer
                         will be named automatically.
        default_initializer(Initialize|None): Where use initializer to initialize the Weight
                         If set None, defaule initializer will be used
P
phlrain 已提交
559
        seed(int): Seed for dropout in LSTM, If it's -1, dropout will use random seed
P
phlrain 已提交
560

561 562

    Returns:
M
minqiyang 已提交
563 564
        rnn_out(Tensor),last_h(Tensor),last_c(Tensor):

565
                        Three tensors, rnn_out, last_h, last_c:
M
minqiyang 已提交
566

567 568 569 570
                        - rnn_out is result of LSTM hidden, shape is (seq_len x batch_size x hidden_size) \
                          if is_bidirec set to True, shape will be ( seq_len x batch_sze x hidden_size*2)
                        - last_h is the hidden state of the last step of LSTM \
                          shape is ( num_layers x batch_size x hidden_size ) \
M
minqiyang 已提交
571
                          if is_bidirec set to True, shape will be ( num_layers*2 x batch_size x hidden_size)
572 573
                        - last_c(Tensor): the cell state of the last step of LSTM \
                          shape is ( num_layers x batch_size x hidden_size ) \
M
minqiyang 已提交
574
                          if is_bidirec set to True, shape will be ( num_layers*2 x batch_size x hidden_size)
575 576 577 578 579 580 581 582 583 584 585 586 587 588 589


    Examples:
        .. code-block:: python

            input = embedding
            batch_size = 20
            max_len = 100
            dropout_prob = 0.2
            input_size = 100
            hidden_size = 150
            num_layers = 1
            init_hidden1 = layers.fill_constant( [num_layers, batch_size, hidden_size], 'float32', 0.0, stop_grad=False)
            init_cell1 = layers.fill_constant( [num_layers, batch_size, hidden_size], 'float32', 0.0, stop_grad=False)

P
phlrain 已提交
590
            rnn_out, last_h, last_c = layers.lstm( input, init_h, init_c, \
591 592 593 594 595 596
                    max_len, dropout_prob, input_size, hidden_size, \
                    num_layers)
    """

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

P
phlrain 已提交
597 598 599
    dtype = input.dtype
    input_shape = list(input.shape)
    input_size = input_shape[-1]
600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658
    weight_size = 0
    for i in range(num_layers):
        if i == 0:
            input_weight_size = (input_size * hidden_size) * 4
        else:
            if is_bidirec:
                input_weight_size = (hidden_size * 2 * hidden_size) * 4
            else:
                input_weight_size = (hidden_size * hidden_size) * 4

        hidden_weight_size = (hidden_size * hidden_size) * 4

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

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

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

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

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


659 660 661 662 663 664 665 666 667 668 669
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',
670 671
                  dtype='float32',
                  name=None):
672 673 674
    """
    **Dynamic LSTMP Layer**

675 676 677 678 679 680
    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).
681 682 683 684 685

    The formula is as follows:

    .. math::

686
        i_t & = \sigma(W_{ix}x_{t} + W_{ir}r_{t-1} + W_{ic}c_{t-1} + b_i)
687

688
        f_t & = \sigma(W_{fx}x_{t} + W_{fr}r_{t-1} + W_{fc}c_{t-1} + b_f)
689

690
        \\tilde{c_t} & = act_g(W_{cx}x_t + W_{cr}r_{t-1} + b_c)
691

692
        o_t & = \sigma(W_{ox}x_{t} + W_{or}r_{t-1} + W_{oc}c_t + b_o)
693

694
        c_t & = f_t \odot c_{t-1} + i_t \odot \\tilde{c_t}
695

696
        h_t & = o_t \odot act_h(c_t)
697

698
        r_t & = \overline{act_h}(W_{rh}h_t)
699

700 701 702 703 704 705
    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, \
706
          we use vectors to reprenset these diagonal weight matrices.
707
    * :math:`b`: Denotes bias vectors (e.g. :math:`b_i` is the input gate \
708
          bias vector).
709 710 711
    * :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 \
712
          the same size as the cell output activation vector :math:`h`.
713
    * :math:`h`: The hidden state.
714
    * :math:`r`: The recurrent projection of the hidden state.
715 716
    * :math:`\\tilde{c_t}`: The candidate hidden state, whose \
          computation is based on the current input and previous hidden state.
717
    * :math:`\odot`: The element-wise product of the vectors.
718
    * :math:`act_g` and :math:`act_h`: The cell input and cell output \
719
          activation functions and `tanh` is usually used for them.
720 721
    * :math:`\overline{act_h}`: The activation function for the projection \
          output, usually using `identity` or same as :math:`act_h`.
722 723 724 725

    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.
726

727 728 729 730 731 732 733 734 735 736 737 738
    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.
739
        param_attr(ParamAttr|None): The parameter attribute for the learnable
740 741
                               hidden-hidden weight and projection weight.

742 743
                               - Hidden-hidden weight = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}.
744 745
                               - The shape of hidden-hidden weight is (P x 4D),
                                 where P is the projection size and D the hidden
746 747
                                 size.
                               - Projection weight = {:math:`W_{rh}`}.
748
                               - The shape of projection weight is (D x P).
749 750 751 752 753

                               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.
754
        bias_attr(ParamAttr|None): The bias attribute for the learnable bias
755 756 757 758 759 760
                              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`}.
761
                                - The shape is (1 x 4D).
762 763 764
                              2. `use_peepholes = True`
                                - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
                                                 W_{fc}, W_{oc}`}.
765
                                - The shape is (1 x 7D).
766 767 768 769 770

                              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.
771 772 773 774 775 776 777 778 779
        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.
780
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
781 782
                              default "tanh".
        proj_activation(str): The activation for projection output.
783
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
784 785
                              default "tanh".
        dtype(str): Data type. Choices = ["float32", "float64"], default "float32".
786 787
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
788 789

    Returns:
790 791 792 793
        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`.
794 795

    Examples:
796

797 798
        .. code-block:: python

799 800 801 802
            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])
803
            hidden_dim, proj_dim = 512, 256
804
            fc_out = fluid.layers.fc(input=emb, size=hidden_dim * 4,
805
                                     act=None, bias_attr=None)
806 807 808
            proj_out, _ = fluid.layers.dynamic_lstmp(input=fc_out,
                                                     size=hidden_dim * 4,
                                                     proj_size=proj_dim,
809 810 811 812
                                                     use_peepholes=False,
                                                     is_reverse=True,
                                                     cell_activation="tanh",
                                                     proj_activation="tanh")
813
    """
814

815
    assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp."
816
    helper = LayerHelper('lstmp', **locals())
817
    size = size // 4
818 819 820 821 822 823 824 825 826 827
    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)

828 829 830 831 832 833
    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)
834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861

    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
862 863 864 865 866 867 868
def dynamic_gru(input,
                size,
                param_attr=None,
                bias_attr=None,
                is_reverse=False,
                gate_activation='sigmoid',
                candidate_activation='tanh',
869 870
                h_0=None,
                origin_mode=False):
G
guosheng 已提交
871
    """
872
    **Gated Recurrent Unit (GRU) Layer**
G
guosheng 已提交
873

874 875 876
    if origin_mode is False, then the equation of a gru step is from paper
    `Empirical Evaluation of Gated Recurrent Neural Networks on Sequence
    Modeling <https://arxiv.org/pdf/1412.3555.pdf>`_ .
877

G
guosheng 已提交
878 879 880 881 882 883 884 885 886
    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)
887

G
guosheng 已提交
888
        h_t & = (1-u_t) \odot h_{t-1} + u_t \odot \\tilde{h_t}
889

890 891 892

    if origin_mode is True then the equation is from paper
    Learning Phrase Representations using RNN Encoder-Decoder for Statistical
893 894 895 896 897 898 899 900 901 902 903 904
    Machine Translation <https://arxiv.org/pdf/1406.1078.pdf>`_

    .. math::

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

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

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

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

G
guosheng 已提交
905
    The :math:`\odot` is the element-wise product of the vectors. :math:`act_g`
906 907
    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
908 909 910 911
    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
912
    to use fully-connect layer before GRU layer.
G
guosheng 已提交
913 914

    Args:
915 916
        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
917
            Variable is a matrix with shape :math:`(T \\times 3D)`, where
918
            :math:`T` is the total time steps in this mini-batch, :math:`D`
G
guosheng 已提交
919 920
            is the hidden size.
        size(int): The dimension of the gru cell.
921
        param_attr(ParamAttr|None): The parameter attribute for the learnable
G
guosheng 已提交
922 923
            hidden-hidden weight matrix. Note:

924
            - The shape of the weight matrix is :math:`(T \\times 3D)`, where
G
guosheng 已提交
925
              :math:`D` is the hidden size.
926
            - All elements in the weight matrix can be divided into two parts.
G
guosheng 已提交
927
              The first part are weights of the update gate and reset gate with
928
              shape :math:`(D \\times 2D)`, and the second part are weights for
G
guosheng 已提交
929
              candidate hidden state with shape :math:`(D \\times D)`.
930 931 932 933 934

            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
935
            of GRU.Note that the bias with :math:`(1 \\times 3D)` concatenates
936
            the bias in the update gate, reset gate and candidate calculations.
937 938 939
            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
940 941
            bias_attr. If the Initializer of the bias_attr is not set, the bias
            is initialized zero. Default: None.
942
        is_reverse(bool): Whether to compute reversed GRU, default
G
guosheng 已提交
943 944 945
            :attr:`False`.
        gate_activation(str): The activation for update gate and reset gate.
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "sigmoid".
946
        candidate_activation(str): The activation for candidate hidden state.
G
guosheng 已提交
947
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh".
948 949 950 951
        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
952 953

    Returns:
954
        Variable: The hidden state of GRU. The shape is :math:`(T \\times D)`, \
955
            and sequence length is the same with the input.
956

G
guosheng 已提交
957
    Examples:
958

G
guosheng 已提交
959 960
        .. code-block:: python

961 962 963 964
            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
965
            hidden_dim = 512
966
            x = fluid.layers.fc(input=emb, size=hidden_dim * 3)
967
            hidden = fluid.layers.dynamic_gru(input=x, size=hidden_dim)
G
guosheng 已提交
968 969 970 971 972 973 974 975 976
    """

    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)
977
    batch_size = input.shape[0]
G
guosheng 已提交
978
    inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
S
sneaxiy 已提交
979
    if h_0:
G
guosheng 已提交
980
        assert h_0.shape == (
981 982 983
            batch_size, size
        ), 'The shape of h0 should be(batch_size, %d)' % size
        inputs['H0'] = h_0
G
guosheng 已提交
984

985 986 987 988
    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
989 990 991 992 993 994 995 996 997 998 999 1000 1001

    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,
1002 1003
            'activation': candidate_activation,
            'origin_mode': origin_mode
G
guosheng 已提交
1004 1005 1006 1007
        })
    return hidden


1008 1009 1010
def gru_unit(input,
             hidden,
             size,
1011 1012
             param_attr=None,
             bias_attr=None,
1013
             activation='tanh',
Q
Qiao Longfei 已提交
1014 1015
             gate_activation='sigmoid',
             origin_mode=False):
1016
    """
1017 1018 1019
    **GRU unit layer**

    if origin_mode is True, then the equation of a gru step is from paper
1020
    `Learning Phrase Representations using RNN Encoder-Decoder for Statistical
1021
    Machine Translation <https://arxiv.org/pdf/1406.1078.pdf>`_
1022

1023 1024
        .. math::
            u_t & = actGate(xu_{t} + W_u h_{t-1} + b_u)
1025

1026
            r_t & = actGate(xr_{t} + W_r h_{t-1} + b_r)
1027

1028
            m_t & = actNode(xm_t + W_c dot(r_t, h_{t-1}) + b_m)
1029

1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044
            h_t & = dot(u_t, h_{t-1}) + dot((1-u_t), m_t)

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

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

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

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

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

1045 1046

    The inputs of gru unit includes :math:`z_t`, :math:`h_{t-1}`. In terms
1047 1048 1049
    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
1050 1051
    connected layer has to be applied, such that :math:`z_t = W_{fc}x_t`.

1052 1053
    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
1054 1055 1056
    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`.
1057 1058 1059

    Args:
        input (Variable): The fc transformed input value of current step.
1060
        hidden (Variable): The hidden value of gru unit from previous step.
1061
        size (integer): The input dimension value.
1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075
        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
1076
            of GRU.Note that the bias with :math:`(1 \\times 3D)` concatenates
1077
            the bias in the update gate, reset gate and candidate calculations.
1078 1079 1080
            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
1081 1082
            bias_attr. If the Initializer of the bias_attr is not set, the bias
            is initialized zero. Default: None.
1083 1084 1085 1086
        activation (string): The activation type for cell (actNode).
                             Default: 'tanh'
        gate_activation (string): The activation type for gates (actGate).
                                  Default: 'sigmoid'
1087

1088 1089 1090 1091 1092 1093
    Returns:
        tuple: The hidden value, reset-hidden value and gate values.

    Examples:

        .. code-block:: python
1094

1095
             # assuming we have x_t_data and prev_hidden of size=10
1096
             x_t = fluid.layers.fc(input=x_t_data, size=30)
1097 1098
             hidden_val, r_h_val, gate_val = fluid.layers.gru_unit(input=x_t,
                                                    hidden = prev_hidden)
1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110

    """
    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()
1111
    size = size // 3
1112 1113

    # create weight
1114 1115
    weight = helper.create_parameter(
        attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype)
1116

1117 1118 1119
    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)
1120
    inputs = {'Input': input, 'HiddenPrev': hidden, 'Weight': weight}
1121
    # create bias
1122
    if helper.bias_attr:
1123 1124 1125
        bias_size = [1, 3 * size]
        bias = helper.create_parameter(
            attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
1126
        inputs['Bias'] = bias
1127 1128 1129

    helper.append_op(
        type='gru_unit',
1130
        inputs=inputs,
1131 1132 1133 1134 1135 1136
        outputs={
            'Gate': gate,
            'ResetHiddenPrev': reset_hidden_pre,
            'Hidden': updated_hidden,
        },
        attrs={
1137 1138
            'activation': 2,  # tanh
            'gate_activation': 1,  # sigmoid
1139 1140 1141 1142 1143
        })

    return updated_hidden, reset_hidden_pre, gate


Y
yuyang18 已提交
1144
@templatedoc()
1145
def linear_chain_crf(input, label, param_attr=None):
Y
yuyang18 已提交
1146 1147 1148 1149 1150 1151 1152
    """
    Linear Chain CRF.

    ${comment}

    Args:
        input(${emission_type}): ${emission_comment}
D
dzhwinter 已提交
1153
        input(${transition_type}): ${transition_comment}
Y
yuyang18 已提交
1154 1155 1156 1157
        label(${label_type}): ${label_comment}
        param_attr(ParamAttr): The attribute of the learnable parameter.

    Returns:
D
dzhwinter 已提交
1158 1159 1160
        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
1161 1162

    """
1163 1164 1165 1166 1167 1168
    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())
1169 1170 1171 1172 1173 1174 1175 1176
    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())
1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191
    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
1192 1193 1194 1195
@templatedoc()
def crf_decoding(input, param_attr, label=None):
    """
    ${comment}
Y
yi.wu 已提交
1196

W
wopeizl 已提交
1197 1198
    Args:
        input(${emission_type}): ${emission_comment}
Y
yi.wu 已提交
1199

W
wopeizl 已提交
1200
        param_attr(ParamAttr): The parameter attribute for training.
Y
yuyang18 已提交
1201

W
wopeizl 已提交
1202
        label(${label_type}): ${label_comment}
1203

W
wopeizl 已提交
1204 1205
    Returns:
        Variable: ${viterbi_path_comment}
Y
yi.wu 已提交
1206

W
wopeizl 已提交
1207 1208
    Examples:
        .. code-block:: python
Y
yi.wu 已提交
1209

W
wopeizl 已提交
1210 1211 1212 1213 1214 1215 1216 1217 1218 1219
           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],
1220
                "Transition": transition,
W
wopeizl 已提交
1221 1222
                "Label": label},
        outputs={"ViterbiPath": [viterbi_path]})
1223

W
wopeizl 已提交
1224
    return viterbi_path
1225 1226


Y
yi.wu 已提交
1227
@templatedoc()
F
fengjiayi 已提交
1228
def cos_sim(X, Y):
1229
    """
Y
yi.wu 已提交
1230 1231 1232
    ${comment}

    Args:
1233 1234
        X (Variable): ${x_comment}.
        Y (Variable): ${y_comment}.
F
fengjiayi 已提交
1235

Y
yi.wu 已提交
1236
    Returns:
1237
        Variable: the output of cosine(X, Y).
1238
    """
F
fengjiayi 已提交
1239
    helper = LayerHelper('cos_sim', **locals())
1240 1241 1242
    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)
1243 1244 1245 1246 1247 1248 1249 1250 1251 1252
    helper.append_op(
        type='cos_sim',
        inputs={'X': [X],
                'Y': [Y]},
        outputs={'Out': [out],
                 'XNorm': [xnorm],
                 'YNorm': [ynorm]})
    return out


1253 1254 1255 1256 1257
def dropout(x,
            dropout_prob,
            is_test=False,
            seed=None,
            name=None,
1258
            dropout_implementation="downgrade_in_infer"):
1259 1260 1261 1262 1263
    """
    Computes dropout.

    Drop or keep each element of `x` independently. Dropout is a regularization
    technique for reducing overfitting by preventing neuron co-adaption during
1264
    training. The dropout operator randomly sets (according to the given dropout
1265 1266 1267
    probability) the outputs of some units to zero, while others are remain
    unchanged.

1268 1269
    dropout op can be removed from the program to make the program more efficient.

1270
    Args:
1271 1272
        x (Variable): The input tensor variable.
        dropout_prob (float): Probability of setting units to zero.
1273 1274 1275 1276 1277 1278 1279
        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.
1280 1281
        dropout_implementation(string): ['downgrade_in_infer'(default)|'upscale_in_train']

1282
                                        1. downgrade_in_infer(default), downgrade the outcome at inference
1283 1284 1285 1286 1287 1288

                                           - train: out = input * mask
                                           - inference: out = input * dropout_prob

                                           (mask is a tensor same shape with input, value is 0 or 1
                                           ratio of 0 is dropout_prob)
1289
                                        2. upscale_in_train, upscale the outcome at training time
1290

1291 1292
                                           - train: out = input * mask / ( 1.0 - dropout_prob )
                                           - inference: out = input
1293

1294 1295
                                           (mask is a tensor same shape with input, value is 0 or 1
                                           ratio of 0 is dropout_prob)
1296

M
minqiyang 已提交
1297

1298
    Returns:
1299
        Variable: A tensor variable is the shape with `x`.
1300 1301

    Examples:
1302

1303 1304
        .. code-block:: python

1305 1306
            x = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
            droped = fluid.layers.dropout(x, dropout_prob=0.5)
1307 1308
    """

F
fengjiayi 已提交
1309
    helper = LayerHelper('dropout', **locals())
1310 1311 1312
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    mask = helper.create_variable_for_type_inference(
        dtype=x.dtype, stop_gradient=True)
1313 1314 1315 1316

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

1317 1318 1319 1320 1321
    helper.append_op(
        type='dropout',
        inputs={'X': [x]},
        outputs={'Out': [out],
                 'Mask': [mask]},
1322 1323 1324 1325
        attrs={
            'dropout_prob': dropout_prob,
            'is_test': is_test,
            'fix_seed': seed is not None,
1326 1327
            'seed': seed if seed is not None else 0,
            'dropout_implementation': dropout_implementation,
1328
        })
1329 1330 1331
    return out


J
jerrywgz 已提交
1332
def cross_entropy(input, label, soft_label=False, ignore_index=kIgnoreIndex):
1333
    """
1334 1335
    **Cross Entropy Layer**

1336 1337 1338
    This layer computes the cross entropy between `input` and `label`. It
    supports both standard cross-entropy and soft-label cross-entropy loss
    computation.
1339 1340

    1) One-hot cross-entropy:
F
fengjiayi 已提交
1341
        `soft_label = False`, `Label[i, 0]` indicates the class index for sample i:
1342

1343
        .. math::
1344

1345 1346 1347
            Y[i] = -\log(X[i, Label[i]])

    2) Soft-label cross-entropy:
F
fengjiayi 已提交
1348 1349
        `soft_label = True`, `Label[i, j]` indicates the soft label of class j
        for sample i:
1350 1351 1352 1353 1354

        .. math::

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

1355
       Please make sure that in this case the summation of each row of `label`
1356 1357 1358
       equals one.

    3) One-hot cross-entropy with vecterized `label`:
F
fengjiayi 已提交
1359 1360
         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
1361
         to a one-hot cross-entropy with one-hot label representation.
1362

1363
    Args:
1364
        input (Variable|list):  a 2-D tensor with shape [N x D], where N is the
1365 1366 1367 1368
                                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.
1369
        label (Variable|list): the ground truth which is a 2-D tensor. When
1370 1371 1372 1373
                               `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
1374
        soft_label (bool): a flag indicating whether to
1375
                                           interpretate the given labels as soft
1376
                                           labels. Default: `False`.
M
minqiyang 已提交
1377 1378
        ignore_index (int): Specifies a target value that is ignored and does
                            not contribute to the input gradient. Only valid
J
jerrywgz 已提交
1379
                            if soft_label is set to False. Default: kIgnoreIndex
1380 1381 1382 1383 1384

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

    Raises:
1385 1386 1387
         ValueError:

                      1. the 1st dimension of ``input`` and ``label`` are not equal.
M
minqiyang 已提交
1388

1389 1390
                      2. when ``soft_label == True``, and the 2nd dimension of
                         ``input`` and ``label`` are not equal.
M
minqiyang 已提交
1391

1392 1393
                      3. when ``soft_label == False``, and the 2nd dimension of
                         ``label`` is not 1.
1394 1395 1396 1397 1398 1399

    Examples:
        .. code-block:: python

          predict = fluid.layers.fc(input=net, size=classdim, act='softmax')
          cost = fluid.layers.cross_entropy(input=predict, label=label)
1400
    """
F
fengjiayi 已提交
1401
    helper = LayerHelper('cross_entropy', **locals())
1402
    out = helper.create_variable_for_type_inference(dtype=input.dtype)
1403 1404 1405 1406 1407
    helper.append_op(
        type='cross_entropy',
        inputs={'X': [input],
                'Label': [label]},
        outputs={'Y': [out]},
1408 1409
        attrs={"soft_label": soft_label,
               "ignore_index": ignore_index})
1410 1411 1412
    return out


F
frankwhzhang 已提交
1413
def bpr_loss(input, label, name=None):
1414 1415 1416
    """
    Bayesian Personalized Ranking Loss Operator.

1417
    This operator belongs to pairwise ranking loss. Label is the desired item.
1418 1419 1420 1421 1422 1423
    The loss at a given point in one session is defined as:
    $Y[i] = -\frac{1}{N_{i}-1} * \sum_{0\le j<N_{i},~ j\neq Label[i]}\log(\sigma(X[i, Label[i]]-X[i, j]))$

    Learn more details by reading paper <session-based recommendations with recurrent
    neural networks>(https://arxiv.org/abs/1511.06939)

1424 1425 1426 1427 1428 1429
    Args:
        input (Variable|list):  a 2-D tensor with shape [N x D], where N is the
                                batch size and D is the number of classes.
                                This input is not probability but logits.
        label (Variable|list):  the ground truth which is a 2-D tensor.  `label`
                                is a tensor<int64> with shape [N x 1].
F
frankwhzhang 已提交
1430 1431
        name (str|None):        A name for this layer(optional). If set None, the
                                layer will be named automatically. Default: None.
1432 1433 1434
    Returns:
        A 2-D tensor with shape [N x 1], the bpr loss.

1435 1436 1437
    Examples:
        .. code-block:: python

1438
          cost = fluid.layers.bpr_loss(input=predict, label=label)
1439
    """
1440 1441 1442 1443 1444 1445

    helper = LayerHelper('bpr_loss', **locals())
    out = helper.create_variable_for_type_inference(dtype=input.dtype)
    helper.append_op(
        type='bpr_loss',
        inputs={'X': [input],
1446
                'Label': [label]},
1447 1448 1449 1450
        outputs={'Y': [out]})
    return out


F
fengjiayi 已提交
1451
def square_error_cost(input, label):
1452
    """
1453 1454
    **Square error cost layer**

1455 1456
    This layer accepts input predictions and target label and returns the
    squared error cost.
1457

1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470
    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:
1471 1472
        input (Variable): Input tensor, has predictions.
        label (Variable): Label tensor, has target labels.
1473 1474

    Returns:
1475
        Variable: The tensor variable storing the element-wise squared error \
1476
                  difference of input and label.
1477 1478 1479 1480 1481 1482 1483 1484

    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)

1485
    """
F
fengjiayi 已提交
1486
    helper = LayerHelper('square_error_cost', **locals())
1487
    minus_out = helper.create_variable_for_type_inference(dtype=input.dtype)
1488 1489 1490 1491 1492 1493
    helper.append_op(
        type='elementwise_sub',
        inputs={'X': [input],
                'Y': [label]},
        outputs={'Out': [minus_out]})

1494
    square_out = helper.create_variable_for_type_inference(dtype=input.dtype)
1495
    helper.append_op(
F
fengjiayi 已提交
1496 1497
        type='square', inputs={'X': [minus_out]},
        outputs={'Out': [square_out]})
1498 1499 1500
    return square_out


Y
yi.wu 已提交
1501
@templatedoc()
1502 1503 1504 1505
def chunk_eval(input,
               label,
               chunk_scheme,
               num_chunk_types,
F
fengjiayi 已提交
1506
               excluded_chunk_types=None):
1507
    """
Y
yi.wu 已提交
1508
    **Chunk Evaluator**
Y
yi.wu 已提交
1509

1510
    This function computes and outputs the precision, recall and
1511
    F1-score of chunk detection.
Y
yi.wu 已提交
1512

M
minqiyang 已提交
1513
    For some basics of chunking, please refer to
1514
    `Chunking with Support Vector Machines <https://aclanthology.info/pdf/N/N01/N01-1025.pdf>`_ .
Y
yi.wu 已提交
1515 1516 1517 1518 1519 1520

    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
1521

Y
yi.wu 已提交
1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546
       ====== ====== ======  =====  ==  ============   =====  ===== =====  ==  =========
              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
1547

Y
yi.wu 已提交
1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571
       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
1572
    Args:
1573 1574 1575 1576 1577
        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
1578

Y
yi.wu 已提交
1579
    Returns:
Y
update  
yi.wu 已提交
1580 1581 1582
        tuple: tuple containing: precision, recall, f1_score,
        num_infer_chunks, num_label_chunks,
        num_correct_chunks
1583

Y
yi.wu 已提交
1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595
    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)
1596
    """
F
fengjiayi 已提交
1597
    helper = LayerHelper("chunk_eval", **locals())
1598 1599

    # prepare output
1600 1601 1602 1603 1604 1605 1606
    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")
1607 1608 1609 1610 1611 1612 1613 1614

    helper.append_op(
        type="chunk_eval",
        inputs={"Inference": [input],
                "Label": [label]},
        outputs={
            "Precision": [precision],
            "Recall": [recall],
1615 1616 1617 1618
            "F1-Score": [f1_score],
            "NumInferChunks": [num_infer_chunks],
            "NumLabelChunks": [num_label_chunks],
            "NumCorrectChunks": [num_correct_chunks]
1619 1620 1621
        },
        attrs={
            "num_chunk_types": num_chunk_types,
1622 1623
            "chunk_scheme": chunk_scheme,
            "excluded_chunk_types": excluded_chunk_types or []
1624
        })
1625 1626
    return (precision, recall, f1_score, num_infer_chunks, num_label_chunks,
            num_correct_chunks)
1627 1628


1629
@templatedoc()
1630 1631 1632 1633 1634 1635 1636
def sequence_conv(input,
                  num_filters,
                  filter_size=3,
                  filter_stride=1,
                  padding=None,
                  bias_attr=None,
                  param_attr=None,
1637 1638
                  act=None,
                  name=None):
1639 1640 1641 1642
    """
    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.
1643 1644 1645 1646 1647 1648 1649

    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.
1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662
        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
1663

1664 1665
    Returns:
        Variable: output of sequence_conv
1666 1667 1668 1669 1670 1671 1672
    """

    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)
1673
    pre_bias = helper.create_variable_for_type_inference(dtype)
1674 1675 1676 1677 1678 1679 1680 1681 1682 1683

    helper.append_op(
        type='sequence_conv',
        inputs={
            'X': [input],
            'Filter': [filter_param],
        },
        outputs={"Out": pre_bias},
        attrs={
            'contextStride': filter_stride,
1684
            'contextStart': -int(filter_size // 2),
1685 1686 1687 1688 1689 1690
            'contextLength': filter_size
        })
    pre_act = helper.append_bias_op(pre_bias)
    return helper.append_activation(pre_act)


1691
def sequence_softmax(input, use_cudnn=False, name=None):
1692 1693 1694
    """
    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
1695
    input Tensor can be either :math:`[N, 1]` or :math:`[N]`, where :math:`N`
1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711
    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 \
1712 1713 1714
            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.
1715

1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726
    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)
    """
1727 1728
    helper = LayerHelper('sequence_softmax', **locals())
    dtype = helper.input_dtype()
1729
    softmax_out = helper.create_variable_for_type_inference(dtype)
1730 1731 1732 1733 1734 1735 1736 1737
    helper.append_op(
        type="sequence_softmax",
        inputs={"X": input},
        outputs={"Out": softmax_out},
        attrs={"use_cudnn": use_cudnn})
    return softmax_out


1738
def softmax(input, use_cudnn=True, name=None):
Q
qiaolongfei 已提交
1739
    """
1740
    The input of the softmax operator is a tensor of any rank. The output tensor
1741
    has the same shape as the input.
Q
qiaolongfei 已提交
1742

1743 1744 1745 1746 1747 1748
    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
1749
    K-dimensional vector of real values in the range [0, 1] that add up to 1.
Q
qiaolongfei 已提交
1750 1751 1752 1753 1754 1755 1756

    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.

1757
    For each row :math:`i` and each column :math:`j` in the matrix, we have:
Q
qiaolongfei 已提交
1758 1759 1760 1761 1762 1763 1764 1765

    .. 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 \
1766 1767 1768
            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
1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780

    Returns:
        Variable: output of softmax

    Examples:

        .. code-block:: python

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

    """
1781 1782
    helper = LayerHelper('softmax', **locals())
    dtype = helper.input_dtype()
1783
    softmax_out = helper.create_variable_for_type_inference(dtype)
1784 1785 1786 1787 1788 1789 1790 1791
    helper.append_op(
        type="softmax",
        inputs={"X": input},
        outputs={"Out": softmax_out},
        attrs={"use_cudnn": use_cudnn})
    return softmax_out


1792 1793 1794
def conv2d(input,
           num_filters,
           filter_size,
C
chengduoZH 已提交
1795 1796
           stride=1,
           padding=0,
1797
           dilation=1,
1798 1799 1800
           groups=None,
           param_attr=None,
           bias_attr=None,
C
chengduoZH 已提交
1801
           use_cudnn=True,
1802 1803
           act=None,
           name=None):
1804
    """
C
chengduoZH 已提交
1805
    The convolution2D layer calculates the output based on the input, filter
T
tensor-tang 已提交
1806 1807
    and strides, paddings, dilations, groups parameters. Input and
    Output are in NCHW format, where N is batch size, C is the number of
1808
    channels, H is the height of the feature, and W is the width of the feature.
T
tensor-tang 已提交
1809 1810 1811 1812 1813 1814 1815
    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.
1816 1817 1818
    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
1819

1820
    For each input :math:`X`, the equation is:
C
refine  
chengduoZH 已提交
1821

C
chengduoZH 已提交
1822 1823
    .. math::

C
refine  
chengduoZH 已提交
1824
        Out = \sigma (W \\ast X + b)
C
chengduoZH 已提交
1825

T
tensor-tang 已提交
1826
    Where:
C
chengduoZH 已提交
1827

1828 1829 1830 1831 1832
    * :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
1833
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
1834 1835 1836

    Example:

1837 1838
        - Input:

W
weixing02 已提交
1839
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
C
refine  
chengduoZH 已提交
1840

W
weixing02 已提交
1841
          Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)`
C
refine  
chengduoZH 已提交
1842

1843
        - Output:
T
tensor-tang 已提交
1844

W
weixing02 已提交
1845
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
C
refine  
chengduoZH 已提交
1846

C
chengduoZH 已提交
1847
        Where
1848 1849

        .. math::
C
chengduoZH 已提交
1850

W
weixing02 已提交
1851 1852
            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
1853 1854

    Args:
1855
        input (Variable): The input image with [N, C, H, W] format.
T
tensor-tang 已提交
1856
        num_filters(int): The number of filter. It is as same as the output
1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873
            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
1874 1875 1876 1877 1878
            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)`,
1879
            and the :math:`std` is :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None.
1880 1881 1882 1883 1884
        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.
1885 1886
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
1887 1888
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None
1889
        name (str|None): A name for this layer(optional). If set None, the layer
1890
            will be named automatically. Default: None
C
chengduoZH 已提交
1891 1892

    Returns:
1893
        Variable: The tensor variable storing the convolution and \
C
chengduoZH 已提交
1894 1895
                  non-linearity activation result.

C
refine  
chengduoZH 已提交
1896
    Raises:
1897 1898
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
C
refine  
chengduoZH 已提交
1899

C
chengduoZH 已提交
1900 1901 1902
    Examples:
        .. code-block:: python

1903 1904
          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")
1905 1906 1907
    """

    num_channels = input.shape[1]
1908
    assert param_attr is not False, "param_attr should not be False here."
1909
    l_type = 'conv2d'
1910 1911
    if (num_channels == groups and num_filters % num_channels == 0 and
            not use_cudnn):
1912
        l_type = 'depthwise_conv2d'
1913 1914 1915 1916

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

1917 1918 1919 1920 1921
    if groups is None:
        num_filter_channels = num_channels
    else:
        if num_channels % groups != 0:
            raise ValueError("num_channels must be divisible by groups.")
1922
        num_filter_channels = num_channels // groups
1923

C
chengduoZH 已提交
1924 1925 1926
    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')
1927
    dilation = utils.convert_to_list(dilation, 2, 'dilation')
C
chengduoZH 已提交
1928

C
chengduoZH 已提交
1929 1930
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
1931 1932

    input_shape = input.shape
M
minqiyang 已提交
1933
    filter_shape = [num_filters, int(num_filter_channels)] + filter_size
1934 1935

    def _get_default_param_initializer():
1936 1937
        filter_elem_num = filter_size[0] * filter_size[1] * num_channels
        std = (2.0 / filter_elem_num)**0.5
1938 1939 1940 1941 1942 1943 1944 1945
        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())

1946
    pre_bias = helper.create_variable_for_type_inference(dtype)
1947

1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961
    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)

1962
    helper.append_op(
1963
        type=l_type,
1964 1965 1966 1967 1968
        inputs={
            'Input': input,
            'Filter': filter_param,
        },
        outputs={"Output": pre_bias},
C
chengduoZH 已提交
1969 1970 1971
        attrs={
            'strides': stride,
            'paddings': padding,
1972
            'dilations': dilation,
C
chengduoZH 已提交
1973
            'groups': groups,
1974
            'use_cudnn': use_cudnn,
1975
            'use_mkldnn': False,
1976
            'fuse_relu_before_depthwise_conv': False
C
chengduoZH 已提交
1977
        })
1978 1979 1980 1981 1982 1983

    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
1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000
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
2001 2002 2003 2004 2005 2006
    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
2007 2008 2009 2010 2011 2012 2013 2014 2015

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

    .. math::

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

    In the above equation:

2016 2017
    * :math:`X`: Input value, a tensor with NCDHW format.
    * :math:`W`: Filter value, a tensor with MCDHW format.
C
chengduoZH 已提交
2018 2019 2020
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
2021
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046

    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,
2047
            it must contain three integers, (filter_size_D, filter_size_H, filter_size_W).
C
chengduoZH 已提交
2048 2049
            Otherwise, the filter will be a square.
        stride (int|tuple): The stride size. If stride is a tuple, it must
2050
            contain three integers, (stride_D, stride_H, stride_W). Otherwise, the
C
chengduoZH 已提交
2051 2052
            stride_D = stride_H = stride_W = stride. Default: stride = 1.
        padding (int|tuple): The padding size. If padding is a tuple, it must
2053
            contain three integers, (padding_D, padding_H, padding_W). Otherwise, the
C
chengduoZH 已提交
2054 2055
            padding_D = padding_H = padding_W = padding. Default: padding = 0.
        dilation (int|tuple): The dilation size. If dilation is a tuple, it must
2056
            contain three integers, (dilation_D, dilation_H, dilation_W). Otherwise, the
C
chengduoZH 已提交
2057 2058 2059 2060 2061 2062
            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
2063 2064 2065 2066 2067 2068 2069 2070 2071 2072
        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
2073 2074
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
2075 2076
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None.
C
chengduoZH 已提交
2077
        name (str|None): A name for this layer(optional). If set None, the layer
2078
            will be named automatically. Default: None.
C
chengduoZH 已提交
2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090

    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

2091 2092
          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
2093 2094 2095
    """

    l_type = 'conv3d'
2096
    assert param_attr is not False, "param_attr should not be False here."
C
chengduoZH 已提交
2097 2098 2099 2100 2101 2102 2103 2104 2105 2106
    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.")
2107
        num_filter_channels = num_channels // groups
C
chengduoZH 已提交
2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120

    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():
2121 2122 2123
        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
2124 2125 2126 2127 2128 2129 2130 2131
        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())

2132
    pre_bias = helper.create_variable_for_type_inference(dtype)
C
chengduoZH 已提交
2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146

    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
2147
            'use_mkldnn': False
C
chengduoZH 已提交
2148 2149
        })

2150
    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
C
chengduoZH 已提交
2151 2152 2153 2154

    return helper.append_activation(pre_act)


2155
def sequence_pool(input, pool_type, is_test=False):
2156
    """
2157 2158 2159
    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.
2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170

    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:
2171
         x.lod = [[2, 3, 2]]
2172 2173 2174 2175 2176
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
2177
         with condition len(x.lod[-1]) == out.dims[0]
2178 2179 2180 2181 2182 2183 2184

       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)
2185 2186
         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
2187

2188 2189
    Args:
        input(variable): The input variable which is a LoDTensor.
2190
        pool_type (string): The pooling type of sequence_pool.
2191
            It supports average, sum, sqrt and max.
2192
        is_test(bool, Default False): Used distinguish training from scoring mode.
2193 2194 2195 2196 2197 2198 2199

    Returns:
        The sequence pooling variable which is a Tensor.

    Examples:

        .. code-block:: python
F
fengjiayi 已提交
2200

2201
             x = fluid.layers.data(name='x', shape=[7, 1],
2202 2203 2204 2205 2206
                              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')
2207 2208
             last_x = fluid.layers.sequence_pool(input=x, pool_type='last')
             first_x = fluid.layers.sequence_pool(input=x, pool_type='first')
2209
    """
F
fengjiayi 已提交
2210
    helper = LayerHelper('sequence_pool', **locals())
2211
    dtype = helper.input_dtype()
2212 2213
    pool_out = helper.create_variable_for_type_inference(dtype)
    max_index = helper.create_variable_for_type_inference(dtype)
2214 2215 2216 2217 2218 2219

    helper.append_op(
        type="sequence_pool",
        inputs={"X": input},
        outputs={"Out": pool_out,
                 "MaxIndex": max_index},
2220 2221
        attrs={"pooltype": pool_type.upper(),
               "is_test": is_test})
2222

2223 2224 2225 2226 2227
    # 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

2228 2229 2230
    return pool_out


C
add doc  
chengduoZH 已提交
2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249
@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())
2250
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
C
add doc  
chengduoZH 已提交
2251 2252 2253 2254 2255
    helper.append_op(
        type='sequence_concat', inputs={'X': input}, outputs={'Out': [out]})
    return out


F
fengjiayi 已提交
2256
def sequence_first_step(input):
2257
    """
L
Luo Tao 已提交
2258
    This function gets the first step of sequence.
2259 2260 2261 2262

    .. code-block:: text

       x is a 1-level LoDTensor:
2263
         x.lod = [[2, 3, 2]]
2264 2265 2266 2267 2268
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
2269
         with condition len(x.lod[-1]) == out.dims[0]
2270
         out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1)
F
fengjiayi 已提交
2271

2272 2273 2274 2275 2276 2277 2278 2279 2280
    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
2281

2282
             x = fluid.layers.data(name='x', shape=[7, 1],
2283 2284 2285
                              dtype='float32', lod_level=1)
             x_first_step = fluid.layers.sequence_first_step(input=x)
    """
2286 2287 2288
    return sequence_pool(input=input, pool_type="first")


F
fengjiayi 已提交
2289
def sequence_last_step(input):
2290
    """
L
Luo Tao 已提交
2291
    This function gets the last step of sequence.
2292 2293 2294 2295

    .. code-block:: text

       x is a 1-level LoDTensor:
2296
         x.lod = [[2, 3, 2]]
2297 2298 2299 2300 2301
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
2302
         with condition len(x.lod[-1]) == out.dims[0]
2303
         out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1)
F
fengjiayi 已提交
2304

2305 2306 2307 2308 2309 2310 2311 2312 2313
    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
2314

2315
             x = fluid.layers.data(name='x', shape=[7, 1],
2316 2317 2318
                              dtype='float32', lod_level=1)
             x_last_step = fluid.layers.sequence_last_step(input=x)
    """
2319 2320 2321
    return sequence_pool(input=input, pool_type="last")


2322 2323 2324 2325
def sequence_slice(input, offset, length, name=None):
    """
    **Sequence Slice Layer**

2326
    The layer crops a subsequence from given sequence with given start
2327 2328 2329 2330 2331
    offset and subsequence length.

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

    .. code-block:: text
2332

2333
              - Case:
2334

2335
            Given the input Variable **input**:
2336

2337 2338 2339
                input.data = [[a1, a2], [b1, b2], [c1, c2], [d1, d2], [e1, e2]],
                input.lod = [[3, 2]],
                input.dims = (5, 2),
2340

2341
            with offset.data = [[0], [1]] and length.data = [[2], [1]],
2342

2343
            the output Variable will be
2344

2345 2346 2347
                out.data = [[a1, a2], [b1, b2], [e1, e2]],
                out.lod = [[2, 1]],
                out.dims = (3, 2).
2348

M
minqiyang 已提交
2349
    Note:
2350
          The first dimension size of **input**, **offset** and **length**
2351
          should be equal. The **offset** should start from 0.
2352

2353
    Args:
2354
        input(Variable): The input Variable which consists of the complete
Y
Yibing Liu 已提交
2355
                         sequences.
2356 2357 2358 2359 2360 2361
        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
2362
        Variable: The output subsequences.
2363 2364 2365 2366 2367 2368 2369 2370 2371 2372

    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"))
2373
             subseqs = fluid.layers.sequence_slice(input=seqs, offset=offset,
2374 2375 2376 2377
                                                   length=length)
    """
    helper = LayerHelper("sequence_slice", **locals())
    dtype = helper.input_dtype()
2378
    out = helper.create_variable_for_type_inference(dtype)
2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392

    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
2393
@templatedoc()
2394
def pool2d(input,
C
chengduoZH 已提交
2395 2396
           pool_size=-1,
           pool_type="max",
C
chengduoZH 已提交
2397 2398
           pool_stride=1,
           pool_padding=0,
2399
           global_pooling=False,
C
chengduoZH 已提交
2400
           use_cudnn=True,
2401
           ceil_mode=False,
2402 2403
           name=None,
           exclusive=True):
2404
    """
F
fengjiayi 已提交
2405
    ${comment}
2406 2407

    Args:
2408 2409 2410
        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
2411
                          feature, and W is the width of the feature.
J
JiabinYang 已提交
2412
        pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list,
J
JiabinYang 已提交
2413 2414
            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
2415
        pool_type: ${pooling_type_comment}
J
JiabinYang 已提交
2416 2417 2418 2419 2420 2421
        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.
2422 2423 2424
        global_pooling (bool): ${global_pooling_comment}
        use_cudnn (bool): ${use_cudnn_comment}
        ceil_mode (bool): ${ceil_mode_comment}
2425
        name (str|None): A name for this layer(optional). If set None, the
F
fengjiayi 已提交
2426
                        layer will be named automatically.
2427
        exclusive (bool): Whether to exclude padding points in average pooling
2428
                          mode, default is true
F
fengjiayi 已提交
2429

2430
    Returns:
F
fengjiayi 已提交
2431
        Variable: The pooling result.
F
fengjiayi 已提交
2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444

    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(
2445 2446 2447 2448
                            input=data,
                            pool_size=2,
                            pool_type='max',
                            pool_stride=1,
F
fengjiayi 已提交
2449
                            global_pooling=False)
2450 2451 2452 2453 2454
    """
    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
2455

C
chengduoZH 已提交
2456 2457 2458 2459 2460
    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
2461 2462 2463 2464
    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
2465 2466
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
2467

C
Add doc  
chengduoZH 已提交
2468
    l_type = 'pool2d'
2469 2470

    helper = LayerHelper(l_type, **locals())
2471
    dtype = helper.input_dtype()
2472
    pool_out = helper.create_variable_for_type_inference(dtype)
2473 2474

    helper.append_op(
2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485
        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,
2486 2487
            "use_mkldnn": False,
            "exclusive": exclusive,
2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500
        })

    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,
2501 2502
           name=None,
           exclusive=True):
2503 2504
    """
    This function adds the operator for pooling in 3-dimensions, using the
2505
    pooling configurations mentioned in input parameters.
2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517

    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.
2518
        exclusive (bool): Whether to exclude padding points in average pooling
2519
                          mode, default is true
2520

2521
    Returns:
2522
        Variable: output of pool3d layer.
2523 2524 2525 2526 2527
    """
    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
2528

C
chengduoZH 已提交
2529 2530 2531 2532 2533
    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))

2534 2535 2536
    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
2537

C
chengduoZH 已提交
2538 2539
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
2540

2541 2542
    l_type = "pool3d"
    helper = LayerHelper(l_type, **locals())
2543
    dtype = helper.input_dtype()
2544
    pool_out = helper.create_variable_for_type_inference(dtype)
2545 2546

    helper.append_op(
2547
        type=l_type,
2548 2549 2550 2551 2552 2553 2554
        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
2555
            "paddings": pool_padding,
2556
            "use_cudnn": use_cudnn,
2557
            "ceil_mode": ceil_mode,
2558 2559
            "use_mkldnn": False,
            "exclusive": exclusive,
2560 2561 2562 2563 2564
        })

    return pool_out


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
@templatedoc(op_type="pool2d")
def adaptive_pool2d(input,
                    pool_size,
                    pool_type="max",
                    require_index=False,
                    name=None):
    """
    ${comment}

    Args:
        input (Variable): The input tensor of pooling operator. The format of
                          input tensor is NCHW, where N is batch size, C is
                          the number of channels, H is the height of the
                          feature, and W is the width of the feature.
        pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list,
            it must contain two integers, (pool_size_Height, pool_size_Width).
        pool_type: ${pooling_type_comment}
        require_index (bool): If true, the index of max pooling point along with outputs.
            it cannot be set in average pooling type.
        name (str|None): A name for this layer(optional). If set None, the
                        layer will be named automatically.

    Returns:
        Variable: The pooling result.

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

    Examples:
        .. code-block:: python

M
minqiyang 已提交
2598
          # suppose input data in shape of [N, C, H, W], `pool_size` is [m, n],
2599
          # output shape is [N, C, m, n], adaptive pool divide H and W dimentions
M
minqiyang 已提交
2600
          # of input data into m * n grids averagely and performs poolings in each
2601 2602
          # grid to get output.
          # adaptive average pool performs calculations as follow:
M
minqiyang 已提交
2603
          #
2604 2605 2606 2607 2608 2609 2610 2611
          #     for i in range(m):
          #         for j in range(n):
          #             hstart = floor(i * H / m)
          #             hend = ceil((i + 1) * H / m)
          #             wstart = floor(i * W / n)
          #             wend = ceil((i + 1) * W / n)
          #             output[:, :, i, j] = avg(input[:, :, hstart: hend, wstart: wend])
          #
2612 2613
          data = fluid.layers.data(
              name='data', shape=[3, 32, 32], dtype='float32')
2614
          pool_out = fluid.layers.adaptive_pool2d(
2615 2616
                            input=data,
                            pool_size=[3, 3],
2617
                            pool_type='avg')
2618 2619 2620 2621 2622 2623 2624 2625 2626 2627
    """
    if pool_type not in ["max", "avg"]:
        raise ValueError(
            "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",
            str(pool_type))

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

2628
    pool_size = utils.convert_to_list(pool_size, 2, 'pool_size')
2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653

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

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

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

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

2654
    return (pool_out, mask) if require_index else pool_out
2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689


@templatedoc(op_type="pool3d")
def adaptive_pool3d(input,
                    pool_size,
                    pool_type="max",
                    require_index=False,
                    name=None):
    """
    ${comment}

    Args:
        input (Variable): The input tensor of pooling operator. The format of
                          input tensor is NCHW, where N is batch size, C is
                          the number of channels, H is the height of the
                          feature, and W is the width of the feature.
        pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list,
            it must contain two integers, (Depth, Height, Width).
        pool_type: ${pooling_type_comment}
        require_index (bool): If true, the index of max pooling point along with outputs.
            it cannot be set in average pooling type.
        name (str|None): A name for this layer(optional). If set None, the
                        layer will be named automatically.

    Returns:
        Variable: The pooling result.

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

    Examples:
        .. code-block:: python

2690 2691
          # suppose input data in shape of [N, C, D, H, W], `pool_size` is [l, m, n],
          # output shape is [N, C, l, m, n], adaptive pool divide D, H and W dimentions
M
minqiyang 已提交
2692
          # of input data into l * m * n grids averagely and performs poolings in each
2693 2694
          # grid to get output.
          # adaptive average pool performs calculations as follow:
M
minqiyang 已提交
2695
          #
2696 2697 2698 2699 2700 2701 2702 2703 2704
          #     for i in range(l):
          #         for j in range(m):
          #             for k in range(n):
          #                 dstart = floor(i * D / l)
          #                 dend = ceil((i + 1) * D / l)
          #                 hstart = floor(j * H / m)
          #                 hend = ceil((j + 1) * H / m)
          #                 wstart = floor(k * W / n)
          #                 wend = ceil((k + 1) * W / n)
M
minqiyang 已提交
2705
          #                 output[:, :, i, j, k] =
2706 2707
          #                     avg(input[:, :, dstart:dend, hstart: hend, wstart: wend])
          #
2708 2709
          data = fluid.layers.data(
              name='data', shape=[3, 32, 32], dtype='float32')
2710
          pool_out, mask = fluid.layers.adaptive_pool3d(
2711 2712
                            input=data,
                            pool_size=[3, 3],
2713
                            pool_type='avg')
2714 2715 2716 2717 2718 2719 2720 2721 2722 2723
    """
    if pool_type not in ["max", "avg"]:
        raise ValueError(
            "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",
            str(pool_type))

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

2724
    pool_size = utils.convert_to_list(pool_size, 3, 'pool_size')
2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749

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

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

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

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

2750
    return (pool_out, mask) if require_index else pool_out
2751 2752


2753 2754 2755 2756 2757 2758 2759
def batch_norm(input,
               act=None,
               is_test=False,
               momentum=0.9,
               epsilon=1e-05,
               param_attr=None,
               bias_attr=None,
2760
               data_layout='NCHW',
Y
Yang Yang 已提交
2761
               in_place=False,
2762 2763
               name=None,
               moving_mean_name=None,
2764
               moving_variance_name=None,
2765
               do_model_average_for_mean_and_var=False,
2766 2767
               fuse_with_relu=False,
               use_global_stats=False):
2768
    """
Q
qiaolongfei 已提交
2769 2770 2771 2772
    **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
2773

Q
qiaolongfei 已提交
2774
    1. NHWC `[batch, in_height, in_width, in_channels]`
Q
qiaolongfei 已提交
2775

Q
qiaolongfei 已提交
2776 2777
    2. NCHW `[batch, in_channels, in_height, in_width]`

Q
qiaolongfei 已提交
2778 2779 2780
    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
2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792

    :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
2793

2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806

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

    ..  math::

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

2807
    Args:
Q
qiaolongfei 已提交
2808
        input(variable): The input variable which is a LoDTensor.
Q
qiaolongfei 已提交
2809 2810 2811 2812
        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):
2813 2814 2815 2816 2817 2818 2819 2820
        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
2821
        data_layout(string, default NCHW): NCHW|NHWC
2822
        in_place(bool, Default False): Make the input and output of batch norm reuse memory.
Q
qiaolongfei 已提交
2823 2824 2825 2826
        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
2827
        do_model_average_for_mean_and_var(bool, Default False): Do model average for mean and variance or not.
2828
        fuse_with_relu (bool): if True, this OP performs relu after batch norm.
2829 2830 2831 2832 2833
        use_global_stats(bool, Default False): Whether to use global mean and
            variance. In inference or test mode, set use_global_stats to true
            or is_test to true, and the behavior is equivalent.
            In train mode, when setting use_global_stats True, the global mean
            and variance are also used during train period.
2834 2835

    Returns:
Q
qiaolongfei 已提交
2836
        Variable: A tensor variable which is the result after applying batch normalization on the input.
Q
qiaolongfei 已提交
2837 2838 2839 2840 2841 2842 2843

    Examples:

        .. code-block:: python

            hidden1 = fluid.layers.fc(input=x, size=200, param_attr='fc1.w')
            hidden2 = fluid.layers.batch_norm(input=hidden1)
2844
    """
2845
    assert bias_attr is not False, "bias_attr should not be False in batch_norm."
2846 2847 2848
    helper = LayerHelper('batch_norm', **locals())
    dtype = helper.input_dtype()

W
Wu Yi 已提交
2849 2850 2851 2852
    # use fp32 for bn parameter
    if dtype == core.VarDesc.VarType.FP16:
        dtype = core.VarDesc.VarType.FP32

2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869
    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))
2870 2871 2872
    # setting stop_gradient=True to reduce computation
    if use_global_stats and helper.param_attr.learning_rate == 0.:
        scale.stop_gradient = True
2873 2874

    bias = helper.create_parameter(
2875
        attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
2876 2877
    # setting stop_gradient=True to reduce computation
    if use_global_stats and helper.bias_attr.learning_rate == 0.:
2878
        bias.stop_gradient = True
2879

2880 2881
    mean = helper.create_parameter(
        attr=ParamAttr(
2882 2883 2884
            name=moving_mean_name,
            initializer=Constant(0.0),
            trainable=False,
W
wanghaoshuang 已提交
2885
            do_model_average=do_model_average_for_mean_and_var),
2886
        shape=param_shape,
W
Wu Yi 已提交
2887
        dtype=dtype)
2888 2889 2890 2891 2892 2893
    mean.stop_gradient = True

    variance = helper.create_parameter(
        attr=ParamAttr(
            name=moving_variance_name,
            initializer=Constant(1.0),
2894
            trainable=False,
W
wanghaoshuang 已提交
2895
            do_model_average=do_model_average_for_mean_and_var),
2896
        shape=param_shape,
W
Wu Yi 已提交
2897
        dtype=dtype)
2898
    variance.stop_gradient = True
2899 2900 2901 2902 2903 2904

    # create output
    # mean and mean_out share the same memory
    mean_out = mean
    # variance and variance out share the same memory
    variance_out = variance
2905 2906 2907 2908
    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)
2909

2910 2911
    batch_norm_out = input if in_place else helper.create_variable_for_type_inference(
        dtype)
2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928

    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
        },
2929 2930 2931 2932
        attrs={
            "momentum": momentum,
            "epsilon": epsilon,
            "is_test": is_test,
2933
            "data_layout": data_layout,
X
Xin Pan 已提交
2934
            "use_mkldnn": False,
2935 2936
            "fuse_with_relu": fuse_with_relu,
            "use_global_stats": use_global_stats
2937
        })
2938 2939 2940 2941

    return helper.append_activation(batch_norm_out)


H
heqiaozhi 已提交
2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068
def data_norm(input,
              act=None,
              epsilon=1e-05,
              param_attr=None,
              data_layout='NCHW',
              in_place=False,
              use_mkldnn=False,
              name=None,
              moving_mean_name=None,
              moving_variance_name=None,
              do_model_average_for_mean_and_var=False):
    """
    **Data Normalization Layer**

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

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

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

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

    ..  math::

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

    Args:
        input(variable): The input variable which is a LoDTensor.
        act(string, Default None): Activation type, linear|relu|prelu|...
        epsilon(float, Default 1e-05):
        param_attr(ParamAttr): The parameter attribute for Parameter `scale`.
        data_layout(string, default NCHW): NCHW|NHWC
        in_place(bool, Default False): Make the input and output of batch norm reuse memory.
        use_mkldnn(bool, Default false): ${use_mkldnn_comment}
        name(string, Default None): A name for this layer(optional). If set None, the layer
            will be named automatically.
        moving_mean_name(string, Default None): The name of moving_mean which store the global Mean.
        moving_variance_name(string, Default None): The name of the moving_variance which store the global Variance.
        do_model_average_for_mean_and_var(bool, Default False): Do model average for mean and variance or not.

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

    Examples:

        .. code-block:: python

            data = fluid.layers.data(input=x, size=200, param_attr='fc1.w')
            hidden2 = fluid.layers.data_norm(input=hidden1)
    """
    helper = LayerHelper('data_norm', **locals())
    dtype = helper.input_dtype()

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

    param_shape = [channel_num]

    batch_size_default = 1e4
    batch_sum_default = 0.0
    batch_square_sum_default = 1e4

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

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

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

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

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

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

    helper.append_op(
        type="data_norm",
        inputs={
            "X": input,
            "BatchSize": batch_size,
            "BatchSum": batch_sum,
            "BatchSquareSum": batch_square_sum
        },
        outputs={"Y": data_norm_out,
                 "Means": means,
                 "Scales": scales},
        attrs={"epsilon": epsilon,
               "use_mkldnn": use_mkldnn})

    return helper.append_activation(data_norm_out)


Y
yuyang18 已提交
3069
@templatedoc()
3070 3071 3072 3073 3074 3075 3076 3077 3078 3079
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
3080
    ${comment}
3081 3082 3083

    The formula is as follows:

Y
yuyang18 已提交
3084
    ..  math::
3085 3086 3087 3088 3089 3090 3091

        \\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
3092 3093 3094 3095 3096 3097 3098 3099
    * :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
3100

3101 3102
    Args:
        input(Variable): The input tensor variable.
3103
        scale(bool): Whether to learn the adaptive gain :math:`g` after
S
sneaxiy 已提交
3104
            normalization. Default True.
3105
        shift(bool): Whether to learn the adaptive bias :math:`b` after
S
sneaxiy 已提交
3106 3107
            normalization. Default True.
        begin_norm_axis(int): The normalization will be performed along
3108
            dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`.
S
sneaxiy 已提交
3109
            Default 1.
3110
        epsilon(float): The small value added to the variance to prevent
S
sneaxiy 已提交
3111
            division by zero. Default 1e-05.
3112
        param_attr(ParamAttr|None): The parameter attribute for the learnable
S
sneaxiy 已提交
3113 3114
            gain :math:`g`. If :attr:`scale` is False, :attr:`param_attr` is
            omitted. If :attr:`scale` is True and :attr:`param_attr` is None,
3115 3116
            a default :code:`ParamAttr` would be added as scale. The
            :attr:`param_attr` is initialized as 1 if it is added. Default None.
3117
        bias_attr(ParamAttr|None): The parameter attribute for the learnable
S
sneaxiy 已提交
3118 3119
            bias :math:`b`. If :attr:`shift` is False, :attr:`bias_attr` is
            omitted. If :attr:`shift` is True and :attr:`param_attr` is None,
3120
            a default :code:`ParamAttr` would be added as bias. The
S
sneaxiy 已提交
3121
            :attr:`bias_attr` is initialized as 0 if it is added. Default None.
3122
        act(str): Activation to be applied to the output of layer normalizaiton.
S
sneaxiy 已提交
3123 3124 3125
                  Default None.
        name(str): The name of this layer. It is optional. Default None, and a
                   unique name would be generated automatically.
3126 3127

    Returns:
Y
yuyang18 已提交
3128
        ${y_comment}
3129 3130 3131

    Examples:

Y
yuyang18 已提交
3132 3133 3134
        >>> data = fluid.layers.data(name='data', shape=[3, 32, 32],
        >>>                          dtype='float32')
        >>> x = fluid.layers.layer_norm(input=data, begin_norm_axis=1)
3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149
    """
    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
3150
    if shift:
3151 3152 3153 3154 3155 3156
        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
3157 3158 3159 3160 3161
    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)
3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176

    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
3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188
@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**

3189
    Refer to `Group Normalization <https://arxiv.org/abs/1803.08494>`_ .
D
Dun 已提交
3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236

    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
H
heqiaozhi 已提交
3237 3238
    mean_out = helper.create_variable(dtype=dtype, stop_gradient=True)
    variance_out = helper.create_variable(dtype=dtype, stop_gradient=True)
D
Dun 已提交
3239
    group_norm_out = helper.create_variable(dtype=dtype)
D
Dun 已提交
3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251 3252 3253 3254

    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)


3255 3256 3257 3258
def conv2d_transpose(input,
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
3259 3260 3261
                     padding=0,
                     stride=1,
                     dilation=1,
3262
                     groups=None,
3263
                     param_attr=None,
3264
                     bias_attr=None,
C
chengduoZH 已提交
3265
                     use_cudnn=True,
3266
                     act=None,
3267
                     name=None):
3268
    """
3269 3270 3271 3272 3273 3274 3275 3276
    **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
3277 3278
    layer, please refer to the following explanation and references
    `therein <http://www.matthewzeiler.com/wp-content/uploads/2017/07/cvpr2010.pdf>`_.
3279 3280 3281
    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.
3282 3283 3284 3285 3286

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

    .. math::

3287
        Out = \sigma (W \\ast X + b)
3288

3289
    Where:
3290 3291 3292

    * :math:`X`: Input value, a tensor with NCHW format.
    * :math:`W`: Filter value, a tensor with MCHW format.
3293 3294 3295 3296
    * :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.
3297

3298 3299 3300 3301
    Example:

        - Input:

3302
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
3303

3304
          Filter shape: :math:`(C_{in}, C_{out}, H_f, W_f)`
3305 3306 3307

        - Output:

3308
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
3309 3310

        Where
3311

3312 3313
        .. math::

3314 3315
           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 \\\\
3316 3317
           H_{out} &\in [ H^\prime_{out}, H^\prime_{out} + strides[0] ) \\\\
           W_{out} &\in [ W^\prime_{out}, W^\prime_{out} + strides[1] )
3318 3319

    Args:
3320 3321 3322 3323
        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
3324 3325 3326 3327
            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.
3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345
        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.
3346 3347 3348 3349 3350 3351 3352 3353 3354 3355
            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.
3356
        use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
3357 3358 3359
            library is installed. Default: True.
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None.
3360
        name(str|None): A name for this layer(optional). If set None, the layer
3361
            will be named automatically. Default: True.
3362 3363

    Returns:
3364
        Variable: The tensor variable storing the convolution transpose result.
3365 3366

    Raises:
3367 3368
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
3369 3370 3371 3372

    Examples:
       .. code-block:: python

3373 3374
          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)
3375
    """
3376
    assert param_attr is not False, "param_attr should not be False in conv2d_transpose."
3377 3378 3379 3380 3381 3382 3383 3384
    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())
3385 3386 3387
    if not isinstance(input, Variable):
        raise TypeError("Input of conv2d_transpose must be Variable")

C
chengduoZH 已提交
3388 3389 3390
    padding = utils.convert_to_list(padding, 2, 'padding')
    stride = utils.convert_to_list(stride, 2, 'stride')
    dilation = utils.convert_to_list(dilation, 2, 'dilation')
3391

C
chengduoZH 已提交
3392 3393
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
3394

3395 3396 3397 3398 3399
    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]
3400

3401 3402
        h_in = input.shape[2]
        w_in = input.shape[3]
3403

C
chengduoZH 已提交
3404
        filter_size_h = (output_size[0] - (h_in - 1) * stride[0] + 2 *
3405
                         padding[0] - 1) // dilation[0] + 1
C
chengduoZH 已提交
3406
        filter_size_w = (output_size[1] - (w_in - 1) * stride[1] + 2 *
3407
                         padding[1] - 1) // dilation[1] + 1
3408
        filter_size = [filter_size_h, filter_size_w]
C
chengduoZH 已提交
3409 3410 3411
    else:
        filter_size = utils.convert_to_list(filter_size, 2,
                                            'conv2d_transpose.filter_size')
3412

3413 3414 3415 3416 3417 3418 3419
    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')
3420
    groups = 1 if groups is None else groups
3421
    filter_shape = [input_channel, num_filters // groups] + filter_size
3422

3423 3424 3425
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

3426
    pre_bias = helper.create_variable_for_type_inference(dtype=input.dtype)
3427
    helper.append_op(
3428
        type=op_type,
3429 3430
        inputs={'Input': [input],
                'Filter': [img_filter]},
3431
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
3432
        attrs={
3433
            'output_size': output_size,
3434 3435 3436 3437 3438
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn
3439 3440
        })

3441 3442 3443
    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
    out = helper.append_activation(pre_act)
    return out
3444 3445


3446
def conv3d_transpose(input,
3447 3448 3449
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
3450 3451 3452
                     padding=0,
                     stride=1,
                     dilation=1,
3453
                     groups=None,
3454
                     param_attr=None,
3455
                     bias_attr=None,
C
chengduoZH 已提交
3456
                     use_cudnn=True,
3457
                     act=None,
3458
                     name=None):
3459
    """
3460
    **Convlution3D transpose layer**
3461

3462
    The convolution3D transpose layer calculates the output based on the input,
3463
    filter, and dilations, strides, paddings. Input(Input) and output(Output)
3464 3465 3466 3467 3468 3469
    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>`_.
3470 3471 3472
    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.
3473 3474 3475 3476 3477

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

    .. math::

3478
        Out = \sigma (W \\ast X + b)
3479 3480 3481

    In the above equation:

3482 3483
    * :math:`X`: Input value, a tensor with NCDHW format.
    * :math:`W`: Filter value, a tensor with MCDHW format.
3484 3485 3486 3487
    * :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.
3488

3489 3490 3491 3492
    Example:

        - Input:

3493
          Input shape: :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})`
3494

3495
          Filter shape: :math:`(C_{in}, C_{out}, D_f, H_f, W_f)`
3496 3497 3498

        - Output:

3499
          Output shape: :math:`(N, C_{out}, D_{out}, H_{out}, W_{out})`
3500 3501

        Where
3502

3503 3504
        .. math::

3505 3506 3507
           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
3508 3509

    Args:
3510
        input(Variable): The input image with [N, C, D, H, W] format.
3511 3512 3513
        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
3514
            tuple, it must contain three integers, (image_D, image_H, image_W). This
3515 3516
            parameter only works when filter_size is None.
        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
3517
            it must contain three integers, (filter_size_D, filter_size_H, filter_size_W).
3518 3519 3520
            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
3521 3522
            contain three integers, (padding_D, padding_H, padding_W). Otherwise, the
            padding_D = padding_H = padding_W = padding. Default: padding = 0.
3523
        stride(int|tuple): The stride size. If stride is a tuple, it must
3524 3525
            contain three integers, (stride_D, stride_H, stride_W). Otherwise, the
            stride_D = stride_H = stride_W = stride. Default: stride = 1.
3526
        dilation(int|tuple): The dilation size. If dilation is a tuple, it must
3527 3528 3529
            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
3530 3531 3532 3533 3534
            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
3535 3536 3537 3538 3539 3540 3541 3542 3543
        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.
3544 3545
        use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
3546 3547
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None.
3548 3549
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.
3550 3551

    Returns:
3552
        Variable: The tensor variable storing the convolution transpose result.
3553 3554

    Raises:
3555 3556
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
3557 3558 3559 3560

    Examples:
       .. code-block:: python

3561 3562
          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)
3563
    """
3564
    assert param_attr is not False, "param_attr should not be False in conv3d_transpose."
3565 3566
    l_type = "conv3d_transpose"
    helper = LayerHelper(l_type, **locals())
3567
    if not isinstance(input, Variable):
3568
        raise TypeError("Input of conv3d_transpose must be Variable")
3569 3570
    input_channel = input.shape[1]

3571 3572 3573
    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
3574

C
chengduoZH 已提交
3575 3576 3577
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")

3578 3579 3580 3581 3582 3583
    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]

3584 3585 3586
        d_in = input.shape[2]
        h_in = input.shape[3]
        w_in = input.shape[4]
C
chengduoZH 已提交
3587

3588
        filter_size_d = (output_size[0] - (d_in - 1) * stride[0] + 2 *
3589
                         padding[0] - 1) // dilation[0] + 1
3590
        filter_size_h = (output_size[1] - (h_in - 1) * stride[1] + 2 *
3591
                         padding[1] - 1) // dilation[1] + 1
3592
        filter_size_w = (output_size[2] - (w_in - 1) * stride[2] + 2 *
3593
                         padding[2] - 1) // dilation[2] + 1
3594
        filter_size = [filter_size_d, filter_size_h, filter_size_w]
C
chengduoZH 已提交
3595
    else:
3596 3597
        filter_size = utils.convert_to_list(filter_size, 3,
                                            'conv3d_transpose.filter_size')
3598

3599
    groups = 1 if groups is None else groups
3600
    filter_shape = [input_channel, num_filters // groups] + filter_size
3601 3602 3603
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

3604
    pre_bias = helper.create_variable_for_type_inference(dtype=input.dtype)
3605
    helper.append_op(
3606
        type=l_type,
3607 3608
        inputs={'Input': [input],
                'Filter': [img_filter]},
3609
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
3610 3611 3612 3613
        attrs={
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
3614
            'groups': groups,
C
chengduoZH 已提交
3615 3616
            'use_cudnn': use_cudnn
        })
3617

3618 3619
    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
    out = helper.append_activation(pre_act)
3620
    return out
3621 3622


Y
yangyaming 已提交
3623
def sequence_expand(x, y, ref_level=-1, name=None):
3624
    """Sequence Expand Layer. This layer will expand the input variable **x**
Y
yangyaming 已提交
3625 3626 3627 3628
    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:
3629 3630 3631 3632 3633

    .. code-block:: text

        * Case 1
            x is a LoDTensor:
3634
                x.lod  = [[2,        2]]
Y
yangyaming 已提交
3635
                x.data = [[a], [b], [c], [d]]
3636 3637 3638
                x.dims = [4, 1]

            y is a LoDTensor:
3639 3640
                y.lod = [[2,    2],
                         [3, 3, 1, 1]]
3641

Y
yangyaming 已提交
3642
            ref_level: 0
3643

Y
yangyaming 已提交
3644
            then output is a 1-level LoDTensor:
3645
                out.lod =  [[2,        2,        2,        2]]
Y
yangyaming 已提交
3646
                out.data = [[a], [b], [a], [b], [c], [d], [c], [d]]
3647 3648 3649 3650
                out.dims = [8, 1]

        * Case 2
            x is a Tensor:
Y
yangyaming 已提交
3651
                x.data = [[a], [b], [c]]
3652 3653 3654
                x.dims = [3, 1]

            y is a LoDTensor:
3655
                y.lod = [[2, 0, 3]]
3656

Y
yangyaming 已提交
3657
            ref_level: -1
3658

Y
yangyaming 已提交
3659 3660 3661
            then output is a Tensor:
                out.data = [[a], [a], [c], [c], [c]]
                out.dims = [5, 1]
3662 3663 3664
    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
3665 3666
        ref_level (int): Lod level of `y` to be referred by `x`. If set to -1,
                         refer the last level of lod.
3667
        name(str|None): A name for this layer(optional). If set None, the layer
Y
yangyaming 已提交
3668
                        will be named automatically.
3669 3670 3671 3672 3673 3674 3675 3676 3677 3678

    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
3679
            out = layers.sequence_expand(x=x, y=y, ref_level=0)
3680
    """
3681
    helper = LayerHelper('sequence_expand', input=x, **locals())
3682
    dtype = helper.input_dtype()
3683
    tmp = helper.create_variable_for_type_inference(dtype)
3684
    helper.append_op(
Y
yangyaming 已提交
3685 3686 3687 3688 3689
        type='sequence_expand',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': tmp},
        attrs={'ref_level': ref_level})
3690
    return tmp
3691 3692


3693 3694 3695 3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713 3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748
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()
3749
    tmp = helper.create_variable_for_type_inference(dtype)
3750 3751 3752 3753 3754 3755 3756 3757
    helper.append_op(
        type='sequence_expand_as',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': tmp})
    return tmp


3758
@templatedoc()
3759
def sequence_pad(x, pad_value, maxlen=None, name=None):
3760 3761 3762 3763 3764
    """
    ${comment}

    Args:
        x(Variable): Input variable which should contain lod information.
M
minqiyang 已提交
3765 3766 3767
        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
3768
            automatically broadcasted to the shape of time step.
M
minqiyang 已提交
3769 3770 3771 3772
        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
3773 3774 3775
            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
3776

3777
    Returns:
M
minqiyang 已提交
3778
        Variable: The padded sequence batch and the original lengths before
3779
                  padding. All sequences has the same length.
M
minqiyang 已提交
3780

3781 3782 3783 3784 3785 3786 3787
    Examples:
        .. code-block:: python

            import numpy

            x = fluid.layers.data(name='y', shape=[10, 5],
                             dtype='float32', lod_level=1)
3788
            pad_value = fluid.layers.assign(
D
dongzhihong 已提交
3789
                input=numpy.array([0.0], dtype=numpy.float32))
3790 3791 3792 3793 3794
            out = fluid.layers.sequence_pad(x=x, pad_value=pad_value)
    """

    helper = LayerHelper('sequence_pad', input=x, **locals())
    dtype = helper.input_dtype()
3795 3796
    out = helper.create_variable_for_type_inference(dtype)
    length = helper.create_variable_for_type_inference(dtype)
3797 3798 3799 3800

    pad_value.stop_gradient = True
    length.stop_gradient = True

3801 3802 3803 3804 3805 3806
    if maxlen is None:
        maxlen = -1
    helper.append_op(
        type='sequence_pad',
        inputs={'X': x,
                'PadValue': pad_value},
3807 3808
        outputs={'Out': out,
                 'Length': length},
3809
        attrs={'padded_length': maxlen})
3810
    return out, length
3811 3812


3813
def sequence_unpad(x, length, name=None):
3814
    """
3815
    **Sequence Unpad Layer**
3816

3817 3818
    This layer removes the padding data in the input sequences and convert
    them into sequences with actual length as output, identitied by lod
3819 3820 3821 3822 3823 3824 3825 3826 3827
    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],
3828 3829 3830
		      [11.0, 12.0, 13.0, 14.0, 15.0]],

	in which there are 3 sequences padded to length 5, and the acutal length
3831
	specified by input Variable **length**:
3832 3833 3834 3835 3836 3837

	    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]]
3838
	    out.lod = [[2, 3, 4]]
3839 3840 3841 3842 3843 3844

    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.
3845 3846
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.
3847 3848 3849 3850 3851 3852 3853 3854 3855 3856 3857 3858 3859 3860

    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()
3861
    out = helper.create_variable_for_type_inference(dtype)
3862 3863 3864 3865 3866 3867 3868 3869 3870 3871 3872

    length.stop_gradient = True

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


3873 3874 3875 3876 3877 3878 3879
def beam_search(pre_ids,
                pre_scores,
                ids,
                scores,
                beam_size,
                end_id,
                level=0,
3880
                is_accumulated=True,
3881 3882
                name=None,
                return_parent_idx=False):
3883
    """
3884 3885
    Beam search is a classical algorithm for selecting candidate words in a
    machine translation task.
3886 3887 3888

    Refer to `Beam search <https://en.wikipedia.org/wiki/Beam_search>`_
    for more details.
M
minqiyang 已提交
3889 3890

    This layer does the search in beams for one time step. Specifically, it
3891 3892 3893
    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
3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904
    computation cell. If :attr:`ids` is not set, it will be calculated out
    according to :attr:`scores`. Additionally, :attr:`pre_ids` and
    :attr:`pre_scores` are the output of beam_search at previous step, they
    are needed for special use to handle ended candidate translations.

    Note that if :attr:`is_accumulated` is :attr:`True`, the :attr:`scores`
    passed in should be accumulated scores. Else, the :attr:`scores` are
    considered as the straightforward scores and will be transformed to the
    log field and accumulated the :attr:`pre_scores` in this operator.
    Length penalty should be done with extra operators before calculating the
    accumulated scores if needed.
3905 3906 3907 3908

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

        fluid/tests/book/test_machine_translation.py
3909

3910
    Args:
3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933
        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.
3934 3935
        is_accumulated(bool, default True): Whether the input :attr:`score` is
             accumulated scores.
3936 3937
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
3938 3939 3940 3941
        return_parent_idx(bool): Whether to return an extra Tensor variable 
                        preserving the selected_ids' parent indice in pre_ids
                        in output, which can be used to gather cell states at
                        the next time step.
F
fengjiayi 已提交
3942

3943
    Returns:
3944 3945 3946 3947
        Variable: The LodTensor tuple containing the selected ids and the \
            corresponding scores. If :attr:`return_parent_idx` is :attr:`True`, \
            an extra Tensor variable preserving the selected_ids' parent indice \
            is included.
3948 3949 3950 3951

    Examples:
        .. code-block:: python

3952 3953 3954 3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967 3968
            # 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
3969
    helper = LayerHelper('beam_search', **locals())
3970 3971 3972 3973 3974 3975
    score_type = pre_scores.dtype
    id_type = pre_ids.dtype

    inputs = {"pre_ids": pre_ids, "pre_scores": pre_scores, "scores": scores}
    if ids is not None:
        inputs["ids"] = ids
Q
Qiao Longfei 已提交
3976

3977 3978 3979
    selected_scores = helper.create_variable_for_type_inference(
        dtype=score_type)
    selected_ids = helper.create_variable_for_type_inference(dtype=id_type)
3980 3981 3982 3983 3984
    # parent_idx is a tensor used to gather cell states at the next time
    # step. Though lod in selected_ids can also be used to gather by
    # sequence_expand, it is not efficient.
    # gather_op's index input only supports int32 dtype currently
    parent_idx = helper.create_variable_for_type_inference(dtype="int32")
Q
Qiao Longfei 已提交
3985 3986 3987

    helper.append_op(
        type='beam_search',
3988
        inputs=inputs,
Q
Qiao Longfei 已提交
3989 3990 3991
        outputs={
            'selected_ids': selected_ids,
            'selected_scores': selected_scores,
3992
            'parent_idx': parent_idx
Q
Qiao Longfei 已提交
3993 3994 3995 3996 3997 3998
        },
        attrs={
            # TODO(ChunweiYan) to assure other value support
            'level': level,
            'beam_size': beam_size,
            'end_id': end_id,
3999
            'is_accumulated': is_accumulated,
Q
Qiao Longfei 已提交
4000
        })
4001 4002 4003 4004
    if return_parent_idx:
        return selected_ids, selected_scores, parent_idx
    else:
        return selected_ids, selected_scores
Q
Qiao Longfei 已提交
4005 4006


4007 4008 4009 4010 4011 4012 4013
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
4014

4015 4016 4017 4018 4019 4020 4021 4022 4023
    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.
4024

4025 4026 4027 4028 4029 4030
    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.
4031

4032 4033
    Examples:
        .. code-block:: python
4034

4035 4036 4037 4038 4039 4040
            # 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())
4041 4042
    sentence_ids = helper.create_variable_for_type_inference(dtype=ids.dtype)
    sentence_scores = helper.create_variable_for_type_inference(dtype=ids.dtype)
4043 4044 4045 4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057

    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


4058 4059 4060 4061
def lstm_unit(x_t,
              hidden_t_prev,
              cell_t_prev,
              forget_bias=0.0,
4062
              param_attr=None,
4063 4064
              bias_attr=None,
              name=None):
4065 4066 4067 4068
    """Lstm unit layer. The equation of a lstm step is:

        .. math::

4069
            i_t & = \sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i)
4070

4071
            f_t & = \sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + b_f)
4072

4073
            c_t & = f_tc_{t-1} + i_t tanh (W_{x_c}x_t + W_{h_c}h_{t-1} + b_c)
4074

4075
            o_t & = \sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + b_o)
4076 4077 4078

            h_t & = o_t tanh(c_t)

4079 4080 4081 4082 4083 4084
    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:
4085 4086 4087

        .. math::

4088
            L_{i_t} = W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i
4089 4090 4091 4092 4093 4094 4095 4096

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

        .. math::

            i_t = \sigma(L_{i_t})

4097
    This layer has two outputs including :math:`h_t` and :math:`o_t`.
4098 4099

    Args:
4100 4101 4102 4103 4104 4105
        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.
4106
        forget_bias (float): The forget bias of lstm unit.
4107 4108 4109 4110 4111 4112 4113 4114 4115 4116 4117 4118
        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.
4119 4120
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
4121 4122

    Returns:
4123
        tuple: The hidden value and cell value of lstm unit.
4124 4125

    Raises:
4126 4127 4128 4129
        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.
4130 4131 4132 4133 4134 4135

    Examples:

        .. code-block:: python

             x_t = fluid.layers.fc(input=x_t_data, size=10)
4136
             prev_hidden = fluid.layers.fc(input=prev_hidden_data, size=30)
4137
             prev_cell = fluid.layers.fc(input=prev_cell_data, size=30)
4138
             hidden_value, cell_value = fluid.layers.lstm_unit(x_t=x_t,
4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149 4150 4151 4152 4153 4154
                                                    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
4155
        raise ValueError("The 1st dimensions of x_t, hidden_t_prev and "
4156 4157 4158 4159
                         "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 "
4160 4161
                         "cell_t_prev must be the same.")

4162 4163 4164
    if bias_attr is None:
        bias_attr = ParamAttr()

4165
    size = cell_t_prev.shape[1]
4166
    concat_out = concat(input=[x_t, hidden_t_prev], axis=1)
4167 4168
    fc_out = fc(input=concat_out,
                size=4 * size,
4169
                param_attr=param_attr,
4170
                bias_attr=bias_attr)
4171
    dtype = x_t.dtype
4172 4173
    c = helper.create_variable_for_type_inference(dtype)
    h = helper.create_variable_for_type_inference(dtype)
4174 4175 4176 4177 4178 4179 4180 4181 4182

    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})

4183
    return h, c
4184 4185


4186
def reduce_sum(input, dim=None, keep_dim=False, name=None):
4187
    """
4188
    Computes the sum of tensor elements over the given dimension.
4189 4190 4191

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
4192
        dim (list|int|None): The dimensions along which the sum is performed. If
4193 4194
            :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
4195 4196
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
4197
        keep_dim (bool|False): Whether to reserve the reduced dimension in the
4198
            output Tensor. The result tensor will have one fewer dimension
4199
            than the :attr:`input` unless :attr:`keep_dim` is true.
4200 4201
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
4202 4203 4204

    Returns:
        Variable: The reduced Tensor variable.
F
fengjiayi 已提交
4205

4206 4207 4208 4209 4210 4211
    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]]
4212
            # Each example is followed by the corresponding output tensor.
4213 4214 4215 4216
            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
4217 4218 4219 4220

            # x is a Tensor variable with shape [2, 2, 2] and elements as below:
            #      [[[1, 2], [3, 4]],
            #      [[5, 6], [7, 8]]]
4221
            # Each example is followed by the corresponding output tensor.
W
whs 已提交
4222 4223 4224
            fluid.layers.reduce_sum(x, dim=[1, 2]) # [10, 26]
            fluid.layers.reduce_sum(x, dim=[0, 1]) # [16, 20]

4225 4226
    """
    helper = LayerHelper('reduce_sum', **locals())
4227
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
4228 4229
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
4230 4231 4232 4233 4234
    helper.append_op(
        type='reduce_sum',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
4235
            'dim': dim if dim != None else [0],
4236 4237 4238 4239
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
4240 4241


4242
def reduce_mean(input, dim=None, keep_dim=False, name=None):
4243
    """
4244
    Computes the mean of the input tensor's elements along the given dimension.
4245 4246 4247

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
4248 4249 4250
        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
4251
            must be in the range :math:`[-rank(input), rank(input))`. If
4252
            :math:`dim[i] < 0`, the dimension to reduce is
4253
            :math:`rank(input) + dim[i]`.
4254 4255
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
4256
            than the :attr:`input` unless :attr:`keep_dim` is true.
4257
        name(str|None): A name for this layer(optional). If set `None`, the layer
4258
                       will be named automatically.
4259 4260

    Returns:
4261
        Variable: The reduced mean Variable.
F
fengjiayi 已提交
4262

4263 4264 4265 4266 4267 4268 4269 4270 4271 4272
    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
4273 4274
            fluid.layers.reduce_mean(
                x, dim=1, keep_dim=True)  # [[0.475], [0.4]]
W
whs 已提交
4275 4276 4277 4278 4279 4280 4281

            # 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]
4282 4283
    """
    helper = LayerHelper('reduce_mean', **locals())
4284
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
4285 4286
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
4287 4288 4289 4290 4291
    helper.append_op(
        type='reduce_mean',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
4292
            'dim': dim if dim != None else [0],
4293 4294 4295 4296
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
4297 4298


4299
def reduce_max(input, dim=None, keep_dim=False, name=None):
4300
    """
4301
    Computes the maximum of tensor elements over the given dimension.
4302 4303 4304

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
4305
        dim (list|int|None): The dimension along which the maximum is computed.
4306 4307 4308
            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
4309
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
4310 4311
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
4312
            than the :attr:`input` unless :attr:`keep_dim` is true.
4313 4314
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
4315 4316 4317

    Returns:
        Variable: The reduced Tensor variable.
4318

4319 4320 4321 4322 4323 4324 4325 4326 4327 4328 4329
    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
4330 4331 4332 4333 4334 4335 4336

            # 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]
4337 4338
    """
    helper = LayerHelper('reduce_max', **locals())
4339
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
4340 4341
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
4342 4343 4344 4345 4346
    helper.append_op(
        type='reduce_max',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
4347
            'dim': dim if dim != None else [0],
4348 4349 4350 4351 4352 4353
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


4354
def reduce_min(input, dim=None, keep_dim=False, name=None):
4355
    """
4356
    Computes the minimum of tensor elements over the given dimension.
4357 4358 4359

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
4360
        dim (list|int|None): The dimensions along which the minimum is computed.
4361 4362 4363
            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
4364
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
4365 4366
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
4367
            than the :attr:`input` unless :attr:`keep_dim` is true.
4368 4369
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
4370 4371 4372

    Returns:
        Variable: The reduced Tensor variable.
4373

4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384
    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
4385 4386 4387 4388 4389 4390 4391

            # 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]
4392 4393
    """
    helper = LayerHelper('reduce_min', **locals())
4394
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
4395 4396
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
4397 4398 4399 4400 4401
    helper.append_op(
        type='reduce_min',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
4402
            'dim': dim if dim != None else [0],
4403 4404 4405 4406
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
4407 4408


4409 4410 4411 4412 4413 4414
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
4415
        dim (list|int|None): The dimensions along which the product is performed. If
4416 4417
            :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
4418 4419
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
4420 4421 4422
        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
4423
        name(str|None): A name for this layer(optional). If set None, the
4424
            layer will be named automatically.
4425 4426 4427 4428 4429 4430 4431 4432 4433 4434 4435 4436 4437 4438

    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
4439
            fluid.layers.reduce_prod(x, dim=1,
4440
                                     keep_dim=True)  # [[0.027], [0.0084]]
W
whs 已提交
4441 4442 4443 4444 4445 4446 4447

            # 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]
4448 4449
    """
    helper = LayerHelper('reduce_prod', **locals())
4450
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
W
whs 已提交
4451 4452
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
4453 4454 4455 4456 4457
    helper.append_op(
        type='reduce_prod',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
4458
            'dim': dim if dim != None else [0],
4459 4460 4461 4462 4463 4464
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


4465
def split(input, num_or_sections, dim=-1, name=None):
G
guosheng 已提交
4466
    """
4467
    Split the input tensor into multiple sub-tensors.
G
guosheng 已提交
4468 4469 4470

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
4471 4472 4473 4474 4475
        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
4476
            :attr:`dim` dimension orderly.
4477
        dim (int): The dimension along which to split. If :math:`dim < 0`, the
G
guosheng 已提交
4478
            dimension to split along is :math:`rank(input) + dim`.
4479 4480
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
4481 4482

    Returns:
D
dzhwinter 已提交
4483
        list(Variable): The list of segmented tensor variables.
G
guosheng 已提交
4484 4485 4486 4487 4488 4489 4490 4491 4492

    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
4493 4494
            x0, x1, x2 = fluid.layers.split(
                x, num_or_sections=[2, 3, 4], dim=1)
G
guosheng 已提交
4495 4496 4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509
            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 = [
4510
        helper.create_variable_for_type_inference(dtype=helper.input_dtype())
G
guosheng 已提交
4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523
        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
4524 4525 4526 4527 4528 4529 4530 4531 4532


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

4533
    .. math::
4534 4535

        y = \\frac{x}{ \sqrt{\sum {x^2} + epsion }}
4536 4537 4538 4539 4540

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

    Args:
4541
        x(Variable|list): The input tensor to l2_normalize layer.
4542
        axis(int): The axis on which to apply normalization. If `axis < 0`, \
4543 4544
            the dimension to normalization is rank(X) + axis. -1 is the
            last dimension.
4545
        epsilon(float): The epsilon value is used to avoid division by zero, \
4546
            the defalut value is 1e-10.
4547
        name(str|None): A name for this layer(optional). If set None, the layer \
4548
            will be named automatically.
4549 4550

    Returns:
4551
        Variable: The output tensor variable is the same shape with `x`.
4552 4553

    Examples:
4554

4555 4556
        .. code-block:: python

4557 4558 4559 4560
            data = fluid.layers.data(name="data",
                                     shape=(3, 17, 13),
                                     dtype="float32")
            normed = fluid.layers.l2_normalize(x=data, axis=1)
4561 4562
    """

F
fengjiayi 已提交
4563 4564
    if len(x.shape) == 1:
        axis = 0
4565 4566
    helper = LayerHelper("l2_normalize", **locals())

4567 4568
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    norm = helper.create_variable_for_type_inference(dtype=x.dtype)
4569
    helper.append_op(
4570 4571 4572 4573
        type="norm",
        inputs={"X": x},
        outputs={"Out": out,
                 "Norm": norm},
4574
        attrs={
4575 4576
            "axis": 1 if axis is None else axis,
            "epsilon": epsilon,
4577 4578
        })
    return out
4579 4580


S
sneaxiy 已提交
4581
def matmul(x, y, transpose_x=False, transpose_y=False, alpha=1.0, name=None):
4582
    """
Y
ying 已提交
4583 4584 4585 4586
    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.
4587

C
chengduoZH 已提交
4588
    The actual behavior depends on the shapes of :math:`x`, :math:`y` and the
4589
    flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically:
4590

4591 4592 4593 4594 4595
    - 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
4596
      :math:`[1, D]` in transposed form.
4597

C
chengduoZH 已提交
4598
    - After transpose, the two tensors are 2-D or n-D and matrix multiplication
4599
      performs in the following way.
4600

4601
      - If both are 2-D, they are multiplied like conventional matrices.
C
chengduoZH 已提交
4602
      - If either is n-D, it is treated as a stack of matrices residing in the
4603
        last two dimensions and a batched matrix multiply supporting broadcast
4604
        applies on the two tensors.
4605

4606 4607
    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
4608
    removed after matrix multiplication.
4609 4610 4611

    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
4612 4613 4614
        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
4615
        alpha (float): The scale of output. Default 1.0.
4616
        name(str|None): A name for this layer(optional). If set None, the layer
4617
            will be named automatically.
4618 4619

    Returns:
4620
        Variable: The product Tensor variable.
4621

G
guosheng 已提交
4622 4623 4624
    Examples:
        .. code-block:: python

4625
            # Examples to clarify shapes of the inputs and output
C
chengduoZH 已提交
4626 4627
            # x: [B, ..., M, K], y: [B, ..., K, N]
            fluid.layers.matmul(x, y)  # out: [B, ..., M, N]
Y
ying 已提交
4628

4629 4630
            # x: [B, M, K], y: [B, K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
4631

4632 4633
            # x: [B, M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
4634

4635 4636
            # x: [M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [M, N]
Y
ying 已提交
4637 4638 4639 4640

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

4641 4642
            # x: [K], y: [K]
            fluid.layers.matmul(x, y)  # out: [1]
4643

Y
ying 已提交
4644
            # x: [M], y: [N]
4645
            fluid.layers.matmul(x, y, True, True)  # out: [M, N]
G
guosheng 已提交
4646
    """
Y
ying 已提交
4647 4648 4649 4650 4651 4652 4653 4654 4655 4656 4657 4658

    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
4659
            y_shape = y_shape + [1]
Y
ying 已提交
4660 4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675

        # 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)

4676
    helper = LayerHelper('matmul', **locals())
4677
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
G
guosheng 已提交
4678
    helper.append_op(
4679 4680 4681 4682
        type='matmul',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
S
sneaxiy 已提交
4683 4684 4685
        attrs={
            'transpose_X': transpose_x,
            'transpose_Y': transpose_y,
S
sneaxiy 已提交
4686
            'alpha': float(alpha),
S
sneaxiy 已提交
4687
        })
4688
    return out
4689 4690


4691
def topk(input, k, name=None):
4692 4693 4694 4695
    """
    This operator is used to find values and indices of the k largest entries
    for the last dimension.

F
fengjiayi 已提交
4696
    If the input is a vector (1-D Tensor), finds the k largest entries in the vector
4697 4698 4699 4700 4701 4702
    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
4703 4704 4705 4706 4707 4708 4709 4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722 4723
    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]]

4724 4725 4726
    Args:
        input(Variable): The input variable which can be a vector or Tensor with
            higher rank.
4727
        k(int | Variable):  The number of top elements to look for along the last dimension
F
fengjiayi 已提交
4728
                 of input.
4729
        name(str|None): A name for this layer(optional). If set None, the layer
4730
                       will be named automatically.
F
fengjiayi 已提交
4731
                       Default: None
4732 4733

    Returns:
4734 4735 4736
        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
4737
        within the last dimension of input.
4738

F
fengjiayi 已提交
4739 4740
    Raises:
        ValueError: If k < 1 or k is not less than the last dimension of input
4741 4742 4743 4744 4745 4746 4747

    Examples:
        .. code-block:: python

            top5_values, top5_indices = layers.topk(input, k=5)
    """
    helper = LayerHelper("top_k", **locals())
4748 4749
    values = helper.create_variable_for_type_inference(dtype=input.dtype)
    indices = helper.create_variable_for_type_inference(dtype="int64")
4750 4751 4752 4753 4754 4755
    inputs = {"X": [input]}
    attrs = None
    if isinstance(k, Variable):
        inputs['K'] = k
    else:
        attrs = {'k': k}
4756 4757
    helper.append_op(
        type="top_k",
4758
        inputs=inputs,
4759 4760
        outputs={"Out": [values],
                 "Indices": [indices]},
4761
        attrs=attrs)
4762 4763 4764 4765 4766
    values.stop_gradient = True
    indices.stop_gradient = True
    return values, indices


4767
def edit_distance(input, label, normalized=True, ignored_tokens=None):
4768
    """
Y
ying 已提交
4769 4770 4771 4772 4773 4774 4775 4776 4777
    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
4778

Y
ying 已提交
4779
    "kitten" -> "sitten" -> "sittin" -> "sitting"
W
wanghaoshuang 已提交
4780

4781
    The input is a LoDTensor consisting of all the hypothesis strings with
Y
ying 已提交
4782 4783
    the total number denoted by `batch_size`, and the separation is specified
    by the LoD information. And the `batch_size` reference strings are arranged
4784
    in order in the same way in the input LoDTensor.
W
wanghaoshuang 已提交
4785

4786
    The output contains the `batch_size` results and each stands for the edit
Y
ying 已提交
4787 4788
    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
4789

4790 4791 4792
    Args:
        input(Variable): The indices for hypothesis strings.
        label(Variable): The indices for reference strings.
4793
        normalized(bool, default True): Indicated whether to normalize the edit distance by
Y
ying 已提交
4794
                          the length of reference string.
4795
        ignored_tokens(list<int>, default None): Tokens that should be removed before
Y
ying 已提交
4796
                                     calculating edit distance.
4797
        name (str): The name of this layer. It is optional.
4798

W
wanghaoshuang 已提交
4799
    Returns:
W
wanghaoshuang 已提交
4800
        Variable: sequence-to-sequence edit distance in shape [batch_size, 1].
W
wanghaoshuang 已提交
4801 4802 4803 4804

    Examples:
        .. code-block:: python

T
tink2123 已提交
4805 4806
            x = fluid.layers.data(name='x', shape=[1], dtype='float32')
            y = fluid.layers.data(name='y', shape=[1], dtype='float32')
4807
            cost = fluid.layers.edit_distance(input=x,label=y)
4808
    """
4809
    helper = LayerHelper("edit_distance", **locals())
4810

4811
    # remove some tokens from input and labels
W
wanghaoshuang 已提交
4812
    if ignored_tokens is not None and len(ignored_tokens) > 0:
4813 4814
        erased_input = helper.create_variable_for_type_inference(dtype="int64")
        erased_label = helper.create_variable_for_type_inference(dtype="int64")
4815 4816 4817 4818 4819

        helper.append_op(
            type="sequence_erase",
            inputs={"X": [input]},
            outputs={"Out": [erased_input]},
W
wanghaoshuang 已提交
4820
            attrs={"tokens": ignored_tokens})
4821 4822 4823 4824 4825
        input = erased_input

        helper.append_op(
            type="sequence_erase",
            inputs={"X": [label]},
4826
            outputs={"Out": [erased_label]},
W
wanghaoshuang 已提交
4827
            attrs={"tokens": ignored_tokens})
4828 4829
        label = erased_label

4830
    # edit distance op
4831 4832
    edit_distance_out = helper.create_variable_for_type_inference(dtype="int64")
    sequence_num = helper.create_variable_for_type_inference(dtype="int64")
4833 4834 4835 4836
    helper.append_op(
        type="edit_distance",
        inputs={"Hyps": [input],
                "Refs": [label]},
4837 4838
        outputs={"Out": [edit_distance_out],
                 "SequenceNum": [sequence_num]},
4839 4840
        attrs={"normalized": normalized})

4841
    return edit_distance_out, sequence_num
4842 4843 4844 4845 4846


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
4847

Y
ying 已提交
4848 4849 4850 4851
    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.
4852 4853 4854 4855 4856 4857 4858 4859 4860 4861 4862 4863 4864 4865 4866 4867 4868

    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]]

4869
        input.lod = [[4, 4]]
M
minqiyang 已提交
4870

4871
        Computation:
4872

4873 4874 4875 4876 4877 4878
        step1: Apply argmax to first input sequence which is input.data[0:4]. Then we get:
               [[0], [2], [1], [0]]
        step2: merge repeated tokens and remove blank which is 0. Then we get first output sequence:
               [[2], [1]]

        Finally:
4879 4880 4881 4882 4883

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

4884
        output.lod = [[2, 1]]
4885

4886

4887 4888
    Args:

Y
ying 已提交
4889 4890 4891 4892 4893 4894 4895 4896 4897
        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).
4898
        name (str): The name of this layer. It is optional.
4899 4900

    Returns:
4901 4902 4903
        Variable: CTC greedy decode result which is a 2-D tensor with shape [Lp, 1]. \
                  'Lp' is the sum if all output sequences' length. If all the sequences \
                  in result were empty, the result LoDTensor will be [-1] with  \
M
minqiyang 已提交
4904
                  LoD [[]] and dims [1, 1].
4905 4906 4907 4908 4909

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
W
wanghaoshuang 已提交
4910

4911
            cost = fluid.layers.ctc_greedy_decoder(input=x, blank=0)
W
wanghaoshuang 已提交
4912
    """
4913
    helper = LayerHelper("ctc_greedy_decoder", **locals())
4914
    _, topk_indices = topk(input, k=1)
4915 4916

    # ctc align op
4917
    ctc_out = helper.create_variable_for_type_inference(dtype="int64")
4918 4919 4920
    helper.append_op(
        type="ctc_align",
        inputs={"Input": [topk_indices]},
W
wanghaoshuang 已提交
4921
        outputs={"Output": [ctc_out]},
4922 4923
        attrs={"merge_repeated": True,
               "blank": blank})
4924
    return ctc_out
4925 4926


W
Wu Yi 已提交
4927
def warpctc(input, label, blank=0, norm_by_times=False, use_cudnn=False):
4928
    """
4929 4930
    An operator integrating the open source Warp-CTC library
    (https://github.com/baidu-research/warp-ctc)
4931
    to compute Connectionist Temporal Classification (CTC) loss.
4932 4933
    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
4934 4935 4936
    input tensor.

    Args:
4937
       input (Variable): The unscaled probabilities of variable-length sequences,
4938 4939 4940 4941
         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).
4942
       label (Variable): The ground truth of variable-length sequence,
4943 4944 4945
         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
4946 4947
         Temporal Classification (CTC) loss, which is in the
         half-opened interval [0, num_classes + 1).
4948 4949 4950
       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
4951
         follewed by a mean_op.
W
Wu Yi 已提交
4952
       use_cudnn (bool, default false): Whether to use cudnn.
4953 4954

    Returns:
4955 4956
        Variable: The Connectionist Temporal Classification (CTC) loss,
        which is a 2-D Tensor of the shape [batch_size, 1].
4957 4958

    Examples:
4959

4960
        .. code-block:: python
4961

4962 4963 4964
            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)
4965 4966

    """
F
fengjiayi 已提交
4967
    helper = LayerHelper('warpctc', **locals())
4968 4969
    loss_out = helper.create_variable_for_type_inference(dtype=input.dtype)
    grad_out = helper.create_variable_for_type_inference(dtype=input.dtype)
4970 4971 4972 4973 4974 4975
    helper.append_op(
        type='warpctc',
        inputs={'Logits': [input],
                'Label': [label]},
        outputs={'WarpCTCGrad': [grad_out],
                 'Loss': [loss_out]},
W
Wu Yi 已提交
4976 4977 4978 4979 4980
        attrs={
            'blank': blank,
            'norm_by_times': norm_by_times,
            'use_cudnn': use_cudnn
        })
4981
    return loss_out
4982 4983 4984 4985 4986 4987 4988 4989 4990 4991 4992 4993 4994 4995 4996


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]]
4997 4998 4999
            x.data = [[1,  2], [3,  4],
                      [5,  6], [7,  8],
                      [9, 10], [11, 12]]
5000 5001 5002 5003 5004
            x.dims = [6, 2]

        set new_dim = 4

        then out is a LoDTensor:
5005

5006
            out.lod  = [[0, 1, 3]]
5007 5008 5009 5010

            out.data = [[1,  2,  3,  4],
                        [5,  6,  7,  8],
                        [9, 10, 11, 12]]
5011 5012 5013 5014 5015 5016 5017
            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:
5018 5019 5020

       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.
5021 5022

    Returns:
5023

5024 5025 5026 5027 5028
        Variable: Reshaped LoDTensor according to new dimension.

    Examples:
        .. code-block:: python

5029
            x = fluid.layers.data(shape=[5, 20], dtype='float32', lod_level=1)
5030
            x_reshaped = fluid.layers.sequence_reshape(input=x, new_dim=10)
5031 5032
    """
    helper = LayerHelper('sequence_reshape', **locals())
5033
    out = helper.create_variable_for_type_inference(helper.input_dtype())
5034 5035 5036 5037 5038 5039
    helper.append_op(
        type='sequence_reshape',
        inputs={'X': [input]},
        outputs={'Out': [out]},
        attrs={'new_dim': new_dim})
    return out
5040 5041


5042 5043 5044 5045
# 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
5046 5047 5048 5049 5050 5051
def nce(input,
        label,
        num_total_classes,
        sample_weight=None,
        param_attr=None,
        bias_attr=None,
5052
        num_neg_samples=None,
5053 5054 5055
        name=None,
        sampler="uniform",
        custom_dist=None,
5056 5057
        seed=0,
        is_sparse=False):
5058 5059 5060 5061 5062 5063 5064
    """
    ${comment}

    Args:
        input (Variable): input variable.
        label (Variable): label.
        num_total_classes (int):${num_total_classes_comment}
5065 5066
        sample_weight (Variable|None): A Variable of shape [batch_size, 1]
            storing a weight for each sample. The default weight for each
5067
            sample is 1.0.
5068 5069 5070 5071 5072 5073 5074 5075 5076
        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.
5077
        num_neg_samples (int): ${num_neg_samples_comment}
5078 5079
        name (str|None): A name for this layer(optional). If set None, the layer
             will be named automatically. Default: None.
5080 5081 5082
        sampler (str): The sampler used to sample class from negtive classes.
                       It can be 'uniform', 'log_uniform' or 'custom_dist'.
                       default: 'uniform'.
5083
        custom_dist (float[]): A float[] with size=num_total_classes.
5084 5085 5086 5087
                       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.
5088
        is_sparse(bool): The flag indicating whether to use sparse update, the weight@GRAD and bias@GRAD will be changed to SelectedRows.
F
fengjiayi 已提交
5089

5090
    Returns:
5091 5092 5093 5094 5095 5096 5097 5098 5099 5100 5101 5102 5103 5104 5105 5106 5107 5108 5109 5110 5111 5112 5113 5114 5115 5116 5117
        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')
5118 5119 5120 5121 5122 5123 5124 5125 5126

            #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)
5127

5128
    """
Y
Yang Yu 已提交
5129 5130 5131
    helper = LayerHelper('nce', **locals())
    assert isinstance(input, Variable)
    assert isinstance(label, Variable)
5132 5133

    dim = input.shape[1]
Y
Yang Yu 已提交
5134 5135 5136 5137 5138 5139
    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)
5140
    inputs = {}
5141 5142 5143 5144 5145 5146 5147
    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
5148 5149 5150
    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
5151

5152 5153 5154 5155
    inputs['Input'] = input
    inputs['Label'] = label
    inputs['Weight'] = w
    inputs['SampleWeight'] = sample_weight if sample_weight is not None else []
5156 5157 5158 5159 5160 5161 5162

    if sampler == "uniform":
        sampler = 0
    elif sampler == "log_uniform":
        sampler = 1
    elif sampler == "custom_dist":
        assert custom_dist is not None
5163 5164 5165 5166 5167 5168 5169 5170 5171
        # assert isinstance(custom_dist, Variable)

        custom_dist_len = len(custom_dist)
        alias_probs_ = [0] * custom_dist_len
        alias_ = [0] * custom_dist_len
        bigs = []
        littles = []
        for i in range(custom_dist_len):
            normal_prob = custom_dist[i] * custom_dist_len
5172
            if normal_prob - 1.0 > 0:
5173
                bigs.append((i, normal_prob))
5174
            elif 1.0 - normal_prob > 0:
5175 5176 5177 5178 5179 5180 5181 5182 5183 5184 5185 5186 5187 5188 5189
                littles.append((i, normal_prob))
            else:
                alias_probs_[i] = normal_prob
                alias_[i] = -1

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

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

            alias_probs_[little[0]] = little[1]
            alias_[little[0]] = big_idx
            big_left = big[1] + little[1] - 1
5190
            if big_left - 1.0 > 0:
5191
                bigs.append((big_idx, big_left))
5192
            elif 1.0 - big_left > 0:
5193 5194 5195 5196 5197 5198 5199 5200 5201 5202 5203 5204 5205 5206
                littles.append((big_idx, big_left))
            else:
                alias_probs_[big_idx] = big_left
                alias_[big_idx] = -1

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

5207 5208 5209 5210 5211 5212 5213 5214 5215 5216 5217 5218 5219 5220 5221
        def _init_by_numpy_array(numpy_array):
            ret = helper.create_parameter(
                attr=ParamAttr(),
                shape=numpy_array.shape,
                dtype=numpy_array.dtype,
                default_initializer=NumpyArrayInitializer(numpy_array))
            ret.stop_gradient = True
            return ret

        inputs['CustomDistProbs'] = _init_by_numpy_array(
            np.array(custom_dist).astype('float32'))
        inputs['CustomDistAlias'] = _init_by_numpy_array(
            np.array(alias_).astype('int32'))
        inputs['CustomDistAliasProbs'] = _init_by_numpy_array(
            np.array(alias_probs_).astype('float32'))
5222 5223 5224 5225
        sampler = 2
    else:
        raise Exception("Unsupported sampler type.")

5226 5227 5228 5229 5230
    if num_neg_samples is None:
        num_neg_samples = 10
    else:
        num_neg_samples = int(num_neg_samples)

5231 5232 5233 5234
    remote_prefetch = is_sparse
    print(
        "With sparse mode, if your models has only small parameter prefetch may cause speed down"
    )
5235

Y
Yang Yu 已提交
5236 5237
    attrs = {
        'num_total_classes': int(num_total_classes),
5238 5239
        'num_neg_samples': num_neg_samples,
        'seed': seed,
5240
        'sampler': sampler,
5241 5242
        'is_sparse': is_sparse,
        'remote_prefetch': remote_prefetch
Y
Yang Yu 已提交
5243
    }
Y
Yang Yu 已提交
5244 5245 5246

    helper.append_op(
        type='nce',
5247
        inputs=inputs,
Y
Yang Yu 已提交
5248 5249 5250 5251 5252 5253
        outputs={
            'Cost': cost,
            'SampleLogits': sample_logits,
            'SampleLabels': sample_labels
        },
        attrs=attrs)
Y
Yang Yu 已提交
5254
    return cost / (num_neg_samples + 1)
5255 5256


5257 5258
def hsigmoid(input,
             label,
5259
             num_classes,
5260 5261
             param_attr=None,
             bias_attr=None,
J
JiabinYang 已提交
5262
             name=None,
5263 5264 5265
             path_table=None,
             path_code=None,
             is_custom=False,
J
JiabinYang 已提交
5266
             is_sparse=False):
W
weixing02 已提交
5267 5268
    """
    The hierarchical sigmoid operator is used to accelerate the training
M
minqiyang 已提交
5269
    process of language model. This operator organizes the classes into a
M
minqiyang 已提交
5270
    complete binary tree, or you can use is_custom to pass your own tree to
5271
    implement hierarchical. Each leaf node represents a class(a word) and each
G
guosheng 已提交
5272 5273 5274 5275 5276 5277
    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.

5278
    Using default tree you can Refer to `Hierarchical Probabilistic Neural Network Language Model
G
guosheng 已提交
5279
    <http://www.iro.umontreal.ca/~lisa/pointeurs/hierarchical-nnlm-aistats05.pdf>`_
M
minqiyang 已提交
5280

5281 5282
    And if you want to use the costumed tree by set 'is_custom' as true you may need to do following things first:

5283 5284 5285 5286
    1. using your word dict to build a binary tree, each leaf node should be an word of your word dict
    2. build a dict to store word_id -> word's leaf to root path, we call it path_table.
    3. build a dict to store word_id -> code of word's leaf to root path, we call it path_code. Code
       means label of each binary classification, using 1 indicate true, 0 indicate false.
M
minqiyang 已提交
5287
    4. now, each word should has its path and code along the path, you can pass a batch of path and code
5288
       related to the same batch of inputs.
5289

W
weixing02 已提交
5290
    Args:
M
minqiyang 已提交
5291
        input (Variable): The input tensor variable with shape
G
guosheng 已提交
5292 5293 5294 5295
            :math:`[N \\times D]`, where :math:`N` is the size of mini-batch,
            and :math:`D` is the feature size.
        label (Variable): The tensor variable contains labels of training data.
            It's a tensor with shape is :math:`[N \\times 1]`.
M
minqiyang 已提交
5296 5297
        num_classes: (int), The number of classes, must not be less than 2. with default tree this has to be set,
            it should never be None under is_custom=False, but while is_custom is true, it should be non leaf num
5298
            which indicates the num of classes using by binary classify.
5299 5300 5301 5302 5303 5304 5305 5306 5307 5308 5309
        param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
             of hsigmoid. If it is set to None or one attribute of ParamAttr, hsigmoid
             will create ParamAttr as param_attr. If the Initializer of the param_attr
             is not set, the parameter is initialized with Xavier. Default: None.
        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of hsigmoid.
             If it is set to False, no bias will be added to the output units.
             If it is set to None or one attribute of ParamAttr, hsigmoid
             will create ParamAttr as bias_attr. If the Initializer of the bias_attr
             is not set, the bias is initialized zero. Default: None.
        name (str|None): A name for this layer(optional). If set None, the layer
             will be named automatically. Default: None.
M
minqiyang 已提交
5310
        path_table: (Variable|None) this variable can store each batch of samples' path to root,
5311
            it should be in leaf -> root order
M
minqiyang 已提交
5312 5313 5314
            path_table should have the same shape with path_code, and for each sample i path_table[i] indicates a np.array like
            structure and each element in this array is indexes in parent nodes' Weight Matrix.
        path_code:  (Variable|None) this variable can store each batch of samples' code,
5315
            each code consist with every code of parent nodes. it should be in leaf -> root order
M
minqiyang 已提交
5316
        is_custom: (bool|False)using user defined binary tree instead of default complete binary tree, if costum is
5317
             set you need to set path_table/path_code/num_classes, otherwise num_classes should be set
M
minqiyang 已提交
5318
        is_sparse: (bool|False)using sparse update instead of dense update, if set, the gradient
5319
             of W and input will be sparse.
W
weixing02 已提交
5320 5321

    Returns:
J
JiabinYang 已提交
5322
        Out: (LodTensor) The cost of hierarchical sigmoid operator. the shape is [N, 1]
W
weixing02 已提交
5323 5324 5325 5326 5327

    Examples:

        .. code-block:: python

G
guosheng 已提交
5328 5329 5330
            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
5331 5332 5333 5334
    """

    helper = LayerHelper('hierarchical_sigmoid', **locals())
    dtype = helper.input_dtype()
5335 5336
    out = helper.create_variable_for_type_inference(dtype)
    pre_out = helper.create_variable_for_type_inference(dtype)
W
weixing02 已提交
5337
    dim = input.shape[1]
5338
    if ((num_classes is None) or (num_classes < 2)) and (not is_custom):
J
JiabinYang 已提交
5339 5340 5341
        raise ValueError(
            "num_classes must not be less than 2 with default tree")

5342 5343 5344 5345
    if (is_custom) and (path_code is None):
        raise ValueError("path_code should not be None with costum tree")
    elif (is_custom) and (path_table is None):
        raise ValueError("path_table should not be None with costum tree")
5346 5347
    elif (is_custom) and (num_classes is None):
        raise ValueError("num_classes should not be None with costum tree")
5348 5349 5350
    else:
        pass

J
JiabinYang 已提交
5351
    weights = None
5352 5353 5354 5355
    remote_prefetch = is_sparse
    print(
        "With sparse mode, if your models has only small parameter prefetch may cause speed down"
    )
5356
    if not is_custom:
J
JiabinYang 已提交
5357 5358 5359 5360 5361 5362 5363 5364
        weights = helper.create_parameter(
            attr=helper.param_attr,
            shape=[num_classes - 1, dim],
            is_bias=False,
            dtype=input.dtype)
    else:
        weights = helper.create_parameter(
            attr=helper.param_attr,
5365
            shape=[num_classes, dim],
J
JiabinYang 已提交
5366 5367
            is_bias=False,
            dtype=input.dtype)
5368 5369 5370
    inputs = {
        "X": input,
        "W": weights,
5371
        "PathTable": path_table,
5372
        "PathCode": path_code,
5373 5374
        "Label": label
    }
W
weixing02 已提交
5375
    if helper.bias_attr:
5376
        if not is_custom:
J
JiabinYang 已提交
5377 5378
            bias = helper.create_parameter(
                attr=helper.bias_attr,
5379
                shape=[num_classes - 1, 1],
J
JiabinYang 已提交
5380 5381 5382 5383 5384 5385
                is_bias=True,
                dtype=input.dtype)
            inputs['Bias'] = bias
        else:
            bias = helper.create_parameter(
                attr=helper.bias_attr,
5386
                shape=[num_classes, 1],
J
JiabinYang 已提交
5387 5388 5389
                is_bias=True,
                dtype=input.dtype)
            inputs['Bias'] = bias
W
weixing02 已提交
5390 5391
    helper.append_op(
        type="hierarchical_sigmoid",
W
weixing02 已提交
5392
        inputs=inputs,
W
weixing02 已提交
5393
        outputs={"Out": out,
5394 5395 5396 5397 5398 5399 5400
                 "PreOut": pre_out,
                 "W_Out": weights},
        attrs={
            "num_classes": num_classes,
            "is_sparse": is_sparse,
            "remote_prefetch": remote_prefetch
        })
W
weixing02 已提交
5401 5402 5403
    return out


Y
fix ci.  
ying 已提交
5404
def transpose(x, perm, name=None):
5405 5406 5407 5408 5409 5410 5411
    """
    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:
5412 5413 5414
        x (Variable): The input Tensor.
        perm (list): A permutation of the dimensions of `input`.
        name (str): The name of this layer. It is optional.
5415 5416 5417 5418 5419 5420 5421

    Returns:
        Variable: A transposed Tensor.

    Examples:
        .. code-block:: python

5422
            # use append_batch_size=False to avoid prepending extra
5423
            # batch size in shape
5424
            x = fluid.layers.data(name='x', shape=[5, 10, 15],
5425
                            dtype='float32', append_batch_size=False)
Y
fix ci.  
ying 已提交
5426
            x_transposed = layers.transpose(x, perm=[1, 0, 2])
5427 5428
    """

Y
fix ci.  
ying 已提交
5429
    if len(perm) != len(x.shape):
5430 5431 5432
        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
5433 5434 5435 5436 5437 5438
    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)))
5439 5440

    helper = LayerHelper('transpose', **locals())
5441 5442
    out = helper.create_variable_for_type_inference(x.dtype)
    x_shape = helper.create_variable_for_type_inference(x.dtype)
5443
    helper.append_op(
5444
        type='transpose2',
Y
fix ci.  
ying 已提交
5445
        inputs={'X': [x]},
5446 5447
        outputs={'Out': [out],
                 'XShape': [x_shape]},
5448 5449
        attrs={'axis': perm})
    return out
5450 5451


5452 5453 5454 5455 5456 5457 5458
def im2sequence(input,
                filter_size=1,
                stride=1,
                padding=0,
                input_image_size=None,
                out_stride=1,
                name=None):
5459
    """
5460 5461 5462 5463 5464 5465 5466
    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:
5467 5468 5469 5470 5471 5472 5473 5474 5475 5476

    .. 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
5477 5478 5479 5480 5481 5482 5483 5484 5485 5486 5487 5488 5489 5490 5491 5492 5493 5494

        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.

5495 5496 5497 5498 5499 5500 5501 5502 5503
        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.

5504 5505 5506
        name (int): The name of this layer. It is optional.

    Returns:
W
wanghaoshuang 已提交
5507 5508 5509 5510 5511
        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.
5512 5513 5514 5515 5516 5517 5518 5519 5520 5521 5522 5523 5524 5525 5526 5527 5528 5529 5530 5531 5532 5533 5534 5535 5536 5537 5538

    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
5539 5540 5541
            filter = [2, 2]
            stride = [1, 1]
            padding = [0, 0]
5542 5543 5544 5545 5546 5547 5548 5549 5550 5551 5552 5553

            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.]]

5554
            output.dims = {8, 8}
5555

5556
            output.lod = [[4, 4]]
5557

5558
    Examples:
5559 5560 5561

        .. code-block:: python

5562 5563
            output = fluid.layers.im2sequence(
                input=layer, stride=[1, 1], filter_size=[2, 2])
5564 5565

    """
W
wanghaoshuang 已提交
5566 5567 5568 5569 5570 5571 5572 5573 5574 5575

    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])
5576 5577 5578 5579 5580 5581 5582
    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
5583
    helper = LayerHelper('im2sequence', **locals())
5584
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
5585
    helper.append_op(
5586
        type='im2sequence', inputs=inputs, outputs={'Out': out}, attrs=attrs)
5587
    return out
5588 5589


Y
yuyang18 已提交
5590
@templatedoc()
5591
def row_conv(input, future_context_size, param_attr=None, act=None):
Y
yuyang18 已提交
5592 5593
    """
    ${comment}
5594 5595

    Args:
Y
yuyang18 已提交
5596
        input (${x_type}): ${x_comment}.
Y
yangyaming 已提交
5597 5598
        future_context_size (int): Future context size. Please note, the shape
            of convolution kernel is [future_context_size + 1, D].
5599 5600 5601 5602 5603
        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
5604
        ${out_comment}.
5605 5606

    Examples:
Y
yuyang18 已提交
5607 5608 5609 5610
        >>> 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)
5611 5612 5613 5614 5615 5616
    """
    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)
5617
    out = helper.create_variable_for_type_inference(dtype)
5618 5619 5620 5621 5622
    helper.append_op(
        type='row_conv',
        inputs={'X': [input],
                'Filter': [filter_param]},
        outputs={'Out': [out]})
Y
yangyaming 已提交
5623
    return helper.append_activation(out)
5624 5625


Y
yuyang18 已提交
5626
@templatedoc()
5627 5628
def multiplex(inputs, index):
    """
Y
yuyang18 已提交
5629 5630 5631 5632 5633 5634 5635
    ${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)
5636 5637

    Args:
Y
yuyang18 已提交
5638 5639
       inputs (list): ${x_comment}.
       index (${ids_type}): ${ids_comment}.
5640 5641

    Returns:
Y
yuyang18 已提交
5642
        ${out_comment}.
5643 5644
    """
    helper = LayerHelper('multiplex', **locals())
Y
yangyaming 已提交
5645 5646 5647 5648 5649

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

5650
    out = helper.create_variable_for_type_inference(inputs[0].dtype)
5651 5652 5653 5654 5655 5656
    helper.append_op(
        type='multiplex',
        inputs={'X': inputs,
                'Ids': index},
        outputs={'Out': [out]})
    return out
5657 5658


5659 5660 5661
def softmax_with_cross_entropy(logits,
                               label,
                               soft_label=False,
J
jerrywgz 已提交
5662
                               ignore_index=kIgnoreIndex,
5663 5664
                               numeric_stable_mode=False,
                               return_softmax=False):
5665 5666
    """
    **Softmax With Cross Entropy Operator.**
5667

5668 5669 5670 5671
    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.
5672

5673 5674 5675
    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.
5676

5677 5678 5679
    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.
5680

5681
    The equation is as follows:
5682

5683
    1) Hard label (one-hot label, so every sample has exactly one class)
5684

5685 5686 5687 5688
    .. math::

        loss_j =  -\\text{logit}_{label_j} +
        \\log\\left(\\sum_{i=0}^{K}\\exp(\\text{logit}_i)\\right), j = 1,..., K
5689

5690 5691 5692
    2) Soft label (each sample can have a distribution over all classes)

    .. math::
5693

5694 5695 5696 5697
        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
5698 5699 5700
    3) If numeric_stable_mode is True, softmax is calculated first by:

    .. math::
5701

5702
        max_j &= \\max_{i=0}^{K}{\\text{logit}_i}
S
sneaxiy 已提交
5703

5704
        log\\_max\\_sum_j &= \\log\\sum_{i=0}^{K}\\exp(logit_i - max_j)
S
sneaxiy 已提交
5705

5706
        softmax_j &= \\exp(logit_j - max_j - {log\\_max\\_sum}_j)
S
sneaxiy 已提交
5707 5708 5709

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

5710 5711 5712 5713 5714 5715 5716 5717
    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
5718 5719
        ignore_index (int): Specifies a target value that is ignored and does
                            not contribute to the input gradient. Only valid
J
jerrywgz 已提交
5720
                            if soft_label is set to False. Default: kIgnoreIndex
S
sneaxiy 已提交
5721 5722 5723
        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.
5724 5725 5726
                                    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
5727
                                    stable algorithm. Default: False
5728
        return_softmax (bool): A flag indicating whether to return the softmax
5729
                               along with the cross entropy loss. Default: False
5730

5731
    Returns:
5732 5733 5734 5735 5736
        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 \
                                            2-D tensor with shape [N x K].
5737 5738 5739 5740 5741 5742 5743

    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
5744 5745
            out = fluid.layers.softmax_with_cross_entropy(
                logits=fc, label=label)
5746 5747
    """
    helper = LayerHelper('softmax_with_cross_entropy', **locals())
5748 5749
    softmax = helper.create_variable_for_type_inference(dtype=logits.dtype)
    loss = helper.create_variable_for_type_inference(dtype=logits.dtype)
5750 5751 5752 5753 5754 5755
    helper.append_op(
        type='softmax_with_cross_entropy',
        inputs={'Logits': logits,
                'Label': label},
        outputs={'Softmax': softmax,
                 'Loss': loss},
S
sneaxiy 已提交
5756 5757 5758 5759 5760
        attrs={
            'soft_label': soft_label,
            'ignore_index': ignore_index,
            'numeric_stable_mode': numeric_stable_mode
        })
5761 5762 5763 5764

    if return_softmax:
        return loss, softmax

5765 5766 5767 5768 5769
    return loss


def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None):
    """
Y
Yibing Liu 已提交
5770 5771
    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.
5772
    For each instance, it computes the smooth L1 loss element by element first
5773
    and then sums all the losses. So the shape of ouput Variable is
5774
    [batch_size, 1].
5775

5776 5777
    Args:
        x (Variable): A tensor with rank at least 2. The input value of smooth
5778
            L1 loss op with shape [batch_size, dim1, ..., dimN].
5779
        y (Variable): A tensor with rank at least 2. The target value of smooth
Y
Yibing Liu 已提交
5780
            L1 loss op with same shape as :attr:`x`.
5781
        inside_weight (Variable|None):  A tensor with rank at least 2. This
5782 5783
            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
5784
            by this tensor element by element.
5785
        outside_weight (Variable|None): A tensor with rank at least 2. This
5786 5787
            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
5788
            element by element.
5789
        sigma (float|None): Hyper parameter of smooth L1 loss layer. A float
5790 5791
           scalar with default value 1.0.

5792
    Returns:
5793
        Variable: The output smooth L1 loss with shape [batch_size, 1].
5794 5795 5796 5797 5798

    Examples:
        .. code-block:: python

            data = fluid.layers.data(name='data', shape=[128], dtype='float32')
F
stash  
fengjiayi 已提交
5799 5800
            label = fluid.layers.data(
                name='label', shape=[100], dtype='float32')
5801
            fc = fluid.layers.fc(input=data, size=100)
F
fengjiayi 已提交
5802
            out = fluid.layers.smooth_l1(x=fc, y=label)
5803
    """
5804

5805
    helper = LayerHelper('smooth_l1_loss', **locals())
5806 5807
    diff = helper.create_variable_for_type_inference(dtype=x.dtype)
    loss = helper.create_variable_for_type_inference(dtype=x.dtype)
5808 5809 5810 5811 5812 5813 5814 5815 5816 5817 5818 5819
    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
5820 5821 5822 5823


def one_hot(input, depth):
    """
Y
Yibing Liu 已提交
5824
    This layer creates the one-hot representations for input indices.
5825 5826

    Args:
Y
Yibing Liu 已提交
5827 5828
        input(Variable): Input indices, last dimension must be 1.
        depth(scalar): An interger defining the depth of the one-hot dimension.
5829 5830

    Returns:
Y
Yibing Liu 已提交
5831
        Variable: The one-hot representations of input.
5832 5833

    Examples:
5834
        .. code-block:: python
5835

Y
Yibing Liu 已提交
5836 5837
            label = layers.data(name="label", shape=[1], dtype="float32")
            one_hot_label = layers.one_hot(input=label, depth=10)
5838 5839
    """
    helper = LayerHelper("one_hot", **locals())
5840
    one_hot_out = helper.create_variable_for_type_inference(dtype='float32')
5841 5842 5843 5844 5845 5846
    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
5847 5848


Y
Yu Yang 已提交
5849
def autoincreased_step_counter(counter_name=None, begin=1, step=1):
Y
Yu Yang 已提交
5850
    """
Y
yi.wu 已提交
5851 5852 5853
    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
5854 5855 5856 5857 5858 5859

    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.

5860 5861
    Returns:
        Variable: The global run counter.
Y
yi.wu 已提交
5862 5863 5864 5865 5866 5867

    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
5868 5869
    """
    helper = LayerHelper('global_step_counter')
Y
Yu Yang 已提交
5870 5871
    if counter_name is None:
        counter_name = '@STEP_COUNTER@'
Y
Yu Yang 已提交
5872 5873 5874 5875 5876
    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
5877
                value=begin - 1, force_cpu=True))
W
Wu Yi 已提交
5878
        helper.main_program.global_block()._prepend_op(
Y
Yu Yang 已提交
5879 5880
            type='increment',
            inputs={'X': [counter]},
Y
Yu Yang 已提交
5881
            outputs={'Out': [counter]},
5882 5883
            attrs={'step': float(step)},
            stop_gradient=True)
Y
Yu Yang 已提交
5884 5885 5886
        counter.stop_gradient = True

    return counter
Y
yangyaming 已提交
5887 5888


5889
def reshape(x, shape, actual_shape=None, act=None, inplace=False, name=None):
5890
    """
C
caoying03 已提交
5891 5892
    Gives a new shape to the input Tensor without changing its data.

5893 5894 5895 5896 5897
    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
5898

5899
    Some tricks exist when specifying the target shape.
C
caoying03 已提交
5900

5901 5902 5903 5904
    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.

5905
    2. 0 means the actual dimension value is going to be copied from the
5906 5907 5908 5909
    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
5910 5911

    1. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
W
wanghaoshuang 已提交
5912
    is [6, 8], the reshape operator will transform x into a 2-D tensor with
5913
    shape [6, 8] and leaving x's data unchanged.
C
caoying03 已提交
5914

5915
    2. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
5916 5917
    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
5918 5919
    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
5920
    dimensions.
C
caoying03 已提交
5921

5922
    3. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
5923 5924 5925 5926
    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.
5927 5928

    Args:
5929
        x(variable): The input tensor.
5930 5931
        shape(list): The new shape. At most one dimension of the new shape can
                     be -1.
5932 5933 5934 5935 5936
        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`.
5937 5938
        act (str): The non-linear activation to be applied to the reshaped tensor
                   variable.
C
chengduozh 已提交
5939 5940 5941
        inplace(bool): If ``inplace`` is `True`, the input and output of ``layers.reshape``
                       are the same variable, otherwise, the input and output of
                       ``layers.reshape`` are different variables. Note that if :attr:`x`
C
chengduozh 已提交
5942
                       is more than one layer's input, ``inplace`` must be :attr:`False`.
5943
        name (str): The name of this layer. It is optional.
5944

5945
    Returns:
G
guosheng 已提交
5946 5947 5948 5949
        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.
5950

5951 5952 5953
    Raises:
        TypeError: if actual_shape is neither Variable nor None.

5954 5955
    Examples:
        .. code-block:: python
5956

5957
            data = fluid.layers.data(
5958
                name='data', shape=[2, 4, 6], dtype='float32')
C
caoying03 已提交
5959
            reshaped = fluid.layers.reshape(
G
guosheng 已提交
5960
                x=data, shape=[-1, 0, 3, 2], inplace=True)
5961 5962 5963
    """

    if not (isinstance(shape, list) or isinstance(shape, tuple)):
L
luotao1 已提交
5964
        raise ValueError("Input shape must be a python list or tuple.")
5965 5966 5967 5968 5969
    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")
5970

5971 5972 5973 5974 5975 5976 5977 5978 5979 5980 5981 5982 5983 5984 5985
    # 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.")

5986
    helper = LayerHelper("reshape2", **locals())
5987 5988
    out = x if inplace else helper.create_variable_for_type_inference(
        dtype=x.dtype)
5989
    x_shape = helper.create_variable_for_type_inference(dtype=x.dtype)
5990
    helper.append_op(
5991
        type="reshape2",
5992
        inputs=inputs,
D
dzhwinter 已提交
5993
        attrs={"shape": shape},
5994 5995
        outputs={"Out": out,
                 "XShape": x_shape})
5996

D
dzhwinter 已提交
5997
    return helper.append_activation(out)
5998

5999

6000
def squeeze(input, axes, name=None):
6001
    """
M
minqiyang 已提交
6002 6003 6004
    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
6005
    selected with shape entry not equal to one, an error is raised.
M
minqiyang 已提交
6006

6007 6008 6009 6010 6011 6012 6013 6014 6015 6016 6017 6018 6019 6020 6021 6022 6023 6024 6025 6026 6027
    For example:

    .. code-block:: text

        Case 1:

          Given
            X.shape = (1, 3, 1, 5)
          and
            axes = [0]
          we get:
            Out.shape = (3, 1, 5)

        Case 2:

          Given
            X.shape = (1, 3, 1, 5)
          and
            axes = []
          we get:
            Out.shape = (3, 5)
M
minqiyang 已提交
6028

6029
    Args:
6030
        input (Variable): The input variable to be squeezed.
6031
        axes (list): List of integers, indicating the dimensions to be squeezed.
6032
        name (str|None): Name for this layer.
6033 6034 6035 6036 6037 6038 6039 6040

    Returns:
        Variable: Output squeezed variable.

    Examples:
        .. code-block:: python

            x = layers.data(name='x', shape=[5, 1, 10])
6041
            y = layers.sequeeze(input=x, axes=[1])
6042 6043
    """
    helper = LayerHelper("squeeze", **locals())
6044 6045
    out = helper.create_variable_for_type_inference(dtype=input.dtype)
    x_shape = helper.create_variable_for_type_inference(dtype=input.dtype)
6046
    helper.append_op(
6047
        type="squeeze2",
6048
        inputs={"X": input},
6049
        attrs={"axes": axes},
6050 6051
        outputs={"Out": out,
                 "XShape": x_shape})
6052

6053 6054 6055
    return out


6056
def unsqueeze(input, axes, name=None):
6057
    """
M
minqiyang 已提交
6058 6059 6060
    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.
6061

M
minqiyang 已提交
6062
    For example:
6063 6064 6065

    .. code-block:: text

M
minqiyang 已提交
6066
      Given a tensor such that tensor with shape [3, 4, 5],
6067
      then Unsqueezed tensor with axes=[0, 4] has shape [1, 3, 4, 5, 1].
M
minqiyang 已提交
6068

6069
    Args:
6070
        input (Variable): The input variable to be unsqueezed.
6071
        axes (list): List of integers, indicating the dimensions to be inserted.
6072
        name (str|None): Name for this layer.
6073 6074 6075 6076 6077 6078 6079 6080

    Returns:
        Variable: Output unsqueezed variable.

    Examples:
        .. code-block:: python

            x = layers.data(name='x', shape=[5, 10])
6081
            y = layers.unsequeeze(input=x, axes=[1])
6082 6083
    """
    helper = LayerHelper("unsqueeze", **locals())
6084 6085
    out = helper.create_variable_for_type_inference(dtype=input.dtype)
    x_shape = helper.create_variable_for_type_inference(dtype=input.dtype)
6086
    helper.append_op(
6087
        type="unsqueeze2",
6088
        inputs={"X": input},
6089
        attrs={"axes": axes},
6090 6091
        outputs={"Out": out,
                 "XShape": x_shape})
6092

6093 6094
    return out

6095

6096
def lod_reset(x, y=None, target_lod=None):
Y
yangyaming 已提交
6097
    """
Y
Yibing Liu 已提交
6098
    Set LoD of :attr:`x` to a new one specified by :attr:`y` or
6099 6100 6101 6102
    :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
6103
    :attr:`Y.data` or :attr:`target_lod`, only one level LoD is supported.
Y
yangyaming 已提交
6104 6105 6106 6107 6108 6109

    .. code-block:: text

        * Example 1:

            Given a 1-level LoDTensor x:
6110
                x.lod =  [[ 2,           3,                   1 ]]
Y
yangyaming 已提交
6111 6112 6113
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

6114
            target_lod: [4, 2]
Y
yangyaming 已提交
6115 6116

            then we get a 1-level LoDTensor:
6117
                out.lod =  [[4,                          2]]
Y
yangyaming 已提交
6118 6119 6120 6121 6122 6123
                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:
6124
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
6125 6126 6127 6128
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a Tensor:
6129
                y.data = [[2, 4]]
Y
yangyaming 已提交
6130 6131 6132
                y.dims = [1, 3]

            then we get a 1-level LoDTensor:
6133
                out.lod =  [[2,            4]]
Y
yangyaming 已提交
6134 6135 6136 6137 6138 6139
                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:
6140
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
6141 6142 6143 6144
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a 2-level LoDTensor:
6145
                y.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
6146 6147 6148 6149
                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:
6150
                out.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
6151 6152 6153 6154 6155
                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.
6156
        y (Variable|None): If provided, output's LoD would be derived
Y
Yibing Liu 已提交
6157
                           from :attr:`y`.
Y
yangyaming 已提交
6158
        target_lod (list|tuple|None): One level LoD which should be considered
Y
Yibing Liu 已提交
6159
                                      as target LoD when :attr:`y` not provided.
Y
yangyaming 已提交
6160 6161

    Returns:
Y
Yibing Liu 已提交
6162
        Variable: Output variable with LoD specified by this layer.
Y
yangyaming 已提交
6163 6164

    Raises:
Y
Yibing Liu 已提交
6165
        ValueError: If :attr:`y` and :attr:`target_lod` are both None.
Y
yangyaming 已提交
6166 6167 6168 6169 6170 6171 6172 6173 6174

    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())
6175
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
Y
yangyaming 已提交
6176 6177 6178 6179 6180 6181 6182 6183 6184 6185 6186 6187 6188 6189
    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
6190 6191 6192 6193 6194 6195 6196 6197 6198 6199 6200


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
6201
      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
6202 6203 6204 6205 6206 6207 6208 6209 6210 6211 6212 6213 6214 6215 6216 6217 6218 6219 6220 6221 6222 6223 6224 6225 6226 6227 6228 6229

    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
6230 6231
          data = fluid.layers.data(
              name="data", shape=[3, 112, 112], dtype="float32")
D
dragonwarrior 已提交
6232 6233 6234 6235 6236 6237 6238 6239 6240 6241 6242 6243
          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))

6244 6245 6246
    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
6247 6248 6249 6250 6251 6252 6253 6254 6255 6256 6257 6258 6259
    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
6260 6261 6262 6263


def pad(x, paddings, pad_value=0., name=None):
    """
6264
    Pads a tensor with a constant value given by :attr:`pad_value`, and the
W
wanghaoshuang 已提交
6265
    padded width is specified by :attr:`paddings`.
6266

6267 6268 6269 6270
    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]`.
6271 6272 6273 6274 6275 6276 6277 6278 6279 6280 6281 6282 6283 6284 6285 6286 6287 6288 6289 6290 6291 6292

    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
6293
                         The length of :attr:paddings must be
6294 6295 6296 6297 6298 6299 6300 6301 6302 6303
                         :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
6304

6305 6306 6307 6308 6309 6310
            # 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()
6311
    out = helper.create_variable_for_type_inference(dtype)
6312 6313 6314 6315 6316 6317 6318
    helper.append_op(
        type='pad',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'paddings': paddings,
               'pad_value': float(pad_value)})
    return out
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
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)
6352 6353
		And
            pad_value = -1,
6354

6355 6356 6357 6358 6359 6360 6361 6362 6363 6364 6365 6366 6367 6368
        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)
6369 6370 6371 6372 6373 6374 6375 6376 6377 6378 6379 6380 6381 6382 6383 6384 6385 6386 6387 6388 6389

    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()
6390
    out = helper.create_variable_for_type_inference(dtype)
6391 6392 6393 6394 6395 6396 6397 6398 6399
    helper.append_op(
        type='pad_constant_like',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
        attrs={'pad_value': float(pad_value)})
    return out


6400 6401 6402 6403 6404 6405 6406
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
6407 6408
    called label-smoothing regularization (LSR).

6409 6410 6411 6412 6413 6414 6415 6416 6417 6418 6419 6420 6421 6422 6423 6424 6425 6426 6427 6428 6429 6430 6431
    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
6432
                              be :math:`(1, class\_num)`.
6433 6434
        epsilon(float): The weight used to mix up the original ground-truth
                        distribution and the fixed distribution.
6435
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32,
6436 6437 6438 6439 6440 6441 6442 6443 6444 6445 6446 6447 6448 6449 6450 6451 6452 6453 6454
                                                  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
6455
    smooth_label = helper.create_variable_for_type_inference(dtype)
6456 6457 6458 6459 6460 6461 6462
    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
6463 6464


W
wopeizl 已提交
6465 6466 6467 6468 6469 6470 6471 6472 6473 6474 6475 6476 6477 6478 6479 6480 6481 6482 6483 6484 6485 6486 6487 6488 6489 6490 6491 6492 6493 6494 6495 6496 6497 6498 6499 6500
@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
6501 6502


J
jerrywgz 已提交
6503 6504 6505 6506 6507 6508
@templatedoc()
def roi_align(input,
              rois,
              pooled_height=1,
              pooled_width=1,
              spatial_scale=1.0,
J
jerrywgz 已提交
6509 6510
              sampling_ratio=-1,
              name=None):
J
jerrywgz 已提交
6511 6512 6513 6514 6515 6516 6517 6518 6519 6520 6521 6522 6523 6524 6525 6526
    """
    ${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

6527 6528 6529
            align_out = fluid.layers.roi_align(input=x,
                                               rois=rois,
                                               pooled_height=7,
J
jerrywgz 已提交
6530 6531 6532 6533 6534 6535
                                               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
6536
    align_out = helper.create_variable_for_type_inference(dtype)
J
jerrywgz 已提交
6537 6538 6539 6540 6541 6542 6543 6544 6545 6546 6547 6548 6549 6550
    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
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
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:
6577 6578
        .. code-block:: python

W
whs 已提交
6579 6580 6581 6582
            predictions = fluid.layers.softmax(x)
            loss = fluid.layers.dice_loss(input=predictions, label=label, 2)
    """
    label = one_hot(label, depth=input.shape[-1])
6583
    reduce_dim = list(range(1, len(input.shape)))
W
whs 已提交
6584 6585 6586 6587 6588 6589
    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)
6590 6591


6592 6593 6594 6595
def image_resize(input,
                 out_shape=None,
                 scale=None,
                 name=None,
6596
                 resample='BILINEAR',
6597 6598
                 actual_shape=None,
                 align_corners=True,
T
tink2123 已提交
6599
                 align_mode=1):
6600
    """
Q
qiaolongfei 已提交
6601
    **Resize a Batch of Images**
F
stash  
fengjiayi 已提交
6602

6603
    The input must be a tensor of the shape (num_batches, channels, in_h, in_w),
6604 6605 6606
    and the resizing only applies on the last two dimensions(hight and width).

    Supporting resample methods:
Q
update  
qiaolongfei 已提交
6607

6608
        'BILINEAR' : Bilinear interpolation
6609

6610
        'NEAREST' : Nearest neighbor interpolation
F
stash  
fengjiayi 已提交
6611

6612 6613 6614 6615 6616 6617 6618 6619 6620 6621
    Nearest neighbor interpolation is to perform nearest neighbor interpolation
    in both the 3rd dimention(in height direction) and the 4th dimention(in width 
    direction) on input tensor.
            
    Bilinear interpolation is an extension of linear interpolation for 
    interpolating functions of two variables (e.g. H-direction and 
    W-direction in this op) on a rectilinear 2D grid. The key idea is 
    to perform linear interpolation first in one direction, and then 
    again in the other direction.

T
tink2123 已提交
6622
    Align_corners and align_mode are optinal parameters,the calculation method 
6623 6624 6625 6626
    of interpolation can be selected by them.

    Example:

T
tink2123 已提交
6627
      For scale:
6628 6629 6630 6631 6632 6633 6634 6635 6636 6637 6638 6639
      
        if align_corners = True && out_size > 1 :

          scale_factor = (in_size-1.0)/(out_size-1.0)
        
        else:
          
          scale_factor = float(in_size/out_size)
        
      
      Nearest neighbor interpolation:
      
T
tink2123 已提交
6640
      if:
6641 6642 6643 6644 6645 6646 6647 6648
          align_corners = False

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

          H_out = \left \lfloor {H_{in} * scale_{}factor}} \right \rfloor
          W_out = \left \lfloor {W_{in} * scale_{}factor}} \right \rfloor

T
tink2123 已提交
6649
      else:
6650 6651 6652 6653 6654 6655 6656 6657 6658 6659
          align_corners = True

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

          H_out = round(H_{in} * scale_{factor})
          W_out = round(W_{in} * scale_{factor})

      Bilinear interpolation:

T
tink2123 已提交
6660
      if:
6661 6662 6663 6664 6665 6666 6667 6668 6669
          align_corners = False , align_mode = 0
          
          input : (N,C,H_in,W_in)
          output: (N,C,H_out,W_out) where:
          
          H_out = (H_{in}+0.5) * scale_{factor} - 0.5
          W_out = (W_{in}+0.5) * scale_{factor} - 0.5


T
tink2123 已提交
6670
      else:
6671 6672 6673 6674 6675 6676 6677 6678 6679 6680 6681 6682 6683 6684 6685
       
          input : (N,C,H_in,W_in)
          output: (N,C,H_out,W_out) where:

          H_out = H_{in} * scale_{factor}
          W_out = W_{in} * scale_{factor}

    For details of nearest neighbor interpolation, please refer to Wikipedia: 
    https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation.

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



6686
    Args:
6687
        input (Variable): The input tensor of image resize layer,
6688 6689
                          This is a 4-D tensor of the shape
                          (num_batches, channels, in_h, in_w).
6690
        out_shape(list|tuple|Variable|None): Output shape of image resize
6691 6692
                                    layer, the shape is (out_h, out_w).
                                    Default: None
6693
        scale(float|None): The multiplier for the input height or width.
6694 6695 6696
                         At least one of out_shape or scale must be set.
                         And out_shape has a higher priority than scale.
                         Default: None
6697 6698
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
6699
        resample(str): The resample method. It supports 'BILINEAR' and 'NEAREST'
6700
                       currently.
6701
                       Default: 'BILINEAR'
6702 6703 6704
        actual_shape(Variable): An optional input to specify output shape
                                dynamically. If provided, image resize
                                according to this given shape rather than
6705
                                :attr:`out_shape` and :attr:`scale` specifying
6706 6707 6708 6709 6710 6711 6712
                                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
6713 6714
                                constructing stage.
                                Default: None
6715 6716 6717 6718
        align_corners(bool) :  An optional bool, If True, the centers of the 4 corner pixels of the 
                               input and output tensors are aligned, preserving the values at the 
                               corner pixels.
                               Default: True
T
tink2123 已提交
6719
        align_mode(int)  :  An optional for bilinear interpolation. can be \'0\' 
T
tink2123 已提交
6720 6721
                            for src_idx = scale*(dst_indx+0.5)-0.5 , can be \'1\' for 
                            src_idx = scale*dst_index .
6722 6723

    Returns:
Q
update  
qiaolongfei 已提交
6724 6725
        Variable: The output is a 4-D tensor of the shape
        (num_batches, channls, out_h, out_w).
F
stash  
fengjiayi 已提交
6726

6727 6728 6729
    Raises:
        TypeError: out_shape should be a list or tuple or Variable.
        TypeError: actual_shape should either be Variable or None.
6730
        ValueError: The 'resample' of image_resize can only be 'BILINEAR'
6731 6732 6733
                    or 'NEAREST' currently.
        ValueError: One of out_shape and scale must not be None.
        ValueError: out_shape length should be 2.
6734 6735
        TypeError: align_corners shoule be a bool value
        ValueError: align_mode can only be '0' or '1'
6736

6737 6738 6739
    Examples:
        .. code-block:: python

6740
            out = fluid.layers.image_resize(input, out_shape=[12, 12], resample="NEAREST")
6741
    """
6742 6743 6744 6745
    resample_methods = {
        'BILINEAR': 'bilinear',
        'NEAREST': 'nearest',
    }
6746 6747
    if resample not in resample_methods:
        raise ValueError(
6748
            "The 'resample' of image_resize can only be 'BILINEAR' or 'NEAREST' currently."
6749
        )
6750
    resample_type = resample_methods[resample]
6751 6752 6753 6754 6755 6756

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

6757
    if out_shape is None and scale is None:
6758
        raise ValueError("One of out_shape and scale must not be None.")
6759
    helper = LayerHelper('{}_interp'.format(resample_type), **locals())
6760
    dtype = helper.input_dtype()
6761 6762 6763 6764

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

6765 6766 6767
    out_h = 0
    out_w = 0
    inputs = {"X": input}
6768
    if out_shape is not None:
6769 6770 6771 6772
        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.")
6773
            inputs['OutSize'] = out_shape
6774 6775 6776 6777 6778 6779 6780 6781
        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]
6782 6783 6784 6785
    else:
        out_h = int(input.shape[2] * scale)
        out_w = int(input.shape[3] * scale)

6786 6787 6788 6789 6790
    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.")

6791
    out = helper.create_variable_for_type_inference(dtype)
6792
    helper.append_op(
6793
        type='{}_interp'.format(resample_type),
6794
        inputs=inputs,
6795
        outputs={"Out": out},
6796 6797 6798 6799 6800 6801 6802
        attrs={
            "out_h": out_h,
            "out_w": out_w,
            "interp_method": resample_type,
            "align_corners": align_corners,
            "align_mode": align_mode
        })
6803
    return out
F
stash  
fengjiayi 已提交
6804 6805


6806
@templatedoc(op_type="bilinear_interp")
6807 6808 6809 6810
def resize_bilinear(input,
                    out_shape=None,
                    scale=None,
                    name=None,
6811 6812
                    actual_shape=None,
                    align_corners=True,
T
tink2123 已提交
6813
                    align_mode=1):
6814
    """
6815 6816
    Resize input by performing bilinear interpolation based on given
    output shape which specified by actual_shape, out_shape and scale
6817 6818
    in priority order.

6819 6820 6821 6822
    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
6823 6824
    again in the other direction.

6825
    For details of bilinear interpolation, please refer to Wikipedia:
6826
    https://en.wikipedia.org/wiki/Bilinear_interpolation
Y
yuyang18 已提交
6827

T
tink2123 已提交
6828
    Align_corners and align_mode are optinal parameters,the calculation 
6829 6830 6831
    method of interpolation can be selected by them.


T
tink2123 已提交
6832
    Align_corners and align_mode are optinal parameters,the calculation method 
6833 6834 6835 6836
    of interpolation can be selected by them.

    Example:

T
tink2123 已提交
6837
      For scale:
6838 6839 6840 6841 6842 6843 6844 6845 6846 6847 6848
      
        if align_corners = True && out_size > 1 :

          scale_factor = (in_size-1.0)/(out_size-1.0)
        
        else:
          
          scale_factor = float(in_size/out_size)     

    Bilinear interpolation:

T
tink2123 已提交
6849
      if:
6850 6851 6852 6853 6854 6855 6856 6857 6858
          align_corners = False , align_mode = 0
          
          input : (N,C,H_in,W_in)
          output: (N,C,H_out,W_out) where:
          
          H_out = (H_{in}+0.5) * scale_{factor} - 0.5
          W_out = (W_{in}+0.5) * scale_{factor} - 0.5


T
tink2123 已提交
6859 6860
      else:

6861 6862 6863 6864 6865 6866 6867 6868
          input : (N,C,H_in,W_in)
          output: (N,C,H_out,W_out) where:

          H_out = H_{in} * scale_{factor}
          W_out = W_{in} * scale_{factor}



Y
yuyang18 已提交
6869 6870 6871 6872
    Args:
        input(${x_type}): ${x_comment}.

        out_shape(${out_size_type}): ${out_size_comment}.
6873

Y
yuyang18 已提交
6874 6875 6876 6877 6878
        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.
6879 6880 6881
        actual_shape(Variable): An optional input to specify output shape
                                dynamically. If provided, image resize
                                according to this given shape rather than
6882
                                :attr:`out_shape` and :attr:`scale` specifying
6883 6884 6885 6886 6887 6888 6889
                                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
6890 6891
                                constructing stage.
                                Default: None
6892 6893
        align_corners(bool): ${align_corners_comment}
        align_mode(bool): ${align_mode_comment}
Y
yuyang18 已提交
6894 6895 6896

    Returns:
        ${out_comment}.
6897 6898 6899 6900 6901

    Examples:
        .. code-block:: python

            out = fluid.layers.resize_bilinear(input, out_shape=[12, 12])
6902 6903
    """

6904 6905
    return image_resize(input, out_shape, scale, name, 'BILINEAR', actual_shape,
                        align_corners, align_mode)
6906 6907


6908
@templatedoc(op_type="nearest_interp")
6909 6910 6911 6912
def resize_nearest(input,
                   out_shape=None,
                   scale=None,
                   name=None,
6913 6914
                   actual_shape=None,
                   align_corners=True):
6915
    """
6916
    Resize input by performing nearest neighbor interpolation in both the
6917 6918
    3rd dimention(in height direction) and the 4th dimention(in width
    direction) based on given output shape which specified by actual_shape,
6919 6920
    out_shape and scale in priority order.

6921 6922
    Example:

T
tink2123 已提交
6923
      For scale:
6924 6925 6926 6927 6928 6929 6930 6931 6932 6933 6934 6935
      
        if align_corners = True && out_size > 1 :

          scale_factor = (in_size-1.0)/(out_size-1.0)
        
        else:
          
          scale_factor = float(in_size/out_size)
        
      
      Nearest neighbor interpolation:
      
T
tink2123 已提交
6936
      if:
6937 6938 6939 6940 6941 6942 6943 6944
          align_corners = False

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

          H_out = \left \lfloor {H_{in} * scale_{}factor}} \right \rfloor
          W_out = \left \lfloor {W_{in} * scale_{}factor}} \right \rfloor

T
tink2123 已提交
6945
      else:
6946 6947 6948 6949 6950 6951 6952 6953 6954
          align_corners = True

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

          H_out = round(H_{in} * scale_{factor})
          W_out = round(W_{in} * scale_{factor})


6955
    For details of nearest neighbor interpolation, please refer to Wikipedia:
6956
    https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation
Y
yuyang18 已提交
6957 6958 6959 6960 6961

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

        out_shape(${out_size_type}): ${out_size_comment}.
6962

Y
yuyang18 已提交
6963 6964 6965 6966 6967
        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.
6968 6969 6970
        actual_shape(Variable): An optional input to specify output shape
                                dynamically. If provided, image resize
                                according to this given shape rather than
6971
                                :attr:`out_shape` and :attr:`scale` specifying
6972 6973 6974 6975 6976 6977 6978
                                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
6979 6980
                                constructing stage.
                                Default: None
6981
        align_corners(bool): ${align_corners_comment}
Y
yuyang18 已提交
6982 6983 6984

    Returns:
        ${out_comment}.
6985 6986 6987 6988 6989

    Examples:
        .. code-block:: python

            out = fluid.layers.resize_nearest(input, out_shape=[12, 12])
6990 6991
    """

6992 6993
    return image_resize(input, out_shape, scale, name, 'NEAREST', actual_shape,
                        align_corners)
6994 6995 6996 6997


def image_resize_short(input, out_short_len, resample='BILINEAR'):
    """
6998 6999 7000
    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
7001 7002 7003 7004 7005 7006 7007
    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.
7008
        resample (str): resample method, default: BILINEAR.
F
fengjiayi 已提交
7009

7010
    Returns:
Q
update  
qiaolongfei 已提交
7011
        Variable: The output is a 4-D tensor of the shape
7012
        (num_batches, channls, out_h, out_w).
7013 7014 7015 7016 7017 7018 7019 7020 7021 7022
    """
    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
7023 7024 7025
    out_shape[long_idx] = int(
        float(out_shape[long_idx]) * (float(out_short_len) / float(hw[
            short_idx])) + 0.5)
7026 7027 7028
    return image_resize(input=input, out_shape=out_shape, resample=resample)


7029 7030
def gather(input, index):
    """
Q
qiaolongfei 已提交
7031 7032
    **Gather Layer**

7033
    Output is obtained by gathering entries of the outer-most dimension
7034 7035 7036 7037
    of X indexed by `index` and concatenate them together.

    .. math::

7038
        Out = X[Index]
7039 7040 7041 7042 7043 7044 7045


    .. code-block:: text


                Given:

7046 7047
                X = [[1, 2],
                     [3, 4],
7048 7049 7050 7051 7052 7053 7054 7055 7056 7057
                     [5, 6]]

                Index = [1, 2]

                Then:

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

    Args:
7058
        input (Variable): The source input with rank>=1.
7059 7060 7061 7062 7063 7064
        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
7065

7066 7067 7068 7069 7070 7071
        .. code-block:: python

            output = fluid.layers.gather(x, index)
    """
    helper = LayerHelper('gather', **locals())
    dtype = helper.input_dtype()
7072
    out = helper.create_variable_for_type_inference(dtype)
7073 7074 7075 7076 7077 7078 7079 7080
    helper.append_op(
        type="gather",
        inputs={"X": input,
                "Index": index},
        outputs={"Out": out})
    return out


7081 7082 7083 7084 7085 7086 7087 7088 7089 7090 7091 7092 7093 7094 7095 7096 7097 7098 7099 7100 7101 7102 7103 7104 7105 7106 7107 7108 7109 7110 7111
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()
7112
    out = helper.create_variable_for_type_inference(dtype)
7113 7114 7115 7116 7117 7118 7119 7120 7121
    helper.append_op(
        type="scatter",
        inputs={"X": input,
                "Ids": index,
                "Updates": updates},
        outputs={"Out": out})
    return out


7122 7123 7124 7125 7126 7127 7128 7129 7130
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:
7131

7132
    Given the following input:
7133

7134
    .. code-block:: text
7135

7136 7137 7138 7139 7140 7141 7142 7143 7144 7145 7146 7147
        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:
7148

7149
    .. code-block:: text
7150

7151 7152 7153 7154 7155 7156 7157 7158 7159 7160 7161 7162 7163 7164 7165
        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:
7166
        Variable: The output is a tensor with the same shape as input.
7167 7168 7169 7170 7171 7172 7173 7174 7175 7176

    Examples:

        .. code-block:: python

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

    """
    helper = LayerHelper('sequence_scatter', **locals())
    dtype = helper.input_dtype()
7177
    out = helper.create_variable_for_type_inference(dtype)
7178 7179 7180 7181 7182 7183 7184 7185 7186
    helper.append_op(
        type="sequence_scatter",
        inputs={"X": input,
                "Ids": index,
                "Updates": updates},
        outputs={"Out": out})
    return out


7187 7188 7189 7190 7191 7192 7193 7194 7195 7196 7197 7198 7199
@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}
7200

7201 7202 7203
    Examples:
        >>> img = fluid.layers.data("img", [3, 256, 256])
        >>> cropped_img = fluid.layers.random_crop(img, shape=[3, 224, 224])
7204
    """
F
stash  
fengjiayi 已提交
7205
    helper = LayerHelper("random_crop", **locals())
F
fengjiayi 已提交
7206
    dtype = x.dtype
7207
    out = helper.create_variable_for_type_inference(dtype)
7208
    if seed is None:
7209
        seed = np.random.randint(-65536, 65536)
F
fengjiayi 已提交
7210
    op_attrs = {"shape": shape}
F
stash  
fengjiayi 已提交
7211
    if isinstance(seed, int):
F
fengjiayi 已提交
7212 7213 7214 7215 7216
        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
7217 7218 7219 7220
    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
7221
        inputs={"X": x,
F
stash  
fengjiayi 已提交
7222 7223
                "Seed": seed},
        outputs={"Out": out,
F
fengjiayi 已提交
7224 7225
                 "SeedOut": seed},
        attrs=op_attrs)
F
stash  
fengjiayi 已提交
7226
    return out
W
whs 已提交
7227 7228


7229
def log(x, name=None):
7230 7231 7232 7233 7234
    """
    Calculates the natural log of the given input tensor, element-wise.

    .. math::

7235
        Out = \\ln(x)
7236 7237

    Args:
7238
        x (Variable): Input tensor.
7239 7240
        name (str|None, default None): A name for this layer If set None,
            the layer will be named automatically.
7241 7242 7243 7244 7245 7246 7247 7248

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

    Examples:

        .. code-block:: python

7249
            output = fluid.layers.log(x)
7250 7251
    """
    helper = LayerHelper('log', **locals())
W
wanghaoshuang 已提交
7252
    dtype = helper.input_dtype(input_param_name='x')
7253
    out = helper.create_variable_for_type_inference(dtype)
W
wanghaoshuang 已提交
7254
    helper.append_op(type="log", inputs={"X": x}, outputs={"Out": out})
7255 7256 7257
    return out


7258
def relu(x, name=None):
7259 7260
    """
    Relu takes one input data (Tensor) and produces one output data (Tensor)
7261
    where the rectified linear function, y = max(0, x), is applied to
7262 7263 7264 7265
    the tensor elementwise.

    .. math::

7266
        Out = \\max(0, x)
7267 7268

    Args:
7269
        x (Variable): The input tensor.
7270 7271
        name (str|None, default None): A name for this layer If set None,
            the layer will be named automatically.
7272 7273 7274 7275 7276 7277 7278 7279

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

    Examples:

        .. code-block:: python

7280
            output = fluid.layers.relu(x)
7281 7282
    """
    helper = LayerHelper('relu', **locals())
W
wanghaoshuang 已提交
7283
    dtype = helper.input_dtype(input_param_name='x')
7284
    out = helper.create_variable_for_type_inference(dtype)
X
Xin Pan 已提交
7285 7286
    helper.append_op(
        type="relu", inputs={"X": helper.input('x')}, outputs={"Out": out})
7287
    return out
7288 7289


C
chengduo 已提交
7290 7291 7292 7293 7294 7295 7296 7297 7298 7299 7300 7301 7302 7303 7304 7305 7306 7307 7308 7309 7310 7311 7312 7313 7314 7315 7316 7317 7318 7319 7320 7321 7322 7323 7324 7325 7326 7327 7328 7329 7330
@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
7331 7332 7333
def mean_iou(input, label, num_classes):
    """
    Mean Intersection-Over-Union is a common evaluation metric for
7334 7335 7336 7337
    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
7338
    .. math::
7339

7340
        IOU = \\frac{true\_positive}{(true\_positive + false\_positive + false\_negative)}.
W
whs 已提交
7341

7342
    The predictions are accumulated in a confusion matrix and mean-IOU
W
whs 已提交
7343 7344 7345 7346 7347
    is then calculated from it.


    Args:
        input (Variable): A Tensor of prediction results for semantic labels with type int32 or int64.
7348
        label (Variable): A Tensor of ground truth labels with type int32 or int64.
W
whs 已提交
7349
                           Its shape should be the same as input.
7350
        num_classes (int): The possible number of labels.
W
whs 已提交
7351 7352

    Returns:
M
minqiyang 已提交
7353 7354
        mean_iou (Variable),out_wrong(Variable),out_correct(Variable):

7355
                     Three variables:
M
minqiyang 已提交
7356

7357 7358 7359
                     - mean_iou : A Tensor representing the mean intersection-over-union with shape [1].
                     - out_wrong: A Tensor with shape [num_classes]. The wrong numbers of each class.
                     - out_correct: A Tensor with shape [num_classes]. The correct numbers of each class.
W
whs 已提交
7360 7361 7362 7363

    Examples:

        .. code-block:: python
7364

W
whs 已提交
7365 7366 7367 7368
            iou, wrongs, corrects = fluid.layers.mean_iou(predict, label, num_classes)
    """
    helper = LayerHelper('mean_iou', **locals())
    dtype = helper.input_dtype()
7369 7370 7371
    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
7372 7373
    helper.append_op(
        type="mean_iou",
7374 7375
        inputs={"Predictions": input,
                "Labels": label},
W
whs 已提交
7376
        outputs={
7377 7378 7379
            "OutMeanIou": out_mean_iou,
            "OutWrong": out_wrong,
            "OutCorrect": out_correct
W
whs 已提交
7380 7381 7382
        },
        attrs={"num_classes": num_classes})
    return out_mean_iou, out_wrong, out_correct
7383 7384 7385 7386 7387 7388 7389 7390 7391 7392 7393 7394 7395 7396 7397 7398 7399 7400 7401 7402 7403 7404 7405 7406 7407 7408 7409 7410 7411 7412 7413 7414 7415 7416 7417 7418 7419 7420 7421 7422 7423 7424 7425 7426 7427 7428 7429 7430 7431 7432 7433 7434 7435 7436 7437 7438 7439 7440 7441 7442 7443 7444 7445 7446 7447 7448 7449 7450


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")
7451
            crop = fluid.layers.crop(z, shape=[-1, 2, 3])
7452 7453 7454 7455 7456

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

    if not (isinstance(shape, list) or isinstance(shape, tuple) or \
7457
            isinstance(shape, Variable)):
7458 7459 7460 7461 7462
        raise ValueError("The shape should be a list, tuple or Variable.")

    if offsets is None:
        offsets = [0] * len(x.shape)

7463
    out = helper.create_variable_for_type_inference(x.dtype)
7464 7465 7466 7467 7468 7469 7470 7471 7472 7473 7474 7475 7476 7477 7478 7479 7480
    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
7481 7482


7483 7484 7485 7486 7487 7488 7489 7490 7491 7492 7493 7494 7495 7496 7497 7498 7499
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]]]
7500

7501
              out_shape = [2, 3, 5, 5]
7502

7503
          Step 1:
7504

7505 7506 7507
              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:
7508

7509 7510 7511 7512 7513 7514 7515 7516 7517 7518 7519 7520 7521 7522 7523 7524 7525 7526 7527 7528 7529 7530 7531 7532 7533 7534 7535 7536 7537 7538 7539 7540 7541 7542 7543 7544 7545 7546 7547 7548 7549 7550 7551 7552 7553
              C = [[[-1.  -1.  -1.  -1.  -1. ]
                    [-0.5 -0.5 -0.5 -0.5 -0.5]
                    [ 0.   0.   0.   0.   0. ]
                    [ 0.5  0.5  0.5  0.5  0.5]
                    [ 1.   1.   1.   1.   1. ]]
                   [[-1.  -0.5  0.   0.5  1. ]
                    [-1.  -0.5  0.   0.5  1. ]
                    [-1.  -0.5  0.   0.5  1. ]
                    [-1.  -0.5  0.   0.5  1. ]
                    [-1.  -0.5  0.   0.5  1. ]]]
              C[0] is the coordinates in height axis and  C[1] is the coordinates in width axis.

          Step2:

              Tanspose and reshape C to shape [H * W, 2] and append ones to last dimension. The we get:
              C_ = [[-1.  -1.   1. ]
                    [-0.5 -1.   1. ]
                    [ 0.  -1.   1. ]
                    [ 0.5 -1.   1. ]
                    [ 1.  -1.   1. ]
                    [-1.  -0.5  1. ]
                    [-0.5 -0.5  1. ]
                    [ 0.  -0.5  1. ]
                    [ 0.5 -0.5  1. ]
                    [ 1.  -0.5  1. ]
                    [-1.   0.   1. ]
                    [-0.5  0.   1. ]
                    [ 0.   0.   1. ]
                    [ 0.5  0.   1. ]
                    [ 1.   0.   1. ]
                    [-1.   0.5  1. ]
                    [-0.5  0.5  1. ]
                    [ 0.   0.5  1. ]
                    [ 0.5  0.5  1. ]
                    [ 1.   0.5  1. ]
                    [-1.   1.   1. ]
                    [-0.5  1.   1. ]
                    [ 0.   1.   1. ]
                    [ 0.5  1.   1. ]
                    [ 1.   1.   1. ]]
          Step3:
              Compute output by equation $$Output[i] = C_ * Theta[i]^T$$

    Args:
        theta (Variable): A batch of affine transform parameters with shape [N, 2, 3].
M
minqiyang 已提交
7554
        out_shape (Variable | list | tuple): The shape of target output with format [N, C, H, W].
7555
                                             ``out_shape`` can be a Variable or a list or tuple.
7556 7557 7558 7559 7560 7561 7562 7563 7564 7565 7566 7567
        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
7568

7569 7570 7571 7572 7573 7574 7575 7576 7577 7578 7579
            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 \
7580
            isinstance(out_shape, Variable)):
7581 7582 7583 7584 7585 7586 7587 7588 7589 7590 7591 7592 7593 7594 7595 7596 7597 7598 7599 7600 7601
        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


7602 7603
def rank_loss(label, left, right, name=None):
    """
7604

7605 7606
    **Rank loss layer for RankNet**

7607
    `RankNet <http://icml.cc/2015/wp-content/uploads/2015/06/icml_ranking.pdf>`_
7608 7609 7610
    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
7611

7612 7613
    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
7614

7615 7616
    Rank loss layer takes three inputs: left ( :math:`o_i` ), right ( :math:`o_j` ) and
    label ( :math:`P_{i,j}` ). The inputs respectively represent RankNet's output scores
7617 7618
    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
7619

7620 7621 7622 7623 7624 7625 7626 7627
    .. math::

      C_{i,j} &= -\\tilde{P_{ij}} * o_{i,j} + \log(1 + e^{o_{i,j}}) \\\\

      o_{i,j} &=  o_i - o_j  \\\\

      \\tilde{P_{i,j}} &= \\left \{0, 0.5, 1 \\right \} \ or \ \\left \{0, 1 \\right \}

M
minqiyang 已提交
7628 7629 7630

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

7631 7632 7633 7634 7635 7636 7637 7638 7639 7640 7641 7642 7643 7644 7645 7646 7647 7648 7649 7650 7651 7652 7653 7654 7655 7656 7657 7658 7659 7660 7661 7662 7663 7664
    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")

7665
    out = helper.create_variable_for_type_inference("float32")
7666 7667 7668 7669 7670 7671 7672 7673

    helper.append_op(
        type='rank_loss',
        inputs={"Label": label,
                "Left": left,
                "Right": right},
        outputs={'Out': out})
    return out
7674 7675


7676 7677
def margin_rank_loss(label, left, right, margin=0.1, name=None):
    """
M
minqiyang 已提交
7678
    Margin Ranking Loss Layer for ranking problem,
M
minqiyang 已提交
7679
    which compares left score and right score passed in.
M
minqiyang 已提交
7680
    The ranking loss can be defined as following equation:
M
minqiyang 已提交
7681 7682 7683

    .. math::

7684
        rank\_loss = max(0, -label * (left - right) + margin)
M
minqiyang 已提交
7685 7686

    Args:
M
minqiyang 已提交
7687
       label (Variable): Indicates whether the left is ranked higher than the right or not.
M
minqiyang 已提交
7688 7689
       left (Variable): Ranking score for left.
       right (Variable): Ranking score for right.
M
minqiyang 已提交
7690
       margin (float): Indicates the given margin.
M
minqiyang 已提交
7691 7692
       name (str|None): A name for this layer (optional). If set None, the layer
                       will be named automatically.
7693

M
minqiyang 已提交
7694
    Returns:
M
minqiyang 已提交
7695
       Variable: The ranking loss.
7696

M
minqiyang 已提交
7697
    Raises:
M
minqiyang 已提交
7698
       ValueError: Any of label, left, and right is not a Variable.
7699

M
minqiyang 已提交
7700
    Examples:
7701

M
minqiyang 已提交
7702
        .. code-block:: python
7703

M
minqiyang 已提交
7704 7705 7706 7707 7708
           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)
    """
7709
    helper = LayerHelper('margin_rank_loss', **locals())
M
minqiyang 已提交
7710 7711 7712 7713 7714 7715
    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.")
7716 7717
    out = helper.create_variable_for_type_inference(left.dtype)
    act = helper.create_variable_for_type_inference(left.dtype)
7718 7719 7720 7721 7722 7723 7724 7725 7726 7727 7728
    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
7729 7730 7731 7732 7733 7734 7735 7736 7737 7738 7739 7740
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:
7741
        .. code-block:: text
W
whs 已提交
7742

7743
	      Given that X is a channel of image from input:
M
minqiyang 已提交
7744

7745 7746
	      X = [[1, 2, 3],
		   [4, 5, 6]]
M
minqiyang 已提交
7747

7748
	      Case 0:
M
minqiyang 已提交
7749

7750 7751 7752
		paddings = [0, 1, 2, 3],
		mode = 'constant'
		pad_value = 0
M
minqiyang 已提交
7753

7754 7755 7756
		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
7757

7758
	      Case 1:
M
minqiyang 已提交
7759

7760 7761
		paddings = [0, 1, 2, 1],
		mode = 'reflect'
M
minqiyang 已提交
7762

7763 7764 7765
		Out = [[3, 2, 1, 2, 3, 2]
		       [6, 5, 4, 5, 6, 5]
		       [3, 2, 1, 2, 3, 2]]
M
minqiyang 已提交
7766

7767
	      Case 2:
M
minqiyang 已提交
7768

7769 7770
		paddings = [0, 1, 2, 1],
		mode = 'edge'
M
minqiyang 已提交
7771

7772 7773 7774
		Out = [[1, 1, 1, 2, 3, 3]
		       [4, 4, 4, 5, 6, 6]
		       [4, 4, 4, 5, 6, 6]]
M
minqiyang 已提交
7775 7776


W
whs 已提交
7777 7778
    Args:
        input (Variable): The input image with [N, C, H, W] format or [N, H, W, C] format.
7779
        paddings (tuple|list|Variable): The padding size. If padding is a tuple, it must
W
whs 已提交
7780 7781 7782 7783 7784 7785 7786 7787 7788 7789 7790 7791 7792 7793 7794 7795 7796 7797 7798 7799 7800 7801 7802
            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')
7803
    out = helper.create_variable_for_type_inference(dtype)
7804 7805 7806 7807 7808 7809 7810 7811 7812
    inputs = {'X': input}
    attrs = {'mode': mode, 'pad_value': pad_value, 'data_format': data_format}

    if isinstance(paddings, Variable):
        inputs['Paddings'] = paddings
        attrs['paddings'] = []
    else:
        attrs['paddings'] = paddings

W
whs 已提交
7813
    helper.append_op(
7814
        type='pad2d', inputs=inputs, outputs={"Out": out}, attrs=attrs)
W
whs 已提交
7815 7816 7817 7818

    return out


7819 7820 7821 7822 7823 7824 7825 7826 7827 7828 7829 7830
@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}
7831 7832 7833 7834 7835

    Examples:

        .. code-block:: python

Z
ZhenWang 已提交
7836 7837
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
            y = fluid.layers.elu(x, alpha=0.2)
7838 7839
    """
    helper = LayerHelper('elu', **locals())
7840
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
7841 7842 7843 7844 7845 7846 7847 7848 7849 7850 7851 7852 7853 7854 7855 7856 7857 7858 7859 7860
    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}
7861 7862 7863 7864 7865

    Examples:

        .. code-block:: python

Z
ZhenWang 已提交
7866 7867
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
            y = fluid.layers.relu6(x, threshold=6.0)
7868 7869
    """
    helper = LayerHelper('relu6', **locals())
7870
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
7871 7872 7873 7874 7875 7876 7877 7878 7879 7880 7881 7882 7883 7884 7885 7886 7887 7888 7889 7890
    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}
7891 7892 7893 7894 7895

    Examples:

        .. code-block:: python

Z
ZhenWang 已提交
7896 7897
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
            y = fluid.layers.pow(x, factor=2.0)
7898 7899
    """
    helper = LayerHelper('pow', **locals())
7900
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
7901 7902 7903 7904 7905 7906 7907 7908 7909 7910 7911 7912 7913 7914 7915 7916 7917 7918 7919 7920 7921
    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}
7922 7923 7924 7925 7926

    Examples:

        .. code-block:: python

Z
ZhenWang 已提交
7927
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
Z
ZhenWang 已提交
7928
            y = fluid.layers.stanh(x, scale_a=0.67, scale_b=1.72)
7929 7930
    """
    helper = LayerHelper('stanh', **locals())
7931
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
7932 7933 7934 7935 7936 7937 7938 7939 7940 7941 7942 7943 7944 7945 7946 7947 7948 7949 7950 7951 7952 7953
    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}
7954 7955 7956 7957 7958

    Examples:

        .. code-block:: python

Z
ZhenWang 已提交
7959 7960
            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)
7961 7962
    """
    helper = LayerHelper('hard_sigmoid', **locals())
7963
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
7964 7965 7966 7967 7968 7969 7970 7971 7972 7973 7974 7975 7976 7977 7978 7979 7980 7981 7982 7983 7984
    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}
7985 7986 7987 7988 7989

    Examples:

        .. code-block:: python

Z
ZhenWang 已提交
7990 7991
            x = fluid.layers.data(name="x", shape=[3,10,32,32], dtype="float32")
            y = fluid.layers.swish(x, beta=2.0)
7992 7993
    """
    helper = LayerHelper('swish', **locals())
7994
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
7995 7996 7997 7998 7999 8000 8001 8002
    helper.append_op(
        type='swish',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'slope': beta})
    return out


8003 8004 8005 8006
def prelu(x, mode, param_attr=None, name=None):
    """
    Equation:

8007 8008
    .. math::
        y = \max(0, x) + \\alpha * \min(0, x)
8009 8010 8011

    Args:
        x (Variable): The input tensor.
8012
        param_attr(ParamAttr|None): The parameter attribute for the learnable
8013
          weight (alpha).
8014
        mode (string): The mode for weight sharing. It supports all, channel
8015 8016 8017
          and element. all: all elements share same weight
          channel:elements in a channel share same weight
          element:each element has a weight
8018
        name(str|None): A name for this layer(optional). If set None, the layer
8019
          will be named automatically.
8020 8021 8022 8023 8024 8025 8026 8027

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

    Examples:

        .. code-block:: python

8028
            x = fluid.layers.data(name="x", shape=[10,10], dtype="float32")
8029 8030 8031 8032 8033 8034 8035 8036 8037 8038 8039 8040 8041
            mode = 'channel'
            output = fluid.layers.prelu(x,mode)
    """
    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
8042
        attr=helper.param_attr,
8043 8044 8045 8046
        shape=alpha_shape,
        dtype='float32',
        is_bias=False,
        default_initializer=Constant(1.0))
8047
    out = helper.create_variable_for_type_inference(dtype)
8048 8049 8050 8051 8052 8053 8054 8055 8056
    helper.append_op(
        type="prelu",
        inputs={"X": x,
                'Alpha': alpha},
        attrs={"mode": mode},
        outputs={"Out": out})
    return out


8057 8058 8059 8060 8061 8062 8063 8064 8065 8066
@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): ${t_max_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
8067
    Returns:
8068
        output(${out_type}): ${out_comment}
8069 8070 8071

    Examples:

8072
    .. code-block:: python
8073

8074 8075
            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)
8076 8077
    """
    helper = LayerHelper('brelu', **locals())
8078
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
8079 8080 8081 8082 8083 8084 8085 8086 8087 8088 8089 8090 8091 8092 8093 8094 8095 8096
    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): ${alpha_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
8097
    Returns:
8098
        output(${out_type}): ${out_comment}
8099 8100 8101 8102 8103

    Examples:

        .. code-block:: python

8104 8105
            x = fluid.layers.data(name="x", shape=[2,3,16,16], dtype="float32")
            y = fluid.layers.leaky_relu(x, alpha=0.01)
8106 8107
    """
    helper = LayerHelper('leaky_relu', **locals())
8108
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
8109 8110 8111 8112 8113 8114 8115 8116 8117 8118 8119 8120 8121 8122 8123 8124 8125
    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): ${threshold_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
8126
    Returns:
8127
        output(${out_type}): ${out_comment}
8128 8129 8130 8131 8132

    Examples:

        .. code-block:: python

8133 8134
            x = fluid.layers.data(name="x", shape=[2,3,16,16], dtype="float32")
            y = fluid.layers.soft_relu(x, threshold=20.0)
8135 8136
    """
    helper = LayerHelper('soft_relu', **locals())
8137
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
8138 8139 8140 8141 8142 8143 8144 8145
    helper.append_op(
        type='soft_relu',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'threshold': threshold})
    return out


8146 8147 8148 8149
def flatten(x, axis=1, name=None):
    """
    **Flatten layer**
    Flattens the input tensor into a 2D matrix.
M
minqiyang 已提交
8150

8151
    For Example:
M
minqiyang 已提交
8152

8153
    .. code-block:: text
8154

8155 8156 8157 8158 8159 8160 8161 8162 8163 8164 8165 8166 8167 8168 8169 8170 8171 8172 8173 8174 8175
        Case 1:

          Given
            X.shape = (3, 100, 100, 4)

          and
            axis = 2

          We get:
            Out.shape = (3 * 100, 4 * 100)

        Case 2:

          Given
            X.shape = (3, 100, 100, 4)

          and
            axis = 0

          We get:
            Out.shape = (1, 3 * 100 * 100 * 4)
8176 8177 8178

    Args:
        x (Variable): A tensor of rank >= axis.
8179 8180
        axis (int): Indicate up to which input dimensions (exclusive) should
                    be flattened to the outer dimension of the output.
8181 8182 8183 8184 8185 8186 8187 8188
                    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:
8189 8190 8191
        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 \
8192 8193 8194 8195
                  inner dimension of the output.

    Raises:
        ValueError: If x is not a variable.
8196
        ValueError: If axis is not in range [0, rank(x)].
8197 8198 8199 8200 8201 8202 8203 8204 8205 8206 8207 8208 8209 8210 8211 8212

    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)]")

8213 8214
    out = helper.create_variable_for_type_inference(x.dtype)
    x_shape = helper.create_variable_for_type_inference(x.dtype)
8215
    helper.append_op(
8216
        type='flatten2',
8217
        inputs={"X": x},
8218 8219
        outputs={'Out': out,
                 'XShape': x_shape},
8220 8221
        attrs={"axis": axis})
    return out
X
Xin Pan 已提交
8222 8223


8224
def sequence_enumerate(input, win_size, pad_value=0, name=None):
8225
    """
8226
    Generate a new sequence for the input index sequence, which enumerates all the
M
minqiyang 已提交
8227
    sub-sequences with length `win_size` of the input.
C
chenweihang 已提交
8228 8229
    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
8230

8231 8232 8233 8234 8235 8236 8237 8238 8239 8240 8241 8242 8243 8244 8245 8246 8247
    .. code-block:: text

        Case 1:

          Input:
            X.lod = [[0, 3, 5]]
            X.data = [[1], [2], [3], [4], [5]]
            X.dims = [5, 1]

          Attrs:
            win_size = 2
            pad_value = 0

          Output:
            Out.lod = [[0, 3, 5]]
            Out.data = [[1, 2], [2, 3], [3, 0], [4, 5], [5, 0]]
            Out.dims = [5, 2]
8248 8249

    Args:
8250 8251 8252
        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.
8253 8254 8255 8256 8257 8258 8259 8260 8261 8262 8263

    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())
8264 8265
    out = helper.create_variable_for_type_inference(
        helper.input_dtype(), stop_gradient=True)
8266 8267 8268 8269 8270 8271
    helper.append_op(
        type='sequence_enumerate',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={'win_size': win_size,
               'pad_value': pad_value})
8272
    return out
8273

8274

8275 8276 8277 8278 8279 8280 8281 8282 8283
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:
8284

8285
    .. math::
8286

8287 8288 8289
        y(i_1, i_2,..., i_n, j) = (j < x(i_1, i_2,..., i_n))

    Args:
8290
        x (Variable): Input tensor of sequence_mask layer,
8291 8292 8293 8294
                      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.
8295 8296 8297
        name (str|None): A name for this layer(optional). If set None, the
                         layer will be named automatically.

8298 8299
    Returns:
        Variable: The output sequence mask.
8300

8301 8302
    """

8303
    helper = LayerHelper('sequence_mask', **locals())
8304
    if name is None:
8305
        out = helper.create_variable_for_type_inference(dtype=dtype)
8306
    else:
8307
        out = helper.create_variable_for_type_inference(dtype=dtype, name=name)
8308

8309 8310 8311
    helper.append_op(
        type='sequence_mask',
        inputs={'X': [x]},
8312 8313
        outputs={'Y': out},
        attrs={
8314
            'maxlen': maxlen if maxlen is not None else -1,
8315 8316 8317
            'out_dtype': out.dtype
        })
    return out
S
sneaxiy 已提交
8318 8319


X
Xin Pan 已提交
8320
def stack(x, axis=0):
S
sneaxiy 已提交
8321 8322 8323 8324
    """
    **Stack Layer**

    This layer stacks all of the input :code:`x` along axis.
8325 8326 8327 8328 8329 8330 8331

    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
8332
    If :code:`axis` < 0, it would be replaced with :code:`axis+rank(x[0])+1`.
8333
    If :code:`axis` is None, it would be replaced with 0.
S
sneaxiy 已提交
8334

C
chengduozh 已提交
8335 8336
    For Example:

C
chengduozh 已提交
8337 8338 8339 8340 8341 8342 8343 8344 8345 8346 8347 8348 8349 8350 8351 8352 8353 8354 8355 8356 8357 8358 8359 8360 8361 8362 8363 8364 8365 8366 8367 8368 8369 8370 8371 8372 8373 8374
    .. code-block:: text

        Case 1:
          Input:
            x[0].data = [ [1.0 , 2.0 ] ]
            x[0].dims = [1, 2]
            x[1].data = [ [3.0 , 4.0 ] ]
            x[1].dims = [1, 2]
            x[2].data = [ [5.0 , 6.0 ] ]
            x[2].dims = [1, 2]

          Attrs:
            axis = 0

          Output:
            Out.data =[ [ [1.0, 2.0] ],
                        [ [3.0, 4.0] ],
                        [ [5.0, 6.0] ] ]
            Out.dims = [3, 1, 2]

        Case 2:
          Given
            x[0].data = [ [1.0 , 2.0 ] ]
            x[0].dims = [1, 2]
            x[1].data = [ [3.0 , 4.0 ] ]
            x[1].dims = [1, 2]
            x[2].data = [ [5.0 , 6.0 ] ]
            x[2].dims = [1, 2]

          Attrs:
            axis = 1 or axis = -2

          Output:
            Out.data =[ [ [1.0, 2.0]
                          [3.0, 4.0]
                          [5.0, 6.0] ] ]
            Out.dims = [1, 3, 2]

S
sneaxiy 已提交
8375
    Args:
8376
        x (Variable|list(Variable)|tuple(Variable)): Input variables.
S
sneaxiy 已提交
8377
        axis (int|None): The axis along which all inputs are stacked.
8378

S
sneaxiy 已提交
8379 8380
    Returns:
        Variable: The stacked variable.
8381

S
sneaxiy 已提交
8382 8383
    """

X
Xin Pan 已提交
8384 8385 8386 8387 8388 8389
    helper = LayerHelper('stack', **locals())
    axis = 0 if axis is None else axis

    if not isinstance(x, list) and not isinstance(x, tuple):
        x = [x]

8390
    out = helper.create_variable_for_type_inference(x[0].dtype)
X
Xin Pan 已提交
8391
    helper.append_op(
S
sneaxiy 已提交
8392 8393
        type='stack', inputs={'X': x}, outputs={'Y': out},
        attrs={'axis': axis})
8394

X
Xin Pan 已提交
8395
    return out
D
dzhwinter 已提交
8396 8397 8398 8399 8400 8401 8402


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
8403

D
dzhwinter 已提交
8404 8405 8406
    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
8407
    raised.
D
dzhwinter 已提交
8408 8409

    Args:
M
minqiyang 已提交
8410
        x (Variable): Input variable.
D
dzhwinter 已提交
8411 8412
        axis (int): The axis along which the input is unstacked.
        num (int|None): The number of output variables.
M
minqiyang 已提交
8413

D
dzhwinter 已提交
8414 8415
    Returns:
        list(Variable): The unstacked variables.
M
minqiyang 已提交
8416

D
dzhwinter 已提交
8417 8418 8419 8420 8421 8422 8423 8424 8425 8426
    """

    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 = []
8427
    for _ in range(num):
8428
        outs.append(helper.create_variable_for_type_inference(x.dtype))
D
dzhwinter 已提交
8429 8430 8431 8432 8433 8434 8435 8436

    helper.append_op(
        type='unstack',
        inputs={'X': [x]},
        outputs={'Y': outs},
        attrs={'axis': axis,
               'num': num})
    return outs
8437 8438 8439 8440 8441 8442 8443 8444 8445 8446 8447 8448


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
8449

8450 8451 8452 8453
                [
                   [[1], [2], [3]],
                   [[4], [5], [6]]
                ]
M
minqiyang 已提交
8454

8455
        Attr(expand_times):  [1, 2, 2]
M
minqiyang 已提交
8456

8457
        Output(Out) is a 3-D tensor with shape [2, 6, 2]:
M
minqiyang 已提交
8458

8459 8460 8461 8462
                [
                    [[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
8463

8464 8465 8466 8467 8468 8469 8470 8471 8472 8473 8474 8475 8476 8477 8478 8479
    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')
8480
    out = helper.create_variable_for_type_inference(dtype)
8481 8482 8483 8484 8485 8486
    helper.append_op(
        type='expand',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'expand_times': expand_times})
    return out
S
sneaxiy 已提交
8487 8488


G
fix  
gongweibao 已提交
8489 8490 8491
from paddle.fluid.framework import convert_np_dtype_to_dtype_


G
gongweibao 已提交
8492
@templatedoc()
G
fix  
gongweibao 已提交
8493 8494 8495 8496 8497 8498 8499 8500 8501
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
8502
    ${comment}
G
fix  
gongweibao 已提交
8503 8504

    Args:
G
gongweibao 已提交
8505 8506 8507
        input (Variable): ${input_comment}
        shape (tuple|list): ${shape_comment}
        input_dim_idx (Int): ${input_dim_idx_comment}
G
gongweibao 已提交
8508
        output_dim_idx (Int): ${output_dim_idx_comment}
G
gongweibao 已提交
8509 8510 8511
        min (Float): ${min_comment}
        max (Float): ${max_comment}
        seed (Int): ${seed_comment}
G
fix  
gongweibao 已提交
8512 8513
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc
    Returns:
G
gongweibao 已提交
8514
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
8515

8516 8517 8518 8519 8520
    Examples:
        .. code-block:: python

            input = layers.data(name="input", shape=[13, 11], dtype='float32')
            out = layers.uniform_random_batch_size_like(input, [-1, 11])
G
fix  
gongweibao 已提交
8521 8522 8523
    """

    helper = LayerHelper('uniform_random_batch_size_like', **locals())
8524
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
8525 8526 8527 8528 8529 8530 8531 8532 8533 8534 8535 8536 8537 8538 8539 8540
    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
8541 8542


G
gongweibao 已提交
8543
@templatedoc()
X
Xin Pan 已提交
8544
def gaussian_random(shape, mean=0.0, std=1.0, seed=0, dtype='float32'):
G
fix  
gongweibao 已提交
8545
    """
G
gongweibao 已提交
8546
    ${comment}
G
fix  
gongweibao 已提交
8547 8548

    Args:
G
gongweibao 已提交
8549 8550 8551 8552
        shape (tuple|list): ${shape_comment}
        mean (Float): ${mean_comment}
        std (Float): ${std_comment}
        seed (Int): ${seed_comment}
G
fix  
gongweibao 已提交
8553 8554 8555
        dtype(np.dtype|core.VarDesc.VarType|str): Output data type.

    Returns:
G
gongweibao 已提交
8556
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
8557

8558 8559 8560 8561
    Examples:
        .. code-block:: python

            out = layers.gaussian_random(shape=[20, 30])
G
fix  
gongweibao 已提交
8562 8563 8564
    """

    helper = LayerHelper('gaussian_random', **locals())
8565
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
8566 8567 8568 8569 8570 8571 8572 8573 8574 8575
    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
8576
            'use_mkldnn': False
G
fix  
gongweibao 已提交
8577 8578 8579 8580 8581
        })

    return out


G
gongweibao 已提交
8582
@templatedoc()
G
fix  
gongweibao 已提交
8583
def sampling_id(x, min=0.0, max=1.0, seed=0, dtype='float32'):
G
fix  
gongweibao 已提交
8584
    """
G
gongweibao 已提交
8585
    ${comment}
G
fix  
gongweibao 已提交
8586 8587

    Args:
G
gongweibao 已提交
8588 8589 8590 8591
        x (Variable): ${x_comment}
        min (Float): ${min_comment}
        max (Float): ${max_comment}
        seed (Float): ${seed_comment}
G
fix  
gongweibao 已提交
8592
        dtype(np.dtype|core.VarDesc.VarType|str): The type of output data : float32, float_16, int etc
G
fix  
gongweibao 已提交
8593 8594

    Returns:
G
gongweibao 已提交
8595
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
8596

8597 8598 8599 8600 8601 8602 8603 8604 8605 8606
    Examples:
        .. code-block:: python

            x = layers.data(
                name="X",
                shape=[13, 11],
                dtype='float32',
                append_batch_size=False)

            out = layers.sampling_id(x)
G
fix  
gongweibao 已提交
8607 8608 8609
    """

    helper = LayerHelper('sampling_id', **locals())
8610
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
8611 8612 8613 8614 8615 8616 8617 8618 8619 8620 8621
    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
8622
@templatedoc()
G
fix  
gongweibao 已提交
8623 8624 8625 8626 8627 8628 8629 8630 8631
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
8632
    ${comment}
G
fix  
gongweibao 已提交
8633 8634

    Args:
G
gongweibao 已提交
8635 8636
        input (Variable): ${input_comment}
        shape (tuple|list): ${shape_comment}
G
gongweibao 已提交
8637
        input_dim_idx (Int): ${input_dim_idx_comment}
G
gongweibao 已提交
8638 8639 8640 8641
        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
8642
        dtype(np.dtype|core.VarDesc.VarType|str): The type of output data : float32, float_16, int etc
G
fix  
gongweibao 已提交
8643 8644

    Returns:
G
gongweibao 已提交
8645
        out (Variable): ${out_comment}
8646 8647 8648 8649 8650 8651 8652 8653

    Examples:
        .. code-block:: python

            input = layers.data(name="input", shape=[13, 11], dtype='float32')

            out = layers.gaussian_random_batch_size_like(
                input, shape=[-1, 11], mean=1.0, std=2.0)
G
fix  
gongweibao 已提交
8654 8655 8656
    """

    helper = LayerHelper('gaussian_random_batch_size_like', **locals())
8657
    out = helper.create_variable_for_type_inference(dtype)
G
fix  
gongweibao 已提交
8658 8659 8660 8661 8662 8663 8664 8665 8666 8667 8668 8669 8670 8671 8672 8673 8674 8675
    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
8676
@templatedoc()
X
Xin Pan 已提交
8677
def sum(x):
G
fix  
gongweibao 已提交
8678
    """
G
gongweibao 已提交
8679
    ${comment}
G
fix  
gongweibao 已提交
8680 8681

    Args:
G
gongweibao 已提交
8682
        x (Variable): ${x_comment}
G
fix  
gongweibao 已提交
8683 8684

    Returns:
G
gongweibao 已提交
8685
        out (Variable): ${out_comment}
8686 8687 8688 8689 8690 8691

    Examples:
        .. code-block:: python

            input = layers.data(name="input", shape=[13, 11], dtype='float32')
            out = layers.sum(input)
G
fix  
gongweibao 已提交
8692 8693 8694
    """

    helper = LayerHelper('sum', **locals())
8695 8696
    out = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype('x'))
G
fix  
gongweibao 已提交
8697 8698 8699 8700
    helper.append_op(
        type='sum',
        inputs={'X': x},
        outputs={'Out': out},
X
Xin Pan 已提交
8701
        attrs={'use_mkldnn': False})
G
fix  
gongweibao 已提交
8702 8703 8704 8705

    return out


G
gongweibao 已提交
8706
@templatedoc()
G
fix  
gongweibao 已提交
8707 8708
def slice(input, axes, starts, ends):
    """
G
gongweibao 已提交
8709
    ${comment}
G
fix  
gongweibao 已提交
8710 8711

    Args:
G
gongweibao 已提交
8712 8713 8714 8715
        input (Variable): ${input_comment}.
        axes (List): ${axes_comment}
        starts (List): ${starts_comment}
        ends (List): ${ends_comment}
G
fix  
gongweibao 已提交
8716 8717

    Returns:
G
gongweibao 已提交
8718
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
8719

8720 8721 8722 8723 8724 8725 8726 8727 8728 8729 8730
    Examples:
        .. code-block:: python

            starts = [1, 0, 2]
            ends = [3, 3, 4]
            axes = [0, 1, 2]

            input = layers.data(
                name="input", shape=[3, 4, 5, 6], dtype='float32')

            out = layers.slice(input, axes=axes, starts=starts, ends=ends)
G
fix  
gongweibao 已提交
8731 8732 8733
    """

    helper = LayerHelper('slice', **locals())
8734 8735
    out = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype('input'))
G
fix  
gongweibao 已提交
8736 8737 8738 8739 8740 8741 8742 8743 8744 8745 8746
    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
8747
@templatedoc()
G
fix  
gongweibao 已提交
8748 8749
def shape(input):
    """
G
gongweibao 已提交
8750
    ${comment}
G
fix  
gongweibao 已提交
8751 8752

    Args:
G
gongweibao 已提交
8753
        input (Variable): ${input_comment}
G
fix  
gongweibao 已提交
8754 8755

    Returns:
G
gongweibao 已提交
8756
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
8757

8758 8759 8760 8761 8762 8763
    Examples:
        .. code-block:: python

            input = layers.data(
                name="input", shape=[3, 100, 100], dtype="float32")
            out = layers.shape(input)
G
fix  
gongweibao 已提交
8764 8765 8766
    """

    helper = LayerHelper('shape', **locals())
8767
    out = helper.create_variable_for_type_inference(dtype='int32')
G
fix  
gongweibao 已提交
8768
    helper.append_op(
G
fix  
gongweibao 已提交
8769
        type='shape', inputs={'Input': input}, outputs={'Out': out})
G
fix  
gongweibao 已提交
8770 8771

    return out
G
merge  
gongweibao 已提交
8772 8773


S
sneaxiy 已提交
8774 8775 8776 8777 8778 8779 8780 8781
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
8782 8783
    name = helper.kwargs.get('name', None)
    if name is None:
8784
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
S
sneaxiy 已提交
8785 8786 8787
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)
S
sneaxiy 已提交
8788

S
sneaxiy 已提交
8789 8790 8791 8792 8793 8794 8795 8796 8797 8798 8799
    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
8800
def scale(x, scale=1.0, bias=0.0, bias_after_scale=True, act=None, name=None):
S
sneaxiy 已提交
8801 8802 8803 8804 8805 8806 8807 8808
    """
    ${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
8809
        act(basestring|None): Activation applied to the output.
M
minqiyang 已提交
8810
        name(basestring|None): Name of the output.
S
sneaxiy 已提交
8811 8812 8813 8814 8815 8816

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

    helper = LayerHelper('scale', **locals())
S
sneaxiy 已提交
8817
    if name is None:
8818
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
S
sneaxiy 已提交
8819 8820 8821
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)
S
sneaxiy 已提交
8822 8823 8824 8825 8826 8827 8828 8829 8830 8831

    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
8832
    return helper.append_activation(out)
S
sneaxiy 已提交
8833 8834


X
Xin Pan 已提交
8835
def elementwise_add(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
8836 8837 8838
    return _elementwise_op(LayerHelper('elementwise_add', **locals()))


X
Xin Pan 已提交
8839
def elementwise_div(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
8840 8841 8842
    return _elementwise_op(LayerHelper('elementwise_div', **locals()))


X
Xin Pan 已提交
8843
def elementwise_sub(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
8844 8845 8846
    return _elementwise_op(LayerHelper('elementwise_sub', **locals()))


X
Xin Pan 已提交
8847
def elementwise_mul(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
8848 8849 8850
    return _elementwise_op(LayerHelper('elementwise_mul', **locals()))


X
Xin Pan 已提交
8851
def elementwise_max(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
8852 8853 8854
    return _elementwise_op(LayerHelper('elementwise_max', **locals()))


X
Xin Pan 已提交
8855
def elementwise_min(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
8856 8857 8858
    return _elementwise_op(LayerHelper('elementwise_min', **locals()))


X
Xin Pan 已提交
8859
def elementwise_pow(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
8860 8861 8862 8863 8864 8865 8866 8867 8868 8869 8870
    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
8871 8872
            "act (basestring|None): Activation applied to the output.",
            "name (basestring|None): Name of the output."
S
sneaxiy 已提交
8873
        ])
M
minqiyang 已提交
8874 8875


8876
def _logical_op(op_name, x, y, out=None, name=None, binary_op=True):
M
minqiyang 已提交
8877 8878
    helper = LayerHelper(op_name, **locals())

M
minqiyang 已提交
8879 8880
    if binary_op:
        assert x.dtype == y.dtype
M
minqiyang 已提交
8881 8882 8883

    if out is None:
        if name is None:
8884
            out = helper.create_variable_for_type_inference(dtype=x.dtype)
M
minqiyang 已提交
8885 8886 8887 8888 8889 8890 8891 8892 8893 8894 8895 8896 8897 8898 8899
        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()
8900
def logical_and(x, y, out=None, name=None):
M
minqiyang 已提交
8901 8902 8903 8904 8905 8906 8907 8908 8909 8910 8911
    """
    ${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}
8912 8913 8914 8915 8916 8917 8918 8919 8920

    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
8921 8922 8923 8924 8925 8926 8927
    """

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


@templatedoc()
8928
def logical_or(x, y, out=None, name=None):
M
minqiyang 已提交
8929 8930 8931 8932 8933 8934 8935 8936 8937 8938 8939
    """
    ${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}
8940 8941 8942 8943 8944 8945 8946 8947 8948

    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
8949 8950 8951 8952 8953 8954 8955
    """

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


@templatedoc()
8956
def logical_xor(x, y, out=None, name=None):
M
minqiyang 已提交
8957 8958 8959 8960 8961 8962 8963 8964 8965 8966 8967
    """
    ${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}
8968 8969 8970 8971 8972 8973 8974 8975 8976

    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
8977 8978 8979 8980 8981 8982 8983
    """

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


@templatedoc()
8984
def logical_not(x, out=None, name=None):
M
minqiyang 已提交
8985 8986 8987 8988 8989 8990 8991 8992 8993 8994
    """
    ${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}
8995 8996 8997 8998 8999 9000 9001

    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
9002 9003 9004 9005
    """

    return _logical_op(
        op_name="logical_not", x=x, y=None, name=name, out=out, binary_op=False)
9006 9007 9008 9009 9010 9011 9012 9013 9014 9015 9016 9017 9018 9019 9020


@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}
9021 9022 9023 9024 9025 9026 9027

    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)
9028 9029 9030 9031 9032
    """

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

    if name is None:
S
sneaxiy 已提交
9033 9034 9035 9036
        name = unique_name.generate(".".join([helper.name, 'tmp']))

    out = helper.create_variable(
        type=x.type, name=name, dtype=x.dtype, persistable=False)
9037 9038 9039 9040 9041 9042 9043 9044 9045 9046 9047 9048 9049 9050 9051 9052 9053 9054 9055 9056 9057 9058 9059

    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}
9060 9061 9062 9063 9064 9065 9066

    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)
9067 9068 9069 9070 9071
    """

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

    if name is None:
S
sneaxiy 已提交
9072 9073 9074 9075
        name = unique_name.generate(".".join([helper.name, 'tmp']))

    out = helper.create_variable(
        type=x.type, name=name, dtype=x.dtype, persistable=False)
9076 9077 9078 9079 9080 9081 9082 9083

    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
9084 9085 9086 9087 9088 9089 9090 9091 9092 9093 9094 9095 9096 9097 9098 9099 9100 9101


@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:
9102
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
9103 9104 9105 9106 9107 9108 9109 9110 9111 9112
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    helper.append_op(
        type="mean", inputs={"X": x}, attrs={}, outputs={"Out": out})

    return out


C
chengduo 已提交
9113 9114 9115 9116 9117 9118 9119 9120 9121 9122 9123 9124 9125 9126 9127 9128 9129 9130 9131 9132 9133 9134 9135
@templatedoc()
def merge_selected_rows(x, name=None):
    """
    ${comment}

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

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

    helper = LayerHelper("merge_selected_rows", **locals())
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
        type="merge_selected_rows",
        inputs={"X": x},
        attrs={},
        outputs={"Out": out})
    return out


X
Xin Pan 已提交
9136 9137 9138 9139 9140 9141 9142 9143 9144 9145 9146 9147 9148 9149 9150 9151 9152 9153 9154
@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:
9155
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
9156 9157 9158 9159 9160 9161 9162 9163 9164
    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
9165 9166
            "x_num_col_dims": x_num_col_dims,
            "y_num_col_dims": y_num_col_dims
X
Xin Pan 已提交
9167 9168 9169 9170 9171 9172
        },
        outputs={"Out": out})
    return out


@templatedoc()
J
jerrywgz 已提交
9173 9174 9175
def sigmoid_cross_entropy_with_logits(x,
                                      label,
                                      ignore_index=kIgnoreIndex,
9176 9177
                                      name=None,
                                      normalize=False):
X
Xin Pan 已提交
9178 9179 9180 9181 9182 9183
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        label(${label_type}): ${label_comment}
9184
        ignore_index(&{ignore_index}): ${ignore_index_comment}
X
Xin Pan 已提交
9185
        name(basestring|None): Name of the output.
9186 9187
        normalize(bool): If true, divide the output by the number of
            targets != ignore_index.
X
Xin Pan 已提交
9188 9189 9190

    Returns:
        out(${out_type}): ${out_comment}
9191 9192 9193 9194 9195 9196 9197 9198 9199 9200 9201 9202 9203 9204

    Examples:
        .. code-block:: python

            input = fluid.layers.data(
                name='data', shape=[10], dtype='float32')
            label = fluid.layers.data(
                name='data', shape=[10], dtype='float32')
            loss = fluid.layers.sigmoid_cross_entropy_with_logits(
                x=input,
                label=label,
                ignore_index=-1,
                normalize=True) # or False
            # loss = fluid.layers.reduce_sum(loss) # summation of loss
X
Xin Pan 已提交
9205 9206 9207 9208 9209
    """

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

    if name is None:
9210
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
9211 9212 9213 9214 9215 9216 9217 9218
    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},
9219 9220
        attrs={"ignore_index": ignore_index,
               'normalize': normalize},
X
Xin Pan 已提交
9221 9222 9223 9224 9225 9226 9227 9228 9229 9230 9231 9232 9233 9234 9235 9236 9237 9238 9239 9240
        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:
9241
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
9242 9243 9244 9245 9246 9247 9248 9249 9250 9251
    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
9252 9253


J
JiabinYang 已提交
9254
def space_to_depth(x, blocksize, name=None):
J
JiabinYang 已提交
9255
    """
J
JiabinYang 已提交
9256
    Gives a blocksize to space_to_depth the input LoDtensor with Layout: [batch, channel, height, width]
9257 9258 9259

    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
9260
    The attr blocksize indicates the input block size.
9261 9262

    space_to_depth will reorgnize the elements of input with shape[batch, channel, height, width] according
J
JiabinYang 已提交
9263
    to blocksize to construct output with shape [batch, channel * blocksize * blocksize, height/blocksize, width/blocksize]:
9264 9265

    space_to_depth is used to This operation is useful for resizing the activations between convolutions
J
JiabinYang 已提交
9266
    (but keeping all data)
J
JiabinYang 已提交
9267

J
JiabinYang 已提交
9268
    - Non-overlapping blocks of size block_size x block size are rearranged into depth at each location.
9269
    - The depth of the output tensor is block_size * block_size * input channel
J
JiabinYang 已提交
9270 9271 9272 9273 9274
    - 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
9275
    Args:
J
JiabinYang 已提交
9276
        x(variable): The input LoDtensor.
J
JiabinYang 已提交
9277
        blocksize(variable): The blocksize to select the element on each feature map should be > 2
J
JiabinYang 已提交
9278 9279

    Returns:
J
JiabinYang 已提交
9280
        Variable: The output LoDtensor.
J
JiabinYang 已提交
9281 9282

    Raises:
J
JiabinYang 已提交
9283
        TypeError: blocksize type must be a long.
J
JiabinYang 已提交
9284 9285 9286 9287 9288 9289

    Examples:
        .. code-block:: python

            data = fluid.layers.data(
                name='data', shape=[1, 4, 2, 2], dtype='float32')
J
JiabinYang 已提交
9290
            space_to_depthed = fluid.layers.space_to_depth(
J
JiabinYang 已提交
9291
                x=data, blocksize=2)
J
JiabinYang 已提交
9292 9293
    """

J
JiabinYang 已提交
9294
    helper = LayerHelper("space_to_depth", **locals())
J
JiabinYang 已提交
9295

J
JiabinYang 已提交
9296 9297
    if not (isinstance(blocksize, int)):
        raise ValueError("blocksize must be a python Int")
J
JiabinYang 已提交
9298 9299

    if name is None:
J
JiabinYang 已提交
9300 9301
        out = helper.create_variable_for_type_inference(
            dtype=x.dtype)  #fix create
J
JiabinYang 已提交
9302 9303 9304 9305 9306
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    helper.append_op(
J
JiabinYang 已提交
9307
        type="space_to_depth",
J
JiabinYang 已提交
9308
        inputs={"X": x},
J
JiabinYang 已提交
9309
        attrs={"blocksize": blocksize},
J
JiabinYang 已提交
9310
        outputs={"Out": out})
J
JiabinYang 已提交
9311 9312
    return out

J
JiabinYang 已提交
9313

S
sneaxiy 已提交
9314 9315
@templatedoc()
def sequence_reverse(x, name=None):
9316
    """
S
sneaxiy 已提交
9317 9318 9319 9320 9321 9322 9323 9324 9325 9326 9327
    ${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
9328
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
S
sneaxiy 已提交
9329 9330 9331 9332 9333 9334 9335 9336 9337 9338
    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
9339 9340


9341 9342 9343 9344 9345 9346
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.
9347

9348 9349 9350 9351 9352 9353 9354 9355 9356 9357 9358 9359 9360 9361 9362 9363 9364 9365 9366
    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
9367
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
9368 9369 9370 9371 9372 9373 9374 9375 9376 9377 9378 9379
    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
9380 9381


B
barrierye 已提交
9382
def similarity_focus(input, axis, indexes, name=None):
9383
    """
9384
    SimilarityFocus Operator
B
barrierye 已提交
9385 9386

    Generate a similarity focus mask with the same shape of input using the following method:
M
minqiyang 已提交
9387

9388 9389 9390
    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
9391
       is (BatchSize, A, B, C), the shape of tensor T is (BatchSize, B, C).
9392 9393 9394 9395 9396 9397 9398
    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
9399
       each index.
B
barrierye 已提交
9400 9401 9402 9403
    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
9404 9405 9406 9407 9408 9409 9410 9411 9412 9413 9414 9415 9416 9417 9418 9419 9420 9421 9422 9423 9424 9425 9426 9427 9428 9429 9430 9431 9432 9433 9434 9435 9436 9437 9438 9439 9440 9441 9442 9443 9444 9445 9446 9447 9448 9449 9450 9451 9452
    .. 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
9453
    Args:
9454
        input(Variable): The input tensor variable(default float). It should
B
barrierye 已提交
9455
            be a 4-D tensor with shape [BatchSize, A, B, C].
B
barrierye 已提交
9456
        axis(int): Indicating the dimension to be selected. It can only be
9457
            1, 2 or 3.
B
barrierye 已提交
9458
        indexes(list): Indicating the indexes of the selected dimension.
B
barrierye 已提交
9459 9460

    Returns:
9461 9462
        Variable: A tensor variable with the same shape and same type \
                  as the input.
9463

B
barrierye 已提交
9464 9465
    Examples:
        .. code-block:: python
9466

B
barrierye 已提交
9467
            data = fluid.layers.data(
B
barrierye 已提交
9468 9469
              name='data', shape=[2, 3, 2, 2], dtype='float32')
            x = fluid.layers.layer_norm(input=data, axis=1, indexes=[0])
9470

B
barrierye 已提交
9471 9472 9473 9474 9475 9476 9477 9478 9479 9480 9481 9482
    """
    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
9483 9484 9485 9486 9487
    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
9488 9489 9490 9491 9492 9493 9494
    helper.append_op(
        type='similarity_focus',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={"axis": axis,
               "indexes": indexes})
    return out
B
barrierye 已提交
9495 9496


M
minqiyang 已提交
9497 9498
def hash(input, hash_size, num_hash=1, name=None):
    """
M
minqiyang 已提交
9499 9500
    Hash the input to an integer whose value is less than the given hash size.

M
minqiyang 已提交
9501 9502
    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
9503 9504 9505 9506 9507 9508 9509 9510 9511 9512 9513 9514 9515 9516 9517 9518 9519 9520 9521 9522 9523 9524 9525 9526 9527 9528 9529 9530 9531 9532 9533 9534 9535 9536 9537 9538 9539 9540

    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
9541
        num_hash (int): The times of hash, default 1.
M
minqiyang 已提交
9542
        name (str, default None): The name of this layer.
M
minqiyang 已提交
9543 9544 9545 9546 9547 9548

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

    Examples:
       .. code-block:: python
9549

M
minqiyang 已提交
9550 9551 9552
           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
9553 9554
    """
    helper = LayerHelper('hash', **locals())
M
minqiyang 已提交
9555 9556
    out = helper.create_variable_for_type_inference(
        helper.input_dtype(), stop_gradient=True)
M
minqiyang 已提交
9557 9558 9559 9560 9561 9562 9563
    helper.append_op(
        type='hash',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={'num_hash': num_hash,
               'mod_by': hash_size})
    return out
9564 9565


9566
@templatedoc()
9567 9568
def grid_sampler(x, grid, name=None):
    """
9569
    This operation samples input X by using bilinear interpolation based on
9570
    flow field grid, which is usually gennerated by :code:`affine_grid` . The grid of
9571 9572 9573 9574
    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
9575
    interpolation value of 4 nearest corner points.
9576

9577
    .. code-block:: text
9578

9579 9580
        Step 1:
        Get (x, y) grid coordinates and scale to [0, H-1/W-1].
9581

9582 9583
        grid_x = 0.5 * (grid[:, :, :, 0] + 1) * (W - 1)
        grid_y = 0.5 * (grid[:, :, :, 1] + 1) * (H - 1)
9584

9585 9586 9587
        Step 2:
        Indices input data X with grid (x, y) in each [H, W] area, and bilinear
        interpolate point value by 4 nearest points.
9588

9589 9590 9591 9592 9593 9594 9595 9596 9597
          wn ------- y_n ------- en
          |           |           |
          |          d_n          |
          |           |           |
         x_w --d_w-- grid--d_e-- x_e
          |           |           |
          |          d_s          |
          |           |           |
          ws ------- y_s ------- wn
9598

9599 9600 9601 9602
        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
9603

9604 9605 9606 9607
        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
9608

9609 9610 9611 9612
        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
9613

9614 9615
        output = wn * d_e * d_s + en * d_w * d_s
               + ws * d_e * d_n + es * d_w * d_n
9616 9617

    Args:
9618 9619 9620
        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.
9621 9622

    Returns:
9623
        Variable: Output of shape [N, C, H, W] data samples input X
9624 9625
        using bilnear interpolation based on input grid.

9626 9627 9628 9629 9630 9631 9632 9633
    Examples:

        .. 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)
9634

9635 9636 9637 9638 9639 9640 9641 9642 9643
    """
    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")

9644
    out = helper.create_variable_for_type_inference(x.dtype)
9645 9646
    ipts = {'X': x, 'Grid': grid}

9647
    helper.append_op(type='grid_sampler', inputs=ipts, outputs={'Output': out})
9648 9649 9650
    return out


9651 9652 9653 9654 9655 9656 9657 9658 9659 9660 9661 9662 9663 9664 9665 9666 9667 9668 9669 9670 9671 9672 9673 9674 9675 9676 9677 9678 9679 9680 9681 9682 9683 9684 9685 9686 9687 9688 9689 9690 9691 9692 9693 9694 9695 9696 9697
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


9698 9699 9700 9701 9702 9703 9704 9705 9706 9707 9708 9709 9710 9711 9712 9713 9714 9715 9716
def teacher_student_sigmoid_loss(input,
                                 label,
                                 soft_max_up_bound=15.0,
                                 soft_max_lower_bound=-15.0):
    """
    **Teacher Student Log Loss Layer**

    This layer accepts input predictions and target label and returns the
    teacher_student loss.

    .. math::
        loss = max(x, 0) - x * z + log(1 + exp(-abs(x))) + max(x, 0) - x * z' + log(1 + exp(-abs(x)))

    Args:
        input (Variable|list):  a 2-D tensor with shape [N x 1], where N is the
                                batch size. This input is a probability computed
                                by the previous operator.
        label (Variable|list):  the ground truth which is a 2-D tensor with
                                shape [N x 1], where N is the batch size.
9717
        soft_max_up_bound  (float):  if input > soft_max_up_bound, will be bound
9718 9719 9720 9721 9722 9723 9724 9725 9726 9727 9728 9729 9730 9731 9732 9733 9734 9735 9736 9737 9738
        soft_max_lower_bound (float): if input < soft_max_lower_bound, will be bound

    Returns:
        Variable: A 2-D tensor with shape [N x 1], the teacher_student_sigmoid_loss.

    Examples:
        .. code-block:: python
          cost = fluid.layers.teacher_student_sigmoid_loss(input=similarity, label=label)
    """
    helper = LayerHelper('teacher_student_sigmoid_loss', **locals())
    out = helper.create_variable(dtype=input.dtype)
    helper.append_op(
        type='teacher_student_sigmoid_loss',
        inputs={'X': [input],
                'Label': [label]},
        outputs={'Y': [out]},
        attrs={"soft_max_lower_bound": float(soft_max_lower_bound), \
                "soft_max_up_bound": float(soft_max_up_bound)})
    return out


9739 9740 9741 9742
def add_position_encoding(input, alpha, beta, name=None):
    """
    **Add Position Encoding Layer**

9743
    This layer accepts an input 3D-Tensor of shape [N x M x P], and returns an
9744 9745
    output Tensor of shape [N x M x P] with positional encoding value.

9746
    Refer to `Attention Is All You Need <http://arxiv.org/pdf/1706.03762.pdf>`_ .
9747 9748

    .. math::
9749 9750 9751
        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)
9752 9753

    Where:
9754 9755
      - :math:`PE(pos, 2i)` : the increment for the number at even position
      - :math:`PE(pos, 2i + 1)` : the increment for the number at odd position
9756 9757 9758 9759 9760 9761 9762 9763 9764 9765 9766 9767 9768 9769

    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)
9770

9771 9772 9773 9774 9775 9776 9777 9778 9779 9780 9781 9782 9783 9784 9785 9786
    """
    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
9787 9788 9789 9790 9791 9792 9793 9794 9795 9796


def bilinear_tensor_product(x,
                            y,
                            size,
                            act=None,
                            name=None,
                            param_attr=None,
                            bias_attr=None):
    """
Q
Qiao Longfei 已提交
9797
    **Add Bilinear Tensor Product Layer**
9798

Q
Qiao Longfei 已提交
9799
    This layer performs bilinear tensor product on two inputs.
9800 9801 9802
    For example:

    .. math::
9803
       out_{i} = x * W_{i} * {y^\mathrm{T}}, i=0,1,...,size-1
9804

9805
    In this formula:
9806 9807
      - :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].
9808
      - :math:`W_{i}`: the i-th learned weight, shape is [M, N]
9809
      - :math:`out_{i}`: the i-th element of out, shape is [batch_size, size].
9810 9811 9812
      - :math:`y^\mathrm{T}`: the transpose of :math:`y_{2}`.

    Args:
9813 9814
        x (Variable): 2-D input tensor with shape [batch_size, M]
        y (Variable): 2-D input tensor with shape [batch_size, N]
9815 9816 9817
        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.
9818
        param_attr (ParamAttr, default None): The parameter attribute for the learnable w.
9819
            parameters/weights of this layer.
9820
        bias_attr (ParamAttr, default None): The parameter attribute for the bias
9821 9822 9823 9824
            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:
9825
        Variable: A 2-D Tensor of shape [batch_size, size].
9826 9827 9828 9829

    Examples:
        .. code-block:: python

9830
          tensor = bilinear_tensor_product(x=layer1, y=layer2, size=1000)
9831 9832
    """
    helper = LayerHelper('bilinear_tensor_product', **locals())
Q
Qiao Longfei 已提交
9833
    dtype = helper.input_dtype('x')
9834 9835 9836 9837

    param_shape = [size, x.shape[1], y.shape[1]]

    w = helper.create_parameter(
Q
Qiao Longfei 已提交
9838
        attr=helper.param_attr, shape=param_shape, dtype=dtype, is_bias=False)
9839 9840 9841 9842 9843 9844 9845 9846 9847 9848 9849 9850 9851 9852 9853 9854 9855

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

    inputs = {"X": x, "Y": y, "Weight": w}
    if helper.bias_attr:
        bias_size = [1, size]
        bias = helper.create_parameter(
            attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
        inputs["Bias"] = bias
    helper.append_op(
        type="bilinear_tensor_product", inputs=inputs, outputs={"Out": out})

    # add activation
    return helper.append_activation(out)
C
chengduo 已提交
9856 9857 9858 9859 9860 9861 9862 9863 9864 9865 9866 9867 9868 9869 9870 9871 9872 9873 9874 9875 9876 9877 9878


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

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

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

    helper = LayerHelper('get_tensor_from_selected_rows', **locals())
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
        type='get_tensor_from_selected_rows',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={})
    return out
9879 9880


S
shippingwang 已提交
9881
def shuffle_channel(x, group, name=None):
S
shippingwang 已提交
9882 9883
    """
    **Shuffle Channel Operator**
9884

9885 9886 9887 9888 9889 9890
    This operator shuffles the channels of input x.
    It divide the input channels in each group into :attr:`group` subgroups,
    and obtain a new order by selecting element from every subgroup one by one.

    Please refer to the paper
    https://arxiv.org/pdf/1707.01083.pdf
S
shippingwang 已提交
9891
    
9892
    .. code-block:: text
9893

9894 9895 9896 9897 9898 9899 9900 9901 9902 9903 9904 9905 9906 9907 9908 9909 9910 9911 9912 9913 9914 9915 9916 9917 9918 9919 9920 9921
        Given a 4-D tensor input with the shape (N, C, H, W):
            input.shape = (1, 4, 2, 2)
            input.data =[[[[0.1, 0.2],
                           [0.2, 0.3]],

                          [[0.3, 0.4],
                           [0.4, 0.5]],

                          [[0.5, 0.6],
                           [0.6, 0.7]],

                          [[0.7, 0.8],
                           [0.8, 0.9]]]]
            Given group: 2
            then we get a 4-D tensor out whth the same shape of input:
            out.shape = (1, 4, 2, 2)
            out.data = [[[[0.1, 0.2],
                          [0.2, 0.3]],
                          
                         [[0.5, 0.6],
                          [0.6, 0.7]],
                          
                         [[0.3, 0.4],
                          [0.4, 0.5]],
                          
                         [[0.7, 0.8],
                          [0.8, 0.9]]]]
                        
S
shippingwang 已提交
9922
    Args: 
9923 9924
        x(Variable): The input tensor variable. It should be a 4-D tensor with shape [N, C, H, W]
        group(int): Indicating the conuts of subgroups, It should divide the number of channels.
S
shippingwang 已提交
9925 9926

    Returns:
9927 9928
        out(Variable): the channels shuffling result is a tensor variable with the 
        same shape and same type as the input.
S
shippingwang 已提交
9929 9930

    Raises:
S
shippingwang 已提交
9931
        ValueError: If group is not an int type variable.
S
shippingwang 已提交
9932 9933 9934

    Examples:
        .. code-block:: python
9935 9936

            input = fluid.layers.data(name='input', shape=[4,2,2], dtype='float32')
9937
            out = fluid.layers.shuffle_channel(x=input, group=2)
S
shippingwang 已提交
9938 9939 9940
    """
    helper = LayerHelper("shuffle_channel", **locals())

9941
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
S
shippingwang 已提交
9942 9943 9944 9945 9946 9947 9948 9949 9950

    if not isinstance(group, int):
        raise TypeError("group must be int type")

    helper.append_op(
        type="shuffle_channel",
        inputs={"X": x},
        outputs={"Out": out},
        attrs={"group": group})
S
shippingwang 已提交
9951
    return out
S
Add  
shippingwang 已提交
9952 9953


S
sneaxiy 已提交
9954
class PyFuncRegistry(object):
S
sneaxiy 已提交
9955 9956 9957
    _register_funcs = []

    def __init__(self, func):
S
sneaxiy 已提交
9958
        if func is None or not callable(func):
S
sneaxiy 已提交
9959 9960 9961
            raise TypeError('func must be a Python function')

        self._func = func
M
minqiyang 已提交
9962
        # find named args using reflection
S
sneaxiy 已提交
9963 9964 9965 9966 9967 9968 9969
        args = inspect.getargspec(self._func)
        if len(args[0]) == 0 and args[1] is None and args[2] is None:
            # Function with no inputs
            self._named_args = None
        else:
            self._named_args = args[0]
        self._id = core._append_python_callable_object_and_return_id(self)
S
sneaxiy 已提交
9970 9971 9972
        '''
        Why record self here?

M
minqiyang 已提交
9973 9974
        1. For debug usage. Users can call
           :code:`py_func.registered_func(idx)` method
S
sneaxiy 已提交
9975
           to find the registered function corresponding
M
minqiyang 已提交
9976
           to :code:`idx`.
S
sneaxiy 已提交
9977

M
minqiyang 已提交
9978 9979
        2. For increasing reference count of self.
           It seems that to release Python object
S
sneaxiy 已提交
9980
           whose reference count is 1 would cause
M
minqiyang 已提交
9981
           segmentation fault error in C++ side.
S
sneaxiy 已提交
9982 9983
           May be lack of Python GC in C++ side?
        '''
S
sneaxiy 已提交
9984
        PyFuncRegistry._register_funcs.append(self)
S
sneaxiy 已提交
9985 9986 9987 9988 9989 9990 9991 9992 9993 9994 9995 9996 9997 9998

    @classmethod
    def registered_func(cls, idx):
        return cls._register_funcs[idx]._func

    @classmethod
    def registered_func_num(cls):
        return len(cls._register_funcs)

    @property
    def id(self):
        return self._id

    def __call__(self, *args):
S
sneaxiy 已提交
9999 10000 10001 10002 10003 10004 10005 10006 10007
        if self._named_args is None:
            func_ret = self._func()
        else:
            kwargs = dict()
            idx = 0
            for arg in self._named_args:
                kwargs[arg] = args[idx]
                idx += 1
            func_ret = self._func(*args[idx:], **kwargs)
S
sneaxiy 已提交
10008

S
sneaxiy 已提交
10009 10010
        if not isinstance(func_ret, (list, tuple)):
            func_ret = (func_ret, )
S
sneaxiy 已提交
10011 10012

        ret = []
S
sneaxiy 已提交
10013 10014 10015
        for each_ret in func_ret:
            if each_ret is None or isinstance(each_ret, core.LoDTensor):
                ret.append(each_ret)
S
sneaxiy 已提交
10016 10017
                continue

S
sneaxiy 已提交
10018 10019
            if not isinstance(each_ret, np.ndarray):
                each_ret = np.array(each_ret)
S
sneaxiy 已提交
10020

S
sneaxiy 已提交
10021 10022 10023
            tensor = core.LoDTensor()
            tensor.set(each_ret, core.CPUPlace())
            ret.append(tensor)
S
sneaxiy 已提交
10024

S
sneaxiy 已提交
10025
        return tuple(ret)
S
sneaxiy 已提交
10026 10027


S
sneaxiy 已提交
10028 10029 10030 10031
@templatedoc()
def py_func(func, x, out, backward_func=None, skip_vars_in_backward_input=None):
    """
    PyFunc Operator.
M
minqiyang 已提交
10032

S
sneaxiy 已提交
10033 10034 10035 10036 10037 10038 10039 10040
    User can use :code:`py_func` to register operators in Python side.
    The inputs of :code:`func` is :code:`LoDTensor` and outputs can be
    numpy array or :code:`LoDTensor`. Paddle would call the registered
    :code:`func` in forward part, and call :code:`backward_func` in
    backward part (if :code:`backward_func` is not None).

    User should set the right data type and shape of :code:`out` before
    calling this function. However, data types and shapes of gradients of
S
sneaxiy 已提交
10041
    :code:`out` and :code:`x` would be inferred automatically.
S
sneaxiy 已提交
10042

S
sneaxiy 已提交
10043 10044
    Input orders of :code:`backward_func` would be: forward inputs
    :code:`x`, forward outputs :code:`out` and backward input gradients of
S
sneaxiy 已提交
10045 10046 10047 10048
    :code:`out`. If some variables of :code:`out` have no gradient, the input
    tensor would be None in Python side. If some variables of :code:`in` have
    no gradient, users should return None.

S
sneaxiy 已提交
10049
    This function can also be used to debug the running network. User can
M
minqiyang 已提交
10050
    add a :code:`py_func` operator without output, and print input
S
sneaxiy 已提交
10051 10052
    :code:`x` inside :code:`func`.

S
sneaxiy 已提交
10053 10054 10055 10056 10057
    Args:
        func (callable): forward Python function.
        x (Variable|list(Variable)|tuple(Variable)): inputs of :code:`func`.
        out (Variable|list(Variable)|tuple(Variable)): outputs of :code:`func`.
            Paddle cannot infer shapes and data types of :code:`out`. Users
M
minqiyang 已提交
10058
            should create :code:`out` beforehand.
S
sneaxiy 已提交
10059
        backward_func (callable|None): backward Python function.
M
minqiyang 已提交
10060
                                       None means no backward. Default None.
S
sneaxiy 已提交
10061
        skip_vars_in_backward_input (Variable|list(Variable)|tuple(Variable)):
M
minqiyang 已提交
10062
            Variables that are not needed in :code:`backward_func` inputs.
S
sneaxiy 已提交
10063 10064
            These variables must be any of :code:`x` and :code:`out`.
            If set, these vars would not be inputs of :code:`backward_func`,
M
minqiyang 已提交
10065
            Only useful when :code:`backward_func` is not None. Default None.
S
sneaxiy 已提交
10066 10067 10068

    Returns:
        out (Variable|list(Variable)|tuple(Variable)): input :code:`out`
S
sneaxiy 已提交
10069 10070

    Examples:
M
minqiyang 已提交
10071

S
sneaxiy 已提交
10072 10073 10074 10075 10076
        >>> import paddle.fluid as fluid
        >>> import six
        >>>
        >>> def create_tmp_var(name, dtype, shape):
        >>>     return fluid.default_main_program().current_block().create_var(
M
minqiyang 已提交
10077
        >>>         name=name, dtype=dtype, shape=shape)
S
sneaxiy 已提交
10078 10079
        >>>
        >>> # tanh activation has been provided by Paddle C++ op
M
minqiyang 已提交
10080
        >>> # Here, we only use tanh to be an example to show the usage
S
sneaxiy 已提交
10081 10082 10083
        >>> # of py_func
        >>> def tanh(x):
        >>>     return np.tanh(x)
M
minqiyang 已提交
10084
        >>>
S
sneaxiy 已提交
10085 10086 10087 10088 10089
        >>> # forward input x is skipped
        >>> def tanh_grad(y, dy):
        >>>     return np.array(dy) * (1 - np.square(np.array(y)))
        >>>
        >>> def debug_func(x):
M
minqiyang 已提交
10090
        >>>     print(x)
S
sneaxiy 已提交
10091 10092 10093 10094 10095 10096
        >>>
        >>> def simple_net(img, label):
        >>>     hidden = img
        >>>     for idx in six.moves.range(4):
        >>>         hidden = fluid.layers.fc(hidden, size=200)
        >>>         new_hidden = create_tmp_var(name='hidden_{}'.format(idx),
M
minqiyang 已提交
10097
        >>>             dtype=hidden.dtype, shape=hidden.shape)
S
sneaxiy 已提交
10098 10099
        >>>
        >>>         # user-defined layers with forward and backward
M
minqiyang 已提交
10100 10101
        >>>         hidden = fluid.layers.py_func(func=tanh, x=hidden,
        >>>             out=new_hidden, backward_func=tanh_grad,
S
sneaxiy 已提交
10102 10103 10104 10105 10106 10107 10108 10109
        >>>             skip_vars_in_backward_input=hidden)
        >>>
        >>>         # user-defined debug layers to print variables
        >>>         fluid.layers.py_func(func=debug_func, x=hidden, out=None)
        >>>
        >>>     prediction = fluid.layers.fc(hidden, size=10, act='softmax')
        >>>     loss = fluid.layers.cross_entropy(input=prediction, label=label)
        >>>     return fluid.layers.mean(loss)
S
sneaxiy 已提交
10110
    """
S
sneaxiy 已提交
10111
    helper = LayerHelper('py_func', **locals())
S
sneaxiy 已提交
10112 10113 10114
    if x is None:
        x = []
    elif isinstance(x, Variable):
S
sneaxiy 已提交
10115
        x = [x]
S
sneaxiy 已提交
10116 10117
    elif not isinstance(x, (list, tuple)):
        raise TypeError('Input must be Variable/list(Variable)/tuple(Variable)')
S
sneaxiy 已提交
10118

S
sneaxiy 已提交
10119 10120 10121
    if out is None:
        out_list = []
    elif isinstance(out, Variable):
S
sneaxiy 已提交
10122
        out_list = [out]
S
sneaxiy 已提交
10123
    elif isinstance(out, (list, tuple)):
S
sneaxiy 已提交
10124
        out_list = out
S
sneaxiy 已提交
10125 10126 10127
    else:
        raise TypeError(
            'Output must be Variable/list(Variable)/tuple(Variable)')
S
sneaxiy 已提交
10128

S
sneaxiy 已提交
10129 10130
    fwd_func_id = PyFuncRegistry(func).id
    bwd_func_id = PyFuncRegistry(
S
sneaxiy 已提交
10131
        backward_func).id if backward_func is not None else -1
S
sneaxiy 已提交
10132 10133

    for each_out in out_list:
S
sneaxiy 已提交
10134 10135
        if len(each_out.shape) == 0:
            raise ValueError(
S
sneaxiy 已提交
10136 10137
                'Output shapes of py_func op should be provided by users manually'
            )
S
sneaxiy 已提交
10138

S
sneaxiy 已提交
10139 10140 10141 10142 10143 10144 10145 10146 10147 10148 10149 10150 10151 10152 10153
    backward_skip_vars = set()
    if backward_func is not None and skip_vars_in_backward_input is not None:
        if isinstance(skip_vars_in_backward_input, Variable):
            skip_vars_in_backward_input = [skip_vars_in_backward_input]

        fwd_in_out = [v.name for v in x]
        fwd_in_out.extend([v.name for v in out_list])
        fwd_in_out = set(fwd_in_out)
        backward_skip_vars = set()
        for v in skip_vars_in_backward_input:
            if not v.name in fwd_in_out:
                raise ValueError(
                    'Variable {} is not found in forward inputs and outputs'
                    .format(v.name))
            backward_skip_vars.add(v.name)
S
sneaxiy 已提交
10154 10155 10156 10157

    helper.append_op(
        type='py_func',
        inputs={'X': x},
S
sneaxiy 已提交
10158 10159
        outputs={'Out': out_list},
        attrs={
S
sneaxiy 已提交
10160 10161 10162
            'forward_callable_id': fwd_func_id,
            'backward_callable_id': bwd_func_id,
            'backward_skip_vars': list(backward_skip_vars)
S
sneaxiy 已提交
10163
        })
S
sneaxiy 已提交
10164
    return out
S
sneaxiy 已提交
10165 10166 10167


# For debug usage
S
sneaxiy 已提交
10168 10169 10170 10171
py_func.registered_func = PyFuncRegistry.registered_func
py_func.registered_func_num = PyFuncRegistry.registered_func_num


10172 10173 10174 10175 10176 10177 10178 10179 10180 10181 10182 10183 10184 10185 10186 10187 10188 10189 10190 10191 10192 10193 10194 10195 10196 10197 10198 10199 10200 10201 10202 10203 10204 10205 10206 10207 10208 10209 10210 10211 10212 10213 10214 10215 10216 10217 10218 10219 10220 10221 10222 10223
@templatedoc()
def psroi_pool(input,
               rois,
               output_channels,
               spatial_scale,
               pooled_height,
               pooled_width,
               name=None):
    """
    ${comment}

    Args:
        input (Variable): ${x_comment}
        rois (Variable): ROIs (Regions of Interest) to pool over.
        output_channels (integer): ${output_channels_comment}
        spatial_scale (float): ${spatial_scale_comment} Default: 1.0
        pooled_height (integer): ${pooled_height_comment} Default: 1
        pooled_width (integer): ${pooled_width_comment} Default: 1
        name (str, default None): The name of this layer.

    Returns:
        Variable: ${out_comment}.

    Examples:
        .. code-block:: python

            pool_out = fluid.layers.psroi_pool(input=x, rois=rois, 490, 1.0, 7, 7)
    """
    helper = LayerHelper('psroi_pool', **locals())
    # check attrs
    if not isinstance(output_channels, int):
        raise TypeError("output_channels must be int type")
    if not isinstance(spatial_scale, float):
        raise TypeError("spatial_scale must be float type")
    if not isinstance(pooled_height, int):
        raise TypeError("pooled_height must be int type")
    if not isinstance(pooled_width, int):
        raise TypeError("pooled_width must be int type")
    dtype = helper.input_dtype()
    out = helper.create_variable_for_type_inference(dtype)
    helper.append_op(
        type='psroi_pool',
        inputs={'X': input,
                'ROIs': rois},
        outputs={'Out': out},
        attrs={
            'output_channels': output_channels,
            'spatial_scale': spatial_scale,
            'pooled_height': pooled_height,
            'pooled_width': pooled_width
        })
    return out
10224

M
minqiyang 已提交
10225

M
minqiyang 已提交
10226
def huber_loss(input, label, delta):
10227
    """
M
minqiyang 已提交
10228 10229 10230
    Huber loss is a loss function used in robust.
    Huber loss can evaluate the fitness of input to label.
    Different from MSE loss, Huber loss is more robust for outliers.
10231 10232 10233 10234

    When the difference between input and label is large than delta
    .. math::

M
minqiyang 已提交
10235
        huber\_loss = delta * (label - input) - 0.5 * delta * delta
10236 10237 10238 10239

    When the difference between input and label is less than delta
    .. math::

M
minqiyang 已提交
10240
        huber\_loss = 0.5 * (label - input) * (label - input)
10241 10242 10243 10244 10245 10246 10247


    Args:
        input (Variable): This input is a probability computed by the previous operator.
                          The first dimension is batch size, and the last dimension is 1.
        label (Variable): The groud truth whose first dimension is batch size
                          and last dimension is 1.
M
minqiyang 已提交
10248
        delta (float): The parameter of huber loss, which controls
10249 10250 10251
                       the range of outliers

    Returns:
M
minqiyang 已提交
10252
        huber\_loss (Variable): The huber loss with shape [batch_size, 1].
10253 10254 10255 10256 10257

    Examples:
        .. code-block:: python

            predictions = fluid.layers.softmax(x)
M
minqiyang 已提交
10258
            loss = fluid.layers.huber_loss(input=predictions, label=label, 1.0)
10259
    """
M
minqiyang 已提交
10260
    helper = LayerHelper('huber_loss', **locals())
10261 10262 10263 10264 10265 10266 10267 10268 10269 10270 10271
    residual = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype())
    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
    helper.append_op(
        type='huber_loss',
        inputs={'X': input,
                'Y': label},
        outputs={'Out': out,
                 'Residual': residual},
        attrs={'delta': delta})
    return out
Z
zhaozhehao 已提交
10272 10273 10274 10275 10276 10277 10278 10279 10280 10281 10282 10283 10284 10285 10286 10287 10288 10289 10290 10291 10292 10293 10294 10295 10296 10297 10298 10299 10300 10301 10302 10303 10304 10305 10306 10307 10308 10309 10310 10311 10312 10313 10314 10315 10316 10317 10318 10319 10320 10321 10322 10323 10324 10325 10326 10327 10328 10329 10330 10331 10332 10333 10334 10335 10336 10337 10338 10339 10340 10341


@templatedoc()
def tree_conv(nodes_vector,
              edge_set,
              output_size,
              num_filters=1,
              max_depth=2,
              act='tanh',
              param_attr=None,
              bias_attr=None,
              name=None):
    """ 
    ${comment}
    		
    Args:
        nodes_vector(${nodes_vector_type}): ${nodes_vector_comment}
        edge_set(${edge_set_type}): ${edge_set_comment}
        output_size(int): output feature width
        num_filters(int): number of filters, Default 1
        max_depth(int): max depth of filters, Default 2
        act(str): activation function, Default tanh
        param_attr(ParamAttr): the parameter attribute for the filters, Default None
        bias_attr(ParamAttr): the parameter attribute for the bias of this layer, Default None
        name(str): a name of this layer(optional). If set None, the layer will be named automatically, Default None

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

    Examples:
        .. code-block:: python

          nodes_vector = layers.data(name='vectors', shape=[None, 10, 5], dtype='float32)
          # None for batch size, 10 for max_node_size of dataset, 5 for vector width
          edge_set = layers.data(name='edge_set', shape=[None, 10, 2], dtype='float32')
          # None for batch size, 10 for max_node_size of dataset, 2 for every edge has two nodes
          # edges must be directional
          out_vector = layers.tree_conv(nodes_vector, edge_set, 6, 1, 2, 'tanh',
              ParamAttr(initializer=Constant(1.0), ParamAttr(initializer=Constant(1.0))
          # the shape of output will be [None, 10, 6, 1],
          # None for batch size, 10 for max_node_size of dataset, 6 for output size, 1 for 1 filter
          out_vector = layers.reshape(out_vector, shape=[None, 10, 6])
          # After reshape, output tensor could be nodes_vector for next tree convolution
          out_vector_2 = layers.tree_conv(out_vector, edge_set, 3, 4, 2, 'tanh',
              ParamAttr(initializer=Constant(1.0), ParamAttr(initializer=Constant(1.0))
          # also output tensor could be pooling(the pooling in paper called global pooling)
          pooled = layers.reduce_max(out_vector, dims=2) # global pooling
    """
    helper = LayerHelper("tree_conv", **locals())
    dtype = helper.input_dtype('nodes_vector')
    feature_size = nodes_vector.shape[2]
    W_shape = [feature_size, 3, output_size, num_filters]
    W = helper.create_parameter(
        attr=param_attr, shape=W_shape, dtype=dtype, is_bias=False)
    if name == None:
        out = helper.create_variable_for_type_inference(dtype=dtype)
    else:
        out = helper.create_variable(name=name, dtype=dtype, persistable=False)
    helper.append_op(
        type='tree_conv',
        inputs={'NodesVector': nodes_vector,
                'EdgeSet': edge_set,
                'Filter': W},
        outputs={'Out': out, },
        attrs={'max_depth': max_depth})
    if helper.bias_attr:
        pre_activation = helper.append_bias_op(out)
    else:
        pre_activation = out
    return helper.append_activation(pre_activation)
反馈
建议
客服 返回
顶部