nn.py 263.3 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
Y
Yu Yang 已提交
14
"""
15
All layers just related to the neural network.
Y
Yu Yang 已提交
16 17
"""

18 19
from __future__ import print_function

20
import numpy as np
Y
Yu Yang 已提交
21 22
from ..layer_helper import LayerHelper
from ..initializer import Normal, Constant
S
sneaxiy 已提交
23
from ..framework import Variable, OpProtoHolder
Y
yangyaming 已提交
24
from ..param_attr import ParamAttr
S
sneaxiy 已提交
25
from .layer_function_generator import autodoc, templatedoc, _generate_doc_string_
26 27
from .tensor import concat
from . import utils
F
fengjiayi 已提交
28
from .. import unique_name
29
from functools import reduce
Y
Yu Yang 已提交
30 31

__all__ = [
X
Xin Pan 已提交
32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58
    'fc',
    'embedding',
    'dynamic_lstm',
    'dynamic_lstmp',
    'dynamic_gru',
    'gru_unit',
    'linear_chain_crf',
    'crf_decoding',
    'cos_sim',
    'cross_entropy',
    'square_error_cost',
    'chunk_eval',
    'sequence_conv',
    'conv2d',
    'conv3d',
    'sequence_pool',
    'sequence_softmax',
    'softmax',
    'pool2d',
    'pool3d',
    'batch_norm',
    'beam_search_decode',
    'conv2d_transpose',
    'conv3d_transpose',
    'sequence_expand',
    'sequence_expand_as',
    'sequence_pad',
Y
Yibing Liu 已提交
59
    'sequence_unpad',
X
Xin Pan 已提交
60 61 62 63 64 65 66 67
    'lstm_unit',
    'reduce_sum',
    'reduce_mean',
    'reduce_max',
    'reduce_min',
    'reduce_prod',
    'sequence_first_step',
    'sequence_last_step',
Y
Yibing Liu 已提交
68
    'sequence_slice',
X
Xin Pan 已提交
69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98
    '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',
    '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
99
    'roi_align',
X
Xin Pan 已提交
100 101 102 103 104 105 106 107 108 109 110 111 112
    'dice_loss',
    'image_resize',
    'image_resize_short',
    'resize_bilinear',
    'gather',
    'scatter',
    'sequence_scatter',
    'random_crop',
    'mean_iou',
    'relu',
    'log',
    'crop',
    'rank_loss',
M
minqiyang 已提交
113
    'margin_rank_loss',
X
Xin Pan 已提交
114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156
    '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',
S
sneaxiy 已提交
157
    'sequence_reverse',
158
    'affine_channel',
Y
Yu Yang 已提交
159 160 161 162 163 164 165 166 167
]


def fc(input,
       size,
       num_flatten_dims=1,
       param_attr=None,
       bias_attr=None,
       act=None,
J
Jacek Czaja 已提交
168
       is_test=False,
169
       name=None):
Y
Yu Yang 已提交
170
    """
171
    **Fully Connected Layer**
Y
Yu Yang 已提交
172

173 174 175 176 177 178 179 180
    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
181
    to the output as well.
C
caoying03 已提交
182

C
caoying03 已提交
183
    This process can be formulated as follows:
184 185 186

    .. math::

187
        Out = Act({\sum_{i=0}^{N-1}X_iW_i + b})
188 189 190

    In the above equation:

C
caoying03 已提交
191 192 193 194
    * :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).
195
    * :math:`Act`: The activation function.
C
caoying03 已提交
196
    * :math:`Out`: The output tensor.
Y
Yu Yang 已提交
197 198

    Args:
R
ranqiu 已提交
199 200 201 202 203 204 205 206 207 208 209 210 211 212 213
        input (Variable|list of Variable): The input tensor(s) of this layer, and the dimension of
            the input tensor(s) is at least 2.
        size(int): The number of output units in this layer.
        num_flatten_dims (int, default 1): The fc layer can accept an input tensor with more than
            two dimensions. If this happens, the multidimensional tensor will first be flattened
            into a 2-dimensional matrix. The parameter `num_flatten_dims` determines how the input
            tensor is flattened: the first `num_flatten_dims` (inclusive, index starts from 1)
            dimensions will be flatten to form the first dimension of the final matrix (height of
            the matrix), and the rest `rank(X) - num_flatten_dims` dimensions are flattened to
            form the second dimension of the final matrix (width of the matrix). For example, suppose
            `X` is a 6-dimensional tensor with a shape [2, 3, 4, 5, 6], and `num_flatten_dims` = 3.
            Then, the flattened matrix will have a shape [2 x 3 x 4, 5 x 6] = [24, 30].
        param_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for learnable
            parameters/weights of this layer.
        bias_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for the bias
214 215
            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
216
        act (str, default None): Activation to be applied to the output of this layer.
J
Jacek Czaja 已提交
217
        is_test(bool): A flag indicating whether execution is in test phase.
R
ranqiu 已提交
218
        name (str, default None): The name of this layer.
Y
Yu Yang 已提交
219

220
    Returns:
F
fengjiayi 已提交
221
        Variable: The transformation result.
222 223

    Raises:
C
caoying03 已提交
224
        ValueError: If rank of the input tensor is less than 2.
225 226 227 228

    Examples:
        .. code-block:: python

F
fengjiayi 已提交
229
          data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
230
          fc = fluid.layers.fc(input=data, size=1000, act="tanh")
Y
Yu Yang 已提交
231
    """
C
caoying03 已提交
232

C
caoying03 已提交
233
    helper = LayerHelper("fc", **locals())
Y
Yu Yang 已提交
234 235 236 237

    dtype = helper.input_dtype()

    mul_results = []
238 239
    for input_var, param_attr in helper.iter_inputs_and_params():
        input_shape = input_var.shape
Y
Yu Yang 已提交
240 241 242
        param_shape = [
            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
        ] + [size]
Y
ying 已提交
243

Y
Yu Yang 已提交
244
        w = helper.create_parameter(
245 246
            attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False)
        tmp = helper.create_tmp_variable(dtype)
247
        helper.append_op(
248 249 250
            type="mul",
            inputs={"X": input_var,
                    "Y": w},
251
            outputs={"Out": tmp},
M
mozga-intel 已提交
252 253
            attrs={"x_num_col_dims": num_flatten_dims,
                   "y_num_col_dims": 1})
254 255 256 257
        mul_results.append(tmp)

    if len(mul_results) == 1:
        pre_bias = mul_results[0]
258
    else:
259 260
        pre_bias = helper.create_tmp_variable(dtype)
        helper.append_op(
261 262 263
            type="sum",
            inputs={"X": mul_results},
            outputs={"Out": pre_bias},
X
Xin Pan 已提交
264
            attrs={"use_mkldnn": False})
265 266 267 268
    # add bias
    pre_activation = helper.append_bias_op(pre_bias, dim_start=num_flatten_dims)
    # add activation
    return helper.append_activation(pre_activation)
Y
Yu Yang 已提交
269 270


271 272 273
def embedding(input,
              size,
              is_sparse=False,
274
              is_distributed=False,
275 276 277
              padding_idx=None,
              param_attr=None,
              dtype='float32'):
Y
Yu Yang 已提交
278
    """
279 280
    **Embedding Layer**

281
    This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in
282 283
    a lookup table. The result of this lookup is the embedding of each ID in the
    :attr:`input`.
284 285 286

    All the input variables are passed in as local variables to the LayerHelper
    constructor.
Y
Yu Yang 已提交
287 288

    Args:
289 290 291 292 293
        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.
294
        is_distributed(bool): Whether to run lookup table from remote parameter server.
295 296
        padding_idx(int|long|None): If :attr:`None`, it makes no effect to lookup.
            Otherwise the given :attr:`padding_idx` indicates padding the output
297
            with zeros whenever lookup encounters it in :attr:`input`. If
298
            :math:`padding_idx < 0`, the :attr:`padding_idx` to use in lookup is
299 300
            :math:`size[0] + dim`.
        param_attr(ParamAttr): Parameters for this layer
301
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc
Y
Yu Yang 已提交
302

303 304 305
    Returns:
        Variable: The tensor variable storing the embeddings of the \
                  supplied inputs.
Y
Yu Yang 已提交
306

307 308
    Examples:
        .. code-block:: python
Y
Yu Yang 已提交
309

C
chengduoZH 已提交
310
          dict_size = len(dataset.ids)
311
          data = fluid.layers.data(name='ids', shape=[32, 32], dtype='float32')
C
chengduoZH 已提交
312
          fc = fluid.layers.embedding(input=data, size=[dict_size, 16])
Y
Yu Yang 已提交
313 314 315 316 317 318
    """

    helper = LayerHelper('embedding', **locals())
    w = helper.create_parameter(
        attr=helper.param_attr, shape=size, dtype=dtype, is_bias=False)
    tmp = helper.create_tmp_variable(dtype)
319 320
    padding_idx = -1 if padding_idx is None else padding_idx if padding_idx >= 0 else (
        size[0] + padding_idx)
Y
Yu Yang 已提交
321 322 323 324 325
    helper.append_op(
        type='lookup_table',
        inputs={'Ids': input,
                'W': w},
        outputs={'Out': tmp},
326 327 328 329 330
        attrs={
            'is_sparse': is_sparse,
            'is_distributed': is_distributed,
            'padding_idx': padding_idx
        })
Y
Yu Yang 已提交
331 332 333
    return tmp


Y
yi.wu 已提交
334
@templatedoc(op_type="lstm")
Y
Yu Yang 已提交
335 336
def dynamic_lstm(input,
                 size,
Y
Yancey 已提交
337 338
                 h_0=None,
                 c_0=None,
Y
Yu Yang 已提交
339 340 341 342 343 344 345
                 param_attr=None,
                 bias_attr=None,
                 use_peepholes=True,
                 is_reverse=False,
                 gate_activation='sigmoid',
                 cell_activation='tanh',
                 candidate_activation='tanh',
346 347
                 dtype='float32',
                 name=None):
Y
Yibing Liu 已提交
348
    """
Y
yi.wu 已提交
349
    ${comment}
Y
Yibing Liu 已提交
350 351

    Args:
Y
yi.wu 已提交
352 353
        input (Variable): ${input_comment}
        size (int): 4 * hidden size.
Y
Yancey 已提交
354 355 356 357 358 359
        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.
360
        param_attr(ParamAttr|None): The parameter attribute for the learnable
361
                               hidden-hidden weights.
Y
Yibing Liu 已提交
362 363 364

                               - Weights = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}
365 366
                               - The shape is (D x 4D), where D is the hidden
                                 size.
C
chengduo 已提交
367 368 369 370 371

                               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.
Y
yi.wu 已提交
372
        bias_attr (ParamAttr|None): The bias attribute for the learnable bias
373 374 375
                              weights, which contains two parts, input-hidden
                              bias weights and peephole connections weights if
                              setting `use_peepholes` to `True`.
Y
Yibing Liu 已提交
376

377
                              1. `use_peepholes = False`
Y
yi.wu 已提交
378 379
                                 - Biases = {:math:`b_c, b_i, b_f, b_o`}.
                                 - The shape is (1 x 4D).
380
                              2. `use_peepholes = True`
Y
yi.wu 已提交
381
                                 - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
Y
Yibing Liu 已提交
382
                                                 W_{fc}, W_{oc}`}.
Y
yi.wu 已提交
383
                                 - The shape is (1 x 7D).
C
chengduo 已提交
384 385 386 387 388

                              If it is set to None or one attribute of ParamAttr,
                              dynamic_lstm will create ParamAttr as bias_attr.
                              If the Initializer of the bias_attr is not set,
                              the bias is initialized zero. Default: None.
Y
yi.wu 已提交
389 390 391 392 393 394 395 396
        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.
Y
Yibing Liu 已提交
397 398

    Returns:
Y
Yibing Liu 已提交
399 400
        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`.
Y
Yibing Liu 已提交
401

Y
Yibing Liu 已提交
402
    Examples:
Y
Yibing Liu 已提交
403 404
        .. code-block:: python

Y
Yibing Liu 已提交
405 406
            hidden_dim = 512
            forward_proj = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
C
chengduo 已提交
407
                                           bias_attr=False)
Y
Yibing Liu 已提交
408 409
            forward, _ = fluid.layers.dynamic_lstm(
                input=forward_proj, size=hidden_dim * 4, use_peepholes=False)
Y
Yibing Liu 已提交
410
    """
C
chengduo 已提交
411
    assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp."
Y
Yu Yang 已提交
412
    helper = LayerHelper('lstm', **locals())
M
minqiyang 已提交
413
    size = size // 4
Y
Yu Yang 已提交
414 415 416 417 418 419 420 421 422 423 424 425
    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_tmp_variable(dtype)
    cell = helper.create_tmp_variable(dtype)
    batch_gate = helper.create_tmp_variable(dtype)
    batch_cell_pre_act = helper.create_tmp_variable(dtype)
Y
Yancey 已提交
426 427 428 429 430 431 432 433 434 435
    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
Y
Yu Yang 已提交
436 437 438

    helper.append_op(
        type='lstm',
Y
Yancey 已提交
439
        inputs=inputs,
Y
Yu Yang 已提交
440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455
        outputs={
            'Hidden': hidden,
            'Cell': cell,
            'BatchGate': batch_gate,
            'BatchCellPreAct': batch_cell_pre_act
        },
        attrs={
            'use_peepholes': use_peepholes,
            'is_reverse': is_reverse,
            'gate_activation': gate_activation,
            'cell_activation': cell_activation,
            'candidate_activation': candidate_activation
        })
    return hidden, cell


Y
Yibing Liu 已提交
456 457 458 459 460 461 462 463 464 465 466
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',
467 468
                  dtype='float32',
                  name=None):
Y
Yibing Liu 已提交
469 470 471
    """
    **Dynamic LSTMP Layer**

472 473 474 475 476 477
    LSTMP (LSTM with recurrent projection) layer has a separate projection
    layer after the LSTM layer, projecting the original hidden state to a
    lower-dimensional one, which is proposed to reduce the number of total
    parameters and furthermore computational complexity for the LSTM,
    espeacially for the case that the size of output units is relative
    large (https://research.google.com/pubs/archive/43905.pdf).
Y
Yibing Liu 已提交
478 479 480 481 482

    The formula is as follows:

    .. math::

483
        i_t & = \sigma(W_{ix}x_{t} + W_{ir}r_{t-1} + W_{ic}c_{t-1} + b_i)
Y
Yibing Liu 已提交
484

485
        f_t & = \sigma(W_{fx}x_{t} + W_{fr}r_{t-1} + W_{fc}c_{t-1} + b_f)
Y
Yibing Liu 已提交
486

487
        \\tilde{c_t} & = act_g(W_{cx}x_t + W_{cr}r_{t-1} + b_c)
Y
Yibing Liu 已提交
488

489
        o_t & = \sigma(W_{ox}x_{t} + W_{or}r_{t-1} + W_{oc}c_t + b_o)
Y
Yibing Liu 已提交
490

491
        c_t & = f_t \odot c_{t-1} + i_t \odot \\tilde{c_t}
Y
Yibing Liu 已提交
492

493
        h_t & = o_t \odot act_h(c_t)
Y
Yibing Liu 已提交
494

495
        r_t & = \overline{act_h}(W_{rh}h_t)
Y
Yibing Liu 已提交
496

Y
Yibing Liu 已提交
497 498 499 500 501 502
    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, \
503
          we use vectors to reprenset these diagonal weight matrices.
Y
Yibing Liu 已提交
504
    * :math:`b`: Denotes bias vectors (e.g. :math:`b_i` is the input gate \
505
          bias vector).
Y
Yibing Liu 已提交
506 507 508
    * :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 \
509
          the same size as the cell output activation vector :math:`h`.
Y
Yibing Liu 已提交
510
    * :math:`h`: The hidden state.
511
    * :math:`r`: The recurrent projection of the hidden state.
Y
Yibing Liu 已提交
512 513
    * :math:`\\tilde{c_t}`: The candidate hidden state, whose \
          computation is based on the current input and previous hidden state.
514
    * :math:`\odot`: The element-wise product of the vectors.
Y
Yibing Liu 已提交
515
    * :math:`act_g` and :math:`act_h`: The cell input and cell output \
516
          activation functions and `tanh` is usually used for them.
Y
Yibing Liu 已提交
517 518
    * :math:`\overline{act_h}`: The activation function for the projection \
          output, usually using `identity` or same as :math:`act_h`.
Y
Yibing Liu 已提交
519 520 521 522

    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.
523

Y
Yibing Liu 已提交
524 525 526 527 528 529 530 531 532 533 534 535
    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.
536
        param_attr(ParamAttr|None): The parameter attribute for the learnable
Y
Yibing Liu 已提交
537 538
                               hidden-hidden weight and projection weight.

539 540
                               - Hidden-hidden weight = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}.
541 542
                               - The shape of hidden-hidden weight is (P x 4D),
                                 where P is the projection size and D the hidden
Y
Yibing Liu 已提交
543 544
                                 size.
                               - Projection weight = {:math:`W_{rh}`}.
545
                               - The shape of projection weight is (D x P).
C
chengduo 已提交
546 547 548 549 550

                               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.
551
        bias_attr(ParamAttr|None): The bias attribute for the learnable bias
Y
Yibing Liu 已提交
552 553 554 555 556 557
                              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`}.
558
                                - The shape is (1 x 4D).
Y
Yibing Liu 已提交
559 560 561
                              2. `use_peepholes = True`
                                - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
                                                 W_{fc}, W_{oc}`}.
562
                                - The shape is (1 x 7D).
C
chengduo 已提交
563 564 565 566 567

                              If it is set to None or one attribute of ParamAttr,
                              dynamic_lstm will create ParamAttr as bias_attr.
                              If the Initializer of the bias_attr is not set,
                              the bias is initialized zero. Default: None.
Y
Yibing Liu 已提交
568 569 570 571 572 573 574 575 576
        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.
577
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
Y
Yibing Liu 已提交
578 579
                              default "tanh".
        proj_activation(str): The activation for projection output.
580
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
Y
Yibing Liu 已提交
581 582
                              default "tanh".
        dtype(str): Data type. Choices = ["float32", "float64"], default "float32".
583 584
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
Y
Yibing Liu 已提交
585 586

    Returns:
587 588 589 590
        tuple: A tuple of two output variable: the projection of hidden state, \
               and cell state of LSTMP. The shape of projection is (T x P), \
               for the cell state which is (T x D), and both LoD is the same \
               with the `input`.
Y
Yibing Liu 已提交
591 592

    Examples:
593

Y
Yibing Liu 已提交
594 595
        .. code-block:: python

596 597 598 599
            dict_dim, emb_dim = 128, 64
            data = fluid.layers.data(name='sequence', shape=[1],
                                     dtype='int32', lod_level=1)
            emb = fluid.layers.embedding(input=data, size=[dict_dim, emb_dim])
Y
Yibing Liu 已提交
600
            hidden_dim, proj_dim = 512, 256
601
            fc_out = fluid.layers.fc(input=emb, size=hidden_dim * 4,
Y
Yibing Liu 已提交
602
                                     act=None, bias_attr=None)
603 604 605
            proj_out, _ = fluid.layers.dynamic_lstmp(input=fc_out,
                                                     size=hidden_dim * 4,
                                                     proj_size=proj_dim,
Y
Yibing Liu 已提交
606 607 608 609
                                                     use_peepholes=False,
                                                     is_reverse=True,
                                                     cell_activation="tanh",
                                                     proj_activation="tanh")
Y
Yibing Liu 已提交
610
    """
611

C
chengduo 已提交
612
    assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp."
Y
Yibing Liu 已提交
613
    helper = LayerHelper('lstmp', **locals())
M
minqiyang 已提交
614
    size = size // 4
Y
Yibing Liu 已提交
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 = 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)

    projection = helper.create_tmp_variable(dtype)
    cell = helper.create_tmp_variable(dtype)
    ordered_proj0 = helper.create_tmp_variable(dtype)
    batch_hidden = helper.create_tmp_variable(dtype)
    batch_gate = helper.create_tmp_variable(dtype)
    batch_cell_pre_act = helper.create_tmp_variable(dtype)

    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
659 660 661 662 663 664 665 666 667
def dynamic_gru(input,
                size,
                param_attr=None,
                bias_attr=None,
                is_reverse=False,
                gate_activation='sigmoid',
                candidate_activation='tanh',
                h_0=None):
    """
668
    **Gated Recurrent Unit (GRU) Layer**
G
guosheng 已提交
669

670
    Refer to `Empirical Evaluation of Gated Recurrent Neural Networks on
671
    Sequence Modeling <https://arxiv.org/abs/1412.3555>`_ .
672

G
guosheng 已提交
673 674 675 676 677 678 679 680 681
    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)
682

G
guosheng 已提交
683
        h_t & = (1-u_t) \odot h_{t-1} + u_t \odot \\tilde{h_t}
684

G
guosheng 已提交
685
    The :math:`\odot` is the element-wise product of the vectors. :math:`act_g`
686 687
    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
688 689 690 691
    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
692
    to use fully-connect layer before GRU layer.
G
guosheng 已提交
693 694

    Args:
695 696
        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
697
            Variable is a matrix with shape :math:`(T \\times 3D)`, where
698
            :math:`T` is the total time steps in this mini-batch, :math:`D`
G
guosheng 已提交
699 700
            is the hidden size.
        size(int): The dimension of the gru cell.
701
        param_attr(ParamAttr|None): The parameter attribute for the learnable
G
guosheng 已提交
702 703
            hidden-hidden weight matrix. Note:

704
            - The shape of the weight matrix is :math:`(T \\times 3D)`, where
G
guosheng 已提交
705
              :math:`D` is the hidden size.
706
            - All elements in the weight matrix can be divided into two parts.
G
guosheng 已提交
707
              The first part are weights of the update gate and reset gate with
708
              shape :math:`(D \\times 2D)`, and the second part are weights for
G
guosheng 已提交
709
              candidate hidden state with shape :math:`(D \\times D)`.
710
        bias_attr(ParamAttr): The parameter attribute for learnable the
G
guosheng 已提交
711
            hidden-hidden bias.
712
        is_reverse(bool): Whether to compute reversed GRU, default
G
guosheng 已提交
713 714 715
            :attr:`False`.
        gate_activation(str): The activation for update gate and reset gate.
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "sigmoid".
716
        candidate_activation(str): The activation for candidate hidden state.
G
guosheng 已提交
717
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh".
718 719 720 721
        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
722 723

    Returns:
G
guosheng 已提交
724
        Variable: The hidden state of GRU. The shape is :math:`(T \\times D)`, \
725
            and sequence length is the same with the input.
726

G
guosheng 已提交
727
    Examples:
728

G
guosheng 已提交
729 730
        .. code-block:: python

731 732 733 734
            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
735
            hidden_dim = 512
736
            x = fluid.layers.fc(input=emb, size=hidden_dim * 3)
G
guosheng 已提交
737 738 739 740 741 742 743 744 745 746
            hidden = fluid.layers.dynamic_gru(input=x, dim=hidden_dim)
    """

    helper = LayerHelper('gru', **locals())
    dtype = helper.input_dtype()

    weight = helper.create_parameter(
        attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype)
    bias = helper.create_parameter(
        attr=helper.bias_attr, shape=[1, 3 * size], dtype=dtype, is_bias=True)
Y
Yancey 已提交
747
    batch_size = input.shape[0]
G
guosheng 已提交
748 749 750
    inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
    if h_0 != None:
        assert h_0.shape == (
Y
Yancey 已提交
751 752 753
            batch_size, size
        ), 'The shape of h0 should be(batch_size, %d)' % size
        inputs['H0'] = h_0
G
guosheng 已提交
754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776

    hidden = helper.create_tmp_variable(dtype)
    batch_gate = helper.create_tmp_variable(dtype)
    batch_reset_hidden_prev = helper.create_tmp_variable(dtype)
    batch_hidden = helper.create_tmp_variable(dtype)

    helper.append_op(
        type='gru',
        inputs=inputs,
        outputs={
            'Hidden': hidden,
            'BatchGate': batch_gate,
            'BatchResetHiddenPrev': batch_reset_hidden_prev,
            'BatchHidden': batch_hidden
        },
        attrs={
            'is_reverse': is_reverse,
            'gate_activation': gate_activation,
            'activation': candidate_activation
        })
    return hidden


Y
Yu Yang 已提交
777 778 779
def gru_unit(input,
             hidden,
             size,
780 781
             param_attr=None,
             bias_attr=None,
Y
Yu Yang 已提交
782
             activation='tanh',
783
             gate_activation='sigmoid'):
Y
Yu Yang 已提交
784
    """
785
    GRU unit layer. The equation of a gru step is:
Y
Yu Yang 已提交
786

787 788
        .. math::
            u_t & = actGate(xu_{t} + W_u h_{t-1} + b_u)
Y
Yu Yang 已提交
789

790
            r_t & = actGate(xr_{t} + W_r h_{t-1} + b_r)
Y
Yu Yang 已提交
791

792
            m_t & = actNode(xm_t + W_c dot(r_t, h_{t-1}) + b_m)
793

794
            h_t & = dot((1-u_t), m_t) + dot(u_t, h_{t-1})
795 796

    The inputs of gru unit includes :math:`z_t`, :math:`h_{t-1}`. In terms
797 798 799
    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
800 801
    connected layer has to be applied, such that :math:`z_t = W_{fc}x_t`.

802 803
    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
804 805 806
    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`.
807 808 809 810 811

    Args:
        input (Variable): The fc transformed input value of current step.
        hidden (Variable): The hidden value of lstm unit from previous step.
        size (integer): The input dimension value.
812 813
        param_attr (ParamAttr): The weight parameters for gru unit. Default: None
        bias_attr (ParamAttr): The bias parameters for gru unit. Default: None
814 815 816 817
        activation (string): The activation type for cell (actNode).
                             Default: 'tanh'
        gate_activation (string): The activation type for gates (actGate).
                                  Default: 'sigmoid'
Y
Yu Yang 已提交
818

819 820 821 822 823 824
    Returns:
        tuple: The hidden value, reset-hidden value and gate values.

    Examples:

        .. code-block:: python
Y
Yu Yang 已提交
825

826
             # assuming we have x_t_data and prev_hidden of size=10
827
             x_t = fluid.layers.fc(input=x_t_data, size=30)
828 829
             hidden_val, r_h_val, gate_val = fluid.layers.gru_unit(input=x_t,
                                                    hidden = prev_hidden)
Y
Yu Yang 已提交
830 831 832 833 834 835 836 837 838 839 840 841

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

    helper = LayerHelper('gru_unit', **locals())
    dtype = helper.input_dtype()
M
minqiyang 已提交
842
    size = size // 3
Y
Yu Yang 已提交
843 844

    # create weight
845 846
    weight = helper.create_parameter(
        attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype)
Y
Yu Yang 已提交
847

848 849 850 851
    gate = helper.create_tmp_variable(dtype)
    reset_hidden_pre = helper.create_tmp_variable(dtype)
    updated_hidden = helper.create_tmp_variable(dtype)
    inputs = {'Input': input, 'HiddenPrev': hidden, 'Weight': weight}
Y
Yu Yang 已提交
852
    # create bias
853
    if helper.bias_attr:
Y
Yu Yang 已提交
854 855 856
        bias_size = [1, 3 * size]
        bias = helper.create_parameter(
            attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
857
        inputs['Bias'] = bias
Y
Yu Yang 已提交
858 859 860

    helper.append_op(
        type='gru_unit',
861
        inputs=inputs,
Y
Yu Yang 已提交
862 863 864 865 866 867
        outputs={
            'Gate': gate,
            'ResetHiddenPrev': reset_hidden_pre,
            'Hidden': updated_hidden,
        },
        attrs={
868 869
            'activation': 2,  # tanh
            'gate_activation': 1,  # sigmoid
Y
Yu Yang 已提交
870 871 872 873 874
        })

    return updated_hidden, reset_hidden_pre, gate


Y
yuyang18 已提交
875
@templatedoc()
876
def linear_chain_crf(input, label, param_attr=None):
Y
yuyang18 已提交
877 878 879 880 881 882 883
    """
    Linear Chain CRF.

    ${comment}

    Args:
        input(${emission_type}): ${emission_comment}
D
dzhwinter 已提交
884
        input(${transition_type}): ${transition_comment}
Y
yuyang18 已提交
885 886 887 888
        label(${label_type}): ${label_comment}
        param_attr(ParamAttr): The attribute of the learnable parameter.

    Returns:
D
dzhwinter 已提交
889 890 891
        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
892 893

    """
Y
Yu Yang 已提交
894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918
    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())
    alpha = helper.create_tmp_variable(dtype=helper.input_dtype())
    emission_exps = helper.create_tmp_variable(dtype=helper.input_dtype())
    transition_exps = helper.create_tmp_variable(dtype=helper.input_dtype())
    log_likelihood = helper.create_tmp_variable(dtype=helper.input_dtype())
    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


Y
yuyang18 已提交
919
@templatedoc()
920
def crf_decoding(input, param_attr, label=None):
Y
yuyang18 已提交
921 922 923 924 925
    """
    ${comment}

    Args:
        input(${emission_type}): ${emission_comment}
Y
yi.wu 已提交
926

Y
yuyang18 已提交
927
        param_attr(ParamAttr): The parameter attribute for training.
Y
yi.wu 已提交
928

Y
yuyang18 已提交
929 930 931
        label(${label_type}): ${label_comment}

    Returns:
Y
update  
yi.wu 已提交
932
        Variable: ${viterbi_path_comment}
933

Y
yi.wu 已提交
934 935 936 937 938
    Examples:
        .. code-block:: python

           crf_decode = layers.crf_decoding(
                input=hidden, param_attr=ParamAttr(name="crfw"))
Y
yuyang18 已提交
939
    """
Y
Yu Yang 已提交
940 941 942 943 944 945 946 947 948 949 950 951 952
    helper = LayerHelper('crf_decoding', **locals())
    transition = helper.get_parameter(param_attr.name)
    viterbi_path = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
        type='crf_decoding',
        inputs={"Emission": [input],
                "Transition": transition,
                "Label": label},
        outputs={"ViterbiPath": [viterbi_path]})

    return viterbi_path


Y
yi.wu 已提交
953
@templatedoc()
F
fengjiayi 已提交
954
def cos_sim(X, Y):
Y
Yu Yang 已提交
955
    """
Y
yi.wu 已提交
956 957 958
    ${comment}

    Args:
959 960
        X (Variable): ${x_comment}.
        Y (Variable): ${y_comment}.
F
fengjiayi 已提交
961

Y
yi.wu 已提交
962
    Returns:
963
        Variable: the output of cosine(X, Y).
Y
Yu Yang 已提交
964
    """
F
fengjiayi 已提交
965
    helper = LayerHelper('cos_sim', **locals())
Y
Yu Yang 已提交
966 967 968 969 970 971 972 973 974 975 976 977 978
    out = helper.create_tmp_variable(dtype=X.dtype)
    xnorm = helper.create_tmp_variable(dtype=X.dtype)
    ynorm = helper.create_tmp_variable(dtype=X.dtype)
    helper.append_op(
        type='cos_sim',
        inputs={'X': [X],
                'Y': [Y]},
        outputs={'Out': [out],
                 'XNorm': [xnorm],
                 'YNorm': [ynorm]})
    return out


979
def dropout(x, dropout_prob, is_test=False, seed=None, name=None):
980 981 982 983 984
    """
    Computes dropout.

    Drop or keep each element of `x` independently. Dropout is a regularization
    technique for reducing overfitting by preventing neuron co-adaption during
985
    training. The dropout operator randomly sets (according to the given dropout
986 987 988 989
    probability) the outputs of some units to zero, while others are remain
    unchanged.

    Args:
990 991
        x (Variable): The input tensor variable.
        dropout_prob (float): Probability of setting units to zero.
992 993 994 995 996 997 998
        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.
999 1000

    Returns:
1001
        Variable: A tensor variable is the shape with `x`.
1002 1003

    Examples:
1004

1005 1006
        .. code-block:: python

1007 1008
            x = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
            droped = fluid.layers.dropout(x, dropout_prob=0.5)
1009 1010
    """

F
fengjiayi 已提交
1011
    helper = LayerHelper('dropout', **locals())
1012 1013
    out = helper.create_tmp_variable(dtype=x.dtype)
    mask = helper.create_tmp_variable(dtype=x.dtype, stop_gradient=True)
C
chengduo 已提交
1014 1015 1016 1017

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

1018 1019 1020 1021 1022
    helper.append_op(
        type='dropout',
        inputs={'X': [x]},
        outputs={'Out': [out],
                 'Mask': [mask]},
1023 1024 1025 1026 1027 1028
        attrs={
            'dropout_prob': dropout_prob,
            'is_test': is_test,
            'fix_seed': seed is not None,
            'seed': seed if seed is not None else 0
        })
1029 1030 1031
    return out


1032
def cross_entropy(input, label, soft_label=False, ignore_index=-100):
Y
Yu Yang 已提交
1033
    """
Y
Yibing Liu 已提交
1034 1035
    **Cross Entropy Layer**

1036 1037 1038
    This layer computes the cross entropy between `input` and `label`. It
    supports both standard cross-entropy and soft-label cross-entropy loss
    computation.
Y
Yibing Liu 已提交
1039 1040

    1) One-hot cross-entropy:
F
fengjiayi 已提交
1041
        `soft_label = False`, `Label[i, 0]` indicates the class index for sample i:
Y
yangyaming 已提交
1042

Y
Yibing Liu 已提交
1043
        .. math::
Y
yangyaming 已提交
1044

Y
Yibing Liu 已提交
1045 1046 1047
            Y[i] = -\log(X[i, Label[i]])

    2) Soft-label cross-entropy:
F
fengjiayi 已提交
1048 1049
        `soft_label = True`, `Label[i, j]` indicates the soft label of class j
        for sample i:
Y
Yibing Liu 已提交
1050 1051 1052 1053 1054

        .. math::

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

Y
Yibing Liu 已提交
1055
       Please make sure that in this case the summation of each row of `label`
Y
Yibing Liu 已提交
1056 1057 1058
       equals one.

    3) One-hot cross-entropy with vecterized `label`:
F
fengjiayi 已提交
1059 1060
         As a special case of 2), when each row of 'label' has only one
         non-zero element which is equal to 1, soft-label cross-entropy degenerates
Y
Yibing Liu 已提交
1061
         to a one-hot cross-entropy with one-hot label representation.
Y
yangyaming 已提交
1062

Y
Yibing Liu 已提交
1063
    Args:
Y
yangyaming 已提交
1064
        input (Variable|list):  a 2-D tensor with shape [N x D], where N is the
1065 1066 1067 1068
                                batch size and D is the number of classes. This
                                input is a probability computed by the previous
                                operator, which is almost always the result of
                                a softmax operator.
Y
yangyaming 已提交
1069
        label (Variable|list): the ground truth which is a 2-D tensor. When
1070 1071 1072 1073
                               `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
1074
        soft_label (bool): a flag indicating whether to
1075
                                           interpretate the given labels as soft
1076
                                           labels. Default: `False`.
M
minqiyang 已提交
1077 1078
        ignore_index (int): Specifies a target value that is ignored and does
                            not contribute to the input gradient. Only valid
1079
                            if soft_label is set to False. Default: -100
Y
Yibing Liu 已提交
1080 1081 1082 1083 1084

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

    Raises:
1085 1086 1087 1088 1089
        `ValueError`: 1) the 1st dimension of `input` and `label` are not equal.
                      2) when `soft_label == True`, and the 2nd dimension of
                         `input` and `label` are not equal.
                      3) when `soft_label == False`, and the 2nd dimension of
                         `label` is not 1.
Y
Yibing Liu 已提交
1090 1091 1092 1093 1094 1095

    Examples:
        .. code-block:: python

          predict = fluid.layers.fc(input=net, size=classdim, act='softmax')
          cost = fluid.layers.cross_entropy(input=predict, label=label)
Y
Yu Yang 已提交
1096
    """
F
fengjiayi 已提交
1097
    helper = LayerHelper('cross_entropy', **locals())
Y
Yu Yang 已提交
1098 1099 1100 1101 1102 1103
    out = helper.create_tmp_variable(dtype=input.dtype)
    helper.append_op(
        type='cross_entropy',
        inputs={'X': [input],
                'Label': [label]},
        outputs={'Y': [out]},
1104 1105
        attrs={"soft_label": soft_label,
               "ignore_index": ignore_index})
Y
Yu Yang 已提交
1106 1107 1108
    return out


F
fengjiayi 已提交
1109
def square_error_cost(input, label):
Y
Yu Yang 已提交
1110
    """
1111 1112
    **Square error cost layer**

1113 1114
    This layer accepts input predictions and target label and returns the
    squared error cost.
Y
ying 已提交
1115

1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128
    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:
1129 1130
        input (Variable): Input tensor, has predictions.
        label (Variable): Label tensor, has target labels.
1131 1132

    Returns:
G
guosheng 已提交
1133
        Variable: The tensor variable storing the element-wise squared error \
1134
                  difference of input and label.
1135 1136 1137 1138 1139 1140 1141 1142

    Examples:
        .. code-block:: python

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

Y
Yu Yang 已提交
1143
    """
F
fengjiayi 已提交
1144
    helper = LayerHelper('square_error_cost', **locals())
Y
Yu Yang 已提交
1145 1146 1147 1148 1149 1150 1151 1152 1153
    minus_out = helper.create_tmp_variable(dtype=input.dtype)
    helper.append_op(
        type='elementwise_sub',
        inputs={'X': [input],
                'Y': [label]},
        outputs={'Out': [minus_out]})

    square_out = helper.create_tmp_variable(dtype=input.dtype)
    helper.append_op(
F
fengjiayi 已提交
1154 1155
        type='square', inputs={'X': [minus_out]},
        outputs={'Out': [square_out]})
Y
Yu Yang 已提交
1156 1157 1158
    return square_out


Y
yi.wu 已提交
1159
@templatedoc()
Y
Yu Yang 已提交
1160 1161 1162 1163
def chunk_eval(input,
               label,
               chunk_scheme,
               num_chunk_types,
F
fengjiayi 已提交
1164
               excluded_chunk_types=None):
Y
Yu Yang 已提交
1165
    """
Y
yi.wu 已提交
1166
    **Chunk Evaluator**
Y
yi.wu 已提交
1167

Y
yangyaming 已提交
1168
    This function computes and outputs the precision, recall and
1169
    F1-score of chunk detection.
Y
yi.wu 已提交
1170

Y
yi.wu 已提交
1171 1172 1173 1174 1175 1176 1177 1178
    For some basics of chunking, please refer to
    'Chunking with Support Vector Machines <https://aclanthology.info/pdf/N/N01/N01-1025.pdf>'.

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

    .. code-block:: python
1179

Y
yi.wu 已提交
1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204
       ====== ====== ======  =====  ==  ============   =====  ===== =====  ==  =========
              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
1205

Y
yi.wu 已提交
1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229
       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
1230
    Args:
1231 1232 1233 1234 1235
        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
1236

Y
yi.wu 已提交
1237
    Returns:
Y
update  
yi.wu 已提交
1238 1239 1240
        tuple: tuple containing: precision, recall, f1_score,
        num_infer_chunks, num_label_chunks,
        num_correct_chunks
1241

Y
yi.wu 已提交
1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253
    Examples:
        .. code-block:: python

            crf = fluid.layers.linear_chain_crf(
                input=hidden, label=label, param_attr=ParamAttr(name="crfw"))
            crf_decode = fluid.layers.crf_decoding(
                input=hidden, param_attr=ParamAttr(name="crfw"))
            fluid.layers.chunk_eval(
                input=crf_decode,
                label=label,
                chunk_scheme="IOB",
                num_chunk_types=(label_dict_len - 1) / 2)
Y
Yu Yang 已提交
1254
    """
F
fengjiayi 已提交
1255
    helper = LayerHelper("chunk_eval", **locals())
Y
Yu Yang 已提交
1256 1257 1258 1259 1260

    # prepare output
    precision = helper.create_tmp_variable(dtype="float32")
    recall = helper.create_tmp_variable(dtype="float32")
    f1_score = helper.create_tmp_variable(dtype="float32")
1261 1262 1263
    num_infer_chunks = helper.create_tmp_variable(dtype="int64")
    num_label_chunks = helper.create_tmp_variable(dtype="int64")
    num_correct_chunks = helper.create_tmp_variable(dtype="int64")
Y
Yu Yang 已提交
1264 1265 1266 1267 1268 1269 1270 1271

    helper.append_op(
        type="chunk_eval",
        inputs={"Inference": [input],
                "Label": [label]},
        outputs={
            "Precision": [precision],
            "Recall": [recall],
1272 1273 1274 1275
            "F1-Score": [f1_score],
            "NumInferChunks": [num_infer_chunks],
            "NumLabelChunks": [num_label_chunks],
            "NumCorrectChunks": [num_correct_chunks]
Y
Yu Yang 已提交
1276 1277 1278
        },
        attrs={
            "num_chunk_types": num_chunk_types,
G
guosheng 已提交
1279 1280
            "chunk_scheme": chunk_scheme,
            "excluded_chunk_types": excluded_chunk_types or []
Y
Yu Yang 已提交
1281
        })
1282 1283
    return (precision, recall, f1_score, num_infer_chunks, num_label_chunks,
            num_correct_chunks)
Y
Yu Yang 已提交
1284 1285


1286
@templatedoc()
Y
Yu Yang 已提交
1287 1288 1289 1290 1291 1292 1293
def sequence_conv(input,
                  num_filters,
                  filter_size=3,
                  filter_stride=1,
                  padding=None,
                  bias_attr=None,
                  param_attr=None,
C
chengduo 已提交
1294 1295
                  act=None,
                  name=None):
Y
Yu Yang 已提交
1296 1297 1298 1299
    """
    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.
1300 1301 1302 1303 1304 1305 1306

    Args:
        input (Variable): ${x_comment}
        num_filters (int): number of filters.
        filter_size (int): the filter size (H and W).
        filter_stride (int): stride of the filter.
        padding (bool): if True, add paddings.
C
chengduo 已提交
1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319
        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
1320

1321 1322
    Returns:
        Variable: output of sequence_conv
Y
Yu Yang 已提交
1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340
    """

    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)
    pre_bias = helper.create_tmp_variable(dtype)

    helper.append_op(
        type='sequence_conv',
        inputs={
            'X': [input],
            'Filter': [filter_param],
        },
        outputs={"Out": pre_bias},
        attrs={
            'contextStride': filter_stride,
M
minqiyang 已提交
1341
            'contextStart': -int(filter_size // 2),
Y
Yu Yang 已提交
1342 1343 1344 1345 1346 1347
            'contextLength': filter_size
        })
    pre_act = helper.append_bias_op(pre_bias)
    return helper.append_activation(pre_act)


C
chengduo 已提交
1348
def sequence_softmax(input, use_cudnn=False, name=None):
1349 1350 1351
    """
    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
1352
    input Tensor can be either :math:`[N, 1]` or :math:`[N]`, where :math:`N`
1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368
    is the sum of the length of all sequences.

    For i-th sequence in a mini-batch:

    .. math::

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

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

    Args:
        input (Variable): The input variable which is a LoDTensor.
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \
C
chengduo 已提交
1369 1370 1371
            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.
1372

1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383
    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)
    """
1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394
    helper = LayerHelper('sequence_softmax', **locals())
    dtype = helper.input_dtype()
    softmax_out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type="sequence_softmax",
        inputs={"X": input},
        outputs={"Out": softmax_out},
        attrs={"use_cudnn": use_cudnn})
    return softmax_out


C
chengduo 已提交
1395
def softmax(input, use_cudnn=True, name=None):
Q
qiaolongfei 已提交
1396
    """
1397
    The input of the softmax operator is a tensor of any rank. The output tensor
F
fengjiayi 已提交
1398
    has the same shape as the input.
Q
qiaolongfei 已提交
1399

1400 1401 1402 1403 1404 1405
    The input tensor will first be logically flattened to a 2-D matrix. The matrix's
    second dimension(row length) is as same as the last dimension of the input
    tensor, and the first dimension(column length) is the product of all other
    dimensions of the input tensor. For each row of the matrix, the softmax operator
    squashes the K-dimensional(K is the width of the matrix, which is also the size
    of the input tensor's last dimension) vector of arbitrary real values to a
F
fengjiayi 已提交
1406
    K-dimensional vector of real values in the range [0, 1] that add up to 1.
Q
qiaolongfei 已提交
1407 1408 1409 1410 1411 1412 1413

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

F
fengjiayi 已提交
1414
    For each row :math:`i` and each column :math:`j` in the matrix, we have:
Q
qiaolongfei 已提交
1415 1416 1417 1418 1419 1420 1421 1422

    .. math::

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

    Args:
        input (Variable): The input variable.
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \
C
chengduo 已提交
1423 1424 1425
            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
1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437

    Returns:
        Variable: output of softmax

    Examples:

        .. code-block:: python

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

    """
1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448
    helper = LayerHelper('softmax', **locals())
    dtype = helper.input_dtype()
    softmax_out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type="softmax",
        inputs={"X": input},
        outputs={"Out": softmax_out},
        attrs={"use_cudnn": use_cudnn})
    return softmax_out


Y
Yu Yang 已提交
1449 1450 1451
def conv2d(input,
           num_filters,
           filter_size,
C
chengduoZH 已提交
1452 1453
           stride=1,
           padding=0,
1454
           dilation=1,
Y
Yu Yang 已提交
1455 1456 1457
           groups=None,
           param_attr=None,
           bias_attr=None,
C
chengduoZH 已提交
1458
           use_cudnn=True,
1459 1460
           act=None,
           name=None):
Y
Yu Yang 已提交
1461
    """
C
chengduoZH 已提交
1462
    The convolution2D layer calculates the output based on the input, filter
T
tensor-tang 已提交
1463 1464
    and strides, paddings, dilations, groups parameters. Input and
    Output are in NCHW format, where N is batch size, C is the number of
1465
    channels, H is the height of the feature, and W is the width of the feature.
T
tensor-tang 已提交
1466 1467 1468 1469 1470 1471 1472
    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.
1473 1474 1475
    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
1476

1477
    For each input :math:`X`, the equation is:
C
refine  
chengduoZH 已提交
1478

C
chengduoZH 已提交
1479 1480
    .. math::

C
refine  
chengduoZH 已提交
1481
        Out = \sigma (W \\ast X + b)
C
chengduoZH 已提交
1482

T
tensor-tang 已提交
1483
    Where:
C
chengduoZH 已提交
1484

1485 1486 1487 1488 1489
    * :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
1490
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
1491 1492 1493

    Example:

1494 1495
        - Input:

W
weixing02 已提交
1496
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
C
refine  
chengduoZH 已提交
1497

W
weixing02 已提交
1498
          Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)`
C
refine  
chengduoZH 已提交
1499

1500
        - Output:
T
tensor-tang 已提交
1501

W
weixing02 已提交
1502
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
C
refine  
chengduoZH 已提交
1503

C
chengduoZH 已提交
1504
        Where
1505 1506

        .. math::
C
chengduoZH 已提交
1507

W
weixing02 已提交
1508 1509
            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
1510 1511

    Args:
1512
        input (Variable): The input image with [N, C, H, W] format.
T
tensor-tang 已提交
1513
        num_filters(int): The number of filter. It is as same as the output
1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530
            image channel.
        filter_size (int|tuple|None): The filter size. If filter_size is a tuple,
            it must contain two integers, (filter_size_H, filter_size_W).
            Otherwise, the filter will be a square.
        stride (int|tuple): The stride size. If stride is a tuple, it must
            contain two integers, (stride_H, stride_W). Otherwise, the
            stride_H = stride_W = stride. Default: stride = 1.
        padding (int|tuple): The padding size. If padding is a tuple, it must
            contain two integers, (padding_H, padding_W). Otherwise, the
            padding_H = padding_W = padding. Default: padding = 0.
        dilation (int|tuple): The dilation size. If dilation is a tuple, it must
            contain two integers, (dilation_H, dilation_W). Otherwise, the
            dilation_H = dilation_W = dilation. Default: dilation = 1.
        groups (int): The groups number of the Conv2d Layer. According to grouped
            convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
            the first half of the filters is only connected to the first half
            of the input channels, while the second half of the filters is only
C
chengduo 已提交
1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541
            connected to the second half of the input channels. Default: groups=1.
        param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
            of conv2d. If it is set to None or one attribute of ParamAttr, conv2d
            will create ParamAttr as param_attr. If the Initializer of the param_attr
            is not set, the parameter is initialized with :math:`Normal(0.0, std)`,
             and the :math:`std` is :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None.
        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv2d.
            If it is set to False, no bias will be added to the output units.
            If it is set to None or one attribute of ParamAttr, conv2d
            will create ParamAttr as bias_attr. If the Initializer of the bias_attr
            is not set, the bias is initialized zero. Default: None.
1542 1543
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
C
chengduo 已提交
1544 1545
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None
1546
        name (str|None): A name for this layer(optional). If set None, the layer
C
chengduo 已提交
1547
            will be named automatically. Default: None
C
chengduoZH 已提交
1548 1549

    Returns:
G
guosheng 已提交
1550
        Variable: The tensor variable storing the convolution and \
C
chengduoZH 已提交
1551 1552
                  non-linearity activation result.

C
refine  
chengduoZH 已提交
1553
    Raises:
1554 1555
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
C
refine  
chengduoZH 已提交
1556

C
chengduoZH 已提交
1557 1558 1559
    Examples:
        .. code-block:: python

1560 1561
          data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32')
          conv2d = fluid.layers.conv2d(input=data, num_filters=2, filter_size=3, act="relu")
Y
Yu Yang 已提交
1562 1563 1564
    """

    num_channels = input.shape[1]
C
chengduo 已提交
1565
    assert param_attr is not False, "param_attr should not be False here."
1566
    l_type = 'conv2d'
X
xzl 已提交
1567 1568
    if (num_channels == groups and num_filters % num_channels == 0 and
            not use_cudnn):
1569
        l_type = 'depthwise_conv2d'
1570 1571 1572 1573

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

Y
Yu Yang 已提交
1574 1575 1576 1577 1578
    if groups is None:
        num_filter_channels = num_channels
    else:
        if num_channels % groups != 0:
            raise ValueError("num_channels must be divisible by groups.")
M
minqiyang 已提交
1579
        num_filter_channels = num_channels // groups
Y
Yu Yang 已提交
1580

C
chengduoZH 已提交
1581 1582 1583
    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')
1584
    dilation = utils.convert_to_list(dilation, 2, 'dilation')
C
chengduoZH 已提交
1585

C
chengduoZH 已提交
1586 1587
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1588 1589

    input_shape = input.shape
M
minqiyang 已提交
1590
    filter_shape = [num_filters, int(num_filter_channels)] + filter_size
Y
Yu Yang 已提交
1591 1592

    def _get_default_param_initializer():
C
chengduo 已提交
1593 1594
        filter_elem_num = filter_size[0] * filter_size[1] * num_channels
        std = (2.0 / filter_elem_num)**0.5
Y
Yu Yang 已提交
1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605
        return Normal(0.0, std, 0)

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

    pre_bias = helper.create_tmp_variable(dtype)

    helper.append_op(
1606
        type=l_type,
Y
Yu Yang 已提交
1607 1608 1609 1610 1611
        inputs={
            'Input': input,
            'Filter': filter_param,
        },
        outputs={"Output": pre_bias},
C
chengduoZH 已提交
1612 1613 1614
        attrs={
            'strides': stride,
            'paddings': padding,
1615
            'dilations': dilation,
C
chengduoZH 已提交
1616
            'groups': groups,
1617
            'use_cudnn': use_cudnn,
X
Xin Pan 已提交
1618
            'use_mkldnn': False
C
chengduoZH 已提交
1619
        })
Y
Yu Yang 已提交
1620 1621 1622 1623 1624 1625

    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
1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642
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
1643 1644 1645 1646 1647 1648
    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
1649 1650 1651 1652 1653 1654 1655 1656 1657

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

    .. math::

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

    In the above equation:

1658 1659
    * :math:`X`: Input value, a tensor with NCDHW format.
    * :math:`W`: Filter value, a tensor with MCDHW format.
C
chengduoZH 已提交
1660 1661 1662
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
1663
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688

    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,
1689
            it must contain three integers, (filter_size_D, filter_size_H, filter_size_W).
C
chengduoZH 已提交
1690 1691
            Otherwise, the filter will be a square.
        stride (int|tuple): The stride size. If stride is a tuple, it must
1692
            contain three integers, (stride_D, stride_H, stride_W). Otherwise, the
C
chengduoZH 已提交
1693 1694
            stride_D = stride_H = stride_W = stride. Default: stride = 1.
        padding (int|tuple): The padding size. If padding is a tuple, it must
1695
            contain three integers, (padding_D, padding_H, padding_W). Otherwise, the
C
chengduoZH 已提交
1696 1697
            padding_D = padding_H = padding_W = padding. Default: padding = 0.
        dilation (int|tuple): The dilation size. If dilation is a tuple, it must
1698
            contain three integers, (dilation_D, dilation_H, dilation_W). Otherwise, the
C
chengduoZH 已提交
1699 1700 1701 1702 1703 1704
            dilation_D = dilation_H = dilation_W = dilation. Default: dilation = 1.
        groups (int): The groups number of the Conv3d Layer. According to grouped
            convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
            the first half of the filters is only connected to the first half
            of the input channels, while the second half of the filters is only
            connected to the second half of the input channels. Default: groups=1
C
chengduo 已提交
1705 1706 1707 1708 1709 1710 1711 1712 1713 1714
        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
1715 1716
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
C
chengduo 已提交
1717 1718
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None.
C
chengduoZH 已提交
1719
        name (str|None): A name for this layer(optional). If set None, the layer
C
chengduo 已提交
1720
            will be named automatically. Default: None.
C
chengduoZH 已提交
1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732

    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

1733 1734
          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
1735 1736 1737
    """

    l_type = 'conv3d'
C
chengduo 已提交
1738
    assert param_attr is not False, "param_attr should not be False here."
C
chengduoZH 已提交
1739 1740 1741 1742 1743 1744 1745 1746 1747 1748
    helper = LayerHelper(l_type, **locals())
    dtype = helper.input_dtype()

    num_channels = input.shape[1]

    if groups is None:
        num_filter_channels = num_channels
    else:
        if num_channels % groups != 0:
            raise ValueError("num_channels must be divisible by groups.")
M
minqiyang 已提交
1749
        num_filter_channels = num_channels // groups
C
chengduoZH 已提交
1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762

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

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

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

    def _get_default_param_initializer():
C
chengduo 已提交
1763 1764 1765
        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
1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788
        return Normal(0.0, std, 0)

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

    pre_bias = helper.create_tmp_variable(dtype)

    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
1789
            'use_mkldnn': False
C
chengduoZH 已提交
1790 1791
        })

1792
    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
C
chengduoZH 已提交
1793 1794 1795 1796

    return helper.append_activation(pre_act)


F
fengjiayi 已提交
1797
def sequence_pool(input, pool_type):
Y
Yu Yang 已提交
1798
    """
Y
yangyaming 已提交
1799 1800 1801
    This function add the operator for sequence pooling.
    It pools features of all time-steps of each instance, and is applied
    on top of the input using pool_type mentioned in the parameters.
L
Luo Tao 已提交
1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812

    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:
1813
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
1814 1815 1816 1817 1818
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
1819
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
1820 1821 1822 1823 1824 1825 1826

       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)
1827 1828
         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
1829

L
Luo Tao 已提交
1830 1831
    Args:
        input(variable): The input variable which is a LoDTensor.
Y
yangyaming 已提交
1832
        pool_type (string): The pooling type of sequence_pool.
L
Luo Tao 已提交
1833 1834 1835 1836 1837 1838 1839 1840
            It supports average, sum, sqrt and max.

    Returns:
        The sequence pooling variable which is a Tensor.

    Examples:

        .. code-block:: python
F
fengjiayi 已提交
1841

Y
yangyaming 已提交
1842
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1843 1844 1845 1846 1847
                              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')
1848 1849
             last_x = fluid.layers.sequence_pool(input=x, pool_type='last')
             first_x = fluid.layers.sequence_pool(input=x, pool_type='first')
Y
Yu Yang 已提交
1850
    """
F
fengjiayi 已提交
1851
    helper = LayerHelper('sequence_pool', **locals())
Y
Yu Yang 已提交
1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862
    dtype = helper.input_dtype()
    pool_out = helper.create_tmp_variable(dtype)
    max_index = helper.create_tmp_variable(dtype)

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

Y
yangyaming 已提交
1863 1864 1865 1866 1867
    # when pool_type is max, variable max_index is initialized,
    # so we stop the gradient explicitly here
    if pool_type == 'max':
        max_index.stop_gradient = True

Y
Yu Yang 已提交
1868 1869 1870
    return pool_out


C
add doc  
chengduoZH 已提交
1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895
@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())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
        type='sequence_concat', inputs={'X': input}, outputs={'Out': [out]})
    return out


F
fengjiayi 已提交
1896
def sequence_first_step(input):
L
Luo Tao 已提交
1897
    """
L
Luo Tao 已提交
1898
    This function gets the first step of sequence.
L
Luo Tao 已提交
1899 1900 1901 1902

    .. code-block:: text

       x is a 1-level LoDTensor:
1903
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
1904 1905 1906 1907 1908
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
1909
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
1910
         out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1)
F
fengjiayi 已提交
1911

L
Luo Tao 已提交
1912 1913 1914 1915 1916 1917 1918 1919 1920
    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
1921

Y
yangyaming 已提交
1922
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1923 1924 1925
                              dtype='float32', lod_level=1)
             x_first_step = fluid.layers.sequence_first_step(input=x)
    """
1926 1927 1928
    return sequence_pool(input=input, pool_type="first")


F
fengjiayi 已提交
1929
def sequence_last_step(input):
L
Luo Tao 已提交
1930
    """
L
Luo Tao 已提交
1931
    This function gets the last step of sequence.
L
Luo Tao 已提交
1932 1933 1934 1935

    .. code-block:: text

       x is a 1-level LoDTensor:
1936
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
1937 1938 1939 1940 1941
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
1942
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
1943
         out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1)
F
fengjiayi 已提交
1944

L
Luo Tao 已提交
1945 1946 1947 1948 1949 1950 1951 1952 1953
    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
1954

Y
yangyaming 已提交
1955
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1956 1957 1958
                              dtype='float32', lod_level=1)
             x_last_step = fluid.layers.sequence_last_step(input=x)
    """
1959 1960 1961
    return sequence_pool(input=input, pool_type="last")


Y
Yibing Liu 已提交
1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974
def sequence_slice(input, offset, length, name=None):
    """
    **Sequence Slice Layer**

    The layer crops a subsequence from given sequence with given start 
    offset and subsequence length.

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

    .. code-block:: text
    
	- Case:

1975 1976 1977 1978 1979
            Given the input Variable **input**:
                
                input.data = [[a1, a2], [b1, b2], [c1, c2], [d1, d2], [e1, e2]],
                input.lod = [[3, 2]],
                input.dims = (5, 2),
Y
Yibing Liu 已提交
1980

1981
            with offset.data = [[0], [1]] and length.data = [[2], [1]],
Y
Yibing Liu 已提交
1982

1983 1984 1985 1986 1987
            the output Variable will be
                
                out.data = [[a1, a2], [b1, b2], [e1, e2]],
                out.lod = [[2, 1]],
                out.dims = (3, 2).
Y
Yibing Liu 已提交
1988
	
1989 1990
    NOTE: The first dimension size of **input**, **offset** and **length** 
          should be equal. The **offset** should start from 0.
Y
Yibing Liu 已提交
1991 1992 1993
    
    Args:
        input(Variable): The input Variable which consists of the complete 
Y
Yibing Liu 已提交
1994
                         sequences.
Y
Yibing Liu 已提交
1995 1996 1997 1998 1999 2000
        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
2001
        Variable: The output subsequences.
Y
Yibing Liu 已提交
2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031

    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"))
             subseqs = fluid.layers.sequence_slice(input=seqs, offset=offset, 
                                                   length=length)
    """
    helper = LayerHelper("sequence_slice", **locals())
    dtype = helper.input_dtype()
    out = helper.create_tmp_variable(dtype)

    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
2032
@templatedoc()
Y
Yu Yang 已提交
2033
def pool2d(input,
C
chengduoZH 已提交
2034 2035
           pool_size=-1,
           pool_type="max",
C
chengduoZH 已提交
2036 2037
           pool_stride=1,
           pool_padding=0,
C
caoying03 已提交
2038
           global_pooling=False,
C
chengduoZH 已提交
2039
           use_cudnn=True,
2040
           ceil_mode=False,
C
caoying03 已提交
2041
           name=None):
Y
Yu Yang 已提交
2042
    """
F
fengjiayi 已提交
2043
    ${comment}
2044 2045

    Args:
2046 2047 2048
        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
2049
                          feature, and W is the width of the feature.
2050
        pool_size (int): The side length of pooling windows. All pooling
F
fengjiayi 已提交
2051
                         windows are squares with pool_size on a side.
F
fengjiayi 已提交
2052
        pool_type: ${pooling_type_comment}
2053 2054
        pool_stride (int): stride of the pooling layer.
        pool_padding (int): padding size.
F
fengjiayi 已提交
2055 2056 2057
        global_pooling: ${global_pooling_comment}
        use_cudnn: ${use_cudnn_comment}
        ceil_mode: ${ceil_mode_comment}
2058
        name (str|None): A name for this layer(optional). If set None, the
F
fengjiayi 已提交
2059 2060
                        layer will be named automatically.

2061
    Returns:
F
fengjiayi 已提交
2062
        Variable: The pooling result.
F
fengjiayi 已提交
2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075

    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(
2076 2077 2078 2079
                            input=data,
                            pool_size=2,
                            pool_type='max',
                            pool_stride=1,
F
fengjiayi 已提交
2080
                            global_pooling=False)
Y
Yu Yang 已提交
2081 2082 2083 2084 2085
    """
    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
2086

C
chengduoZH 已提交
2087 2088 2089 2090 2091
    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
2092 2093 2094 2095
    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
2096 2097
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
2098

C
Add doc  
chengduoZH 已提交
2099
    l_type = 'pool2d'
2100 2101

    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
2102 2103 2104 2105
    dtype = helper.input_dtype()
    pool_out = helper.create_tmp_variable(dtype)

    helper.append_op(
2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116
        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,
X
Xin Pan 已提交
2117
            "use_mkldnn": False
2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133
        })

    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,
           name=None):
    """
    This function adds the operator for pooling in 3-dimensions, using the
Y
Yu Yang 已提交
2134
    pooling configurations mentioned in input parameters.
2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146

    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.
2147

2148
    Returns:
2149
        Variable: output of pool3d layer.
Y
Yu Yang 已提交
2150 2151 2152 2153 2154
    """
    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
2155

C
chengduoZH 已提交
2156 2157 2158 2159 2160
    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))

2161 2162 2163
    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
2164

C
chengduoZH 已提交
2165 2166
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
2167

2168 2169
    l_type = "pool3d"
    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
2170 2171 2172 2173
    dtype = helper.input_dtype()
    pool_out = helper.create_tmp_variable(dtype)

    helper.append_op(
2174
        type=l_type,
Y
Yu Yang 已提交
2175 2176 2177 2178 2179 2180 2181
        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
2182
            "paddings": pool_padding,
2183
            "use_cudnn": use_cudnn,
2184
            "ceil_mode": ceil_mode,
X
Xin Pan 已提交
2185
            "use_mkldnn": False
Y
Yu Yang 已提交
2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197
        })

    return pool_out


def batch_norm(input,
               act=None,
               is_test=False,
               momentum=0.9,
               epsilon=1e-05,
               param_attr=None,
               bias_attr=None,
C
caoying03 已提交
2198
               data_layout='NCHW',
Y
Yang Yang 已提交
2199
               in_place=False,
2200 2201
               name=None,
               moving_mean_name=None,
W
wanghaoshuang 已提交
2202
               moving_variance_name=None,
2203 2204
               do_model_average_for_mean_and_var=False,
               fuse_with_relu=False):
Y
Yu Yang 已提交
2205
    """
Q
qiaolongfei 已提交
2206 2207 2208 2209
    **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
2210

Q
qiaolongfei 已提交
2211
    1. NHWC `[batch, in_height, in_width, in_channels]`
Q
qiaolongfei 已提交
2212

Q
qiaolongfei 已提交
2213 2214
    2. NCHW `[batch, in_channels, in_height, in_width]`

Q
qiaolongfei 已提交
2215 2216 2217
    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
2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229

    :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
2230 2231

    Args:
Q
qiaolongfei 已提交
2232
        input(variable): The input variable which is a LoDTensor.
Q
qiaolongfei 已提交
2233 2234 2235 2236
        act(string, Default None): Activation type, linear|relu|prelu|...
        is_test(bool, Default False): Used for training or training.
        momentum(float, Default 0.9):
        epsilon(float, Default 1e-05):
C
chengduo 已提交
2237 2238 2239 2240 2241 2242 2243 2244
        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
2245
        data_layout(string, default NCHW): NCHW|NHWC
2246
        in_place(bool, Default False): Make the input and output of batch norm reuse memory.
Q
qiaolongfei 已提交
2247 2248 2249 2250
        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
2251
        do_model_average_for_mean_and_var(bool, Default False): Do model average for mean and variance or not.
2252
        fuse_with_relu (bool): if True, this OP performs relu after batch norm.
2253 2254

    Returns:
Q
qiaolongfei 已提交
2255
        Variable: A tensor variable which is the result after applying batch normalization on the input.
Q
qiaolongfei 已提交
2256 2257 2258 2259 2260 2261 2262

    Examples:

        .. code-block:: python

            hidden1 = fluid.layers.fc(input=x, size=200, param_attr='fc1.w')
            hidden2 = fluid.layers.batch_norm(input=hidden1)
Y
Yu Yang 已提交
2263
    """
C
chengduo 已提交
2264
    assert bias_attr is not False, "bias_attr should not be False in batch_norm."
Y
Yu Yang 已提交
2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286
    helper = LayerHelper('batch_norm', **locals())
    dtype = helper.input_dtype()

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

    param_shape = [channel_num]

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

    bias = helper.create_parameter(
2287
        attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
Y
Yu Yang 已提交
2288

2289 2290
    mean = helper.create_parameter(
        attr=ParamAttr(
W
wanghaoshuang 已提交
2291 2292 2293
            name=moving_mean_name,
            initializer=Constant(0.0),
            trainable=False,
W
wanghaoshuang 已提交
2294
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
2295
        shape=param_shape,
2296 2297 2298 2299 2300 2301 2302
        dtype=input.dtype)
    mean.stop_gradient = True

    variance = helper.create_parameter(
        attr=ParamAttr(
            name=moving_variance_name,
            initializer=Constant(1.0),
W
wanghaoshuang 已提交
2303
            trainable=False,
W
wanghaoshuang 已提交
2304
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
2305
        shape=param_shape,
2306 2307
        dtype=input.dtype)
    variance.stop_gradient = True
Y
Yu Yang 已提交
2308 2309 2310 2311 2312 2313

    # create output
    # mean and mean_out share the same memory
    mean_out = mean
    # variance and variance out share the same memory
    variance_out = variance
Q
QI JUN 已提交
2314 2315
    saved_mean = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
    saved_variance = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
Y
Yu Yang 已提交
2316

2317
    batch_norm_out = input if in_place else helper.create_tmp_variable(dtype)
Y
Yu Yang 已提交
2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334

    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
        },
2335 2336 2337 2338
        attrs={
            "momentum": momentum,
            "epsilon": epsilon,
            "is_test": is_test,
X
Xin Pan 已提交
2339
            "use_mkldnn": False,
2340
            "fuse_with_relu": fuse_with_relu
2341
        })
Y
Yu Yang 已提交
2342 2343 2344 2345

    return helper.append_activation(batch_norm_out)


Y
yuyang18 已提交
2346
@templatedoc()
G
guosheng 已提交
2347 2348 2349 2350 2351 2352 2353 2354 2355 2356
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
2357
    ${comment}
G
guosheng 已提交
2358 2359 2360

    The formula is as follows:

Y
yuyang18 已提交
2361
    ..  math::
G
guosheng 已提交
2362 2363 2364 2365 2366 2367 2368

        \\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
2369 2370 2371 2372 2373 2374 2375 2376
    * :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
2377

G
guosheng 已提交
2378 2379
    Args:
        input(Variable): The input tensor variable.
2380
        scale(bool): Whether to learn the adaptive gain :math:`g` after
S
sneaxiy 已提交
2381
            normalization. Default True.
2382
        shift(bool): Whether to learn the adaptive bias :math:`b` after
S
sneaxiy 已提交
2383 2384
            normalization. Default True.
        begin_norm_axis(int): The normalization will be performed along
G
guosheng 已提交
2385
            dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`.
S
sneaxiy 已提交
2386
            Default 1.
2387
        epsilon(float): The small value added to the variance to prevent
S
sneaxiy 已提交
2388
            division by zero. Default 1e-05.
G
guosheng 已提交
2389
        param_attr(ParamAttr|None): The parameter attribute for the learnable
S
sneaxiy 已提交
2390 2391 2392 2393
            gain :math:`g`. If :attr:`scale` is False, :attr:`param_attr` is
            omitted. If :attr:`scale` is True and :attr:`param_attr` is None,
            a default :code:`ParamAttr` would be added as scale. The 
            :attr:`param_attr` is initialized as 1 if it is added. Default None. 
G
guosheng 已提交
2394
        bias_attr(ParamAttr|None): The parameter attribute for the learnable
S
sneaxiy 已提交
2395 2396 2397 2398
            bias :math:`b`. If :attr:`shift` is False, :attr:`bias_attr` is
            omitted. If :attr:`shift` is True and :attr:`param_attr` is None,
            a default :code:`ParamAttr` would be added as bias. The 
            :attr:`bias_attr` is initialized as 0 if it is added. Default None.
G
guosheng 已提交
2399
        act(str): Activation to be applied to the output of layer normalizaiton.
S
sneaxiy 已提交
2400 2401 2402
                  Default None.
        name(str): The name of this layer. It is optional. Default None, and a
                   unique name would be generated automatically.
G
guosheng 已提交
2403 2404

    Returns:
Y
yuyang18 已提交
2405
        ${y_comment}
G
guosheng 已提交
2406 2407 2408

    Examples:

Y
yuyang18 已提交
2409 2410 2411
        >>> data = fluid.layers.data(name='data', shape=[3, 32, 32],
        >>>                          dtype='float32')
        >>> x = fluid.layers.layer_norm(input=data, begin_norm_axis=1)
G
guosheng 已提交
2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426
    """
    helper = LayerHelper('layer_norm', **locals())
    dtype = helper.input_dtype()

    # create intput and parameters
    inputs = {'X': input}
    input_shape = input.shape
    param_shape = [reduce(lambda x, y: x * y, input_shape[begin_norm_axis:])]
    if scale:
        scale = helper.create_parameter(
            attr=helper.param_attr,
            shape=param_shape,
            dtype=dtype,
            default_initializer=Constant(1.0))
        inputs['Scale'] = scale
G
guosheng 已提交
2427
    if shift:
G
guosheng 已提交
2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451
        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
    mean_out = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
    variance_out = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
    layer_norm_out = helper.create_tmp_variable(dtype)

    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)


Y
Yu Yang 已提交
2452 2453 2454 2455
def conv2d_transpose(input,
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
2456 2457 2458
                     padding=0,
                     stride=1,
                     dilation=1,
2459
                     groups=None,
C
caoying03 已提交
2460
                     param_attr=None,
2461
                     bias_attr=None,
C
chengduoZH 已提交
2462
                     use_cudnn=True,
2463
                     act=None,
C
caoying03 已提交
2464
                     name=None):
Y
Yu Yang 已提交
2465
    """
2466 2467 2468 2469 2470 2471 2472 2473
    **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
2474 2475
    layer, please refer to the following explanation and references
    `therein <http://www.matthewzeiler.com/wp-content/uploads/2017/07/cvpr2010.pdf>`_.
2476 2477 2478
    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.
2479 2480 2481 2482 2483

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

    .. math::

2484
        Out = \sigma (W \\ast X + b)
2485

2486
    Where:
2487 2488 2489

    * :math:`X`: Input value, a tensor with NCHW format.
    * :math:`W`: Filter value, a tensor with MCHW format.
2490 2491 2492 2493
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
Y
Yu Yang 已提交
2494

2495 2496 2497 2498
    Example:

        - Input:

2499
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
2500

2501
          Filter shape: :math:`(C_{in}, C_{out}, H_f, W_f)`
2502 2503 2504

        - Output:

2505
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
2506 2507

        Where
Y
Yu Yang 已提交
2508

2509 2510
        .. math::

2511 2512 2513 2514
           H^\prime_{out} &= (H_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (H_f - 1) + 1 \\\\
           W^\prime_{out} &= (W_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (W_f - 1) + 1 \\\\
           H_{out} \in [ H^\prime_{out}, H^\prime_{out} + strides[0] ) \\\\
           W_{out} \in [ W^\prime_{out}, W^\prime_{out} + strides[1] )
Y
Yu Yang 已提交
2515 2516

    Args:
2517 2518 2519 2520
        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
2521 2522 2523 2524
            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.
2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542
        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
            it must contain two integers, (filter_size_H, filter_size_W).
            Otherwise, the filter will be a square. None if use output size to
            calculate filter_size.
        padding(int|tuple): The padding size. If padding is a tuple, it must
            contain two integers, (padding_H, padding_W). Otherwise, the
            padding_H = padding_W = padding. Default: padding = 0.
        stride(int|tuple): The stride size. If stride is a tuple, it must
            contain two integers, (stride_H, stride_W). Otherwise, the
            stride_H = stride_W = stride. Default: stride = 1.
        dilation(int|tuple): The dilation size. If dilation is a tuple, it must
            contain two integers, (dilation_H, dilation_W). Otherwise, the
            dilation_H = dilation_W = dilation. Default: dilation = 1.
        groups(int): The groups number of the Conv2d transpose layer. Inspired by
            grouped convolution in Alex Krizhevsky's Deep CNN paper, in which
            when group=2, the first half of the filters is only connected to the
            first half of the input channels, while the second half of the
            filters is only connected to the second half of the input channels.
C
chengduo 已提交
2543 2544 2545 2546 2547 2548 2549 2550 2551 2552
            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.
2553
        use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
C
chengduo 已提交
2554 2555 2556
            library is installed. Default: True.
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None.
2557
        name(str|None): A name for this layer(optional). If set None, the layer
C
chengduo 已提交
2558
            will be named automatically. Default: True.
Y
Yu Yang 已提交
2559 2560

    Returns:
2561
        Variable: The tensor variable storing the convolution transpose result.
2562 2563

    Raises:
2564 2565
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
2566 2567 2568 2569

    Examples:
       .. code-block:: python

2570 2571
          data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32')
          conv2d_transpose = fluid.layers.conv2d_transpose(input=data, num_filters=2, filter_size=3)
Y
Yu Yang 已提交
2572
    """
C
chengduo 已提交
2573
    assert param_attr is not False, "param_attr should not be False in conv2d_transpose."
2574 2575 2576 2577 2578 2579 2580 2581
    input_channel = input.shape[1]

    op_type = 'conv2d_transpose'
    if (input_channel == groups and num_filters == input_channel and
            not use_cudnn):
        op_type = 'depthwise_conv2d_transpose'

    helper = LayerHelper(op_type, **locals())
Y
Yu Yang 已提交
2582 2583 2584
    if not isinstance(input, Variable):
        raise TypeError("Input of conv2d_transpose must be Variable")

C
chengduoZH 已提交
2585 2586 2587
    padding = utils.convert_to_list(padding, 2, 'padding')
    stride = utils.convert_to_list(stride, 2, 'stride')
    dilation = utils.convert_to_list(dilation, 2, 'dilation')
G
guosheng 已提交
2588

C
chengduoZH 已提交
2589 2590
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
G
guosheng 已提交
2591

Y
Yu Yang 已提交
2592 2593 2594 2595 2596
    if filter_size is None:
        if output_size is None:
            raise ValueError("output_size must be set when filter_size is None")
        if isinstance(output_size, int):
            output_size = [output_size, output_size]
G
guosheng 已提交
2597

Y
Yu Yang 已提交
2598 2599
        h_in = input.shape[2]
        w_in = input.shape[3]
G
guosheng 已提交
2600

C
chengduoZH 已提交
2601
        filter_size_h = (output_size[0] - (h_in - 1) * stride[0] + 2 *
M
minqiyang 已提交
2602
                         padding[0] - 1) // dilation[0] + 1
C
chengduoZH 已提交
2603
        filter_size_w = (output_size[1] - (w_in - 1) * stride[1] + 2 *
M
minqiyang 已提交
2604
                         padding[1] - 1) // dilation[1] + 1
Y
Yu Yang 已提交
2605
        filter_size = [filter_size_h, filter_size_w]
C
chengduoZH 已提交
2606 2607 2608
    else:
        filter_size = utils.convert_to_list(filter_size, 2,
                                            'conv2d_transpose.filter_size')
C
chengduo 已提交
2609

2610 2611 2612 2613 2614 2615 2616
    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')
2617
    groups = 1 if groups is None else groups
M
minqiyang 已提交
2618
    filter_shape = [input_channel, num_filters // groups] + filter_size
C
chengduo 已提交
2619

Y
Yu Yang 已提交
2620 2621 2622
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

2623
    pre_bias = helper.create_tmp_variable(dtype=input.dtype)
Y
Yu Yang 已提交
2624
    helper.append_op(
2625
        type=op_type,
Y
Yu Yang 已提交
2626 2627
        inputs={'Input': [input],
                'Filter': [img_filter]},
2628
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
2629
        attrs={
2630
            'output_size': output_size,
2631 2632 2633 2634 2635
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn
Y
Yu Yang 已提交
2636 2637
        })

2638 2639 2640
    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
    out = helper.append_activation(pre_act)
    return out
Y
Yu Yang 已提交
2641 2642


2643
def conv3d_transpose(input,
Y
Yu Yang 已提交
2644 2645 2646
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
2647 2648 2649
                     padding=0,
                     stride=1,
                     dilation=1,
2650
                     groups=None,
C
caoying03 已提交
2651
                     param_attr=None,
2652
                     bias_attr=None,
C
chengduoZH 已提交
2653
                     use_cudnn=True,
2654
                     act=None,
C
caoying03 已提交
2655
                     name=None):
Y
Yu Yang 已提交
2656
    """
2657
    **Convlution3D transpose layer**
2658

2659
    The convolution3D transpose layer calculates the output based on the input,
2660
    filter, and dilations, strides, paddings. Input(Input) and output(Output)
2661 2662 2663 2664 2665 2666
    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>`_.
2667 2668 2669
    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.
2670 2671 2672 2673 2674

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

    .. math::

2675
        Out = \sigma (W \\ast X + b)
2676 2677 2678

    In the above equation:

2679 2680
    * :math:`X`: Input value, a tensor with NCDHW format.
    * :math:`W`: Filter value, a tensor with MCDHW format.
2681 2682 2683 2684
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
Y
Yu Yang 已提交
2685

2686 2687 2688 2689
    Example:

        - Input:

2690
          Input shape: :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})`
2691

2692
          Filter shape: :math:`(C_{in}, C_{out}, D_f, H_f, W_f)`
2693 2694 2695

        - Output:

2696
          Output shape: :math:`(N, C_{out}, D_{out}, H_{out}, W_{out})`
2697 2698

        Where
Y
Yu Yang 已提交
2699

2700 2701
        .. math::

2702 2703 2704
           D_{out} &= (D_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (D_f - 1) + 1 \\\\
           H_{out} &= (H_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (H_f - 1) + 1 \\\\
           W_{out} &= (W_{in} - 1) * strides[2] - 2 * paddings[2] + dilations[2] * (W_f - 1) + 1
Y
Yu Yang 已提交
2705 2706

    Args:
2707
        input(Variable): The input image with [N, C, D, H, W] format.
2708 2709 2710
        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
2711
            tuple, it must contain three integers, (image_D, image_H, image_W). This
2712 2713
            parameter only works when filter_size is None.
        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
2714
            it must contain three integers, (filter_size_D, filter_size_H, filter_size_W).
2715 2716 2717
            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
2718 2719
            contain three integers, (padding_D, padding_H, padding_W). Otherwise, the
            padding_D = padding_H = padding_W = padding. Default: padding = 0.
2720
        stride(int|tuple): The stride size. If stride is a tuple, it must
2721 2722
            contain three integers, (stride_D, stride_H, stride_W). Otherwise, the
            stride_D = stride_H = stride_W = stride. Default: stride = 1.
2723
        dilation(int|tuple): The dilation size. If dilation is a tuple, it must
2724 2725 2726
            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
2727 2728 2729 2730 2731
            grouped convolution in Alex Krizhevsky's Deep CNN paper, in which
            when group=2, the first half of the filters is only connected to the
            first half of the input channels, while the second half of the
            filters is only connected to the second half of the input channels.
            Default: groups=1
C
chengduo 已提交
2732 2733 2734 2735 2736 2737 2738 2739 2740
        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.
2741 2742
        use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
C
chengduo 已提交
2743 2744
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None.
2745 2746
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.
Y
Yu Yang 已提交
2747 2748

    Returns:
2749
        Variable: The tensor variable storing the convolution transpose result.
2750 2751

    Raises:
2752 2753
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
2754 2755 2756 2757

    Examples:
       .. code-block:: python

2758 2759
          data = fluid.layers.data(name='data', shape=[3, 12, 32, 32], dtype='float32')
          conv3d_transpose = fluid.layers.conv3d_transpose(input=data, num_filters=2, filter_size=3)
Y
Yu Yang 已提交
2760
    """
C
chengduo 已提交
2761
    assert param_attr is not False, "param_attr should not be False in conv3d_transpose."
2762 2763
    l_type = "conv3d_transpose"
    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
2764
    if not isinstance(input, Variable):
2765
        raise TypeError("Input of conv3d_transpose must be Variable")
Y
Yu Yang 已提交
2766 2767
    input_channel = input.shape[1]

2768 2769 2770
    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
2771

C
chengduoZH 已提交
2772 2773 2774
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")

Y
Yu Yang 已提交
2775 2776 2777 2778 2779 2780
    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]

2781 2782 2783
        d_in = input.shape[2]
        h_in = input.shape[3]
        w_in = input.shape[4]
C
chengduoZH 已提交
2784

2785
        filter_size_d = (output_size[0] - (d_in - 1) * stride[0] + 2 *
M
minqiyang 已提交
2786
                         padding[0] - 1) // dilation[0] + 1
2787
        filter_size_h = (output_size[1] - (h_in - 1) * stride[1] + 2 *
M
minqiyang 已提交
2788
                         padding[1] - 1) // dilation[1] + 1
2789
        filter_size_w = (output_size[2] - (w_in - 1) * stride[2] + 2 *
M
minqiyang 已提交
2790
                         padding[2] - 1) // dilation[2] + 1
2791
        filter_size = [filter_size_d, filter_size_h, filter_size_w]
C
chengduoZH 已提交
2792
    else:
2793 2794
        filter_size = utils.convert_to_list(filter_size, 3,
                                            'conv3d_transpose.filter_size')
Y
Yu Yang 已提交
2795

2796
    groups = 1 if groups is None else groups
M
minqiyang 已提交
2797
    filter_shape = [input_channel, num_filters // groups] + filter_size
Y
Yu Yang 已提交
2798 2799 2800
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

2801
    pre_bias = helper.create_tmp_variable(dtype=input.dtype)
Y
Yu Yang 已提交
2802
    helper.append_op(
2803
        type=l_type,
Y
Yu Yang 已提交
2804 2805
        inputs={'Input': [input],
                'Filter': [img_filter]},
2806
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
2807 2808 2809 2810
        attrs={
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
2811
            'groups': groups,
C
chengduoZH 已提交
2812 2813
            'use_cudnn': use_cudnn
        })
Y
Yu Yang 已提交
2814

2815 2816
    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
    out = helper.append_activation(pre_act)
Y
Yu Yang 已提交
2817
    return out
Y
yangyaming 已提交
2818 2819


Y
yangyaming 已提交
2820
def sequence_expand(x, y, ref_level=-1, name=None):
2821
    """Sequence Expand Layer. This layer will expand the input variable **x**
Y
yangyaming 已提交
2822 2823 2824 2825
    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:
2826 2827 2828 2829 2830

    .. code-block:: text

        * Case 1
            x is a LoDTensor:
2831
                x.lod  = [[2,        2]]
Y
yangyaming 已提交
2832
                x.data = [[a], [b], [c], [d]]
2833 2834 2835
                x.dims = [4, 1]

            y is a LoDTensor:
2836 2837
                y.lod = [[2,    2],
                         [3, 3, 1, 1]]
2838

Y
yangyaming 已提交
2839
            ref_level: 0
2840

Y
yangyaming 已提交
2841
            then output is a 1-level LoDTensor:
2842
                out.lod =  [[2,        2,        2,        2]]
Y
yangyaming 已提交
2843
                out.data = [[a], [b], [a], [b], [c], [d], [c], [d]]
2844 2845 2846 2847
                out.dims = [8, 1]

        * Case 2
            x is a Tensor:
Y
yangyaming 已提交
2848
                x.data = [[a], [b], [c]]
2849 2850 2851
                x.dims = [3, 1]

            y is a LoDTensor:
2852
                y.lod = [[2, 0, 3]]
2853

Y
yangyaming 已提交
2854
            ref_level: -1
2855

Y
yangyaming 已提交
2856 2857 2858
            then output is a Tensor:
                out.data = [[a], [a], [c], [c], [c]]
                out.dims = [5, 1]
2859 2860 2861
    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
2862 2863
        ref_level (int): Lod level of `y` to be referred by `x`. If set to -1,
                         refer the last level of lod.
C
caoying03 已提交
2864
        name(str|None): A name for this layer(optional). If set None, the layer
Y
yangyaming 已提交
2865
                        will be named automatically.
2866 2867 2868 2869 2870 2871 2872 2873 2874 2875

    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
2876
            out = layers.sequence_expand(x=x, y=y, ref_level=0)
2877
    """
Y
yangyaming 已提交
2878
    helper = LayerHelper('sequence_expand', input=x, **locals())
2879 2880 2881
    dtype = helper.input_dtype()
    tmp = helper.create_tmp_variable(dtype)
    helper.append_op(
Y
yangyaming 已提交
2882 2883 2884 2885 2886
        type='sequence_expand',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': tmp},
        attrs={'ref_level': ref_level})
2887
    return tmp
2888 2889


C
chengduo 已提交
2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954
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()
    tmp = helper.create_tmp_variable(dtype)
    helper.append_op(
        type='sequence_expand_as',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': tmp})
    return tmp


F
fengjiayi 已提交
2955
@templatedoc()
2956
def sequence_pad(x, pad_value, maxlen=None, name=None):
F
fengjiayi 已提交
2957 2958 2959 2960 2961
    """
    ${comment}

    Args:
        x(Variable): Input variable which should contain lod information.
M
minqiyang 已提交
2962 2963 2964
        pad_value(Variable): The Variable that holds values that will be fill
            into padded steps. It can be a scalar or a tensor whose shape
            equals to time steps in sequences. If it's a scalar, it will be
F
fengjiayi 已提交
2965
            automatically broadcasted to the shape of time step.
M
minqiyang 已提交
2966 2967 2968 2969
        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
2970 2971 2972
            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
2973

F
fengjiayi 已提交
2974
    Returns:
M
minqiyang 已提交
2975
        Variable: The padded sequence batch and the original lengths before
2976
                  padding. All sequences has the same length.
M
minqiyang 已提交
2977

F
fengjiayi 已提交
2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991
    Examples:
        .. code-block:: python

            import numpy

            x = fluid.layers.data(name='y', shape=[10, 5],
                             dtype='float32', lod_level=1)
            pad_value = fluid.layers.assign(input=numpy.array([0]))
            out = fluid.layers.sequence_pad(x=x, pad_value=pad_value)
    """

    helper = LayerHelper('sequence_pad', input=x, **locals())
    dtype = helper.input_dtype()
    out = helper.create_tmp_variable(dtype)
2992 2993 2994 2995 2996
    length = helper.create_tmp_variable(dtype)

    pad_value.stop_gradient = True
    length.stop_gradient = True

F
fengjiayi 已提交
2997 2998 2999 3000 3001 3002
    if maxlen is None:
        maxlen = -1
    helper.append_op(
        type='sequence_pad',
        inputs={'X': x,
                'PadValue': pad_value},
3003 3004
        outputs={'Out': out,
                 'Length': length},
F
fengjiayi 已提交
3005
        attrs={'padded_length': maxlen})
3006
    return out, length
F
fengjiayi 已提交
3007 3008


3009
def sequence_unpad(x, length, name=None):
Y
Yibing Liu 已提交
3010
    """
3011
    **Sequence Unpad Layer**
Y
Yibing Liu 已提交
3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026

    This layer removes the padding data in the input sequences and convert 
    them into sequences with actual length as output, identitied by lod 
    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],
		      [11.0, 12.0, 13.0, 14.0, 15.0]], 
     
	in which there are 3 sequences padded to length 5, and the acutal length 
3027
	specified by input Variable **length**:
Y
Yibing Liu 已提交
3028 3029 3030 3031 3032 3033

	    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]]
3034
	    out.lod = [[2, 3, 4]]      
Y
Yibing Liu 已提交
3035 3036 3037 3038 3039 3040

    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.
3041 3042
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.
Y
Yibing Liu 已提交
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

    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()
    out = helper.create_tmp_variable(dtype)

    length.stop_gradient = True

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


3069 3070 3071 3072 3073 3074 3075 3076 3077
def beam_search(pre_ids,
                pre_scores,
                ids,
                scores,
                beam_size,
                end_id,
                level=0,
                name=None):
    """
3078 3079
    Beam search is a classical algorithm for selecting candidate words in a
    machine translation task.
Y
Yan Chunwei 已提交
3080 3081 3082

    Refer to `Beam search <https://en.wikipedia.org/wiki/Beam_search>`_
    for more details.
M
minqiyang 已提交
3083 3084

    This layer does the search in beams for one time step. Specifically, it
3085 3086 3087 3088 3089 3090
    selects the top-K candidate word ids of current step from :attr:`ids`
    according to their :attr:`scores` for all source sentences, where K is
    :attr:`beam_size` and :attr:`ids, scores` are predicted results from the
    computation cell. Additionally, :attr:`pre_ids` and :attr:`pre_scores` are
    the output of beam_search at previous step, they are needed for special use
    to handle ended candidate translations.
M
minqiyang 已提交
3091

3092 3093 3094 3095 3096 3097 3098 3099
    Note that the :attr:`scores` passed in should be accumulated scores, and
    length penalty should be done with extra operators before calculating the
    accumulated scores if needed, also suggest finding top-K before it and
    using the top-K candidates following.

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

        fluid/tests/book/test_machine_translation.py
Y
Yan Chunwei 已提交
3100

3101
    Args:
3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126
        pre_ids(Variable): The LodTensor variable which is the output of
            beam_search at previous step. It should be a LodTensor with shape
            :math:`(batch_size, 1)` and lod
            :math:`[[0, 1, ... , batch_size], [0, 1, ..., batch_size]]` at the
            first step.
        pre_scores(Variable): The LodTensor variable which is the output of
            beam_search at previous step.
        ids(Variable): The LodTensor variable containing the candidates ids.
            Its shape should be :math:`(batch_size \\times beam_size, K)`,
            where :math:`K` supposed to be :attr:`beam_size`.
        scores(Variable): The LodTensor variable containing the accumulated
            scores corresponding to :attr:`ids` and its shape is the same as
            the shape of :attr:`ids`.
        beam_size(int): The beam width used in beam search.
        end_id(int): The id of end token.
        level(int, default 0): It can be ignored and mustn't change currently.
            It means the source level of lod, which is explained as following.
            The lod level of :attr:`ids` should be 2. The first level is source
            level which describes how many prefixes (branchs) for each source
            sentece (beam), and the second level is sentence level which
            describes how these candidates belong to the prefix. The paths
            linking prefixes and selected candidates are organized and reserved
            in lod.
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
F
fengjiayi 已提交
3127

3128
    Returns:
3129 3130
        Variable: The LodTensor pair containing the selected ids and the \
            corresponding scores.
Y
Yan Chunwei 已提交
3131 3132 3133 3134

    Examples:
        .. code-block:: python

3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151
            # 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
3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162
    helper = LayerHelper('beam_search', **locals())
    score_type = scores.dtype
    id_type = ids.dtype

    selected_scores = helper.create_tmp_variable(dtype=score_type)
    selected_ids = helper.create_tmp_variable(dtype=id_type)

    helper.append_op(
        type='beam_search',
        inputs={
            'pre_ids': pre_ids,
3163
            'pre_scores': pre_scores,
Q
Qiao Longfei 已提交
3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180
            'ids': ids,
            'scores': scores,
        },
        outputs={
            'selected_ids': selected_ids,
            'selected_scores': selected_scores,
        },
        attrs={
            # TODO(ChunweiYan) to assure other value support
            'level': level,
            'beam_size': beam_size,
            'end_id': end_id,
        })

    return selected_ids, selected_scores


3181 3182 3183 3184 3185 3186 3187
def beam_search_decode(ids, scores, beam_size, end_id, name=None):
    """
    Beam Search Decode Layer. This layer constructs the full hypotheses for
    each source sentence by walking back along the LoDTensorArray :attr:`ids`
    whose lods can be used to restore the path in the beam search tree.
    Please see the following demo for a fully beam search usage example:
        fluid/tests/book/test_machine_translation.py
G
guosheng 已提交
3188

3189 3190 3191 3192 3193 3194 3195 3196 3197
    Args:
        ids(Variable): The LodTensorArray variable containing the selected ids
            of all steps.
        scores(Variable): The LodTensorArray variable containing the selected
            scores of all steps.
        beam_size(int): The beam width used in beam search.
        end_id(int): The id of end token.
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
G
guosheng 已提交
3198

3199 3200 3201 3202 3203 3204
    Returns:
        Variable: The LodTensor pair containing the generated id sequences \
            and the corresponding scores. The shapes and lods of the two \
            LodTensor are same. The lod level is 2 and the two levels \
            separately indicate how many hypotheses each source sentence has \
            and how many ids each hypothesis has.
G
guosheng 已提交
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
    Examples:
        .. code-block:: python
            # Suppose `ids` and `scores` are LodTensorArray variables reserving
            # the selected ids and scores of all steps
            finished_ids, finished_scores = layers.beam_search_decode(
                ids, scores, beam_size=5, end_id=0)
    """
    helper = LayerHelper('beam_search_decode', **locals())
    sentence_ids = helper.create_tmp_variable(dtype=ids.dtype)
    sentence_scores = helper.create_tmp_variable(dtype=ids.dtype)

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

    return sentence_ids, sentence_scores


Y
yangyaming 已提交
3231 3232 3233 3234
def lstm_unit(x_t,
              hidden_t_prev,
              cell_t_prev,
              forget_bias=0.0,
Y
yangyaming 已提交
3235
              param_attr=None,
C
caoying03 已提交
3236 3237
              bias_attr=None,
              name=None):
Y
yangyaming 已提交
3238 3239 3240 3241
    """Lstm unit layer. The equation of a lstm step is:

        .. math::

3242
            i_t & = \sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i)
Y
yangyaming 已提交
3243

3244
            f_t & = \sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + b_f)
Y
yangyaming 已提交
3245

3246
            c_t & = f_tc_{t-1} + i_t tanh (W_{x_c}x_t + W_{h_c}h_{t-1} + b_c)
Y
yangyaming 已提交
3247

3248
            o_t & = \sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + b_o)
Y
yangyaming 已提交
3249 3250 3251

            h_t & = o_t tanh(c_t)

3252 3253 3254 3255 3256 3257
    The inputs of lstm unit include :math:`x_t`, :math:`h_{t-1}` and
    :math:`c_{t-1}`. The 2nd dimensions of :math:`h_{t-1}` and :math:`c_{t-1}`
    should be same. The implementation separates the linear transformation and
    non-linear transformation apart. Here, we take :math:`i_t` as an example.
    The linear transformation is applied by calling a `fc` layer and the
    equation is:
Y
yangyaming 已提交
3258 3259 3260

        .. math::

3261
            L_{i_t} = W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i
Y
yangyaming 已提交
3262 3263 3264 3265 3266 3267 3268 3269

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

        .. math::

            i_t = \sigma(L_{i_t})

Y
yangyaming 已提交
3270
    This layer has two outputs including :math:`h_t` and :math:`o_t`.
Y
yangyaming 已提交
3271 3272

    Args:
Y
yangyaming 已提交
3273 3274 3275 3276 3277 3278
        x_t (Variable): The input value of current step, a 2-D tensor with shape
            M x N, M for batch size and N for input size.
        hidden_t_prev (Variable): The hidden value of lstm unit, a 2-D tensor
            with shape M x S, M for batch size and S for size of lstm unit.
        cell_t_prev (Variable): The cell value of lstm unit, a 2-D tensor with
            shape M x S, M for batch size and S for size of lstm unit.
Y
yangyaming 已提交
3279
        forget_bias (float): The forget bias of lstm unit.
C
chengduo 已提交
3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291
        param_attr(ParamAttr|None): The parameter attribute for the learnable
                               hidden-hidden weights.
                               If it is set to None or one attribute of ParamAttr,
                               lstm_unit will create ParamAttr as param_attr.
                               If the Initializer of the param_attr is not set, the
                               parameter is initialized with Xavier. Default: None.
        bias_attr (ParamAttr|None): The bias attribute for the learnable bias
                              weights. If it is set to False, no bias will be added
                              to the output units. If it is set to None or one attribute of ParamAttr,
                              lstm_unit will create ParamAttr as bias_attr.
                              If the Initializer of the bias_attr is not set,
                              the bias is initialized zero. Default: None.
C
caoying03 已提交
3292 3293
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
Y
yangyaming 已提交
3294 3295

    Returns:
Y
yangyaming 已提交
3296
        tuple: The hidden value and cell value of lstm unit.
Y
yangyaming 已提交
3297 3298

    Raises:
3299 3300 3301 3302
        ValueError: The ranks of **x_t**, **hidden_t_prev** and **cell_t_prev**
                    not be 2 or the 1st dimensions of **x_t**, **hidden_t_prev**
                    and **cell_t_prev** not be the same or the 2nd dimensions of
                    **hidden_t_prev** and **cell_t_prev** not be the same.
Y
yangyaming 已提交
3303 3304 3305 3306 3307 3308

    Examples:

        .. code-block:: python

             x_t = fluid.layers.fc(input=x_t_data, size=10)
3309
             prev_hidden = fluid.layers.fc(input=prev_hidden_data, size=30)
Y
yangyaming 已提交
3310
             prev_cell = fluid.layers.fc(input=prev_cell_data, size=30)
Y
yangyaming 已提交
3311
             hidden_value, cell_value = fluid.layers.lstm_unit(x_t=x_t,
Y
yangyaming 已提交
3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327
                                                    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
3328
        raise ValueError("The 1st dimensions of x_t, hidden_t_prev and "
3329 3330 3331 3332
                         "cell_t_prev must be the same.")

    if hidden_t_prev.shape[1] != cell_t_prev.shape[1]:
        raise ValueError("The 2nd dimensions of hidden_t_prev and "
Y
yangyaming 已提交
3333 3334
                         "cell_t_prev must be the same.")

Y
yangyaming 已提交
3335 3336 3337
    if bias_attr is None:
        bias_attr = ParamAttr()

Y
yangyaming 已提交
3338
    size = cell_t_prev.shape[1]
3339
    concat_out = concat(input=[x_t, hidden_t_prev], axis=1)
Y
yangyaming 已提交
3340 3341
    fc_out = fc(input=concat_out,
                size=4 * size,
Y
yangyaming 已提交
3342
                param_attr=param_attr,
3343
                bias_attr=bias_attr)
Y
yangyaming 已提交
3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355
    dtype = x_t.dtype
    c = helper.create_tmp_variable(dtype)
    h = helper.create_tmp_variable(dtype)

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

Y
yangyaming 已提交
3356
    return h, c
G
guosheng 已提交
3357 3358


C
caoying03 已提交
3359
def reduce_sum(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
3360
    """
Y
yangyaming 已提交
3361
    Computes the sum of tensor elements over the given dimension.
G
guosheng 已提交
3362 3363 3364

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
3365
        dim (list|int|None): The dimensions along which the sum is performed. If
Y
yangyaming 已提交
3366 3367
            :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
3368 3369
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
3370
        keep_dim (bool|False): Whether to reserve the reduced dimension in the
Y
yangyaming 已提交
3371
            output Tensor. The result tensor will have one fewer dimension
G
guosheng 已提交
3372
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
3373 3374
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
3375 3376 3377

    Returns:
        Variable: The reduced Tensor variable.
F
fengjiayi 已提交
3378

G
guosheng 已提交
3379 3380 3381 3382 3383 3384
    Examples:
        .. code-block:: python

            # x is a Tensor variable with following elements:
            #    [[0.2, 0.3, 0.5, 0.9]
            #     [0.1, 0.2, 0.6, 0.7]]
Q
qiaolongfei 已提交
3385
            # Each example is followed by the corresponding output tensor.
G
guosheng 已提交
3386 3387 3388 3389
            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
3390 3391 3392 3393

            # x is a Tensor variable with shape [2, 2, 2] and elements as below:
            #      [[[1, 2], [3, 4]],
            #      [[5, 6], [7, 8]]]
Q
qiaolongfei 已提交
3394
            # Each example is followed by the corresponding output tensor.
W
whs 已提交
3395 3396 3397
            fluid.layers.reduce_sum(x, dim=[1, 2]) # [10, 26]
            fluid.layers.reduce_sum(x, dim=[0, 1]) # [16, 20]

G
guosheng 已提交
3398 3399 3400
    """
    helper = LayerHelper('reduce_sum', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3401 3402
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
G
guosheng 已提交
3403 3404 3405 3406 3407
    helper.append_op(
        type='reduce_sum',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3408
            'dim': dim if dim != None else [0],
G
guosheng 已提交
3409 3410 3411 3412
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
3413 3414


C
caoying03 已提交
3415
def reduce_mean(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
3416
    """
Y
Yibing Liu 已提交
3417
    Computes the mean of the input tensor's elements along the given dimension.
G
guosheng 已提交
3418 3419 3420

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
Yibing Liu 已提交
3421 3422 3423
        dim (list|int|None): The dimension along which the mean is computed. If
            `None`, compute the mean over all elements of :attr:`input`
            and return a variable with a single element, otherwise it
Y
yangyaming 已提交
3424
            must be in the range :math:`[-rank(input), rank(input))`. If
3425
            :math:`dim[i] < 0`, the dimension to reduce is
Y
Yibing Liu 已提交
3426
            :math:`rank(input) + dim[i]`.
Y
yangyaming 已提交
3427 3428
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
G
guosheng 已提交
3429
            than the :attr:`input` unless :attr:`keep_dim` is true.
Y
Yibing Liu 已提交
3430
        name(str|None): A name for this layer(optional). If set `None`, the layer
C
caoying03 已提交
3431
                       will be named automatically.
G
guosheng 已提交
3432 3433

    Returns:
Y
Yibing Liu 已提交
3434
        Variable: The reduced mean Variable.
F
fengjiayi 已提交
3435

G
guosheng 已提交
3436 3437 3438 3439 3440 3441 3442 3443 3444 3445
    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
3446 3447
            fluid.layers.reduce_mean(
                x, dim=1, keep_dim=True)  # [[0.475], [0.4]]
W
whs 已提交
3448 3449 3450 3451 3452 3453 3454

            # x is a Tensor variable with shape [2, 2, 2] and elements as below:
            #      [[[1.0, 2.0], [3.0, 4.0]],
            #      [[5.0, 6.0], [7.0, 8.0]]]
            # Each example is followed by the correspending output tensor.
            fluid.layers.reduce_mean(x, dim=[1, 2]) # [2.5, 6.5]
            fluid.layers.reduce_mean(x, dim=[0, 1]) # [4.0, 5.0]
G
guosheng 已提交
3455 3456 3457
    """
    helper = LayerHelper('reduce_mean', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3458 3459
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
G
guosheng 已提交
3460 3461 3462 3463 3464
    helper.append_op(
        type='reduce_mean',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3465
            'dim': dim if dim != None else [0],
G
guosheng 已提交
3466 3467 3468 3469
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
3470 3471


C
caoying03 已提交
3472
def reduce_max(input, dim=None, keep_dim=False, name=None):
3473
    """
Y
yangyaming 已提交
3474
    Computes the maximum of tensor elements over the given dimension.
3475 3476 3477

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
3478
        dim (list|int|None): The dimension along which the maximum is computed.
Y
yangyaming 已提交
3479 3480 3481
            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
3482
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
Y
yangyaming 已提交
3483 3484
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
3485
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
3486 3487
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
3488 3489 3490

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
3491

3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502
    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
3503 3504 3505 3506 3507 3508 3509

            # 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]
3510 3511 3512
    """
    helper = LayerHelper('reduce_max', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3513 3514
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3515 3516 3517 3518 3519
    helper.append_op(
        type='reduce_max',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3520
            'dim': dim if dim != None else [0],
3521 3522 3523 3524 3525 3526
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


C
caoying03 已提交
3527
def reduce_min(input, dim=None, keep_dim=False, name=None):
3528
    """
Y
yangyaming 已提交
3529
    Computes the minimum of tensor elements over the given dimension.
3530 3531 3532

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
3533
        dim (list|int|None): The dimensions along which the minimum is computed.
Y
yangyaming 已提交
3534 3535 3536
            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
3537
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
Y
yangyaming 已提交
3538 3539
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
3540
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
3541 3542
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
3543 3544 3545

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
3546

3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557
    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
3558 3559 3560 3561 3562 3563 3564

            # 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]
3565 3566 3567
    """
    helper = LayerHelper('reduce_min', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3568 3569
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3570 3571 3572 3573 3574
    helper.append_op(
        type='reduce_min',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3575
            'dim': dim if dim != None else [0],
3576 3577 3578 3579
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
3580 3581


3582 3583 3584 3585 3586 3587
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
3588
        dim (list|int|None): The dimensions along which the product is performed. If
3589 3590
            :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
3591 3592
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
3593 3594 3595
        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
3596
        name(str|None): A name for this layer(optional). If set None, the
Z
zhouhanqing 已提交
3597
            layer will be named automatically.
3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611

    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
3612
            fluid.layers.reduce_prod(x, dim=1,
Z
zhouhanqing 已提交
3613
                                     keep_dim=True)  # [[0.027], [0.0084]]
W
whs 已提交
3614 3615 3616 3617 3618 3619 3620

            # 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]
3621 3622 3623
    """
    helper = LayerHelper('reduce_prod', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3624 3625
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3626 3627 3628 3629 3630
    helper.append_op(
        type='reduce_prod',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3631
            'dim': dim if dim != None else [0],
3632 3633 3634 3635 3636 3637
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


C
caoying03 已提交
3638
def split(input, num_or_sections, dim=-1, name=None):
G
guosheng 已提交
3639
    """
C
caoying03 已提交
3640
    Split the input tensor into multiple sub-tensors.
G
guosheng 已提交
3641 3642 3643

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
C
caoying03 已提交
3644 3645 3646 3647 3648
        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
3649
            :attr:`dim` dimension orderly.
C
caoying03 已提交
3650
        dim (int): The dimension along which to split. If :math:`dim < 0`, the
G
guosheng 已提交
3651
            dimension to split along is :math:`rank(input) + dim`.
C
caoying03 已提交
3652 3653
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
3654 3655

    Returns:
D
dzhwinter 已提交
3656
        list(Variable): The list of segmented tensor variables.
G
guosheng 已提交
3657 3658 3659 3660 3661 3662 3663 3664 3665

    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
3666 3667
            x0, x1, x2 = fluid.layers.split(
                x, num_or_sections=[2, 3, 4], dim=1)
G
guosheng 已提交
3668 3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696
            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 = [
        helper.create_tmp_variable(dtype=helper.input_dtype())
        for i in range(num)
    ]
    helper.append_op(
        type='split',
        inputs={'X': input},
        outputs={'Out': outs},
        attrs={
            'num': num_or_sections if isinstance(num_or_sections, int) else 0,
            'sections': num_or_sections
            if isinstance(num_or_sections, list) else [],
            'axis': dim
        })
    return outs
C
caoying03 已提交
3697 3698 3699 3700 3701 3702 3703 3704 3705


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

3706
    .. math::
3707 3708

        y = \\frac{x}{ \sqrt{\sum {x^2} + epsion }}
C
caoying03 已提交
3709 3710 3711 3712 3713

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

    Args:
3714
        x(Variable|list): The input tensor to l2_normalize layer.
3715
        axis(int): The axis on which to apply normalization. If `axis < 0`, \
3716 3717
            the dimension to normalization is rank(X) + axis. -1 is the
            last dimension.
3718
        epsilon(float): The epsilon value is used to avoid division by zero, \
3719
            the defalut value is 1e-10.
3720
        name(str|None): A name for this layer(optional). If set None, the layer \
3721
            will be named automatically.
C
caoying03 已提交
3722 3723

    Returns:
3724
        Variable: The output tensor variable is the same shape with `x`.
C
caoying03 已提交
3725 3726

    Examples:
3727

C
caoying03 已提交
3728 3729
        .. code-block:: python

3730 3731 3732 3733
            data = fluid.layers.data(name="data",
                                     shape=(3, 17, 13),
                                     dtype="float32")
            normed = fluid.layers.l2_normalize(x=data, axis=1)
C
caoying03 已提交
3734 3735
    """

F
fengjiayi 已提交
3736 3737
    if len(x.shape) == 1:
        axis = 0
C
caoying03 已提交
3738 3739
    helper = LayerHelper("l2_normalize", **locals())

3740 3741
    out = helper.create_tmp_variable(dtype=x.dtype)
    norm = helper.create_tmp_variable(dtype=x.dtype)
C
caoying03 已提交
3742
    helper.append_op(
3743 3744 3745 3746
        type="norm",
        inputs={"X": x},
        outputs={"Out": out,
                 "Norm": norm},
C
caoying03 已提交
3747
        attrs={
3748 3749
            "axis": 1 if axis is None else axis,
            "epsilon": epsilon,
C
caoying03 已提交
3750 3751
        })
    return out
3752 3753


S
sneaxiy 已提交
3754
def matmul(x, y, transpose_x=False, transpose_y=False, alpha=1.0, name=None):
G
guosheng 已提交
3755
    """
Y
ying 已提交
3756 3757 3758 3759
    Applies matrix multiplication to two tensors.

    Currently, the input tensors' rank can be any, but when the rank of any
    inputs is bigger than 3, this two inputs' rank should be equal.
G
guosheng 已提交
3760

C
chengduoZH 已提交
3761
    The actual behavior depends on the shapes of :math:`x`, :math:`y` and the
3762
    flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically:
G
guosheng 已提交
3763

3764 3765 3766 3767 3768
    - 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
3769
      :math:`[1, D]` in transposed form.
G
guosheng 已提交
3770

C
chengduoZH 已提交
3771
    - After transpose, the two tensors are 2-D or n-D and matrix multiplication
3772
      performs in the following way.
G
guosheng 已提交
3773

3774
      - If both are 2-D, they are multiplied like conventional matrices.
C
chengduoZH 已提交
3775
      - If either is n-D, it is treated as a stack of matrices residing in the
Y
ying 已提交
3776
        last two dimensions and a batched matrix multiply supporting broadcast
3777
        applies on the two tensors.
G
guosheng 已提交
3778

Y
ying 已提交
3779 3780
    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
3781
    removed after matrix multiplication.
G
guosheng 已提交
3782 3783 3784

    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
3785 3786 3787
        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
3788
        alpha (float): The scale of output. Default 1.0.
3789
        name(str|None): A name for this layer(optional). If set None, the layer
3790
            will be named automatically.
G
guosheng 已提交
3791 3792

    Returns:
3793
        Variable: The product Tensor variable.
G
guosheng 已提交
3794

G
guosheng 已提交
3795 3796 3797
    Examples:
        .. code-block:: python

3798
            # Examples to clarify shapes of the inputs and output
C
chengduoZH 已提交
3799 3800
            # x: [B, ..., M, K], y: [B, ..., K, N]
            fluid.layers.matmul(x, y)  # out: [B, ..., M, N]
Y
ying 已提交
3801

3802 3803
            # x: [B, M, K], y: [B, K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
3804

3805 3806
            # x: [B, M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
3807

3808 3809
            # x: [M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [M, N]
Y
ying 已提交
3810 3811 3812 3813

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

3814 3815
            # x: [K], y: [K]
            fluid.layers.matmul(x, y)  # out: [1]
3816

Y
ying 已提交
3817
            # x: [M], y: [N]
3818
            fluid.layers.matmul(x, y, True, True)  # out: [M, N]
G
guosheng 已提交
3819
    """
Y
ying 已提交
3820 3821 3822 3823 3824 3825 3826 3827 3828 3829 3830 3831

    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
3832
            y_shape = y_shape + [1]
Y
ying 已提交
3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848

        # 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)

3849
    helper = LayerHelper('matmul', **locals())
Y
ying 已提交
3850
    out = helper.create_tmp_variable(dtype=x.dtype)
G
guosheng 已提交
3851
    helper.append_op(
3852 3853 3854 3855
        type='matmul',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
S
sneaxiy 已提交
3856 3857 3858
        attrs={
            'transpose_X': transpose_x,
            'transpose_Y': transpose_y,
S
sneaxiy 已提交
3859
            'alpha': float(alpha),
S
sneaxiy 已提交
3860
        })
3861
    return out
3862 3863


3864
def topk(input, k, name=None):
Q
qingqing01 已提交
3865 3866 3867 3868
    """
    This operator is used to find values and indices of the k largest entries
    for the last dimension.

F
fengjiayi 已提交
3869
    If the input is a vector (1-D Tensor), finds the k largest entries in the vector
Q
qingqing01 已提交
3870 3871 3872 3873 3874 3875
    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
3876 3877 3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892 3893 3894 3895 3896
    For example:

    .. code-block:: text

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

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

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

Q
qingqing01 已提交
3897 3898 3899
    Args:
        input(Variable): The input variable which can be a vector or Tensor with
            higher rank.
3900
        k(int):  The number of top elements to look for along the last dimension
F
fengjiayi 已提交
3901
                 of input.
3902
        name(str|None): A name for this layer(optional). If set None, the layer
3903
                       will be named automatically.
F
fengjiayi 已提交
3904
                       Default: None
Q
qingqing01 已提交
3905 3906

    Returns:
3907 3908 3909
        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
3910
        within the last dimension of input.
Q
qingqing01 已提交
3911

F
fengjiayi 已提交
3912 3913
    Raises:
        ValueError: If k < 1 or k is not less than the last dimension of input
Q
qingqing01 已提交
3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933

    Examples:
        .. code-block:: python

            top5_values, top5_indices = layers.topk(input, k=5)
    """
    helper = LayerHelper("top_k", **locals())
    values = helper.create_tmp_variable(dtype=input.dtype)
    indices = helper.create_tmp_variable(dtype="int64")
    helper.append_op(
        type="top_k",
        inputs={"X": [input]},
        outputs={"Out": [values],
                 "Indices": [indices]},
        attrs={"k": k})
    values.stop_gradient = True
    indices.stop_gradient = True
    return values, indices


3934
def edit_distance(input, label, normalized=True, ignored_tokens=None):
3935
    """
Y
ying 已提交
3936 3937 3938 3939 3940 3941 3942 3943 3944
    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
3945

Y
ying 已提交
3946
    "kitten" -> "sitten" -> "sittin" -> "sitting"
W
wanghaoshuang 已提交
3947

3948
    The input is a LoDTensor consisting of all the hypothesis strings with
Y
ying 已提交
3949 3950
    the total number denoted by `batch_size`, and the separation is specified
    by the LoD information. And the `batch_size` reference strings are arranged
3951
    in order in the same way in the input LoDTensor.
W
wanghaoshuang 已提交
3952

3953
    The output contains the `batch_size` results and each stands for the edit
Y
ying 已提交
3954 3955
    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
3956

3957 3958 3959
    Args:
        input(Variable): The indices for hypothesis strings.
        label(Variable): The indices for reference strings.
3960
        normalized(bool, default True): Indicated whether to normalize the edit distance by
Y
ying 已提交
3961
                          the length of reference string.
3962
        ignored_tokens(list<int>, default None): Tokens that should be removed before
Y
ying 已提交
3963
                                     calculating edit distance.
3964
        name (str): The name of this layer. It is optional.
3965

W
wanghaoshuang 已提交
3966
    Returns:
W
wanghaoshuang 已提交
3967
        Variable: sequence-to-sequence edit distance in shape [batch_size, 1].
W
wanghaoshuang 已提交
3968 3969 3970 3971 3972

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
3973
            y = fluid.layers.data(name='y', shape=[7], dtype='float32')
3974
            cost = fluid.layers.edit_distance(input=x,label=y)
3975
    """
3976
    helper = LayerHelper("edit_distance", **locals())
3977

3978
    # remove some tokens from input and labels
W
wanghaoshuang 已提交
3979
    if ignored_tokens is not None and len(ignored_tokens) > 0:
3980 3981 3982 3983 3984 3985 3986
        erased_input = helper.create_tmp_variable(dtype="int64")
        erased_label = helper.create_tmp_variable(dtype="int64")

        helper.append_op(
            type="sequence_erase",
            inputs={"X": [input]},
            outputs={"Out": [erased_input]},
W
wanghaoshuang 已提交
3987
            attrs={"tokens": ignored_tokens})
3988 3989 3990 3991 3992
        input = erased_input

        helper.append_op(
            type="sequence_erase",
            inputs={"X": [label]},
W
whs 已提交
3993
            outputs={"Out": [erased_label]},
W
wanghaoshuang 已提交
3994
            attrs={"tokens": ignored_tokens})
3995 3996
        label = erased_label

3997 3998
    # edit distance op
    edit_distance_out = helper.create_tmp_variable(dtype="int64")
3999
    sequence_num = helper.create_tmp_variable(dtype="int64")
4000 4001 4002 4003
    helper.append_op(
        type="edit_distance",
        inputs={"Hyps": [input],
                "Refs": [label]},
4004 4005
        outputs={"Out": [edit_distance_out],
                 "SequenceNum": [sequence_num]},
4006 4007
        attrs={"normalized": normalized})

4008
    return edit_distance_out, sequence_num
4009 4010 4011 4012 4013


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
4014

Y
ying 已提交
4015 4016 4017 4018
    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.
4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031 4032 4033 4034 4035

    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]]

4036
        input.lod = [[4, 4]]
4037 4038 4039 4040 4041 4042 4043

        Then:

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

4044
        output.lod = [[2, 1]]
4045 4046 4047

    Args:

Y
ying 已提交
4048 4049 4050 4051 4052 4053 4054 4055 4056
        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).
4057
        name (str): The name of this layer. It is optional.
4058 4059

    Returns:
4060
        Variable: CTC greedy decode result. If all the sequences in result were
4061
        empty, the result LoDTensor will be [-1] with LoD [[]] and dims [1, 1].
4062 4063 4064 4065 4066

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
W
wanghaoshuang 已提交
4067

4068
            cost = fluid.layers.ctc_greedy_decoder(input=x, blank=0)
W
wanghaoshuang 已提交
4069
    """
4070
    helper = LayerHelper("ctc_greedy_decoder", **locals())
Q
qingqing01 已提交
4071
    _, topk_indices = topk(input, k=1)
4072 4073 4074 4075 4076 4077

    # ctc align op
    ctc_out = helper.create_tmp_variable(dtype="int64")
    helper.append_op(
        type="ctc_align",
        inputs={"Input": [topk_indices]},
W
wanghaoshuang 已提交
4078
        outputs={"Output": [ctc_out]},
4079 4080
        attrs={"merge_repeated": True,
               "blank": blank})
4081
    return ctc_out
4082 4083


F
fengjiayi 已提交
4084
def warpctc(input, label, blank=0, norm_by_times=False):
W
wanghaoshuang 已提交
4085
    """
4086 4087
    An operator integrating the open source Warp-CTC library
    (https://github.com/baidu-research/warp-ctc)
W
wanghaoshuang 已提交
4088
    to compute Connectionist Temporal Classification (CTC) loss.
4089 4090
    It can be aliased as softmax with CTC, since a native softmax activation is
    interated to the Warp-CTC library, to to normlize values for each row of the
W
wanghaoshuang 已提交
4091 4092 4093
    input tensor.

    Args:
4094
       input (Variable): The unscaled probabilities of variable-length sequences,
W
wanghaoshuang 已提交
4095 4096 4097 4098
         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).
4099
       label (Variable): The ground truth of variable-length sequence,
4100 4101 4102
         which is a 2-D Tensor with LoD information. It is of the shape [Lg, 1],
         where Lg is th sum of all labels' length.
       blank (int, default 0): The blank label index of Connectionist
W
wanghaoshuang 已提交
4103 4104
         Temporal Classification (CTC) loss, which is in the
         half-opened interval [0, num_classes + 1).
4105 4106 4107
       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
4108
         follewed by a mean_op.
W
wanghaoshuang 已提交
4109 4110

    Returns:
4111 4112
        Variable: The Connectionist Temporal Classification (CTC) loss,
        which is a 2-D Tensor of the shape [batch_size, 1].
W
wanghaoshuang 已提交
4113 4114

    Examples:
4115

W
wanghaoshuang 已提交
4116
        .. code-block:: python
4117

4118 4119 4120
            label = fluid.layers.data(shape=[11, 8], dtype='float32', lod_level=1)
            predict = fluid.layers.data(shape=[11, 1], dtype='float32')
            cost = fluid.layers.warpctc(input=predict, label=label)
W
wanghaoshuang 已提交
4121 4122

    """
F
fengjiayi 已提交
4123
    helper = LayerHelper('warpctc', **locals())
W
wanghaoshuang 已提交
4124 4125 4126 4127 4128 4129 4130 4131 4132 4133 4134
    loss_out = helper.create_tmp_variable(dtype=input.dtype)
    grad_out = helper.create_tmp_variable(dtype=input.dtype)
    helper.append_op(
        type='warpctc',
        inputs={'Logits': [input],
                'Label': [label]},
        outputs={'WarpCTCGrad': [grad_out],
                 'Loss': [loss_out]},
        attrs={'blank': blank,
               'norm_by_times': norm_by_times})
    return loss_out
4135 4136 4137 4138 4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149


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]]
4150 4151 4152
            x.data = [[1,  2], [3,  4],
                      [5,  6], [7,  8],
                      [9, 10], [11, 12]]
4153 4154 4155 4156 4157
            x.dims = [6, 2]

        set new_dim = 4

        then out is a LoDTensor:
4158

4159
            out.lod  = [[0, 1, 3]]
4160 4161 4162 4163

            out.data = [[1,  2,  3,  4],
                        [5,  6,  7,  8],
                        [9, 10, 11, 12]]
4164 4165 4166 4167 4168 4169 4170
            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:
4171 4172 4173

       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.
4174 4175

    Returns:
4176

4177 4178 4179 4180 4181
        Variable: Reshaped LoDTensor according to new dimension.

    Examples:
        .. code-block:: python

4182
            x = fluid.layers.data(shape=[5, 20], dtype='float32', lod_level=1)
4183
            x_reshaped = fluid.layers.sequence_reshape(input=x, new_dim=10)
4184 4185 4186 4187 4188 4189 4190 4191 4192
    """
    helper = LayerHelper('sequence_reshape', **locals())
    out = helper.create_tmp_variable(helper.input_dtype())
    helper.append_op(
        type='sequence_reshape',
        inputs={'X': [input]},
        outputs={'Out': [out]},
        attrs={'new_dim': new_dim})
    return out
Y
ying 已提交
4193 4194


4195 4196 4197 4198
# 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
4199 4200 4201 4202 4203 4204
def nce(input,
        label,
        num_total_classes,
        sample_weight=None,
        param_attr=None,
        bias_attr=None,
C
chengduo 已提交
4205 4206
        num_neg_samples=None,
        name=None):
4207 4208 4209 4210 4211 4212 4213
    """
    ${comment}

    Args:
        input (Variable): input variable.
        label (Variable): label.
        num_total_classes (int):${num_total_classes_comment}
4214 4215
        sample_weight (Variable|None): A Variable of shape [batch_size, 1]
            storing a weight for each sample. The default weight for each
4216
            sample is 1.0.
C
chengduo 已提交
4217 4218 4219 4220 4221 4222 4223 4224 4225
        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.
4226
        num_neg_samples (int): ${num_neg_samples_comment}
C
chengduo 已提交
4227 4228
        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
4229

4230
    Returns:
Y
Yibing Liu 已提交
4231 4232 4233 4234 4235 4236 4237 4238 4239 4240 4241 4242 4243 4244 4245 4246 4247 4248 4249 4250 4251 4252 4253 4254 4255 4256 4257
        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')
4258
    """
Y
Yang Yu 已提交
4259 4260 4261
    helper = LayerHelper('nce', **locals())
    assert isinstance(input, Variable)
    assert isinstance(label, Variable)
C
chengduo 已提交
4262 4263

    dim = input.shape[1]
Y
Yang Yu 已提交
4264 4265 4266 4267 4268 4269
    num_true_class = label.shape[1]
    w = helper.create_parameter(
        attr=helper.param_attr,
        shape=[num_total_classes, dim],
        is_bias=False,
        dtype=input.dtype)
C
chengduo 已提交
4270 4271 4272 4273 4274 4275 4276 4277 4278 4279 4280 4281 4282
    inputs = {
        'Input': input,
        'Label': label,
        'Weight': w,
        'SampleWeight': sample_weight if sample_weight is not None else []
    }
    if helper.bias_attr:
        b = helper.create_parameter(
            attr=helper.bias_attr,
            shape=[num_total_classes, 1],
            is_bias=True,
            dtype=input.dtype)
        inputs['Bias'] = b
Y
Yang Yu 已提交
4283 4284 4285 4286
    cost = helper.create_tmp_variable(dtype=input.dtype)
    sample_logits = helper.create_tmp_variable(dtype=input.dtype)
    sample_labels = helper.create_tmp_variable(dtype=label.dtype)

Y
Yang Yu 已提交
4287 4288 4289 4290 4291 4292 4293 4294 4295
    if num_neg_samples is None:
        num_neg_samples = 10
    else:
        num_neg_samples = int(num_neg_samples)

    attrs = {
        'num_total_classes': int(num_total_classes),
        'num_neg_samples': num_neg_samples
    }
Y
Yang Yu 已提交
4296 4297 4298

    helper.append_op(
        type='nce',
C
chengduo 已提交
4299
        inputs=inputs,
Y
Yang Yu 已提交
4300 4301 4302 4303 4304 4305
        outputs={
            'Cost': cost,
            'SampleLogits': sample_logits,
            'SampleLabels': sample_labels
        },
        attrs=attrs)
Y
Yang Yu 已提交
4306
    return cost / (num_neg_samples + 1)
4307 4308


C
chengduo 已提交
4309 4310 4311 4312 4313 4314
def hsigmoid(input,
             label,
             num_classes,
             param_attr=None,
             bias_attr=None,
             name=None):
W
weixing02 已提交
4315 4316
    """
    The hierarchical sigmoid operator is used to accelerate the training
M
minqiyang 已提交
4317
    process of language model. This operator organizes the classes into a
G
guosheng 已提交
4318 4319 4320 4321 4322 4323 4324 4325 4326
    complete binary tree, each leaf node represents a class(a word) and each
    internal node acts as a binary classifier. For each word there's a unique
    path from root to it's leaf node, hsigmoid calculate the cost for each
    internal node on the path, and sum them to get a total cost. hsigmoid can
    achive a acceleration from :math:`O(N)` to :math:`O(logN)`, where :math:`N`
    represents the size of word dict.

    Refer to `Hierarchical Probabilistic Neural Network Language Model
    <http://www.iro.umontreal.ca/~lisa/pointeurs/hierarchical-nnlm-aistats05.pdf>`_
M
minqiyang 已提交
4327

W
weixing02 已提交
4328
    Args:
M
minqiyang 已提交
4329
        input (Variable): The input tensor variable with shape
G
guosheng 已提交
4330 4331 4332 4333 4334
            :math:`[N \\times D]`, where :math:`N` is the size of mini-batch,
            and :math:`D` is the feature size.
        label (Variable): The tensor variable contains labels of training data.
            It's a tensor with shape is :math:`[N \\times 1]`.
        num_classes: (int), The number of classes, must not be less than 2.
C
chengduo 已提交
4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345
        param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
             of hsigmoid. If it is set to None or one attribute of ParamAttr, hsigmoid
             will create ParamAttr as param_attr. If the Initializer of the param_attr
             is not set, the parameter is initialized with Xavier. Default: None.
        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of hsigmoid.
             If it is set to False, no bias will be added to the output units.
             If it is set to None or one attribute of ParamAttr, hsigmoid
             will create ParamAttr as bias_attr. If the Initializer of the bias_attr
             is not set, the bias is initialized zero. Default: None.
        name (str|None): A name for this layer(optional). If set None, the layer
             will be named automatically. Default: None.
W
weixing02 已提交
4346 4347 4348 4349 4350 4351 4352 4353

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

    Examples:

        .. code-block:: python

G
guosheng 已提交
4354 4355 4356
            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
4357 4358 4359 4360 4361 4362 4363 4364
    """

    helper = LayerHelper('hierarchical_sigmoid', **locals())
    dtype = helper.input_dtype()
    out = helper.create_tmp_variable(dtype)
    pre_out = helper.create_tmp_variable(dtype)
    dim = input.shape[1]
    if num_classes < 2:
G
guosheng 已提交
4365
        raise ValueError("num_classes must not be less than 2.")
W
weixing02 已提交
4366 4367 4368 4369 4370
    weights = helper.create_parameter(
        attr=helper.param_attr,
        shape=[num_classes - 1, dim],
        is_bias=False,
        dtype=input.dtype)
W
weixing02 已提交
4371 4372 4373 4374 4375 4376 4377 4378
    inputs = {"X": input, "W": weights, "Label": label}
    if helper.bias_attr:
        bias = helper.create_parameter(
            attr=helper.bias_attr,
            shape=[1, num_classes - 1],
            is_bias=True,
            dtype=input.dtype)
        inputs['Bias'] = bias
W
weixing02 已提交
4379 4380
    helper.append_op(
        type="hierarchical_sigmoid",
W
weixing02 已提交
4381
        inputs=inputs,
W
weixing02 已提交
4382 4383 4384 4385 4386 4387
        outputs={"Out": out,
                 "PreOut": pre_out},
        attrs={"num_classes": num_classes})
    return out


Y
fix ci.  
ying 已提交
4388
def transpose(x, perm, name=None):
Y
ying 已提交
4389 4390 4391 4392 4393 4394 4395
    """
    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:
4396 4397 4398
        x (Variable): The input Tensor.
        perm (list): A permutation of the dimensions of `input`.
        name (str): The name of this layer. It is optional.
Y
ying 已提交
4399 4400 4401 4402 4403 4404 4405 4406

    Returns:
        Variable: A transposed Tensor.

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[5, 10, 15], dtype='float32')
Y
fix ci.  
ying 已提交
4407
            x_transposed = layers.transpose(x, perm=[1, 0, 2])
Y
ying 已提交
4408 4409
    """

Y
fix ci.  
ying 已提交
4410
    if len(perm) != len(x.shape):
Y
ying 已提交
4411 4412 4413
        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
4414 4415 4416 4417 4418 4419
    for idx, dim in enumerate(perm):
        if dim >= len(x.shape):
            raise ValueError(
                "Each element in perm should be less than x's rank. "
                "%d-th element in perm is %d which accesses x's rank %d." %
                (idx, perm[idx], len(x.shape)))
Y
ying 已提交
4420 4421

    helper = LayerHelper('transpose', **locals())
Y
fix ci.  
ying 已提交
4422
    out = helper.create_tmp_variable(x.dtype)
4423
    x_shape = helper.create_tmp_variable(x.dtype)
Y
ying 已提交
4424
    helper.append_op(
4425
        type='transpose2',
Y
fix ci.  
ying 已提交
4426
        inputs={'X': [x]},
4427 4428
        outputs={'Out': [out],
                 'XShape': [x_shape]},
Y
ying 已提交
4429 4430
        attrs={'axis': perm})
    return out
4431 4432


4433 4434 4435 4436 4437 4438 4439
def im2sequence(input,
                filter_size=1,
                stride=1,
                padding=0,
                input_image_size=None,
                out_stride=1,
                name=None):
4440
    """
4441 4442 4443 4444 4445 4446 4447
    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:
4448 4449 4450 4451 4452 4453 4454 4455 4456 4457

    .. 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
4458 4459 4460 4461 4462 4463 4464 4465 4466 4467 4468 4469 4470 4471 4472 4473 4474 4475

        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.

4476 4477 4478 4479 4480 4481 4482 4483 4484
        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.

4485 4486 4487
        name (int): The name of this layer. It is optional.

    Returns:
W
wanghaoshuang 已提交
4488 4489 4490 4491 4492
        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.
4493 4494 4495 4496 4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519

    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
4520 4521 4522
            filter = [2, 2]
            stride = [1, 1]
            padding = [0, 0]
4523 4524 4525 4526 4527 4528 4529 4530 4531 4532 4533 4534

            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.]]

4535
            output.dims = {8, 8}
4536

4537
            output.lod = [[4, 4]]
4538

D
dzhwinter 已提交
4539
     Examples:
4540 4541 4542

        .. code-block:: python

4543 4544
            output = fluid.layers.im2sequence(
                input=layer, stride=[1, 1], filter_size=[2, 2])
4545 4546

    """
W
wanghaoshuang 已提交
4547 4548 4549 4550 4551 4552 4553 4554 4555 4556

    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])
4557 4558 4559 4560 4561 4562 4563
    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
4564
    helper = LayerHelper('im2sequence', **locals())
4565 4566
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
4567
        type='im2sequence', inputs=inputs, outputs={'Out': out}, attrs=attrs)
4568
    return out
4569 4570


Y
yuyang18 已提交
4571
@templatedoc()
4572
def row_conv(input, future_context_size, param_attr=None, act=None):
Y
yuyang18 已提交
4573 4574
    """
    ${comment}
4575 4576

    Args:
Y
yuyang18 已提交
4577
        input (${x_type}): ${x_comment}.
Y
yangyaming 已提交
4578 4579
        future_context_size (int): Future context size. Please note, the shape
            of convolution kernel is [future_context_size + 1, D].
4580 4581 4582 4583 4584
        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
4585
        ${out_comment}.
4586 4587

    Examples:
Y
yuyang18 已提交
4588 4589 4590 4591
        >>> 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)
4592 4593 4594 4595 4596 4597 4598 4599 4600 4601 4602 4603
    """
    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)
    out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type='row_conv',
        inputs={'X': [input],
                'Filter': [filter_param]},
        outputs={'Out': [out]})
Y
yangyaming 已提交
4604
    return helper.append_activation(out)
4605 4606


Y
yuyang18 已提交
4607
@templatedoc()
4608 4609
def multiplex(inputs, index):
    """
Y
yuyang18 已提交
4610 4611 4612 4613 4614 4615 4616
    ${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)
4617 4618

    Args:
Y
yuyang18 已提交
4619 4620
       inputs (list): ${x_comment}.
       index (${ids_type}): ${ids_comment}.
4621 4622

    Returns:
Y
yuyang18 已提交
4623
        ${out_comment}.
4624 4625
    """
    helper = LayerHelper('multiplex', **locals())
Y
yangyaming 已提交
4626 4627 4628 4629 4630 4631

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

    out = helper.create_tmp_variable(inputs[0].dtype)
4632 4633 4634 4635 4636 4637
    helper.append_op(
        type='multiplex',
        inputs={'X': inputs,
                'Ids': index},
        outputs={'Out': [out]})
    return out
4638 4639


4640 4641 4642 4643
def softmax_with_cross_entropy(logits,
                               label,
                               soft_label=False,
                               ignore_index=-100):
4644 4645
    """
    **Softmax With Cross Entropy Operator.**
4646

4647 4648 4649 4650
    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.
4651

4652 4653 4654
    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.
4655

4656 4657 4658
    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.
4659

4660
    The equation is as follows:
4661

4662
    1) Hard label (one-hot label, so every sample has exactly one class)
4663

4664 4665 4666 4667
    .. math::

        loss_j =  -\\text{logit}_{label_j} +
        \\log\\left(\\sum_{i=0}^{K}\\exp(\\text{logit}_i)\\right), j = 1,..., K
4668

4669 4670 4671
    2) Soft label (each sample can have a distribution over all classes)

    .. math::
4672

4673 4674 4675 4676 4677 4678 4679 4680 4681 4682 4683 4684
        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

    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
4685 4686
        ignore_index (int): Specifies a target value that is ignored and does
                            not contribute to the input gradient. Only valid
4687 4688
                            if soft_label is set to False. Default: -100

4689 4690 4691 4692 4693 4694 4695 4696 4697
    Returns:
        Variable: The cross entropy loss is a 2-D tensor with shape [N x 1].

    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
4698 4699
            out = fluid.layers.softmax_with_cross_entropy(
                logits=fc, label=label)
4700 4701 4702 4703 4704 4705 4706 4707 4708 4709
    """
    helper = LayerHelper('softmax_with_cross_entropy', **locals())
    softmax = helper.create_tmp_variable(dtype=logits.dtype)
    loss = helper.create_tmp_variable(dtype=logits.dtype)
    helper.append_op(
        type='softmax_with_cross_entropy',
        inputs={'Logits': logits,
                'Label': label},
        outputs={'Softmax': softmax,
                 'Loss': loss},
4710 4711
        attrs={'soft_label': soft_label,
               'ignore_index': ignore_index})
4712 4713 4714 4715 4716
    return loss


def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None):
    """
Y
Yibing Liu 已提交
4717 4718
    This layer computes the smooth L1 loss for Variable :attr:`x` and :attr:`y`.
    It takes the first dimension of :attr:`x` and :attr:`y` as batch size.
Q
qingqing01 已提交
4719
    For each instance, it computes the smooth L1 loss element by element first
4720
    and then sums all the losses. So the shape of ouput Variable is
4721
    [batch_size, 1].
4722

4723 4724
    Args:
        x (Variable): A tensor with rank at least 2. The input value of smooth
Q
qingqing01 已提交
4725
            L1 loss op with shape [batch_size, dim1, ..., dimN].
4726
        y (Variable): A tensor with rank at least 2. The target value of smooth
Y
Yibing Liu 已提交
4727
            L1 loss op with same shape as :attr:`x`.
4728
        inside_weight (Variable|None):  A tensor with rank at least 2. This
4729 4730
            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
4731
            by this tensor element by element.
4732
        outside_weight (Variable|None): A tensor with rank at least 2. This
4733 4734
            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
4735
            element by element.
4736
        sigma (float|None): Hyper parameter of smooth L1 loss layer. A float
4737 4738
           scalar with default value 1.0.

4739
    Returns:
4740
        Variable: The output smooth L1 loss with shape [batch_size, 1].
4741 4742 4743 4744 4745

    Examples:
        .. code-block:: python

            data = fluid.layers.data(name='data', shape=[128], dtype='float32')
F
stash  
fengjiayi 已提交
4746 4747
            label = fluid.layers.data(
                name='label', shape=[100], dtype='float32')
4748
            fc = fluid.layers.fc(input=data, size=100)
F
fengjiayi 已提交
4749
            out = fluid.layers.smooth_l1(x=fc, y=label)
4750
    """
4751

4752 4753 4754 4755 4756 4757 4758 4759 4760 4761 4762 4763 4764 4765 4766
    helper = LayerHelper('smooth_l1_loss', **locals())
    diff = helper.create_tmp_variable(dtype=x.dtype)
    loss = helper.create_tmp_variable(dtype=x.dtype)
    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
4767 4768 4769 4770


def one_hot(input, depth):
    """
Y
Yibing Liu 已提交
4771
    This layer creates the one-hot representations for input indices.
4772 4773

    Args:
Y
Yibing Liu 已提交
4774 4775
        input(Variable): Input indices, last dimension must be 1.
        depth(scalar): An interger defining the depth of the one-hot dimension.
4776 4777

    Returns:
Y
Yibing Liu 已提交
4778
        Variable: The one-hot representations of input.
4779 4780

    Examples:
C
caoying03 已提交
4781
        .. code-block:: python
4782

Y
Yibing Liu 已提交
4783 4784
            label = layers.data(name="label", shape=[1], dtype="float32")
            one_hot_label = layers.one_hot(input=label, depth=10)
4785 4786 4787 4788 4789 4790 4791 4792 4793
    """
    helper = LayerHelper("one_hot", **locals())
    one_hot_out = helper.create_tmp_variable(dtype='float32')
    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
4794 4795


Y
Yu Yang 已提交
4796
def autoincreased_step_counter(counter_name=None, begin=1, step=1):
Y
Yu Yang 已提交
4797
    """
Y
yi.wu 已提交
4798 4799 4800
    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
4801 4802 4803 4804 4805 4806

    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.

4807 4808
    Returns:
        Variable: The global run counter.
Y
yi.wu 已提交
4809 4810 4811 4812 4813 4814

    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
4815 4816
    """
    helper = LayerHelper('global_step_counter')
Y
Yu Yang 已提交
4817 4818
    if counter_name is None:
        counter_name = '@STEP_COUNTER@'
Y
Yu Yang 已提交
4819 4820 4821 4822 4823
    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
4824
                value=begin - 1, force_cpu=True))
W
Wu Yi 已提交
4825
        helper.main_program.global_block()._prepend_op(
Y
Yu Yang 已提交
4826 4827
            type='increment',
            inputs={'X': [counter]},
Y
Yu Yang 已提交
4828 4829
            outputs={'Out': [counter]},
            attrs={'step': float(step)})
Y
Yu Yang 已提交
4830 4831 4832
        counter.stop_gradient = True

    return counter
Y
yangyaming 已提交
4833 4834


4835
def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None):
C
caoying03 已提交
4836
    """
C
caoying03 已提交
4837 4838
    Gives a new shape to the input Tensor without changing its data.

4839 4840 4841 4842 4843
    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
4844

4845
    Some tricks exist when specifying the target shape.
C
caoying03 已提交
4846

4847 4848 4849 4850
    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.

4851
    2. 0 means the actual dimension value is going to be copied from the
4852 4853 4854 4855
    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
4856 4857

    1. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
W
wanghaoshuang 已提交
4858
    is [6, 8], the reshape operator will transform x into a 2-D tensor with
4859
    shape [6, 8] and leaving x's data unchanged.
C
caoying03 已提交
4860

4861
    2. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
4862 4863
    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
4864 4865
    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
4866
    dimensions.
C
caoying03 已提交
4867

4868
    3. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
4869 4870 4871 4872
    is [-1, 0, 3, 2], the reshape operator will transform x into a 4-D tensor
    with shape [2, 4, 3, 2] and leaving x's data unchanged. In this case,
    besides -1, 0 means the actual dimension value is going to be copied from
    the corresponding dimension of x.
C
caoying03 已提交
4873 4874

    Args:
4875
        x(variable): The input tensor.
C
caoying03 已提交
4876 4877
        shape(list): The new shape. At most one dimension of the new shape can
                     be -1.
4878 4879 4880 4881 4882
        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`.
C
caoying03 已提交
4883
        act (str): The non-linear activation to be applied to output variable.
X
Xin Pan 已提交
4884 4885 4886 4887
        inplace(bool): If this flag is set true, the output
                       shares data with input without copying, otherwise
                       a new output tensor is created
                       whose data is copied from input x.
4888
        name (str): The name of this layer. It is optional.
C
caoying03 已提交
4889

4890 4891
    Returns:
        Variable: The output tensor.
C
caoying03 已提交
4892

X
Xin Pan 已提交
4893 4894 4895
    Raises:
        TypeError: if actual_shape is neither Variable nor None.

C
caoying03 已提交
4896 4897
    Examples:
        .. code-block:: python
G
guosheng 已提交
4898

4899
            data = fluid.layers.data(
4900
                name='data', shape=[2, 4, 6], dtype='float32')
C
caoying03 已提交
4901
            reshaped = fluid.layers.reshape(
4902
                x=data, shape=[-1, 0, 3, 2], act='tanh', inplace=True)
C
caoying03 已提交
4903 4904 4905
    """

    if not (isinstance(shape, list) or isinstance(shape, tuple)):
L
luotao1 已提交
4906
        raise ValueError("Input shape must be a python list or tuple.")
X
Xin Pan 已提交
4907 4908 4909 4910 4911
    inputs = {"X": x}
    if isinstance(actual_shape, Variable):
        inputs["Shape"] = actual_shape
    elif actual_shape is not None:
        raise TypeError("actual_shape should either be Variable or None")
C
caoying03 已提交
4912

4913 4914 4915 4916 4917 4918 4919 4920 4921 4922 4923 4924 4925 4926 4927
    # 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.")

4928
    helper = LayerHelper("reshape2", **locals())
D
dzhwinter 已提交
4929
    out = helper.create_tmp_variable(dtype=x.dtype)
4930
    x_shape = helper.create_tmp_variable(dtype=x.dtype)
C
caoying03 已提交
4931
    helper.append_op(
4932
        type="reshape2",
X
Xin Pan 已提交
4933
        inputs=inputs,
D
dzhwinter 已提交
4934
        attrs={"shape": shape},
4935 4936
        outputs={"Out": out,
                 "XShape": x_shape})
C
caoying03 已提交
4937

D
dzhwinter 已提交
4938
    return helper.append_activation(out)
4939

4940

4941
def squeeze(input, axes, name=None):
Y
Yibing Liu 已提交
4942
    """
M
minqiyang 已提交
4943 4944 4945
    Remove single-dimensional entries from the shape of a tensor. Takes a
    parameter axes with a list of axes to squeeze. If axes is not provided, all
    the single dimensions will be removed from the shape. If an axis is
Y
Yibing Liu 已提交
4946
    selected with shape entry not equal to one, an error is raised.
M
minqiyang 已提交
4947

Y
Yibing Liu 已提交
4948 4949
    Examples:
    Case 1:
M
minqiyang 已提交
4950
      Given
Y
Yibing Liu 已提交
4951 4952 4953 4954 4955 4956 4957 4958
        X.shape = (1, 3, 1, 5)
      and
        axes = [0]
      we get:
        Out.shape = (3, 1, 5)
      Case 2:
        Given
          X.shape = (1, 3, 1, 5)
M
minqiyang 已提交
4959
        and
Y
Yibing Liu 已提交
4960 4961 4962
          axes = []
        we get:
          Out.shape = (3, 5)
M
minqiyang 已提交
4963

Y
Yibing Liu 已提交
4964
    Args:
4965
        input (Variable): The input variable to be squeezed.
Y
Yibing Liu 已提交
4966
        axes (list): List of integers, indicating the dimensions to be squeezed.
4967
        name (str|None): Name for this layer.
Y
Yibing Liu 已提交
4968 4969 4970 4971 4972 4973 4974 4975

    Returns:
        Variable: Output squeezed variable.

    Examples:
        .. code-block:: python

            x = layers.data(name='x', shape=[5, 1, 10])
4976
            y = layers.sequeeze(input=x, axes=[1])
Y
Yibing Liu 已提交
4977 4978
    """
    helper = LayerHelper("squeeze", **locals())
4979
    out = helper.create_tmp_variable(dtype=input.dtype)
4980
    x_shape = helper.create_tmp_variable(dtype=input.dtype)
Y
Yibing Liu 已提交
4981
    helper.append_op(
4982
        type="squeeze2",
4983
        inputs={"X": input},
Y
Yibing Liu 已提交
4984
        attrs={"axes": axes},
4985 4986
        outputs={"Out": out,
                 "XShape": x_shape})
Y
Yibing Liu 已提交
4987

4988 4989 4990
    return out


4991
def unsqueeze(input, axes, name=None):
Y
Yibing Liu 已提交
4992
    """
M
minqiyang 已提交
4993 4994 4995
    Insert single-dimensional entries to the shape of a tensor. Takes one
    required argument axes, a list of dimensions that will be inserted.
    Dimension indices in axes are as seen in the output tensor.
Y
Yibing Liu 已提交
4996

M
minqiyang 已提交
4997 4998
    For example:
      Given a tensor such that tensor with shape [3, 4, 5],
Y
Yibing Liu 已提交
4999
      then Unsqueezed tensor with axes=[0, 4] has shape [1, 3, 4, 5, 1].
M
minqiyang 已提交
5000

Y
Yibing Liu 已提交
5001
    Args:
5002
        input (Variable): The input variable to be unsqueezed.
Y
Yibing Liu 已提交
5003
        axes (list): List of integers, indicating the dimensions to be inserted.
5004
        name (str|None): Name for this layer.
Y
Yibing Liu 已提交
5005 5006 5007 5008 5009 5010 5011 5012

    Returns:
        Variable: Output unsqueezed variable.

    Examples:
        .. code-block:: python

            x = layers.data(name='x', shape=[5, 10])
5013
            y = layers.unsequeeze(input=x, axes=[1])
Y
Yibing Liu 已提交
5014 5015
    """
    helper = LayerHelper("unsqueeze", **locals())
5016
    out = helper.create_tmp_variable(dtype=input.dtype)
5017
    x_shape = helper.create_tmp_variable(dtype=input.dtype)
Y
Yibing Liu 已提交
5018
    helper.append_op(
5019
        type="unsqueeze2",
5020
        inputs={"X": input},
Y
Yibing Liu 已提交
5021
        attrs={"axes": axes},
5022 5023
        outputs={"Out": out,
                 "XShape": x_shape})
Y
Yibing Liu 已提交
5024

5025 5026
    return out

5027

Y
yangyaming 已提交
5028
def lod_reset(x, y=None, target_lod=None):
Y
yangyaming 已提交
5029
    """
Y
Yibing Liu 已提交
5030
    Set LoD of :attr:`x` to a new one specified by :attr:`y` or
5031 5032 5033 5034
    :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
5035
    :attr:`Y.data` or :attr:`target_lod`, only one level LoD is supported.
Y
yangyaming 已提交
5036 5037 5038 5039 5040 5041

    .. code-block:: text

        * Example 1:

            Given a 1-level LoDTensor x:
5042
                x.lod =  [[ 2,           3,                   1 ]]
Y
yangyaming 已提交
5043 5044 5045
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

5046
            target_lod: [4, 2]
Y
yangyaming 已提交
5047 5048

            then we get a 1-level LoDTensor:
5049
                out.lod =  [[4,                          2]]
Y
yangyaming 已提交
5050 5051 5052 5053 5054 5055
                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:
5056
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
5057 5058 5059 5060
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a Tensor:
5061
                y.data = [[2, 4]]
Y
yangyaming 已提交
5062 5063 5064
                y.dims = [1, 3]

            then we get a 1-level LoDTensor:
5065
                out.lod =  [[2,            4]]
Y
yangyaming 已提交
5066 5067 5068 5069 5070 5071
                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:
5072
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
5073 5074 5075 5076
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a 2-level LoDTensor:
5077
                y.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
5078 5079 5080 5081
                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:
5082
                out.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
5083 5084 5085 5086 5087
                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.
5088
        y (Variable|None): If provided, output's LoD would be derived
Y
Yibing Liu 已提交
5089
                           from :attr:`y`.
Y
yangyaming 已提交
5090
        target_lod (list|tuple|None): One level LoD which should be considered
Y
Yibing Liu 已提交
5091
                                      as target LoD when :attr:`y` not provided.
Y
yangyaming 已提交
5092 5093

    Returns:
Y
Yibing Liu 已提交
5094
        Variable: Output variable with LoD specified by this layer.
Y
yangyaming 已提交
5095 5096

    Raises:
Y
Yibing Liu 已提交
5097
        ValueError: If :attr:`y` and :attr:`target_lod` are both None.
Y
yangyaming 已提交
5098 5099 5100 5101 5102 5103 5104 5105 5106 5107 5108 5109 5110 5111 5112 5113 5114 5115 5116 5117 5118 5119 5120 5121

    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())
    out = helper.create_tmp_variable(dtype=x.dtype)
    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
5122 5123 5124 5125 5126 5127 5128 5129 5130 5131 5132


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
5133
      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
5134 5135 5136 5137 5138 5139 5140 5141 5142 5143 5144 5145 5146 5147 5148 5149 5150 5151 5152 5153 5154 5155 5156 5157 5158 5159 5160 5161

    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
5162 5163
          data = fluid.layers.data(
              name="data", shape=[3, 112, 112], dtype="float32")
D
dragonwarrior 已提交
5164 5165 5166 5167 5168 5169 5170 5171 5172 5173 5174 5175 5176 5177 5178 5179 5180 5181 5182 5183 5184 5185 5186 5187 5188 5189 5190
          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))

    mid_out = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
    lrn_out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type="lrn",
        inputs={"X": input},
        outputs={
            "Out": lrn_out,
            "MidOut": mid_out,
        },
        attrs={"n": n,
               "k": k,
               "alpha": alpha,
               "beta": beta})

    return lrn_out
G
guosheng 已提交
5191 5192 5193 5194


def pad(x, paddings, pad_value=0., name=None):
    """
G
guosheng 已提交
5195
    Pads a tensor with a constant value given by :attr:`pad_value`, and the
W
wanghaoshuang 已提交
5196
    padded width is specified by :attr:`paddings`.
G
guosheng 已提交
5197

G
guosheng 已提交
5198 5199 5200 5201
    Specifically, the number of values padded before the contents of :attr:`x`
    in dimension :attr:`i` is indicated by :attr:`paddings[i]`, and the number
    of values padded after the contents of :attr:`x` in dimension :attr:`i` is
    indicated by :attr:`paddings[i+1]`.
G
guosheng 已提交
5202 5203 5204 5205 5206 5207 5208 5209 5210 5211 5212 5213 5214 5215 5216 5217 5218 5219 5220 5221 5222 5223

    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
5224
                         The length of :attr:paddings must be
G
guosheng 已提交
5225 5226 5227 5228 5229 5230 5231 5232 5233 5234
                         :math:`rank(x) \\times 2`.
        pad_value (float): The constant value used to pad.
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        Variable: The padded tensor variable.

    Examples:
        .. code-block:: python
G
guosheng 已提交
5235

G
guosheng 已提交
5236 5237 5238 5239 5240 5241 5242 5243 5244 5245 5246 5247 5248 5249
            # 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()
    out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type='pad',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'paddings': paddings,
               'pad_value': float(pad_value)})
    return out
5250 5251


C
chengduo 已提交
5252 5253 5254 5255 5256 5257 5258 5259 5260 5261 5262 5263 5264 5265 5266 5267 5268 5269 5270 5271 5272 5273 5274 5275 5276 5277 5278 5279 5280 5281 5282 5283 5284 5285 5286 5287 5288 5289 5290 5291 5292 5293 5294 5295 5296 5297 5298 5299 5300 5301 5302 5303 5304 5305 5306 5307 5308 5309 5310 5311 5312 5313 5314 5315 5316 5317 5318 5319 5320 5321 5322 5323 5324 5325 5326 5327 5328 5329 5330 5331
def pad_constant_like(x, y, pad_value=0., name=None):
    """
    Pad input(Y) with :attr:`pad_value`, the number of values padded to
    the edges of each axis is specified by the difference of the shape
    of X and Y. ((0, shape_x_0 - shape_y_0), ... (0, shape_x_n - shape_y_n))
    unique pad widths for each axis. The input should be a k-D
    tensor(k > 0 and k < 7).

    See below for an example.

    .. code-block:: text

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

            Y = [[[[35, 36, 37]],
                  [[38, 39, 40]],
                  [[41, 42, 43]]]]
            Y.shape = (1, 3, 1, 3)

    And
        pad_value = -1,

    Return:
        Out = [[[[35, 36, 37],
                  [-1, -1, -1]],
                [[38, 39, 40],
                  [-1, -1, -1]],
                 [[41, 42, 43],
                  [-1, -1, -1]]],
                [[[-1, -1, -1],
                  [-1, -1, -1]],
                 [[-1, -1, -1],
                  [-1, -1, -1]],
                 [[-1, -1, -1],
                  [-1, -1, -1]]]]
        Out.shape = (2, 3, 2, 3)

    Args:
        x (Variable): The input tensor variable.
        y (Variable): The input tensor variable.
        pad_value (float): The constant value used to pad.
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        Variable: The padded tensor variable.

    Examples:
        .. code-block:: python

            # x is a rank 4 tensor variable, x.shape = (2, 3, 2, 3)
            # y is a rank 4 tensor variable, y.shape = (1, 3, 1, 3)
            out = fluid.layers.pad_constant_like(x=x, y=y, pad_value=0.)
            # out is a rank 4 tensor variable, and out.shape = [2, 3 ,2 , 3]
    """
    helper = LayerHelper('pad_constant_like', input=x, **locals())
    dtype = helper.input_dtype()
    out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type='pad_constant_like',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
        attrs={'pad_value': float(pad_value)})
    return out


5332 5333 5334 5335 5336 5337 5338
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
5339 5340
    called label-smoothing regularization (LSR).

5341 5342 5343 5344 5345 5346 5347 5348 5349 5350 5351 5352 5353 5354 5355 5356 5357 5358 5359 5360 5361 5362 5363
    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
5364
                              be :math:`(1, class\_num)`.
5365 5366
        epsilon(float): The weight used to mix up the original ground-truth
                        distribution and the fixed distribution.
5367
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32,
5368 5369 5370 5371 5372 5373 5374 5375 5376 5377 5378 5379 5380 5381 5382 5383 5384 5385 5386 5387 5388 5389 5390 5391 5392 5393 5394
                                                  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
    smooth_label = helper.create_tmp_variable(dtype)
    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
5395 5396


Y
yi.wu 已提交
5397
@templatedoc()
5398 5399
def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0):
    """
Y
yi.wu 已提交
5400
    ${comment}
5401 5402

    Args:
Y
yi.wu 已提交
5403 5404
        input (Variable): ${x_comment}
        rois (Variable): ROIs (Regions of Interest) to pool over.
Y
yi.wu 已提交
5405 5406 5407
        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
5408 5409

    Returns:
Y
update  
yi.wu 已提交
5410
        Variable: ${out_comment}.
5411 5412

    Examples:
5413 5414
        .. code-block:: python

5415
            pool_out = fluid.layers.roi_pool(input=x, rois=rois, 7, 7, 1.0)
5416 5417 5418 5419 5420 5421 5422 5423 5424 5425 5426 5427 5428 5429 5430 5431 5432
    """
    helper = LayerHelper('roi_pool', **locals())
    dtype = helper.input_dtype()
    pool_out = helper.create_tmp_variable(dtype)
    argmaxes = helper.create_tmp_variable(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
5433 5434


J
jerrywgz 已提交
5435 5436 5437 5438 5439 5440
@templatedoc()
def roi_align(input,
              rois,
              pooled_height=1,
              pooled_width=1,
              spatial_scale=1.0,
J
jerrywgz 已提交
5441 5442
              sampling_ratio=-1,
              name=None):
J
jerrywgz 已提交
5443 5444 5445 5446 5447 5448 5449 5450 5451 5452 5453 5454 5455 5456 5457 5458 5459 5460 5461 5462 5463 5464 5465 5466 5467 5468 5469 5470 5471 5472 5473 5474 5475 5476 5477 5478 5479 5480 5481 5482
    """
    ${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

            align_out = fluid.layers.roi_align(input=x, 
                                               rois=rois, 
                                               pooled_height=7, 
                                               pooled_width=7,
                                               spatial_scale=0.5,
                                               sampling_ratio=-1)
    """
    helper = LayerHelper('roi_align', **locals())
    dtype = helper.input_dtype()
    align_out = helper.create_tmp_variable(dtype)
    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
5483 5484 5485 5486 5487 5488 5489 5490 5491 5492 5493 5494 5495 5496 5497 5498 5499 5500 5501 5502 5503 5504 5505 5506 5507 5508
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:
5509 5510
        .. code-block:: python

W
whs 已提交
5511 5512 5513 5514
            predictions = fluid.layers.softmax(x)
            loss = fluid.layers.dice_loss(input=predictions, label=label, 2)
    """
    label = one_hot(label, depth=input.shape[-1])
5515
    reduce_dim = list(range(1, len(input.shape)))
W
whs 已提交
5516 5517 5518 5519 5520 5521
    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)
5522 5523


5524 5525 5526 5527 5528
def image_resize(input,
                 out_shape=None,
                 scale=None,
                 name=None,
                 resample='BILINEAR'):
5529
    """
Q
qiaolongfei 已提交
5530
    **Resize a Batch of Images**
F
stash  
fengjiayi 已提交
5531

5532
    The input must be a tensor of the shape (num_batches, channels, in_h, in_w),
5533 5534 5535
    and the resizing only applies on the last two dimensions(hight and width).

    Supporting resample methods:
Q
update  
qiaolongfei 已提交
5536

5537
        'BILINEAR' : Bilinear interpolation
F
stash  
fengjiayi 已提交
5538

5539
    Args:
5540
        input (Variable): The input tensor of image resize layer,
5541 5542
                          This is a 4-D tensor of the shape
                          (num_batches, channels, in_h, in_w).
5543
        out_shape(list|tuple|Variable|None): Output shape of image resize
5544 5545
                                    layer, the shape is (out_h, out_w).
                                    Default: None
B
baiyf 已提交
5546
        scale(float|None): The multiplier for the input height or width.
5547 5548 5549
                         At least one of out_shape or scale must be set.
                         And out_shape has a higher priority than scale.
                         Default: None
5550 5551
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
5552 5553
        resample(str): The resample method. It can only be 'BILINEAR' currently.
                       Default: 'BILINEAR'
5554 5555

    Returns:
Q
update  
qiaolongfei 已提交
5556 5557
        Variable: The output is a 4-D tensor of the shape
        (num_batches, channls, out_h, out_w).
F
stash  
fengjiayi 已提交
5558

5559 5560 5561
    Examples:
        .. code-block:: python

5562
            out = fluid.layers.image_resize(input, out_shape=[12, 12])
5563
    """
5564 5565 5566 5567
    resample_methods = {'BILINEAR': 'bilinear_interp'}
    if resample not in resample_methods:
        raise ValueError(
            "The 'resample' of image_resize can only be 'BILINEAR' currently.")
5568 5569
    if out_shape is None and scale is None:
        raise ValueError("One of out_shape and scale must not be None")
5570 5571
    helper = LayerHelper('bilinear_interp', **locals())
    dtype = helper.input_dtype()
5572 5573 5574 5575

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

5576 5577 5578
    out_h = 0
    out_w = 0
    inputs = {"X": input}
5579
    if out_shape is not None:
B
baiyf 已提交
5580 5581 5582
        if not (_is_list_or_turple_(out_shape) and
                len(out_shape) == 2) and not isinstance(out_shape, Variable):
            raise ValueError('out_shape should be a list or tuple or variable')
5583 5584 5585 5586 5587 5588
        if _is_list_or_turple_(out_shape):
            out_shape = list(map(int, out_shape))
            out_h = out_shape[0]
            out_w = out_shape[1]
        else:
            inputs['OutSize'] = out_shape
5589 5590 5591 5592
    else:
        out_h = int(input.shape[2] * scale)
        out_w = int(input.shape[3] * scale)

5593 5594
    out = helper.create_tmp_variable(dtype)
    helper.append_op(
5595
        type=resample_methods[resample],
5596
        inputs=inputs,
5597 5598 5599 5600
        outputs={"Out": out},
        attrs={"out_h": out_h,
               "out_w": out_w})
    return out
F
stash  
fengjiayi 已提交
5601 5602


Y
yuyang18 已提交
5603
@templatedoc(op_type="bilinear_interp")
5604 5605
def resize_bilinear(input, out_shape=None, scale=None, name=None):
    """
Y
yuyang18 已提交
5606 5607 5608 5609 5610 5611
    ${comment}

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

        out_shape(${out_size_type}): ${out_size_comment}.
5612

Y
yuyang18 已提交
5613 5614 5615 5616 5617 5618 5619 5620
        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.

    Returns:
        ${out_comment}.
5621 5622 5623 5624 5625 5626 5627
    """

    return image_resize(input, out_shape, scale, name, 'BILINEAR')


def image_resize_short(input, out_short_len, resample='BILINEAR'):
    """
5628 5629 5630
    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
5631 5632 5633 5634 5635 5636 5637
    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.
5638
        resample (str): resample method, default: BILINEAR.
F
fengjiayi 已提交
5639

5640
    Returns:
Q
update  
qiaolongfei 已提交
5641
        Variable: The output is a 4-D tensor of the shape
5642
        (num_batches, channls, out_h, out_w).
5643 5644 5645 5646 5647 5648 5649 5650 5651 5652
    """
    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
5653 5654 5655
    out_shape[long_idx] = int(
        float(out_shape[long_idx]) * (float(out_short_len) / float(hw[
            short_idx])) + 0.5)
5656 5657 5658
    return image_resize(input=input, out_shape=out_shape, resample=resample)


W
whs 已提交
5659 5660
def gather(input, index):
    """
Q
qiaolongfei 已提交
5661 5662
    **Gather Layer**

5663
    Output is obtained by gathering entries of the outer-most dimension
W
whs 已提交
5664 5665 5666 5667
    of X indexed by `index` and concatenate them together.

    .. math::

5668
        Out = X[Index]
W
whs 已提交
5669 5670 5671 5672 5673 5674 5675


    .. code-block:: text


                Given:

5676 5677
                X = [[1, 2],
                     [3, 4],
W
whs 已提交
5678 5679 5680 5681 5682 5683 5684 5685 5686 5687
                     [5, 6]]

                Index = [1, 2]

                Then:

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

    Args:
5688
        input (Variable): The source input with rank>=1.
W
whs 已提交
5689 5690 5691 5692 5693 5694
        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
5695

W
whs 已提交
5696 5697 5698 5699 5700 5701 5702 5703 5704 5705 5706 5707 5708 5709 5710
        .. code-block:: python

            output = fluid.layers.gather(x, index)
    """
    helper = LayerHelper('gather', **locals())
    dtype = helper.input_dtype()
    out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type="gather",
        inputs={"X": input,
                "Index": index},
        outputs={"Out": out})
    return out


5711 5712 5713 5714 5715 5716 5717 5718 5719 5720 5721 5722 5723 5724 5725 5726 5727 5728 5729 5730 5731 5732 5733 5734 5735 5736 5737 5738 5739 5740 5741 5742 5743 5744 5745 5746 5747 5748 5749 5750 5751
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()
    out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type="scatter",
        inputs={"X": input,
                "Ids": index,
                "Updates": updates},
        outputs={"Out": out})
    return out


Q
Qingsheng Li 已提交
5752 5753 5754 5755 5756 5757 5758 5759 5760 5761 5762 5763 5764 5765 5766 5767 5768 5769 5770 5771 5772 5773 5774 5775 5776 5777 5778 5779 5780 5781 5782 5783 5784 5785 5786 5787 5788 5789 5790 5791 5792 5793 5794 5795 5796 5797 5798 5799 5800 5801 5802 5803 5804 5805 5806 5807 5808 5809 5810 5811
def sequence_scatter(input, index, updates, name=None):
    """
    **Sequence Scatter Layer**

    This operator scatters the Updates tensor to the input X. It uses the LoD
    information of Ids to select the rows to update, and use the values in Ids as
    the columns to update in each row of X.

    Here is an example:
    Given the following input:
    .. code-block:: text
        input.data = [[1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
                      [1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
                      [1.0, 1.0, 1.0, 1.0, 1.0, 1.0]]
        input.dims = [3, 6]

        index.data = [[0], [1], [2], [5], [4], [3], [2], [1], [3], [2], [5], [4]]
        index.lod =  [[0,        3,                       8,                 12]]

        updates.data = [[0.3], [0.3], [0.4], [0.1], [0.2], [0.3], [0.4], [0.0], [0.2], [0.3], [0.1], [0.4]]
        updates.lod =  [[  0,            3,                                 8,                         12]]

    Then we have the output:
    .. code-block:: text
        out.data = [[1.3, 1.3, 1.4, 1.0, 1.0, 1.0],
                    [1.0, 1.0, 1.4, 1.3, 1.2, 1.1],
                    [1.0, 1.0, 1.3, 1.2, 1.4, 1.1]]
        out.dims = X.dims = [3, 6]

    Args:
        input (Variable): The source input with rank>=1.
        index (Variable): A LoD Tensor. The index input of sequence scatter op
            where input will be  updated. The index input with rank=1. Its dtype
            should be int32 or int64 as it is used as indexes.
        updates (Variable): A LoD Tensor. The values to scatter to the input
            tensor X, must be a LoDTensor with the same LoD information as index.
        name (str|None): The output variable name. Default None.

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

    Examples:

        .. code-block:: python

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

    """
    helper = LayerHelper('sequence_scatter', **locals())
    dtype = helper.input_dtype()
    out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type="sequence_scatter",
        inputs={"X": input,
                "Ids": index,
                "Updates": updates},
        outputs={"Out": out})
    return out


Y
yuyang18 已提交
5812 5813 5814 5815 5816 5817 5818 5819 5820 5821 5822 5823 5824
@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}
5825

5826 5827 5828
    Examples:
        >>> img = fluid.layers.data("img", [3, 256, 256])
        >>> cropped_img = fluid.layers.random_crop(img, shape=[3, 224, 224])
Y
yuyang18 已提交
5829
    """
F
stash  
fengjiayi 已提交
5830
    helper = LayerHelper("random_crop", **locals())
F
fengjiayi 已提交
5831
    dtype = x.dtype
F
stash  
fengjiayi 已提交
5832
    out = helper.create_tmp_variable(dtype)
Y
yuyang18 已提交
5833
    if seed is None:
5834
        seed = np.random.randint(-65536, 65536)
F
fengjiayi 已提交
5835
    op_attrs = {"shape": shape}
F
stash  
fengjiayi 已提交
5836
    if isinstance(seed, int):
F
fengjiayi 已提交
5837 5838 5839 5840 5841
        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
5842 5843 5844 5845
    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
5846
        inputs={"X": x,
F
stash  
fengjiayi 已提交
5847 5848
                "Seed": seed},
        outputs={"Out": out,
F
fengjiayi 已提交
5849 5850
                 "SeedOut": seed},
        attrs=op_attrs)
F
stash  
fengjiayi 已提交
5851
    return out
W
whs 已提交
5852 5853


5854
def log(x, name=None):
W
wanghaoshuang 已提交
5855 5856 5857 5858 5859
    """
    Calculates the natural log of the given input tensor, element-wise.

    .. math::

5860
        Out = \\ln(x)
W
wanghaoshuang 已提交
5861 5862

    Args:
5863
        x (Variable): Input tensor.
5864 5865
        name (str|None, default None): A name for this layer If set None,
            the layer will be named automatically.
W
wanghaoshuang 已提交
5866 5867 5868 5869 5870 5871 5872 5873

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

    Examples:

        .. code-block:: python

5874
            output = fluid.layers.log(x)
W
wanghaoshuang 已提交
5875 5876
    """
    helper = LayerHelper('log', **locals())
W
wanghaoshuang 已提交
5877
    dtype = helper.input_dtype(input_param_name='x')
W
wanghaoshuang 已提交
5878
    out = helper.create_tmp_variable(dtype)
W
wanghaoshuang 已提交
5879
    helper.append_op(type="log", inputs={"X": x}, outputs={"Out": out})
W
wanghaoshuang 已提交
5880 5881 5882
    return out


5883
def relu(x, name=None):
W
wanghaoshuang 已提交
5884 5885
    """
    Relu takes one input data (Tensor) and produces one output data (Tensor)
5886
    where the rectified linear function, y = max(0, x), is applied to
W
wanghaoshuang 已提交
5887 5888 5889 5890
    the tensor elementwise.

    .. math::

5891
        Out = \\max(0, x)
W
wanghaoshuang 已提交
5892 5893

    Args:
5894
        x (Variable): The input tensor.
5895 5896
        name (str|None, default None): A name for this layer If set None,
            the layer will be named automatically.
W
wanghaoshuang 已提交
5897 5898 5899 5900 5901 5902 5903 5904

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

    Examples:

        .. code-block:: python

5905
            output = fluid.layers.relu(x)
W
wanghaoshuang 已提交
5906 5907
    """
    helper = LayerHelper('relu', **locals())
W
wanghaoshuang 已提交
5908
    dtype = helper.input_dtype(input_param_name='x')
W
wanghaoshuang 已提交
5909
    out = helper.create_tmp_variable(dtype)
W
wanghaoshuang 已提交
5910
    helper.append_op(type="relu", inputs={"X": x}, outputs={"Out": out})
W
wanghaoshuang 已提交
5911
    return out
5912 5913


W
whs 已提交
5914 5915 5916
def mean_iou(input, label, num_classes):
    """
    Mean Intersection-Over-Union is a common evaluation metric for
5917 5918 5919 5920
    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
5921
    .. math::
5922 5923

        IOU = \\frac{true\_positiv}{(true\_positive + false\_positive + false\_negative)}.
W
whs 已提交
5924

5925
    The predictions are accumulated in a confusion matrix and mean-IOU
W
whs 已提交
5926 5927 5928 5929 5930
    is then calculated from it.


    Args:
        input (Variable): A Tensor of prediction results for semantic labels with type int32 or int64.
5931
        label (Variable): A Tensor of ground truth labels with type int32 or int64.
W
whs 已提交
5932
                           Its shape should be the same as input.
5933
        num_classes (int): The possible number of labels.
W
whs 已提交
5934 5935 5936 5937

    Returns:
        mean_iou (Variable): A Tensor representing the mean intersection-over-union with shape [1].
        out_wrong(Variable): A Tensor with shape [num_classes]. The wrong numbers of each class.
5938
        out_correct(Variable): A Tensor with shape [num_classes]. The correct numbers of each class.
W
whs 已提交
5939 5940 5941 5942

    Examples:

        .. code-block:: python
5943

W
whs 已提交
5944 5945 5946 5947 5948 5949 5950 5951 5952
            iou, wrongs, corrects = fluid.layers.mean_iou(predict, label, num_classes)
    """
    helper = LayerHelper('mean_iou', **locals())
    dtype = helper.input_dtype()
    out_mean_iou = helper.create_tmp_variable(dtype='float32')
    out_wrong = helper.create_tmp_variable(dtype='int32')
    out_correct = helper.create_tmp_variable(dtype='int32')
    helper.append_op(
        type="mean_iou",
W
whs 已提交
5953 5954
        inputs={"Predictions": input,
                "Labels": label},
W
whs 已提交
5955
        outputs={
W
whs 已提交
5956 5957 5958
            "OutMeanIou": out_mean_iou,
            "OutWrong": out_wrong,
            "OutCorrect": out_correct
W
whs 已提交
5959 5960 5961
        },
        attrs={"num_classes": num_classes})
    return out_mean_iou, out_wrong, out_correct
5962 5963 5964 5965 5966 5967 5968 5969 5970 5971 5972 5973 5974 5975 5976 5977 5978 5979 5980 5981 5982 5983 5984 5985 5986 5987 5988 5989 5990 5991 5992 5993 5994 5995 5996 5997 5998 5999 6000 6001 6002 6003 6004 6005 6006 6007 6008 6009 6010 6011 6012 6013 6014 6015 6016 6017 6018 6019 6020 6021 6022 6023 6024 6025 6026 6027 6028 6029 6030 6031 6032 6033 6034 6035


def crop(x, shape=None, offsets=None, name=None):
    """
    Crop input into output, as specified by offsets and shape.

    .. code-block:: text

        * Case 1:
            Given
                X = [[0, 1, 2, 0, 0]
                     [0, 3, 4, 0, 0]
                     [0, 0, 0, 0, 0]],
            and
                shape = [2, 2],
                offsets = [0, 1],
            output is:
                Out = [[1, 2],
                       [3, 4]].
        * Case 2:
            Given
                X = [[0, 1, 2, 5, 0]
                     [0, 3, 4, 6, 0]
                     [0, 0, 0, 0, 0]],
            and shape is tensor
                shape = [[0, 0, 0]
                         [0, 0, 0]]
            and
                offsets = [0, 1],

            output is:
                Out = [[1, 2, 5],
                       [3, 4, 6]].

    Args:
        x (Variable): The input tensor variable.
        shape (Variable|list/tuple of integer): The output shape is specified
            by `shape`, which can a Variable or a list/tupe of integer.
            If a tensor Variable, it's rank must be the same as `x`. This way
            is suitable for the case that the output shape may be changed each
            iteration. If a list/tupe of integer, it's length must be the same
            as the rank of `x`
        offsets (Variable|list/tuple of integer|None): Specifies the copping
            offsets at each dimension. It can be a Variable or or a list/tupe
            of integer. If a tensor Variable, it's rank must be the same as `x`.
            This way is suitable for the case that the offsets may be changed
            each iteration. If a list/tupe of integer, it's length must be the
            same as the rank of `x`. If None, the offsets are 0 at each
            dimension.
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        Variable: The cropped tensor variable.

    Raises:
        ValueError: If shape is not a list, tuple or Variable.

    Examples:

        .. code-block:: python

            x = fluid.layers.data(name="x", shape=[3, 5], dtype="float32")
            y = fluid.layers.data(name="y", shape=[2, 3], dtype="float32")
            crop = fluid.layers.crop(x, shape=y)

            # or
            z = fluid.layers.data(name="z", shape=[3, 5], dtype="float32")
            crop = fluid.layers.crop(z, shape=[2, 3])

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

    if not (isinstance(shape, list) or isinstance(shape, tuple) or \
C
chengduo 已提交
6036
                    isinstance(shape, Variable)):
6037 6038 6039 6040 6041 6042 6043 6044 6045 6046 6047 6048 6049 6050 6051 6052 6053 6054 6055 6056 6057 6058 6059
        raise ValueError("The shape should be a list, tuple or Variable.")

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

    out = helper.create_tmp_variable(x.dtype)
    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
6060 6061 6062 6063 6064 6065 6066 6067 6068 6069


def rank_loss(label, left, right, name=None):
    """
    **Rank loss layer for RankNet**

    RankNet(http://icml.cc/2015/wp-content/uploads/2015/06/icml_ranking.pdf)
    is a pairwise ranking model with a training sample consisting of a pair
    of documents, A and B. Label P indicates whether A is ranked higher than B
    or not:
M
minqiyang 已提交
6070

6071 6072
    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
6073

6074 6075 6076 6077
    Rank loss layer takes three inputs: left (o_i), right (o_j) and
    label (P_{i,j}). The inputs respectively represent RankNet's output scores
    for documents A and B and the value of label P. The following equation
    computes rank loss C_{i,j} from the inputs:
M
minqiyang 已提交
6078

6079 6080 6081 6082 6083
    $$
      C_{i,j} = -\tilde{P_{ij}} * o_{i,j} + \log(1 + e^{o_{i,j}}) \\
      o_{i,j} =  o_i - o_j  \\
      \tilde{P_{i,j}} = \left \{0, 0.5, 1 \right \} \ or \ \left \{0, 1 \right \}
    $$
M
minqiyang 已提交
6084 6085 6086

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

6087 6088 6089 6090 6091 6092 6093 6094 6095 6096 6097 6098 6099 6100 6101 6102 6103 6104 6105 6106 6107 6108 6109 6110 6111 6112 6113 6114 6115 6116 6117 6118 6119 6120 6121 6122 6123 6124 6125 6126 6127 6128 6129 6130
    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")

    out = helper.create_tmp_variable("float32")

    helper.append_op(
        type='rank_loss',
        inputs={"Label": label,
                "Left": left,
                "Right": right},
        outputs={'Out': out})
    return out
6131 6132


M
minqiyang 已提交
6133 6134
def margin_rank_loss(label, left, right, margin=0.1, name=None):
    """
M
minqiyang 已提交
6135
    Margin Ranking Loss Layer for ranking problem,
M
minqiyang 已提交
6136
    which compares left score and right score passed in.
M
minqiyang 已提交
6137
    The ranking loss can be defined as following equation:
M
minqiyang 已提交
6138 6139 6140 6141 6142 6143

    .. math::

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

    Args:
M
minqiyang 已提交
6144
       label (Variable): Indicates whether the left is ranked higher than the right or not.
M
minqiyang 已提交
6145 6146
       left (Variable): Ranking score for left.
       right (Variable): Ranking score for right.
M
minqiyang 已提交
6147
       margin (float): Indicates the given margin.
M
minqiyang 已提交
6148 6149 6150
       name (str|None): A name for this layer (optional). If set None, the layer
                       will be named automatically.
    Returns:
M
minqiyang 已提交
6151
       Variable: The ranking loss.
M
minqiyang 已提交
6152
    Raises:
M
minqiyang 已提交
6153
       ValueError: Any of label, left, and right is not a Variable.
M
minqiyang 已提交
6154 6155 6156 6157 6158 6159 6160
    Examples:
        .. code-block:: python
           label = fluid.layers.data(name="label", shape=[4, 1], dtype="float32")
           left = fluid.layers.data(name="left", shape=[4, 1], dtype="float32")
           right = fluid.layers.data(name="right", shape=[4, 1], dtype="float32")
           out = fluid.layers.margin_rank_loss(label, left, right)
    """
M
minqiyang 已提交
6161
    helper = LayerHelper('margin_rank_loss', **locals())
M
minqiyang 已提交
6162 6163 6164 6165 6166 6167
    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.")
M
minqiyang 已提交
6168 6169
    out = helper.create_tmp_variable(left.dtype)
    act = helper.create_tmp_variable(left.dtype)
M
minqiyang 已提交
6170 6171 6172 6173 6174 6175 6176 6177 6178 6179 6180
    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
6181 6182 6183 6184 6185 6186 6187 6188 6189 6190 6191 6192 6193 6194
def pad2d(input,
          paddings=[0, 0, 0, 0],
          mode='constant',
          pad_value=0.0,
          data_format="NCHW",
          name=None):
    """
    Pad 2-d images accordding to 'paddings' and 'mode'.
    If mode is 'reflect', paddings[0] and paddings[1] must be no greater
    than height-1. And the width dimension has the same condition.

    Example:

      Given that X is a channel of image from input:
M
minqiyang 已提交
6195

W
whs 已提交
6196 6197
      X = [[1, 2, 3],
           [4, 5, 6]]
M
minqiyang 已提交
6198

W
whs 已提交
6199
      Case 0:
M
minqiyang 已提交
6200

W
whs 已提交
6201 6202 6203
        paddings = [0, 1, 2, 3],
        mode = 'constant'
        pad_value = 0
M
minqiyang 已提交
6204

W
whs 已提交
6205 6206 6207
        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
6208

W
whs 已提交
6209
      Case 1:
M
minqiyang 已提交
6210

W
whs 已提交
6211 6212
        paddings = [0, 1, 2, 1],
        mode = 'reflect'
M
minqiyang 已提交
6213

W
whs 已提交
6214 6215 6216
        Out = [[3, 2, 1, 2, 3, 2]
               [6, 5, 4, 5, 6, 5]
               [3, 2, 1, 2, 3, 2]]
M
minqiyang 已提交
6217

W
whs 已提交
6218
      Case 2:
M
minqiyang 已提交
6219

W
whs 已提交
6220 6221
        paddings = [0, 1, 2, 1],
        mode = 'edge'
M
minqiyang 已提交
6222

W
whs 已提交
6223 6224 6225
        Out = [[1, 1, 1, 2, 3, 3]
               [4, 4, 4, 5, 6, 6]
               [4, 4, 4, 5, 6, 6]]
M
minqiyang 已提交
6226 6227


W
whs 已提交
6228 6229 6230 6231 6232 6233 6234 6235 6236 6237 6238 6239 6240 6241 6242 6243 6244 6245 6246 6247 6248 6249 6250 6251 6252 6253 6254 6255 6256 6257 6258 6259 6260 6261 6262 6263 6264 6265 6266 6267 6268
    Args:
        input (Variable): The input image with [N, C, H, W] format or [N, H, W, C] format.
        paddings (tuple|list): The padding size. If padding is a tuple, it must
            contain four integers, (padding_top, padding_bottom, padding_left, padding_right).
            Default: padding = [0, 0, 0, 0].
        mode (str): Three modes: constant(default), reflect, edge. Default: constant
        pad_value (float32): The value to fill the padded areas in constant mode. Default: 0
        data_format (str): An optional string from: "NHWC", "NCHW". Specify the data format of
                           the input data.
                           Default: "NCHW"
        name (str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.

    Returns:
        Variable: The tensor variable padded accordding to paddings and mode.


    Examples:
        .. code-block:: python

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

    helper = LayerHelper('pad2d', **locals())
    dtype = helper.input_dtype(input_param_name='input')
    out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type='pad2d',
        inputs={'X': input},
        outputs={"Out": out},
        attrs={
            'paddings': paddings,
            'mode': mode,
            'pad_value': pad_value,
            'data_frmat': data_format
        })

    return out


6269 6270 6271 6272 6273 6274 6275 6276 6277 6278 6279 6280 6281 6282 6283 6284 6285 6286 6287 6288 6289 6290 6291 6292 6293 6294 6295 6296 6297 6298 6299 6300 6301 6302 6303 6304 6305 6306 6307 6308 6309 6310 6311 6312 6313 6314 6315 6316 6317 6318 6319 6320 6321 6322 6323 6324 6325 6326 6327 6328 6329 6330 6331 6332 6333 6334 6335 6336 6337 6338 6339 6340 6341 6342 6343 6344 6345 6346 6347 6348 6349 6350 6351 6352 6353 6354 6355 6356 6357 6358 6359 6360 6361 6362 6363 6364 6365 6366 6367 6368 6369 6370 6371 6372 6373 6374 6375 6376 6377 6378 6379 6380 6381 6382 6383 6384 6385 6386 6387 6388 6389 6390 6391 6392 6393 6394 6395 6396 6397 6398 6399 6400 6401 6402 6403 6404 6405 6406 6407 6408 6409 6410
@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}
    """
    helper = LayerHelper('elu', **locals())
    out = helper.create_tmp_variable(dtype=x.dtype)
    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}
    """
    helper = LayerHelper('relu6', **locals())
    out = helper.create_tmp_variable(dtype=x.dtype)
    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}
    """
    helper = LayerHelper('pow', **locals())
    out = helper.create_tmp_variable(dtype=x.dtype)
    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}
    """
    helper = LayerHelper('stanh', **locals())
    out = helper.create_tmp_variable(dtype=x.dtype)
    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}
    """
    helper = LayerHelper('hard_sigmoid', **locals())
    out = helper.create_tmp_variable(dtype=x.dtype)
    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}
    """
    helper = LayerHelper('swish', **locals())
    out = helper.create_tmp_variable(dtype=x.dtype)
    helper.append_op(
        type='swish',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'slope': beta})
    return out


J
jerrywgz 已提交
6411 6412 6413 6414 6415 6416 6417 6418 6419 6420 6421 6422 6423 6424
def prelu(x, mode, param_attr=None, name=None):
    """
    Equation:

        y = \max(0, x) + alpha \min(0, x)

    Args:
        x (Variable): The input tensor.
	  param_attr(ParamAttr|None): The parameter attribute for the learnable
                                    weight (alpha).
        mode (string): The mode for weight sharing
		       all: all elements share same weight
 		       channel:elements in a channel share same weight
 		       element:each element has a weight
W
whs 已提交
6425
	name(str|None): A name for this layer(optional). If set None, the layer
M
minqiyang 已提交
6426
                        will be named automatically.
J
jerrywgz 已提交
6427 6428 6429 6430 6431 6432 6433 6434 6435 6436 6437 6438 6439 6440 6441 6442 6443 6444 6445 6446 6447 6448 6449 6450 6451 6452 6453 6454 6455 6456 6457 6458 6459 6460 6461 6462 6463

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

    Examples:

        .. code-block:: python

         x = fluid.layers.data(name="x", shape=[10,10], dtype="float32")
            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(
        attr=param_attr,
        shape=alpha_shape,
        dtype='float32',
        is_bias=False,
        default_initializer=Constant(1.0))
    out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type="prelu",
        inputs={"X": x,
                'Alpha': alpha},
        attrs={"mode": mode},
        outputs={"Out": out})
    return out


6464 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 6501 6502 6503 6504 6505 6506 6507 6508 6509 6510 6511 6512 6513 6514 6515 6516 6517 6518 6519 6520 6521 6522 6523 6524 6525 6526 6527 6528 6529 6530 6531
@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): ${t_max_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
     Returns:
        output(${out_type}): ${out_comment}
    """
    helper = LayerHelper('brelu', **locals())
    out = helper.create_tmp_variable(dtype=x.dtype)
    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): ${alpha_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
     Returns:
        output(${out_type}): ${out_comment}
    """
    helper = LayerHelper('leaky_relu', **locals())
    out = helper.create_tmp_variable(dtype=x.dtype)
    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): ${threshold_comment}
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
     Returns:
        output(${out_type}): ${out_comment}
    """
    helper = LayerHelper('soft_relu', **locals())
    out = helper.create_tmp_variable(dtype=x.dtype)
    helper.append_op(
        type='soft_relu',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'threshold': threshold})
    return out


6532 6533 6534 6535 6536 6537 6538 6539 6540 6541 6542 6543 6544
def flatten(x, axis=1, name=None):
    """
    **Flatten layer**
    Flattens the input tensor into a 2D matrix.

    Examples:
    Case 1:
      Given
        X.shape = (3, 100, 100, 4)
      and
        axis = 2
      We get:
        Out.shape = (3 * 100, 4 * 100)
6545

6546 6547 6548 6549 6550 6551 6552 6553 6554 6555
    Case 2:
      Given
        X.shape = (3, 100, 100, 4)
      and
        axis = 0
      We get:
        Out.shape = (1, 3 * 100 * 100 * 4)

    Args:
        x (Variable): A tensor of rank >= axis.
6556 6557
        axis (int): Indicate up to which input dimensions (exclusive) should
                    be flattened to the outer dimension of the output.
6558 6559 6560 6561 6562 6563 6564 6565 6566 6567 6568 6569 6570 6571 6572
                    The value for axis must be in the range [0, R], where R
                    is the rank of the input tensor. When axis = 0, the shape
                    of the output tensor is (1, (d_0 X d_1 ... d_n), where the
                    shape of the input tensor is (d_0, d_1, ... d_n).
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        Variable: A 2D tensor with the contents of the input tensor, with input
                  dimensions up to axis flattened to the outer dimension of
                  the output and remaining input dimensions flattened into the
                  inner dimension of the output.

    Raises:
        ValueError: If x is not a variable.
6573
        ValueError: If axis is not in range [0, rank(x)].
6574 6575 6576 6577 6578 6579 6580 6581 6582 6583 6584 6585 6586 6587 6588 6589 6590

    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)]")

    out = helper.create_tmp_variable(x.dtype)
6591
    x_shape = helper.create_tmp_variable(x.dtype)
6592
    helper.append_op(
6593
        type='flatten2',
6594
        inputs={"X": x},
6595 6596
        outputs={'Out': out,
                 'XShape': x_shape},
6597 6598
        attrs={"axis": axis})
    return out
X
Xin Pan 已提交
6599 6600


C
chenweihang 已提交
6601
def sequence_enumerate(input, win_size, pad_value=0, name=None):
C
chenweihang 已提交
6602
    """
C
chenweihang 已提交
6603
    Generate a new sequence for the input index sequence, which enumerates all the
M
minqiyang 已提交
6604
    sub-sequences with length `win_size` of the input.
C
chenweihang 已提交
6605 6606
    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
6607

C
chenweihang 已提交
6608 6609 6610 6611
    Examples:
    Case 1:
      Input:
        X.lod = [[0, 3, 5]]
6612
        X.data = [[1], [2], [3], [4], [5]]
C
chenweihang 已提交
6613 6614 6615 6616 6617 6618
        X.dims = [5, 1]
      Attrs:
        win_size = 2
        pad_value = 0
      Output:
        Out.lod = [[0, 3, 5]]
6619
        Out.data = [[1, 2], [2, 3], [3, 0], [4, 5], [5, 0]]
C
chenweihang 已提交
6620 6621 6622
        Out.dims = [5, 2]

    Args:
C
chenweihang 已提交
6623 6624 6625
        input (Variable): The input variable which is a index sequence.
        win_size (int): The window size for enumerating all sub-sequences.
        pad_value (int): The padding value, default 0.
C
chenweihang 已提交
6626 6627 6628 6629 6630 6631 6632 6633 6634 6635 6636

    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())
C
chenweihang 已提交
6637
    out = helper.create_tmp_variable(helper.input_dtype(), stop_gradient=True)
C
chenweihang 已提交
6638 6639 6640 6641 6642 6643
    helper.append_op(
        type='sequence_enumerate',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={'win_size': win_size,
               'pad_value': pad_value})
M
minqiyang 已提交
6644
    return out
6645

6646

S
sneaxiy 已提交
6647 6648 6649 6650 6651 6652 6653 6654 6655
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:
6656

S
sneaxiy 已提交
6657
    .. math::
6658

S
sneaxiy 已提交
6659 6660 6661
        y(i_1, i_2,..., i_n, j) = (j < x(i_1, i_2,..., i_n))

    Args:
6662
        x (Variable): Input tensor of sequence_mask layer,
S
sneaxiy 已提交
6663 6664 6665 6666
                      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.
6667 6668 6669
        name (str|None): A name for this layer(optional). If set None, the
                         layer will be named automatically.

S
sneaxiy 已提交
6670 6671
    Returns:
        Variable: The output sequence mask.
6672

S
sneaxiy 已提交
6673 6674
    """

Q
qingqing01 已提交
6675
    helper = LayerHelper('sequence_mask', **locals())
S
sneaxiy 已提交
6676 6677 6678 6679 6680
    if name is None:
        out = helper.create_tmp_variable(dtype=dtype)
    else:
        out = helper.create_tmp_variable(dtype=dtype, name=name)

Q
qingqing01 已提交
6681 6682 6683
    helper.append_op(
        type='sequence_mask',
        inputs={'X': [x]},
S
sneaxiy 已提交
6684 6685
        outputs={'Y': out},
        attrs={
6686
            'maxlen': maxlen if maxlen is not None else -1,
S
sneaxiy 已提交
6687 6688 6689
            'out_dtype': out.dtype
        })
    return out
S
sneaxiy 已提交
6690 6691


X
Xin Pan 已提交
6692
def stack(x, axis=0):
S
sneaxiy 已提交
6693 6694 6695 6696
    """
    **Stack Layer**

    This layer stacks all of the input :code:`x` along axis.
6697 6698 6699 6700 6701 6702 6703

    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
6704
    If :code:`axis` < 0, it would be replaced with :code:`axis+rank(x[0])+1`.
6705
    If :code:`axis` is None, it would be replaced with 0.
S
sneaxiy 已提交
6706 6707

    Args:
6708
        x (Variable|list(Variable)|tuple(Variable)): Input variables.
S
sneaxiy 已提交
6709
        axis (int|None): The axis along which all inputs are stacked.
6710

S
sneaxiy 已提交
6711 6712
    Returns:
        Variable: The stacked variable.
6713

S
sneaxiy 已提交
6714 6715
    """

X
Xin Pan 已提交
6716 6717 6718 6719 6720 6721 6722 6723
    helper = LayerHelper('stack', **locals())
    axis = 0 if axis is None else axis

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

    out = helper.create_tmp_variable(x[0].dtype)
    helper.append_op(
S
sneaxiy 已提交
6724 6725
        type='stack', inputs={'X': x}, outputs={'Y': out},
        attrs={'axis': axis})
6726

X
Xin Pan 已提交
6727
    return out
D
dzhwinter 已提交
6728 6729 6730 6731 6732 6733 6734


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
6735

D
dzhwinter 已提交
6736 6737 6738
    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
6739
    raised.
D
dzhwinter 已提交
6740 6741

    Args:
M
minqiyang 已提交
6742
        x (Variable): Input variable.
D
dzhwinter 已提交
6743 6744
        axis (int): The axis along which the input is unstacked.
        num (int|None): The number of output variables.
M
minqiyang 已提交
6745

D
dzhwinter 已提交
6746 6747
    Returns:
        list(Variable): The unstacked variables.
M
minqiyang 已提交
6748

D
dzhwinter 已提交
6749 6750 6751 6752 6753 6754 6755 6756 6757 6758 6759 6760 6761 6762 6763 6764 6765 6766 6767 6768
    """

    helper = LayerHelper('unstack', **locals())
    if num is None:
        if axis is None or x.shape[axis] <= 0:
            raise ValueError('unknown unstack number')
        else:
            num = x.shape[axis]

    outs = []
    for _ in num:
        outs.append(helper.create_tmp_variable(x.dtype))

    helper.append_op(
        type='unstack',
        inputs={'X': [x]},
        outputs={'Y': outs},
        attrs={'axis': axis,
               'num': num})
    return outs
W
whs 已提交
6769 6770 6771 6772 6773 6774 6775 6776 6777 6778 6779 6780


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
6781

W
whs 已提交
6782 6783 6784 6785
                [
                   [[1], [2], [3]],
                   [[4], [5], [6]]
                ]
M
minqiyang 已提交
6786

W
whs 已提交
6787
        Attr(expand_times):  [1, 2, 2]
M
minqiyang 已提交
6788

W
whs 已提交
6789
        Output(Out) is a 3-D tensor with shape [2, 6, 2]:
M
minqiyang 已提交
6790

W
whs 已提交
6791 6792 6793 6794
                [
                    [[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
6795

W
whs 已提交
6796 6797 6798 6799 6800 6801 6802 6803 6804 6805 6806 6807 6808 6809 6810 6811 6812 6813 6814 6815 6816 6817 6818
    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')
    out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type='expand',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'expand_times': expand_times})
    return out
S
sneaxiy 已提交
6819 6820


G
fix  
gongweibao 已提交
6821 6822 6823
from paddle.fluid.framework import convert_np_dtype_to_dtype_


G
gongweibao 已提交
6824
@templatedoc()
G
fix  
gongweibao 已提交
6825 6826 6827 6828 6829 6830 6831 6832 6833
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
6834
    ${comment}
G
fix  
gongweibao 已提交
6835 6836

    Args:
G
gongweibao 已提交
6837 6838 6839
        input (Variable): ${input_comment}
        shape (tuple|list): ${shape_comment}
        input_dim_idx (Int): ${input_dim_idx_comment}
G
gongweibao 已提交
6840
        output_dim_idx (Int): ${output_dim_idx_comment}
G
gongweibao 已提交
6841 6842 6843
        min (Float): ${min_comment}
        max (Float): ${max_comment}
        seed (Int): ${seed_comment}
G
fix  
gongweibao 已提交
6844 6845
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc
    Returns:
G
gongweibao 已提交
6846
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
6847 6848 6849 6850 6851 6852 6853 6854 6855 6856 6857 6858 6859 6860 6861 6862 6863 6864 6865 6866 6867

    """

    helper = LayerHelper('uniform_random_batch_size_like', **locals())
    out = helper.create_tmp_variable(dtype)
    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
6868 6869


G
gongweibao 已提交
6870
@templatedoc()
X
Xin Pan 已提交
6871
def gaussian_random(shape, mean=0.0, std=1.0, seed=0, dtype='float32'):
G
fix  
gongweibao 已提交
6872
    """
G
gongweibao 已提交
6873
    ${comment}
G
fix  
gongweibao 已提交
6874 6875

    Args:
G
gongweibao 已提交
6876 6877 6878 6879
        shape (tuple|list): ${shape_comment}
        mean (Float): ${mean_comment}
        std (Float): ${std_comment}
        seed (Int): ${seed_comment}
G
fix  
gongweibao 已提交
6880 6881 6882
        dtype(np.dtype|core.VarDesc.VarType|str): Output data type.

    Returns:
G
gongweibao 已提交
6883
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
6884 6885 6886 6887 6888 6889 6890 6891 6892 6893 6894 6895 6896 6897 6898

    """

    helper = LayerHelper('gaussian_random', **locals())
    out = helper.create_tmp_variable(dtype)
    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
6899
            'use_mkldnn': False
G
fix  
gongweibao 已提交
6900 6901 6902 6903 6904
        })

    return out


G
gongweibao 已提交
6905
@templatedoc()
G
fix  
gongweibao 已提交
6906
def sampling_id(x, min=0.0, max=1.0, seed=0, dtype='float32'):
G
fix  
gongweibao 已提交
6907
    """
G
gongweibao 已提交
6908
    ${comment}
G
fix  
gongweibao 已提交
6909 6910

    Args:
G
gongweibao 已提交
6911 6912 6913 6914
        x (Variable): ${x_comment}
        min (Float): ${min_comment}
        max (Float): ${max_comment}
        seed (Float): ${seed_comment}
G
fix  
gongweibao 已提交
6915
        dtype(np.dtype|core.VarDesc.VarType|str): The type of output data : float32, float_16, int etc
G
fix  
gongweibao 已提交
6916 6917

    Returns:
G
gongweibao 已提交
6918
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
6919 6920 6921 6922

    """

    helper = LayerHelper('sampling_id', **locals())
G
fix  
gongweibao 已提交
6923
    out = helper.create_tmp_variable(dtype)
G
fix  
gongweibao 已提交
6924 6925 6926 6927 6928 6929 6930 6931 6932 6933 6934
    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
6935
@templatedoc()
G
fix  
gongweibao 已提交
6936 6937 6938 6939 6940 6941 6942 6943 6944
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
6945
    ${comment}
G
fix  
gongweibao 已提交
6946 6947

    Args:
G
gongweibao 已提交
6948 6949
        input (Variable): ${input_comment}
        shape (tuple|list): ${shape_comment}
G
gongweibao 已提交
6950
        input_dim_idx (Int): ${input_dim_idx_comment}
G
gongweibao 已提交
6951 6952 6953 6954
        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
6955
        dtype(np.dtype|core.VarDesc.VarType|str): The type of output data : float32, float_16, int etc
G
fix  
gongweibao 已提交
6956 6957

    Returns:
G
gongweibao 已提交
6958
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
6959 6960 6961 6962 6963 6964 6965 6966 6967 6968 6969 6970 6971 6972 6973 6974 6975 6976 6977 6978 6979 6980
    """

    helper = LayerHelper('gaussian_random_batch_size_like', **locals())
    out = helper.create_tmp_variable(dtype)
    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
6981
@templatedoc()
X
Xin Pan 已提交
6982
def sum(x):
G
fix  
gongweibao 已提交
6983
    """
G
gongweibao 已提交
6984
    ${comment}
G
fix  
gongweibao 已提交
6985 6986

    Args:
G
gongweibao 已提交
6987
        x (Variable): ${x_comment}
G
fix  
gongweibao 已提交
6988 6989

    Returns:
G
gongweibao 已提交
6990
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
6991 6992 6993
    """

    helper = LayerHelper('sum', **locals())
G
fix  
gongweibao 已提交
6994
    out = helper.create_tmp_variable(dtype=helper.input_dtype('x'))
G
fix  
gongweibao 已提交
6995 6996 6997 6998
    helper.append_op(
        type='sum',
        inputs={'X': x},
        outputs={'Out': out},
X
Xin Pan 已提交
6999
        attrs={'use_mkldnn': False})
G
fix  
gongweibao 已提交
7000 7001 7002 7003

    return out


G
gongweibao 已提交
7004
@templatedoc()
G
fix  
gongweibao 已提交
7005 7006
def slice(input, axes, starts, ends):
    """
G
gongweibao 已提交
7007
    ${comment}
G
fix  
gongweibao 已提交
7008 7009

    Args:
G
gongweibao 已提交
7010 7011 7012 7013
        input (Variable): ${input_comment}.
        axes (List): ${axes_comment}
        starts (List): ${starts_comment}
        ends (List): ${ends_comment}
G
fix  
gongweibao 已提交
7014 7015

    Returns:
G
gongweibao 已提交
7016
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
7017 7018 7019 7020

    """

    helper = LayerHelper('slice', **locals())
G
fix  
gongweibao 已提交
7021
    out = helper.create_tmp_variable(dtype=helper.input_dtype('input'))
G
fix  
gongweibao 已提交
7022 7023 7024 7025 7026 7027 7028 7029 7030 7031 7032
    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
7033
@templatedoc()
G
fix  
gongweibao 已提交
7034 7035
def shape(input):
    """
G
gongweibao 已提交
7036
    ${comment}
G
fix  
gongweibao 已提交
7037 7038

    Args:
G
gongweibao 已提交
7039
        input (Variable): ${input_comment}
G
fix  
gongweibao 已提交
7040 7041

    Returns:
G
gongweibao 已提交
7042
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
7043 7044 7045 7046

    """

    helper = LayerHelper('shape', **locals())
G
fix  
gongweibao 已提交
7047
    out = helper.create_tmp_variable(dtype=helper.input_dtype('input'))
G
fix  
gongweibao 已提交
7048
    helper.append_op(
G
fix  
gongweibao 已提交
7049
        type='shape', inputs={'Input': input}, outputs={'Out': out})
G
fix  
gongweibao 已提交
7050 7051

    return out
G
merge  
gongweibao 已提交
7052 7053


S
sneaxiy 已提交
7054 7055 7056 7057 7058 7059 7060 7061
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
7062 7063 7064 7065 7066 7067
    name = helper.kwargs.get('name', None)
    if name is None:
        out = helper.create_tmp_variable(dtype=x.dtype)
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)
S
sneaxiy 已提交
7068

S
sneaxiy 已提交
7069 7070 7071 7072 7073 7074 7075 7076 7077 7078 7079
    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
7080
def scale(x, scale=1.0, bias=0.0, bias_after_scale=True, act=None, name=None):
S
sneaxiy 已提交
7081 7082 7083 7084 7085 7086 7087 7088
    """
    ${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
7089
        act(basestring|None): Activation applied to the output.
M
minqiyang 已提交
7090
        name(basestring|None): Name of the output.
S
sneaxiy 已提交
7091 7092 7093 7094 7095 7096

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

    helper = LayerHelper('scale', **locals())
S
sneaxiy 已提交
7097 7098 7099 7100 7101
    if name is None:
        out = helper.create_tmp_variable(dtype=x.dtype)
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)
S
sneaxiy 已提交
7102 7103 7104 7105 7106 7107 7108 7109 7110 7111

    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
7112
    return helper.append_activation(out)
S
sneaxiy 已提交
7113 7114


X
Xin Pan 已提交
7115
def elementwise_add(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
7116 7117 7118
    return _elementwise_op(LayerHelper('elementwise_add', **locals()))


X
Xin Pan 已提交
7119
def elementwise_div(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
7120 7121 7122
    return _elementwise_op(LayerHelper('elementwise_div', **locals()))


X
Xin Pan 已提交
7123
def elementwise_sub(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
7124 7125 7126
    return _elementwise_op(LayerHelper('elementwise_sub', **locals()))


X
Xin Pan 已提交
7127
def elementwise_mul(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
7128 7129 7130
    return _elementwise_op(LayerHelper('elementwise_mul', **locals()))


X
Xin Pan 已提交
7131
def elementwise_max(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
7132 7133 7134
    return _elementwise_op(LayerHelper('elementwise_max', **locals()))


X
Xin Pan 已提交
7135
def elementwise_min(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
7136 7137 7138
    return _elementwise_op(LayerHelper('elementwise_min', **locals()))


X
Xin Pan 已提交
7139
def elementwise_pow(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
7140 7141 7142 7143 7144 7145 7146 7147 7148 7149 7150
    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
7151 7152
            "act (basestring|None): Activation applied to the output.",
            "name (basestring|None): Name of the output."
S
sneaxiy 已提交
7153
        ])
M
minqiyang 已提交
7154 7155


7156
def _logical_op(op_name, x, y, out=None, name=None, binary_op=True):
M
minqiyang 已提交
7157 7158
    helper = LayerHelper(op_name, **locals())

M
minqiyang 已提交
7159 7160
    if binary_op:
        assert x.dtype == y.dtype
M
minqiyang 已提交
7161 7162 7163 7164 7165 7166 7167 7168 7169 7170 7171 7172 7173 7174 7175 7176 7177 7178 7179

    if out is None:
        if name is None:
            out = helper.create_tmp_variable(dtype=x.dtype)
        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()
7180
def logical_and(x, y, out=None, name=None):
M
minqiyang 已提交
7181 7182 7183 7184 7185 7186 7187 7188 7189 7190 7191 7192 7193 7194 7195 7196 7197 7198
    """
    ${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}
    """

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


@templatedoc()
7199
def logical_or(x, y, out=None, name=None):
M
minqiyang 已提交
7200 7201 7202 7203 7204 7205 7206 7207 7208 7209 7210 7211 7212 7213 7214 7215 7216 7217
    """
    ${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}
    """

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


@templatedoc()
7218
def logical_xor(x, y, out=None, name=None):
M
minqiyang 已提交
7219 7220 7221 7222 7223 7224 7225 7226 7227 7228 7229 7230 7231 7232 7233 7234 7235 7236
    """
    ${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}
    """

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


@templatedoc()
7237
def logical_not(x, out=None, name=None):
M
minqiyang 已提交
7238 7239 7240 7241 7242 7243 7244 7245 7246 7247 7248 7249 7250 7251
    """
    ${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}
    """

    return _logical_op(
        op_name="logical_not", x=x, y=None, name=name, out=out, binary_op=False)
7252 7253 7254 7255 7256 7257 7258 7259 7260 7261 7262 7263 7264 7265 7266 7267 7268 7269 7270 7271 7272 7273 7274 7275 7276 7277 7278 7279 7280 7281 7282 7283 7284 7285 7286 7287 7288 7289 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


@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}
    """

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

    if name is None:
        out = helper.create_tmp_variable(dtype=x.dtype)
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    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}
    """

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

    if name is None:
        out = helper.create_tmp_variable(dtype=x.dtype)
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    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
7316 7317 7318 7319 7320 7321 7322 7323 7324 7325 7326 7327 7328 7329 7330 7331 7332 7333 7334 7335 7336 7337 7338 7339 7340 7341 7342 7343 7344 7345 7346 7347 7348 7349 7350 7351 7352 7353 7354 7355 7356 7357 7358 7359 7360 7361 7362 7363 7364 7365 7366 7367 7368 7369 7370 7371 7372 7373


@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:
        out = helper.create_tmp_variable(dtype=x.dtype)
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    helper.append_op(
        type="mean", inputs={"X": x}, attrs={}, outputs={"Out": out})

    return out


@templatedoc()
def mul(x, y, x_num_col_dims=1, y_num_col_dims=1, name=None):
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        y(${y_type}): ${y_comment}
        x_num_col_dims(${x_num_col_dims_type}): ${x_num_col_dims_comment}
        y_num_col_dims(${y_num_col_dims_type}): ${y_num_col_dims_comment}
        name(basestring|None): Name of the output.

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

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

    if name is None:
        out = helper.create_tmp_variable(dtype=x.dtype)
    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
7374 7375
            "x_num_col_dims": x_num_col_dims,
            "y_num_col_dims": y_num_col_dims
X
Xin Pan 已提交
7376 7377 7378 7379 7380 7381 7382 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
        },
        outputs={"Out": out})
    return out


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

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

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

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

    if name is None:
        out = helper.create_tmp_variable(dtype=x.dtype)
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    helper.append_op(
        type="sigmoid_cross_entropy_with_logits",
        inputs={"X": x,
                "Label": label},
        attrs={},
        outputs={"Out": out})
    return out


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

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

    Returns:
        out(${out_type}): ${out_comment}
    """
    helper = LayerHelper("maxout", **locals())

    if name is None:
        out = helper.create_tmp_variable(dtype=x.dtype)
    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
S
sneaxiy 已提交
7439 7440 7441 7442 7443 7444 7445 7446 7447 7448 7449 7450 7451 7452 7453 7454 7455 7456 7457 7458 7459 7460 7461 7462 7463 7464 7465


@templatedoc()
def sequence_reverse(x, name=None):
    """ 
    ${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:
        out = helper.create_tmp_variable(dtype=x.dtype)
    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
7466 7467


7468 7469 7470 7471 7472 7473 7474 7475 7476 7477 7478 7479 7480 7481 7482 7483 7484 7485 7486 7487 7488 7489 7490 7491 7492 7493 7494 7495 7496 7497 7498 7499 7500 7501 7502 7503 7504 7505 7506
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.
    
    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:
        out = helper.create_tmp_variable(dtype=x.dtype)
    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