nn.py 244.6 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 59 60 61 62 63 64 65 66 67 68 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152
    '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',
    'lstm_unit',
    'reduce_sum',
    'reduce_mean',
    'reduce_max',
    'reduce_min',
    'reduce_prod',
    'sequence_first_step',
    'sequence_last_step',
    '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',
    'dice_loss',
    'image_resize',
    'image_resize_short',
    'resize_bilinear',
    'gather',
    'scatter',
    'sequence_scatter',
    'random_crop',
    'mean_iou',
    'relu',
    'log',
    'crop',
    'rank_loss',
    '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',
Y
Yu Yang 已提交
153 154 155 156 157 158 159 160 161
]


def fc(input,
       size,
       num_flatten_dims=1,
       param_attr=None,
       bias_attr=None,
       act=None,
J
Jacek Czaja 已提交
162
       is_test=False,
163
       name=None):
Y
Yu Yang 已提交
164
    """
165
    **Fully Connected Layer**
Y
Yu Yang 已提交
166

167 168 169 170 171 172 173 174
    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
175
    to the output as well.
C
caoying03 已提交
176

C
caoying03 已提交
177
    This process can be formulated as follows:
178 179 180

    .. math::

181
        Out = Act({\sum_{i=0}^{N-1}X_iW_i + b})
182 183 184

    In the above equation:

C
caoying03 已提交
185 186 187 188
    * :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).
189
    * :math:`Act`: The activation function.
C
caoying03 已提交
190
    * :math:`Out`: The output tensor.
Y
Yu Yang 已提交
191 192

    Args:
R
ranqiu 已提交
193 194 195 196 197 198 199 200 201 202 203 204 205 206 207
        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
208 209
            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
210
        act (str, default None): Activation to be applied to the output of this layer.
J
Jacek Czaja 已提交
211
        is_test(bool): A flag indicating whether execution is in test phase.
R
ranqiu 已提交
212
        name (str, default None): The name of this layer.
Y
Yu Yang 已提交
213

214
    Returns:
F
fengjiayi 已提交
215
        Variable: The transformation result.
216 217

    Raises:
C
caoying03 已提交
218
        ValueError: If rank of the input tensor is less than 2.
219 220 221 222

    Examples:
        .. code-block:: python

F
fengjiayi 已提交
223
          data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
224
          fc = fluid.layers.fc(input=data, size=1000, act="tanh")
Y
Yu Yang 已提交
225
    """
C
caoying03 已提交
226

C
caoying03 已提交
227
    helper = LayerHelper("fc", **locals())
Y
Yu Yang 已提交
228 229 230 231

    dtype = helper.input_dtype()

    mul_results = []
232 233
    for input_var, param_attr in helper.iter_inputs_and_params():
        input_shape = input_var.shape
Y
Yu Yang 已提交
234 235 236
        param_shape = [
            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
        ] + [size]
Y
ying 已提交
237

Y
Yu Yang 已提交
238
        w = helper.create_parameter(
239 240
            attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False)
        tmp = helper.create_tmp_variable(dtype)
241
        helper.append_op(
242 243 244
            type="mul",
            inputs={"X": input_var,
                    "Y": w},
245
            outputs={"Out": tmp},
M
mozga-intel 已提交
246 247
            attrs={"x_num_col_dims": num_flatten_dims,
                   "y_num_col_dims": 1})
248 249 250 251
        mul_results.append(tmp)

    if len(mul_results) == 1:
        pre_bias = mul_results[0]
252
    else:
253 254
        pre_bias = helper.create_tmp_variable(dtype)
        helper.append_op(
255 256 257
            type="sum",
            inputs={"X": mul_results},
            outputs={"Out": pre_bias},
X
Xin Pan 已提交
258
            attrs={"use_mkldnn": False})
259 260 261 262
    # 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
263 264


265 266 267
def embedding(input,
              size,
              is_sparse=False,
268
              is_distributed=False,
269 270 271
              padding_idx=None,
              param_attr=None,
              dtype='float32'):
Y
Yu Yang 已提交
272
    """
273 274
    **Embedding Layer**

275
    This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in
276 277
    a lookup table. The result of this lookup is the embedding of each ID in the
    :attr:`input`.
278 279 280

    All the input variables are passed in as local variables to the LayerHelper
    constructor.
Y
Yu Yang 已提交
281 282

    Args:
283 284 285 286 287
        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.
288
        is_distributed(bool): Whether to run lookup table from remote parameter server.
289 290
        padding_idx(int|long|None): If :attr:`None`, it makes no effect to lookup.
            Otherwise the given :attr:`padding_idx` indicates padding the output
291
            with zeros whenever lookup encounters it in :attr:`input`. If
292
            :math:`padding_idx < 0`, the :attr:`padding_idx` to use in lookup is
293 294
            :math:`size[0] + dim`.
        param_attr(ParamAttr): Parameters for this layer
295
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc
Y
Yu Yang 已提交
296

297 298 299
    Returns:
        Variable: The tensor variable storing the embeddings of the \
                  supplied inputs.
Y
Yu Yang 已提交
300

301 302
    Examples:
        .. code-block:: python
Y
Yu Yang 已提交
303

C
chengduoZH 已提交
304
          dict_size = len(dataset.ids)
305
          data = fluid.layers.data(name='ids', shape=[32, 32], dtype='float32')
C
chengduoZH 已提交
306
          fc = fluid.layers.embedding(input=data, size=[dict_size, 16])
Y
Yu Yang 已提交
307 308 309 310 311 312
    """

    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)
313 314
    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
315 316 317 318 319
    helper.append_op(
        type='lookup_table',
        inputs={'Ids': input,
                'W': w},
        outputs={'Out': tmp},
320 321 322 323 324
        attrs={
            'is_sparse': is_sparse,
            'is_distributed': is_distributed,
            'padding_idx': padding_idx
        })
Y
Yu Yang 已提交
325 326 327
    return tmp


Y
yi.wu 已提交
328
@templatedoc(op_type="lstm")
Y
Yu Yang 已提交
329 330
def dynamic_lstm(input,
                 size,
Y
Yancey 已提交
331 332
                 h_0=None,
                 c_0=None,
Y
Yu Yang 已提交
333 334 335 336 337 338 339
                 param_attr=None,
                 bias_attr=None,
                 use_peepholes=True,
                 is_reverse=False,
                 gate_activation='sigmoid',
                 cell_activation='tanh',
                 candidate_activation='tanh',
340 341
                 dtype='float32',
                 name=None):
Y
Yibing Liu 已提交
342
    """
Y
yi.wu 已提交
343
    ${comment}
Y
Yibing Liu 已提交
344 345

    Args:
Y
yi.wu 已提交
346 347
        input (Variable): ${input_comment}
        size (int): 4 * hidden size.
Y
Yancey 已提交
348 349 350 351 352 353 354
        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.

355
        param_attr(ParamAttr|None): The parameter attribute for the learnable
356
                               hidden-hidden weights.
Y
Yibing Liu 已提交
357 358 359

                               - Weights = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}
360 361
                               - The shape is (D x 4D), where D is the hidden
                                 size.
Y
yi.wu 已提交
362
        bias_attr (ParamAttr|None): The bias attribute for the learnable bias
363 364 365
                              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
366

367
                              1. `use_peepholes = False`
Y
yi.wu 已提交
368 369
                                 - Biases = {:math:`b_c, b_i, b_f, b_o`}.
                                 - The shape is (1 x 4D).
370
                              2. `use_peepholes = True`
Y
yi.wu 已提交
371
                                 - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
Y
Yibing Liu 已提交
372
                                                 W_{fc}, W_{oc}`}.
Y
yi.wu 已提交
373
                                 - The shape is (1 x 7D).
Y
yi.wu 已提交
374 375 376 377 378 379 380 381
        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
382 383

    Returns:
Y
Yibing Liu 已提交
384 385
        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
386

Y
Yibing Liu 已提交
387
    Examples:
Y
Yibing Liu 已提交
388 389
        .. code-block:: python

Y
Yibing Liu 已提交
390 391
            hidden_dim = 512
            forward_proj = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
392
                                           act=None, bias_attr=None)
Y
Yibing Liu 已提交
393 394
            forward, _ = fluid.layers.dynamic_lstm(
                input=forward_proj, size=hidden_dim * 4, use_peepholes=False)
Y
Yibing Liu 已提交
395
    """
396

Y
Yu Yang 已提交
397
    helper = LayerHelper('lstm', **locals())
M
minqiyang 已提交
398
    size = size // 4
Y
Yu Yang 已提交
399 400 401 402 403 404 405 406 407 408 409 410
    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
411 412 413 414 415 416 417 418 419 420
    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
421 422 423

    helper.append_op(
        type='lstm',
Y
Yancey 已提交
424
        inputs=inputs,
Y
Yu Yang 已提交
425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440
        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
441 442 443 444 445 446 447 448 449 450 451
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',
452 453
                  dtype='float32',
                  name=None):
Y
Yibing Liu 已提交
454 455 456
    """
    **Dynamic LSTMP Layer**

457 458 459 460 461 462
    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
463 464 465 466 467

    The formula is as follows:

    .. math::

468
        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
469

470
        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
471

472
        \\tilde{c_t} & = act_g(W_{cx}x_t + W_{cr}r_{t-1} + b_c)
Y
Yibing Liu 已提交
473

474
        o_t & = \sigma(W_{ox}x_{t} + W_{or}r_{t-1} + W_{oc}c_t + b_o)
Y
Yibing Liu 已提交
475

476
        c_t & = f_t \odot c_{t-1} + i_t \odot \\tilde{c_t}
Y
Yibing Liu 已提交
477

478
        h_t & = o_t \odot act_h(c_t)
Y
Yibing Liu 已提交
479

480
        r_t & = \overline{act_h}(W_{rh}h_t)
Y
Yibing Liu 已提交
481

Y
Yibing Liu 已提交
482 483 484 485 486 487
    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, \
488
          we use vectors to reprenset these diagonal weight matrices.
Y
Yibing Liu 已提交
489
    * :math:`b`: Denotes bias vectors (e.g. :math:`b_i` is the input gate \
490
          bias vector).
Y
Yibing Liu 已提交
491 492 493
    * :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 \
494
          the same size as the cell output activation vector :math:`h`.
Y
Yibing Liu 已提交
495
    * :math:`h`: The hidden state.
496
    * :math:`r`: The recurrent projection of the hidden state.
Y
Yibing Liu 已提交
497 498
    * :math:`\\tilde{c_t}`: The candidate hidden state, whose \
          computation is based on the current input and previous hidden state.
499
    * :math:`\odot`: The element-wise product of the vectors.
Y
Yibing Liu 已提交
500
    * :math:`act_g` and :math:`act_h`: The cell input and cell output \
501
          activation functions and `tanh` is usually used for them.
Y
Yibing Liu 已提交
502 503
    * :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
504 505 506 507

    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.
508

Y
Yibing Liu 已提交
509 510 511 512 513 514 515 516 517 518 519 520
    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.
521
        param_attr(ParamAttr|None): The parameter attribute for the learnable
Y
Yibing Liu 已提交
522 523
                               hidden-hidden weight and projection weight.

524 525
                               - Hidden-hidden weight = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}.
526 527
                               - 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
528 529
                                 size.
                               - Projection weight = {:math:`W_{rh}`}.
530 531
                               - The shape of projection weight is (D x P).
        bias_attr(ParamAttr|None): The bias attribute for the learnable bias
Y
Yibing Liu 已提交
532 533 534 535 536 537
                              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`}.
538
                                - The shape is (1 x 4D).
Y
Yibing Liu 已提交
539 540 541
                              2. `use_peepholes = True`
                                - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
                                                 W_{fc}, W_{oc}`}.
542
                                - The shape is (1 x 7D).
Y
Yibing Liu 已提交
543 544 545 546 547 548 549 550 551
        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.
552
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
Y
Yibing Liu 已提交
553 554
                              default "tanh".
        proj_activation(str): The activation for projection output.
555
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
Y
Yibing Liu 已提交
556 557
                              default "tanh".
        dtype(str): Data type. Choices = ["float32", "float64"], default "float32".
558 559
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
Y
Yibing Liu 已提交
560 561

    Returns:
562 563 564 565
        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
566 567

    Examples:
568

Y
Yibing Liu 已提交
569 570
        .. code-block:: python

571 572 573 574
            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
575
            hidden_dim, proj_dim = 512, 256
576
            fc_out = fluid.layers.fc(input=emb, size=hidden_dim * 4,
Y
Yibing Liu 已提交
577
                                     act=None, bias_attr=None)
578 579 580
            proj_out, _ = fluid.layers.dynamic_lstmp(input=fc_out,
                                                     size=hidden_dim * 4,
                                                     proj_size=proj_dim,
Y
Yibing Liu 已提交
581 582 583 584
                                                     use_peepholes=False,
                                                     is_reverse=True,
                                                     cell_activation="tanh",
                                                     proj_activation="tanh")
Y
Yibing Liu 已提交
585
    """
586

Y
Yibing Liu 已提交
587
    helper = LayerHelper('lstmp', **locals())
M
minqiyang 已提交
588
    size = size // 4
Y
Yibing Liu 已提交
589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632
    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
633 634 635 636 637 638 639 640 641
def dynamic_gru(input,
                size,
                param_attr=None,
                bias_attr=None,
                is_reverse=False,
                gate_activation='sigmoid',
                candidate_activation='tanh',
                h_0=None):
    """
642
    **Gated Recurrent Unit (GRU) Layer**
G
guosheng 已提交
643

644
    Refer to `Empirical Evaluation of Gated Recurrent Neural Networks on
645
    Sequence Modeling <https://arxiv.org/abs/1412.3555>`_ .
646

G
guosheng 已提交
647 648 649 650 651 652 653 654 655
    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)
656

G
guosheng 已提交
657
        h_t & = (1-u_t) \odot h_{t-1} + u_t \odot \\tilde{h_t}
658

G
guosheng 已提交
659
    The :math:`\odot` is the element-wise product of the vectors. :math:`act_g`
660 661
    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
662 663 664 665
    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
666
    to use fully-connect layer before GRU layer.
G
guosheng 已提交
667 668

    Args:
669 670
        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
671
            Variable is a matrix with shape :math:`(T \\times 3D)`, where
672
            :math:`T` is the total time steps in this mini-batch, :math:`D`
G
guosheng 已提交
673 674
            is the hidden size.
        size(int): The dimension of the gru cell.
675
        param_attr(ParamAttr|None): The parameter attribute for the learnable
G
guosheng 已提交
676 677
            hidden-hidden weight matrix. Note:

678
            - The shape of the weight matrix is :math:`(T \\times 3D)`, where
G
guosheng 已提交
679
              :math:`D` is the hidden size.
680
            - All elements in the weight matrix can be divided into two parts.
G
guosheng 已提交
681
              The first part are weights of the update gate and reset gate with
682
              shape :math:`(D \\times 2D)`, and the second part are weights for
G
guosheng 已提交
683
              candidate hidden state with shape :math:`(D \\times D)`.
684
        bias_attr(ParamAttr): The parameter attribute for learnable the
G
guosheng 已提交
685
            hidden-hidden bias.
686
        is_reverse(bool): Whether to compute reversed GRU, default
G
guosheng 已提交
687 688 689
            :attr:`False`.
        gate_activation(str): The activation for update gate and reset gate.
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "sigmoid".
690
        candidate_activation(str): The activation for candidate hidden state.
G
guosheng 已提交
691
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh".
692 693 694 695
        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
696 697

    Returns:
G
guosheng 已提交
698
        Variable: The hidden state of GRU. The shape is :math:`(T \\times D)`, \
699
            and sequence length is the same with the input.
700

G
guosheng 已提交
701
    Examples:
702

G
guosheng 已提交
703 704
        .. code-block:: python

705 706 707 708
            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
709
            hidden_dim = 512
710
            x = fluid.layers.fc(input=emb, size=hidden_dim * 3)
G
guosheng 已提交
711 712 713 714 715 716 717 718 719 720
            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
721
    batch_size = input.shape[0]
G
guosheng 已提交
722 723 724
    inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
    if h_0 != None:
        assert h_0.shape == (
Y
Yancey 已提交
725 726 727
            batch_size, size
        ), 'The shape of h0 should be(batch_size, %d)' % size
        inputs['H0'] = h_0
G
guosheng 已提交
728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750

    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
751 752 753
def gru_unit(input,
             hidden,
             size,
754 755
             param_attr=None,
             bias_attr=None,
Y
Yu Yang 已提交
756
             activation='tanh',
757
             gate_activation='sigmoid'):
Y
Yu Yang 已提交
758
    """
759
    GRU unit layer. The equation of a gru step is:
Y
Yu Yang 已提交
760

761 762
        .. math::
            u_t & = actGate(xu_{t} + W_u h_{t-1} + b_u)
Y
Yu Yang 已提交
763

764
            r_t & = actGate(xr_{t} + W_r h_{t-1} + b_r)
Y
Yu Yang 已提交
765

766
            m_t & = actNode(xm_t + W_c dot(r_t, h_{t-1}) + b_m)
767

768
            h_t & = dot((1-u_t), m_t) + dot(u_t, h_{t-1})
769 770

    The inputs of gru unit includes :math:`z_t`, :math:`h_{t-1}`. In terms
771 772 773
    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
774 775
    connected layer has to be applied, such that :math:`z_t = W_{fc}x_t`.

776 777
    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
778 779 780
    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`.
781 782 783 784 785

    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.
786 787
        param_attr (ParamAttr): The weight parameters for gru unit. Default: None
        bias_attr (ParamAttr): The bias parameters for gru unit. Default: None
788 789 790 791
        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
792

793 794 795 796 797 798
    Returns:
        tuple: The hidden value, reset-hidden value and gate values.

    Examples:

        .. code-block:: python
Y
Yu Yang 已提交
799

800
             # assuming we have x_t_data and prev_hidden of size=10
801
             x_t = fluid.layers.fc(input=x_t_data, size=30)
802 803
             hidden_val, r_h_val, gate_val = fluid.layers.gru_unit(input=x_t,
                                                    hidden = prev_hidden)
Y
Yu Yang 已提交
804 805 806 807 808 809 810 811 812 813 814 815

    """
    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
816
    size = size // 3
Y
Yu Yang 已提交
817 818

    # create weight
819 820
    weight = helper.create_parameter(
        attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype)
Y
Yu Yang 已提交
821

822 823 824 825
    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
826
    # create bias
827
    if helper.bias_attr:
Y
Yu Yang 已提交
828 829 830
        bias_size = [1, 3 * size]
        bias = helper.create_parameter(
            attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
831
        inputs['Bias'] = bias
Y
Yu Yang 已提交
832 833 834

    helper.append_op(
        type='gru_unit',
835
        inputs=inputs,
Y
Yu Yang 已提交
836 837 838 839 840 841
        outputs={
            'Gate': gate,
            'ResetHiddenPrev': reset_hidden_pre,
            'Hidden': updated_hidden,
        },
        attrs={
842 843
            'activation': 2,  # tanh
            'gate_activation': 1,  # sigmoid
Y
Yu Yang 已提交
844 845 846 847 848
        })

    return updated_hidden, reset_hidden_pre, gate


Y
yuyang18 已提交
849
@templatedoc()
850
def linear_chain_crf(input, label, param_attr=None):
Y
yuyang18 已提交
851 852 853 854 855 856 857
    """
    Linear Chain CRF.

    ${comment}

    Args:
        input(${emission_type}): ${emission_comment}
D
dzhwinter 已提交
858
        input(${transition_type}): ${transition_comment}
Y
yuyang18 已提交
859 860 861 862
        label(${label_type}): ${label_comment}
        param_attr(ParamAttr): The attribute of the learnable parameter.

    Returns:
D
dzhwinter 已提交
863 864 865
        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
866 867

    """
Y
Yu Yang 已提交
868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892
    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
893
@templatedoc()
894
def crf_decoding(input, param_attr, label=None):
Y
yuyang18 已提交
895 896 897 898 899
    """
    ${comment}

    Args:
        input(${emission_type}): ${emission_comment}
Y
yi.wu 已提交
900

Y
yuyang18 已提交
901
        param_attr(ParamAttr): The parameter attribute for training.
Y
yi.wu 已提交
902

Y
yuyang18 已提交
903 904 905
        label(${label_type}): ${label_comment}

    Returns:
Y
update  
yi.wu 已提交
906
        Variable: ${viterbi_path_comment}
907

Y
yi.wu 已提交
908 909 910 911 912
    Examples:
        .. code-block:: python

           crf_decode = layers.crf_decoding(
                input=hidden, param_attr=ParamAttr(name="crfw"))
Y
yuyang18 已提交
913
    """
Y
Yu Yang 已提交
914 915 916 917 918 919 920 921 922 923 924 925 926
    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
927
@templatedoc()
F
fengjiayi 已提交
928
def cos_sim(X, Y):
Y
Yu Yang 已提交
929
    """
Y
yi.wu 已提交
930 931 932
    ${comment}

    Args:
933 934
        X (Variable): ${x_comment}.
        Y (Variable): ${y_comment}.
F
fengjiayi 已提交
935

Y
yi.wu 已提交
936
    Returns:
937
        Variable: the output of cosine(X, Y).
Y
Yu Yang 已提交
938
    """
F
fengjiayi 已提交
939
    helper = LayerHelper('cos_sim', **locals())
Y
Yu Yang 已提交
940 941 942 943 944 945 946 947 948 949 950 951 952
    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


953
def dropout(x, dropout_prob, is_test=False, seed=None, name=None):
954 955 956 957 958
    """
    Computes dropout.

    Drop or keep each element of `x` independently. Dropout is a regularization
    technique for reducing overfitting by preventing neuron co-adaption during
959
    training. The dropout operator randomly sets (according to the given dropout
960 961 962 963
    probability) the outputs of some units to zero, while others are remain
    unchanged.

    Args:
964 965
        x (Variable): The input tensor variable.
        dropout_prob (float): Probability of setting units to zero.
966 967 968 969 970 971 972
        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.
973 974

    Returns:
975
        Variable: A tensor variable is the shape with `x`.
976 977

    Examples:
978

979 980
        .. code-block:: python

981 982
            x = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
            droped = fluid.layers.dropout(x, dropout_prob=0.5)
983 984
    """

F
fengjiayi 已提交
985
    helper = LayerHelper('dropout', **locals())
986 987
    out = helper.create_tmp_variable(dtype=x.dtype)
    mask = helper.create_tmp_variable(dtype=x.dtype, stop_gradient=True)
C
chengduo 已提交
988 989 990 991

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

992 993 994 995 996
    helper.append_op(
        type='dropout',
        inputs={'X': [x]},
        outputs={'Out': [out],
                 'Mask': [mask]},
997 998 999 1000 1001 1002
        attrs={
            'dropout_prob': dropout_prob,
            'is_test': is_test,
            'fix_seed': seed is not None,
            'seed': seed if seed is not None else 0
        })
1003 1004 1005
    return out


1006
def cross_entropy(input, label, soft_label=False, ignore_index=-100):
Y
Yu Yang 已提交
1007
    """
Y
Yibing Liu 已提交
1008 1009
    **Cross Entropy Layer**

1010 1011 1012
    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
1013 1014

    1) One-hot cross-entropy:
F
fengjiayi 已提交
1015
        `soft_label = False`, `Label[i, 0]` indicates the class index for sample i:
Y
yangyaming 已提交
1016

Y
Yibing Liu 已提交
1017
        .. math::
Y
yangyaming 已提交
1018

Y
Yibing Liu 已提交
1019 1020 1021
            Y[i] = -\log(X[i, Label[i]])

    2) Soft-label cross-entropy:
F
fengjiayi 已提交
1022 1023
        `soft_label = True`, `Label[i, j]` indicates the soft label of class j
        for sample i:
Y
Yibing Liu 已提交
1024 1025 1026 1027 1028

        .. math::

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

Y
Yibing Liu 已提交
1029
       Please make sure that in this case the summation of each row of `label`
Y
Yibing Liu 已提交
1030 1031 1032
       equals one.

    3) One-hot cross-entropy with vecterized `label`:
F
fengjiayi 已提交
1033 1034
         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
1035
         to a one-hot cross-entropy with one-hot label representation.
Y
yangyaming 已提交
1036

Y
Yibing Liu 已提交
1037
    Args:
Y
yangyaming 已提交
1038
        input (Variable|list):  a 2-D tensor with shape [N x D], where N is the
1039 1040 1041 1042
                                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
1043
        label (Variable|list): the ground truth which is a 2-D tensor. When
1044 1045 1046 1047
                               `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
1048
        soft_label (bool): a flag indicating whether to
1049
                                           interpretate the given labels as soft
1050
                                           labels. Default: `False`.
M
minqiyang 已提交
1051 1052
        ignore_index (int): Specifies a target value that is ignored and does
                            not contribute to the input gradient. Only valid
1053
                            if soft_label is set to False. Default: -100
Y
Yibing Liu 已提交
1054 1055 1056 1057 1058

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

    Raises:
1059 1060 1061 1062 1063
        `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
1064 1065 1066 1067 1068 1069

    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
1070
    """
F
fengjiayi 已提交
1071
    helper = LayerHelper('cross_entropy', **locals())
Y
Yu Yang 已提交
1072 1073 1074 1075 1076 1077
    out = helper.create_tmp_variable(dtype=input.dtype)
    helper.append_op(
        type='cross_entropy',
        inputs={'X': [input],
                'Label': [label]},
        outputs={'Y': [out]},
1078 1079
        attrs={"soft_label": soft_label,
               "ignore_index": ignore_index})
Y
Yu Yang 已提交
1080 1081 1082
    return out


F
fengjiayi 已提交
1083
def square_error_cost(input, label):
Y
Yu Yang 已提交
1084
    """
1085 1086
    **Square error cost layer**

1087 1088
    This layer accepts input predictions and target label and returns the
    squared error cost.
Y
ying 已提交
1089

1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102
    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:
1103 1104
        input (Variable): Input tensor, has predictions.
        label (Variable): Label tensor, has target labels.
1105 1106

    Returns:
G
guosheng 已提交
1107
        Variable: The tensor variable storing the element-wise squared error \
1108
                  difference of input and label.
1109 1110 1111 1112 1113 1114 1115 1116

    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
1117
    """
F
fengjiayi 已提交
1118
    helper = LayerHelper('square_error_cost', **locals())
Y
Yu Yang 已提交
1119 1120 1121 1122 1123 1124 1125 1126 1127
    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
1128 1129
        type='square', inputs={'X': [minus_out]},
        outputs={'Out': [square_out]})
Y
Yu Yang 已提交
1130 1131 1132
    return square_out


Y
yi.wu 已提交
1133
@templatedoc()
Y
Yu Yang 已提交
1134 1135 1136 1137
def chunk_eval(input,
               label,
               chunk_scheme,
               num_chunk_types,
F
fengjiayi 已提交
1138
               excluded_chunk_types=None):
Y
Yu Yang 已提交
1139
    """
Y
yi.wu 已提交
1140
    **Chunk Evaluator**
Y
yi.wu 已提交
1141

Y
yangyaming 已提交
1142
    This function computes and outputs the precision, recall and
1143
    F1-score of chunk detection.
Y
yi.wu 已提交
1144

Y
yi.wu 已提交
1145 1146 1147 1148 1149 1150 1151 1152
    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
1153

Y
yi.wu 已提交
1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178
       ====== ====== ======  =====  ==  ============   =====  ===== =====  ==  =========
              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
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
       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
1204
    Args:
1205 1206 1207 1208 1209
        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
1210

Y
yi.wu 已提交
1211
    Returns:
Y
update  
yi.wu 已提交
1212 1213 1214
        tuple: tuple containing: precision, recall, f1_score,
        num_infer_chunks, num_label_chunks,
        num_correct_chunks
1215

Y
yi.wu 已提交
1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227
    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
1228
    """
F
fengjiayi 已提交
1229
    helper = LayerHelper("chunk_eval", **locals())
Y
Yu Yang 已提交
1230 1231 1232 1233 1234

    # prepare output
    precision = helper.create_tmp_variable(dtype="float32")
    recall = helper.create_tmp_variable(dtype="float32")
    f1_score = helper.create_tmp_variable(dtype="float32")
1235 1236 1237
    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
1238 1239 1240 1241 1242 1243 1244 1245

    helper.append_op(
        type="chunk_eval",
        inputs={"Inference": [input],
                "Label": [label]},
        outputs={
            "Precision": [precision],
            "Recall": [recall],
1246 1247 1248 1249
            "F1-Score": [f1_score],
            "NumInferChunks": [num_infer_chunks],
            "NumLabelChunks": [num_label_chunks],
            "NumCorrectChunks": [num_correct_chunks]
Y
Yu Yang 已提交
1250 1251 1252
        },
        attrs={
            "num_chunk_types": num_chunk_types,
G
guosheng 已提交
1253 1254
            "chunk_scheme": chunk_scheme,
            "excluded_chunk_types": excluded_chunk_types or []
Y
Yu Yang 已提交
1255
        })
1256 1257
    return (precision, recall, f1_score, num_infer_chunks, num_label_chunks,
            num_correct_chunks)
Y
Yu Yang 已提交
1258 1259


1260
@templatedoc()
Y
Yu Yang 已提交
1261 1262 1263 1264 1265 1266 1267
def sequence_conv(input,
                  num_filters,
                  filter_size=3,
                  filter_stride=1,
                  padding=None,
                  bias_attr=None,
                  param_attr=None,
1268
                  act=None):
Y
Yu Yang 已提交
1269 1270 1271 1272
    """
    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.
1273 1274 1275 1276 1277 1278 1279 1280 1281 1282

    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.
        bias_attr (ParamAttr|None): attributes for bias
        param_attr (ParamAttr|None): attributes for parameter
        act (str): the activation type
F
fengjiayi 已提交
1283

1284 1285
    Returns:
        Variable: output of sequence_conv
Y
Yu Yang 已提交
1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303
    """

    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
1304
            'contextStart': -int(filter_size // 2),
Y
Yu Yang 已提交
1305 1306 1307 1308 1309 1310
            'contextLength': filter_size
        })
    pre_act = helper.append_bias_op(pre_bias)
    return helper.append_activation(pre_act)


1311
def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=False):
1312 1313 1314
    """
    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
1315
    input Tensor can be either :math:`[N, 1]` or :math:`[N]`, where :math:`N`
1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333
    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.
        bias_attr (ParamAttr|None): attributes for bias
        param_attr (ParamAttr|None): attributes for parameter
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \
1334
        library is installed. Default: False
1335

1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346
    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)
    """
1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357
    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


1358
def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True, name=None):
Q
qiaolongfei 已提交
1359
    """
1360
    The input of the softmax operator is a tensor of any rank. The output tensor
F
fengjiayi 已提交
1361
    has the same shape as the input.
Q
qiaolongfei 已提交
1362

1363 1364 1365 1366 1367 1368
    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
1369
    K-dimensional vector of real values in the range [0, 1] that add up to 1.
Q
qiaolongfei 已提交
1370 1371 1372 1373 1374 1375 1376

    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
1377
    For each row :math:`i` and each column :math:`j` in the matrix, we have:
Q
qiaolongfei 已提交
1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400

    .. math::

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

    Args:
        input (Variable): The input variable.
        bias_attr (ParamAttr): attributes for bias
        param_attr (ParamAttr): attributes for parameter
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \
        library is installed.

    Returns:
        Variable: output of softmax

    Examples:

        .. code-block:: python

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

    """
1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411
    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
1412 1413 1414
def conv2d(input,
           num_filters,
           filter_size,
C
chengduoZH 已提交
1415 1416
           stride=1,
           padding=0,
1417
           dilation=1,
Y
Yu Yang 已提交
1418 1419 1420
           groups=None,
           param_attr=None,
           bias_attr=None,
C
chengduoZH 已提交
1421
           use_cudnn=True,
1422 1423
           act=None,
           name=None):
Y
Yu Yang 已提交
1424
    """
C
chengduoZH 已提交
1425
    The convolution2D layer calculates the output based on the input, filter
T
tensor-tang 已提交
1426 1427
    and strides, paddings, dilations, groups parameters. Input and
    Output are in NCHW format, where N is batch size, C is the number of
1428
    channels, H is the height of the feature, and W is the width of the feature.
T
tensor-tang 已提交
1429 1430 1431 1432 1433 1434 1435
    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.
1436 1437 1438
    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
1439

1440
    For each input :math:`X`, the equation is:
C
refine  
chengduoZH 已提交
1441

C
chengduoZH 已提交
1442 1443
    .. math::

C
refine  
chengduoZH 已提交
1444
        Out = \sigma (W \\ast X + b)
C
chengduoZH 已提交
1445

T
tensor-tang 已提交
1446
    Where:
C
chengduoZH 已提交
1447

1448 1449 1450 1451 1452
    * :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
1453
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
1454 1455 1456

    Example:

1457 1458
        - Input:

W
weixing02 已提交
1459
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
C
refine  
chengduoZH 已提交
1460

W
weixing02 已提交
1461
          Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)`
C
refine  
chengduoZH 已提交
1462

1463
        - Output:
T
tensor-tang 已提交
1464

W
weixing02 已提交
1465
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
C
refine  
chengduoZH 已提交
1466

C
chengduoZH 已提交
1467
        Where
1468 1469

        .. math::
C
chengduoZH 已提交
1470

W
weixing02 已提交
1471 1472
            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
1473 1474

    Args:
1475
        input (Variable): The input image with [N, C, H, W] format.
T
tensor-tang 已提交
1476
        num_filters(int): The number of filter. It is as same as the output
1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501
            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
            connected to the second half of the input channels. Default: groups=1
        param_attr (ParamAttr): The parameters to the Conv2d Layer. Default: None
        bias_attr (ParamAttr): Bias parameter for the Conv2d layer. Default: None
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
        act (str): Activation type. Default: None
        name (str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.
C
chengduoZH 已提交
1502 1503

    Returns:
G
guosheng 已提交
1504
        Variable: The tensor variable storing the convolution and \
C
chengduoZH 已提交
1505 1506
                  non-linearity activation result.

C
refine  
chengduoZH 已提交
1507
    Raises:
1508 1509
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
C
refine  
chengduoZH 已提交
1510

C
chengduoZH 已提交
1511 1512 1513
    Examples:
        .. code-block:: python

1514 1515
          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
1516 1517 1518
    """

    num_channels = input.shape[1]
1519 1520

    l_type = 'conv2d'
X
xzl 已提交
1521 1522
    if (num_channels == groups and num_filters % num_channels == 0 and
            not use_cudnn):
1523
        l_type = 'depthwise_conv2d'
1524 1525 1526 1527

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

Y
Yu Yang 已提交
1528 1529 1530 1531 1532
    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
1533
        num_filter_channels = num_channels // groups
Y
Yu Yang 已提交
1534

C
chengduoZH 已提交
1535 1536 1537
    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')
1538
    dilation = utils.convert_to_list(dilation, 2, 'dilation')
C
chengduoZH 已提交
1539

C
chengduoZH 已提交
1540 1541
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1542 1543

    input_shape = input.shape
M
minqiyang 已提交
1544
    filter_shape = [num_filters, int(num_filter_channels)] + filter_size
Y
Yu Yang 已提交
1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558

    def _get_default_param_initializer():
        std = (2.0 / (filter_size[0]**2 * num_channels))**0.5
        return Normal(0.0, std, 0)

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

    pre_bias = helper.create_tmp_variable(dtype)

    helper.append_op(
1559
        type=l_type,
Y
Yu Yang 已提交
1560 1561 1562 1563 1564
        inputs={
            'Input': input,
            'Filter': filter_param,
        },
        outputs={"Output": pre_bias},
C
chengduoZH 已提交
1565 1566 1567
        attrs={
            'strides': stride,
            'paddings': padding,
1568
            'dilations': dilation,
C
chengduoZH 已提交
1569
            'groups': groups,
1570
            'use_cudnn': use_cudnn,
X
Xin Pan 已提交
1571
            'use_mkldnn': False
C
chengduoZH 已提交
1572
        })
Y
Yu Yang 已提交
1573 1574 1575 1576 1577 1578

    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
1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595
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
1596 1597 1598 1599 1600 1601
    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
1602 1603 1604 1605 1606 1607 1608 1609 1610

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

    .. math::

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

    In the above equation:

1611 1612
    * :math:`X`: Input value, a tensor with NCDHW format.
    * :math:`W`: Filter value, a tensor with MCDHW format.
C
chengduoZH 已提交
1613 1614 1615
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
1616
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641

    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,
1642
            it must contain three integers, (filter_size_D, filter_size_H, filter_size_W).
C
chengduoZH 已提交
1643 1644
            Otherwise, the filter will be a square.
        stride (int|tuple): The stride size. If stride is a tuple, it must
1645
            contain three integers, (stride_D, stride_H, stride_W). Otherwise, the
C
chengduoZH 已提交
1646 1647
            stride_D = stride_H = stride_W = stride. Default: stride = 1.
        padding (int|tuple): The padding size. If padding is a tuple, it must
1648
            contain three integers, (padding_D, padding_H, padding_W). Otherwise, the
C
chengduoZH 已提交
1649 1650
            padding_D = padding_H = padding_W = padding. Default: padding = 0.
        dilation (int|tuple): The dilation size. If dilation is a tuple, it must
1651
            contain three integers, (dilation_D, dilation_H, dilation_W). Otherwise, the
C
chengduoZH 已提交
1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676
            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
        param_attr (ParamAttr): The parameters to the Conv3d Layer. Default: None
        bias_attr (ParamAttr): Bias parameter for the Conv3d layer. Default: None
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
        act (str): Activation type. Default: None
        name (str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.

    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

1677 1678
          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
1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692
    """

    l_type = 'conv3d'

    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
1693
        num_filter_channels = num_channels // groups
C
chengduoZH 已提交
1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730

    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():
        std = (2.0 / (filter_size[0]**3 * num_channels))**0.5
        return Normal(0.0, std, 0)

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

    pre_bias = helper.create_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
1731
            'use_mkldnn': False
C
chengduoZH 已提交
1732 1733
        })

1734
    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
C
chengduoZH 已提交
1735 1736 1737 1738

    return helper.append_activation(pre_act)


F
fengjiayi 已提交
1739
def sequence_pool(input, pool_type):
Y
Yu Yang 已提交
1740
    """
Y
yangyaming 已提交
1741 1742 1743
    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
1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754

    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:
1755
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
1756 1757 1758 1759 1760
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
1761
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
1762 1763 1764 1765 1766 1767 1768

       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)
1769 1770
         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
1771

L
Luo Tao 已提交
1772 1773
    Args:
        input(variable): The input variable which is a LoDTensor.
Y
yangyaming 已提交
1774
        pool_type (string): The pooling type of sequence_pool.
L
Luo Tao 已提交
1775 1776 1777 1778 1779 1780 1781 1782
            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
1783

Y
yangyaming 已提交
1784
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1785 1786 1787 1788 1789
                              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')
1790 1791
             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
1792
    """
F
fengjiayi 已提交
1793
    helper = LayerHelper('sequence_pool', **locals())
Y
Yu Yang 已提交
1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804
    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
1805 1806 1807 1808 1809
    # 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
1810 1811 1812
    return pool_out


C
add doc  
chengduoZH 已提交
1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837
@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
1838
def sequence_first_step(input):
L
Luo Tao 已提交
1839
    """
L
Luo Tao 已提交
1840
    This function gets the first step of sequence.
L
Luo Tao 已提交
1841 1842 1843 1844

    .. code-block:: text

       x is a 1-level LoDTensor:
1845
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
1846 1847 1848 1849 1850
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
1851
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
1852
         out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1)
F
fengjiayi 已提交
1853

L
Luo Tao 已提交
1854 1855 1856 1857 1858 1859 1860 1861 1862
    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
1863

Y
yangyaming 已提交
1864
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1865 1866 1867
                              dtype='float32', lod_level=1)
             x_first_step = fluid.layers.sequence_first_step(input=x)
    """
1868 1869 1870
    return sequence_pool(input=input, pool_type="first")


F
fengjiayi 已提交
1871
def sequence_last_step(input):
L
Luo Tao 已提交
1872
    """
L
Luo Tao 已提交
1873
    This function gets the last step of sequence.
L
Luo Tao 已提交
1874 1875 1876 1877

    .. code-block:: text

       x is a 1-level LoDTensor:
1878
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
1879 1880 1881 1882 1883
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
1884
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
1885
         out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1)
F
fengjiayi 已提交
1886

L
Luo Tao 已提交
1887 1888 1889 1890 1891 1892 1893 1894 1895
    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
1896

Y
yangyaming 已提交
1897
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1898 1899 1900
                              dtype='float32', lod_level=1)
             x_last_step = fluid.layers.sequence_last_step(input=x)
    """
1901 1902 1903
    return sequence_pool(input=input, pool_type="last")


F
fengjiayi 已提交
1904
@templatedoc()
Y
Yu Yang 已提交
1905
def pool2d(input,
C
chengduoZH 已提交
1906 1907
           pool_size=-1,
           pool_type="max",
C
chengduoZH 已提交
1908 1909
           pool_stride=1,
           pool_padding=0,
C
caoying03 已提交
1910
           global_pooling=False,
C
chengduoZH 已提交
1911
           use_cudnn=True,
1912
           ceil_mode=False,
C
caoying03 已提交
1913
           name=None):
Y
Yu Yang 已提交
1914
    """
F
fengjiayi 已提交
1915
    ${comment}
1916 1917

    Args:
1918 1919 1920
        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
1921
                          feature, and W is the width of the feature.
1922
        pool_size (int): The side length of pooling windows. All pooling
F
fengjiayi 已提交
1923
                         windows are squares with pool_size on a side.
F
fengjiayi 已提交
1924
        pool_type: ${pooling_type_comment}
1925 1926
        pool_stride (int): stride of the pooling layer.
        pool_padding (int): padding size.
F
fengjiayi 已提交
1927 1928 1929
        global_pooling: ${global_pooling_comment}
        use_cudnn: ${use_cudnn_comment}
        ceil_mode: ${ceil_mode_comment}
1930
        name (str|None): A name for this layer(optional). If set None, the
F
fengjiayi 已提交
1931 1932
                        layer will be named automatically.

1933
    Returns:
F
fengjiayi 已提交
1934
        Variable: The pooling result.
F
fengjiayi 已提交
1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947

    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(
1948 1949 1950 1951
                            input=data,
                            pool_size=2,
                            pool_type='max',
                            pool_stride=1,
F
fengjiayi 已提交
1952
                            global_pooling=False)
Y
Yu Yang 已提交
1953 1954 1955 1956 1957
    """
    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
1958

C
chengduoZH 已提交
1959 1960 1961 1962 1963
    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
1964 1965 1966 1967
    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
1968 1969
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1970

C
Add doc  
chengduoZH 已提交
1971
    l_type = 'pool2d'
1972 1973

    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
1974 1975 1976 1977
    dtype = helper.input_dtype()
    pool_out = helper.create_tmp_variable(dtype)

    helper.append_op(
1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988
        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
1989
            "use_mkldnn": False
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005
        })

    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
2006
    pooling configurations mentioned in input parameters.
2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018

    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.
2019

2020
    Returns:
2021
        Variable: output of pool3d layer.
Y
Yu Yang 已提交
2022 2023 2024 2025 2026
    """
    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
2027

C
chengduoZH 已提交
2028 2029 2030 2031 2032
    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))

2033 2034 2035
    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
2036

C
chengduoZH 已提交
2037 2038
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
2039

2040 2041
    l_type = "pool3d"
    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
2042 2043 2044 2045
    dtype = helper.input_dtype()
    pool_out = helper.create_tmp_variable(dtype)

    helper.append_op(
2046
        type=l_type,
Y
Yu Yang 已提交
2047 2048 2049 2050 2051 2052 2053
        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
2054
            "paddings": pool_padding,
2055
            "use_cudnn": use_cudnn,
2056
            "ceil_mode": ceil_mode,
X
Xin Pan 已提交
2057
            "use_mkldnn": False
Y
Yu Yang 已提交
2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069
        })

    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
2070
               data_layout='NCHW',
Y
Yang Yang 已提交
2071
               in_place=False,
2072 2073
               name=None,
               moving_mean_name=None,
W
wanghaoshuang 已提交
2074
               moving_variance_name=None,
2075 2076
               do_model_average_for_mean_and_var=False,
               fuse_with_relu=False):
Y
Yu Yang 已提交
2077
    """
Q
qiaolongfei 已提交
2078 2079 2080 2081
    **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
2082

Q
qiaolongfei 已提交
2083
    1. NHWC `[batch, in_height, in_width, in_channels]`
Q
qiaolongfei 已提交
2084

Q
qiaolongfei 已提交
2085 2086
    2. NCHW `[batch, in_channels, in_height, in_width]`

Q
qiaolongfei 已提交
2087 2088 2089
    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
2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101

    :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
2102 2103

    Args:
Q
qiaolongfei 已提交
2104
        input(variable): The input variable which is a LoDTensor.
Q
qiaolongfei 已提交
2105 2106 2107 2108
        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):
Q
qiaolongfei 已提交
2109 2110 2111
        param_attr(ParamAttr): The parameter attribute for Parameter `scale`.
        bias_attr(ParamAttr): The parameter attribute for Parameter `bias`.
        data_layout(string, default NCHW): NCHW|NHWC
2112
        in_place(bool, Default False): Make the input and output of batch norm reuse memory.
Q
qiaolongfei 已提交
2113 2114 2115 2116
        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
2117
        do_model_average_for_mean_and_var(bool, Default False): Do model average for mean and variance or not.
2118
        fuse_with_relu (bool): if True, this OP performs relu after batch norm.
2119 2120

    Returns:
Q
qiaolongfei 已提交
2121
        Variable: A tensor variable which is the result after applying batch normalization on the input.
Q
qiaolongfei 已提交
2122 2123 2124 2125 2126 2127 2128

    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
2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151
    """
    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(
2152
        attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
Y
Yu Yang 已提交
2153

2154 2155
    mean = helper.create_parameter(
        attr=ParamAttr(
W
wanghaoshuang 已提交
2156 2157 2158
            name=moving_mean_name,
            initializer=Constant(0.0),
            trainable=False,
W
wanghaoshuang 已提交
2159
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
2160
        shape=param_shape,
2161 2162 2163 2164 2165 2166 2167
        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
2168
            trainable=False,
W
wanghaoshuang 已提交
2169
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
2170
        shape=param_shape,
2171 2172
        dtype=input.dtype)
    variance.stop_gradient = True
Y
Yu Yang 已提交
2173 2174 2175 2176 2177 2178

    # 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
2179 2180
    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
2181

2182
    batch_norm_out = input if in_place else helper.create_tmp_variable(dtype)
Y
Yu Yang 已提交
2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199

    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
        },
2200 2201 2202 2203
        attrs={
            "momentum": momentum,
            "epsilon": epsilon,
            "is_test": is_test,
X
Xin Pan 已提交
2204
            "use_mkldnn": False,
2205
            "fuse_with_relu": fuse_with_relu
2206
        })
Y
Yu Yang 已提交
2207 2208 2209 2210

    return helper.append_activation(batch_norm_out)


Y
yuyang18 已提交
2211
@templatedoc()
G
guosheng 已提交
2212 2213 2214 2215 2216 2217 2218 2219 2220 2221
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
2222
    ${comment}
G
guosheng 已提交
2223 2224 2225

    The formula is as follows:

Y
yuyang18 已提交
2226
    ..  math::
G
guosheng 已提交
2227 2228 2229 2230 2231 2232 2233

        \\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
2234 2235 2236 2237 2238 2239 2240 2241
    * :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
2242

G
guosheng 已提交
2243 2244
    Args:
        input(Variable): The input tensor variable.
2245
        scale(bool): Whether to learn the adaptive gain :math:`g` after
G
guosheng 已提交
2246
            normalization.
2247
        shift(bool): Whether to learn the adaptive bias :math:`b` after
G
guosheng 已提交
2248
            normalization.
2249
        begin_norm_axis(bool): The normalization will be performed along
G
guosheng 已提交
2250
            dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`.
2251
        epsilon(float): The small value added to the variance to prevent
G
guosheng 已提交
2252 2253 2254 2255 2256 2257
            division by zero.
        param_attr(ParamAttr|None): The parameter attribute for the learnable
            gain :math:`g`.
        bias_attr(ParamAttr|None): The parameter attribute for the learnable
            bias :math:`b`.
        act(str): Activation to be applied to the output of layer normalizaiton.
2258
        name (str): The name of this layer. It is optional.
G
guosheng 已提交
2259 2260

    Returns:
Y
yuyang18 已提交
2261
        ${y_comment}
G
guosheng 已提交
2262 2263 2264

    Examples:

Y
yuyang18 已提交
2265 2266 2267
        >>> 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
2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282
    """
    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
2283
    if shift:
G
guosheng 已提交
2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307
        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
2308 2309 2310 2311
def conv2d_transpose(input,
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
2312 2313 2314
                     padding=0,
                     stride=1,
                     dilation=1,
2315
                     groups=None,
C
caoying03 已提交
2316
                     param_attr=None,
2317
                     bias_attr=None,
C
chengduoZH 已提交
2318
                     use_cudnn=True,
2319
                     act=None,
C
caoying03 已提交
2320
                     name=None):
Y
Yu Yang 已提交
2321
    """
2322 2323 2324 2325 2326 2327 2328 2329
    **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
2330 2331
    layer, please refer to the following explanation and references
    `therein <http://www.matthewzeiler.com/wp-content/uploads/2017/07/cvpr2010.pdf>`_.
2332 2333 2334
    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.
2335 2336 2337 2338 2339

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

    .. math::

2340
        Out = \sigma (W \\ast X + b)
2341

2342
    Where:
2343 2344 2345

    * :math:`X`: Input value, a tensor with NCHW format.
    * :math:`W`: Filter value, a tensor with MCHW format.
2346 2347 2348 2349
    * :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
2350

2351 2352 2353 2354
    Example:

        - Input:

2355
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
2356

2357
          Filter shape: :math:`(C_{in}, C_{out}, H_f, W_f)`
2358 2359 2360

        - Output:

2361
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
2362 2363

        Where
Y
Yu Yang 已提交
2364

2365 2366
        .. math::

2367 2368 2369 2370
           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
2371 2372

    Args:
2373 2374 2375 2376
        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
2377 2378 2379 2380
            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.
2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407
        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.
            Default: groups=1
        param_attr(ParamAttr): The parameters to the Conv2d_transpose Layer.
                               Default: None
        bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None
        use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
        act(str): Activation type. Default: None
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.
Y
Yu Yang 已提交
2408 2409

    Returns:
2410
        Variable: The tensor variable storing the convolution transpose result.
2411 2412

    Raises:
2413 2414
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
2415 2416 2417 2418

    Examples:
       .. code-block:: python

2419 2420
          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
2421
    """
2422 2423 2424 2425 2426 2427 2428 2429 2430

    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
2431 2432 2433
    if not isinstance(input, Variable):
        raise TypeError("Input of conv2d_transpose must be Variable")

C
chengduoZH 已提交
2434 2435 2436
    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
2437

C
chengduoZH 已提交
2438 2439
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
G
guosheng 已提交
2440

Y
Yu Yang 已提交
2441 2442 2443 2444 2445
    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
2446

Y
Yu Yang 已提交
2447 2448
        h_in = input.shape[2]
        w_in = input.shape[3]
G
guosheng 已提交
2449

C
chengduoZH 已提交
2450
        filter_size_h = (output_size[0] - (h_in - 1) * stride[0] + 2 *
M
minqiyang 已提交
2451
                         padding[0] - 1) // dilation[0] + 1
C
chengduoZH 已提交
2452
        filter_size_w = (output_size[1] - (w_in - 1) * stride[1] + 2 *
M
minqiyang 已提交
2453
                         padding[1] - 1) // dilation[1] + 1
Y
Yu Yang 已提交
2454
        filter_size = [filter_size_h, filter_size_w]
C
chengduoZH 已提交
2455 2456 2457
    else:
        filter_size = utils.convert_to_list(filter_size, 2,
                                            'conv2d_transpose.filter_size')
2458 2459 2460 2461 2462 2463 2464
    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')
2465
    groups = 1 if groups is None else groups
M
minqiyang 已提交
2466
    filter_shape = [input_channel, num_filters // groups] + filter_size
Y
Yu Yang 已提交
2467 2468 2469
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

2470
    pre_bias = helper.create_tmp_variable(dtype=input.dtype)
Y
Yu Yang 已提交
2471
    helper.append_op(
2472
        type=op_type,
Y
Yu Yang 已提交
2473 2474
        inputs={'Input': [input],
                'Filter': [img_filter]},
2475
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
2476
        attrs={
2477
            'output_size': output_size,
2478 2479 2480 2481 2482
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn
Y
Yu Yang 已提交
2483 2484
        })

2485 2486 2487
    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
2488 2489


2490
def conv3d_transpose(input,
Y
Yu Yang 已提交
2491 2492 2493
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
2494 2495 2496
                     padding=0,
                     stride=1,
                     dilation=1,
2497
                     groups=None,
C
caoying03 已提交
2498
                     param_attr=None,
2499
                     bias_attr=None,
C
chengduoZH 已提交
2500
                     use_cudnn=True,
2501
                     act=None,
C
caoying03 已提交
2502
                     name=None):
Y
Yu Yang 已提交
2503
    """
2504
    **Convlution3D transpose layer**
2505

2506
    The convolution3D transpose layer calculates the output based on the input,
2507
    filter, and dilations, strides, paddings. Input(Input) and output(Output)
2508 2509 2510 2511 2512 2513
    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>`_.
2514 2515 2516
    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.
2517 2518 2519 2520 2521

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

    .. math::

2522
        Out = \sigma (W \\ast X + b)
2523 2524 2525

    In the above equation:

2526 2527
    * :math:`X`: Input value, a tensor with NCDHW format.
    * :math:`W`: Filter value, a tensor with MCDHW format.
2528 2529 2530 2531
    * :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
2532

2533 2534 2535 2536
    Example:

        - Input:

2537
          Input shape: :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})`
2538

2539
          Filter shape: :math:`(C_{in}, C_{out}, D_f, H_f, W_f)`
2540 2541 2542

        - Output:

2543
          Output shape: :math:`(N, C_{out}, D_{out}, H_{out}, W_{out})`
2544 2545

        Where
Y
Yu Yang 已提交
2546

2547 2548
        .. math::

2549 2550 2551
           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
2552 2553

    Args:
2554
        input(Variable): The input image with [N, C, D, H, W] format.
2555 2556 2557
        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
2558
            tuple, it must contain three integers, (image_D, image_H, image_W). This
2559 2560
            parameter only works when filter_size is None.
        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
2561
            it must contain three integers, (filter_size_D, filter_size_H, filter_size_W).
2562 2563 2564
            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
2565 2566
            contain three integers, (padding_D, padding_H, padding_W). Otherwise, the
            padding_D = padding_H = padding_W = padding. Default: padding = 0.
2567
        stride(int|tuple): The stride size. If stride is a tuple, it must
2568 2569
            contain three integers, (stride_D, stride_H, stride_W). Otherwise, the
            stride_D = stride_H = stride_W = stride. Default: stride = 1.
2570
        dilation(int|tuple): The dilation size. If dilation is a tuple, it must
2571 2572 2573
            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
2574 2575 2576 2577 2578
            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
2579 2580 2581
        param_attr(ParamAttr): The parameters to the Conv3d_transpose Layer.
            Default: None
        bias_attr(ParamAttr): Bias parameter for the Conv3d layer. Default: None
2582 2583 2584 2585 2586
        use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
        act(str): Activation type. Default: None
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.
Y
Yu Yang 已提交
2587 2588

    Returns:
2589
        Variable: The tensor variable storing the convolution transpose result.
2590 2591

    Raises:
2592 2593
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
2594 2595 2596 2597

    Examples:
       .. code-block:: python

2598 2599
          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
2600
    """
2601 2602
    l_type = "conv3d_transpose"
    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
2603
    if not isinstance(input, Variable):
2604
        raise TypeError("Input of conv3d_transpose must be Variable")
Y
Yu Yang 已提交
2605 2606
    input_channel = input.shape[1]

2607 2608 2609
    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
2610

C
chengduoZH 已提交
2611 2612 2613
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")

Y
Yu Yang 已提交
2614 2615 2616 2617 2618 2619
    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]

2620 2621 2622
        d_in = input.shape[2]
        h_in = input.shape[3]
        w_in = input.shape[4]
C
chengduoZH 已提交
2623

2624
        filter_size_d = (output_size[0] - (d_in - 1) * stride[0] + 2 *
M
minqiyang 已提交
2625
                         padding[0] - 1) // dilation[0] + 1
2626
        filter_size_h = (output_size[1] - (h_in - 1) * stride[1] + 2 *
M
minqiyang 已提交
2627
                         padding[1] - 1) // dilation[1] + 1
2628
        filter_size_w = (output_size[2] - (w_in - 1) * stride[2] + 2 *
M
minqiyang 已提交
2629
                         padding[2] - 1) // dilation[2] + 1
2630
        filter_size = [filter_size_d, filter_size_h, filter_size_w]
C
chengduoZH 已提交
2631
    else:
2632 2633
        filter_size = utils.convert_to_list(filter_size, 3,
                                            'conv3d_transpose.filter_size')
Y
Yu Yang 已提交
2634

2635
    groups = 1 if groups is None else groups
M
minqiyang 已提交
2636
    filter_shape = [input_channel, num_filters // groups] + filter_size
Y
Yu Yang 已提交
2637 2638 2639
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

2640
    pre_bias = helper.create_tmp_variable(dtype=input.dtype)
Y
Yu Yang 已提交
2641
    helper.append_op(
2642
        type=l_type,
Y
Yu Yang 已提交
2643 2644
        inputs={'Input': [input],
                'Filter': [img_filter]},
2645
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
2646 2647 2648 2649
        attrs={
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
2650
            'groups': groups,
C
chengduoZH 已提交
2651 2652
            'use_cudnn': use_cudnn
        })
Y
Yu Yang 已提交
2653

2654 2655
    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
2656
    return out
Y
yangyaming 已提交
2657 2658


Y
yangyaming 已提交
2659
def sequence_expand(x, y, ref_level=-1, name=None):
2660
    """Sequence Expand Layer. This layer will expand the input variable **x**
Y
yangyaming 已提交
2661 2662 2663 2664
    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:
2665 2666 2667 2668 2669

    .. code-block:: text

        * Case 1
            x is a LoDTensor:
2670
                x.lod  = [[2,        2]]
Y
yangyaming 已提交
2671
                x.data = [[a], [b], [c], [d]]
2672 2673 2674
                x.dims = [4, 1]

            y is a LoDTensor:
2675 2676
                y.lod = [[2,    2],
                         [3, 3, 1, 1]]
2677

Y
yangyaming 已提交
2678
            ref_level: 0
2679

Y
yangyaming 已提交
2680
            then output is a 1-level LoDTensor:
2681
                out.lod =  [[2,        2,        2,        2]]
Y
yangyaming 已提交
2682
                out.data = [[a], [b], [a], [b], [c], [d], [c], [d]]
2683 2684 2685 2686
                out.dims = [8, 1]

        * Case 2
            x is a Tensor:
Y
yangyaming 已提交
2687
                x.data = [[a], [b], [c]]
2688 2689 2690
                x.dims = [3, 1]

            y is a LoDTensor:
2691
                y.lod = [[2, 0, 3]]
2692

Y
yangyaming 已提交
2693
            ref_level: -1
2694

Y
yangyaming 已提交
2695 2696 2697
            then output is a Tensor:
                out.data = [[a], [a], [c], [c], [c]]
                out.dims = [5, 1]
2698 2699 2700
    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
2701 2702
        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
2703
        name(str|None): A name for this layer(optional). If set None, the layer
Y
yangyaming 已提交
2704
                        will be named automatically.
2705 2706 2707 2708 2709 2710 2711 2712 2713 2714

    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
2715
            out = layers.sequence_expand(x=x, y=y, ref_level=0)
2716
    """
Y
yangyaming 已提交
2717
    helper = LayerHelper('sequence_expand', input=x, **locals())
2718 2719 2720
    dtype = helper.input_dtype()
    tmp = helper.create_tmp_variable(dtype)
    helper.append_op(
Y
yangyaming 已提交
2721 2722 2723 2724 2725
        type='sequence_expand',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': tmp},
        attrs={'ref_level': ref_level})
2726
    return tmp
2727 2728


C
chengduo 已提交
2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793
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
2794 2795 2796 2797 2798 2799 2800
@templatedoc()
def sequence_pad(x, pad_value, maxlen=None):
    """
    ${comment}

    Args:
        x(Variable): Input variable which should contain lod information.
M
minqiyang 已提交
2801 2802 2803
        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
2804
            automatically broadcasted to the shape of time step.
M
minqiyang 已提交
2805 2806 2807 2808
        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
F
fengjiayi 已提交
2809
            longest original sequence."
M
minqiyang 已提交
2810

F
fengjiayi 已提交
2811
    Returns:
M
minqiyang 已提交
2812
        Variable: The padded sequence batch and the original lengths before
2813
                  padding. All sequences has the same length.
M
minqiyang 已提交
2814

F
fengjiayi 已提交
2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828
    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)
2829 2830 2831 2832 2833
    length = helper.create_tmp_variable(dtype)

    pad_value.stop_gradient = True
    length.stop_gradient = True

F
fengjiayi 已提交
2834 2835 2836 2837 2838 2839
    if maxlen is None:
        maxlen = -1
    helper.append_op(
        type='sequence_pad',
        inputs={'X': x,
                'PadValue': pad_value},
2840 2841
        outputs={'Out': out,
                 'Length': length},
F
fengjiayi 已提交
2842
        attrs={'padded_length': maxlen})
2843
    return out, length
F
fengjiayi 已提交
2844 2845


2846 2847 2848 2849 2850 2851 2852 2853 2854
def beam_search(pre_ids,
                pre_scores,
                ids,
                scores,
                beam_size,
                end_id,
                level=0,
                name=None):
    """
2855 2856
    Beam search is a classical algorithm for selecting candidate words in a
    machine translation task.
Y
Yan Chunwei 已提交
2857 2858 2859

    Refer to `Beam search <https://en.wikipedia.org/wiki/Beam_search>`_
    for more details.
M
minqiyang 已提交
2860 2861

    This layer does the search in beams for one time step. Specifically, it
2862 2863 2864 2865 2866 2867
    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
2868

2869 2870 2871 2872 2873 2874 2875 2876
    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
2877

2878
    Args:
2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903
        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
2904

2905
    Returns:
2906 2907
        Variable: The LodTensor pair containing the selected ids and the \
            corresponding scores.
Y
Yan Chunwei 已提交
2908 2909 2910 2911

    Examples:
        .. code-block:: python

2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928
            # 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
2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939
    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,
2940
            'pre_scores': pre_scores,
Q
Qiao Longfei 已提交
2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957
            '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


2958 2959 2960 2961 2962 2963 2964
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
2965

2966 2967 2968 2969 2970 2971 2972 2973 2974
    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
2975

2976 2977 2978 2979 2980 2981
    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
2982

2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007
    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
3008 3009 3010 3011
def lstm_unit(x_t,
              hidden_t_prev,
              cell_t_prev,
              forget_bias=0.0,
Y
yangyaming 已提交
3012
              param_attr=None,
C
caoying03 已提交
3013 3014
              bias_attr=None,
              name=None):
Y
yangyaming 已提交
3015 3016 3017 3018
    """Lstm unit layer. The equation of a lstm step is:

        .. math::

3019
            i_t & = \sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i)
Y
yangyaming 已提交
3020

3021
            f_t & = \sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + b_f)
Y
yangyaming 已提交
3022

3023
            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
3024

3025
            o_t & = \sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + b_o)
Y
yangyaming 已提交
3026 3027 3028

            h_t & = o_t tanh(c_t)

3029 3030 3031 3032 3033 3034
    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
3035 3036 3037

        .. math::

3038
            L_{i_t} = W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i
Y
yangyaming 已提交
3039 3040 3041 3042 3043 3044 3045 3046

    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
3047
    This layer has two outputs including :math:`h_t` and :math:`o_t`.
Y
yangyaming 已提交
3048 3049

    Args:
Y
yangyaming 已提交
3050 3051 3052 3053 3054 3055
        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
3056
        forget_bias (float): The forget bias of lstm unit.
Y
yangyaming 已提交
3057 3058
        param_attr (ParamAttr): The attributes of parameter weights, used to set
            initializer, name etc.
Y
yangyaming 已提交
3059 3060
        bias_attr (ParamAttr): The attributes of bias weights, if not False,
            bias weights will be created and be set to default value.
C
caoying03 已提交
3061 3062
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
Y
yangyaming 已提交
3063 3064

    Returns:
Y
yangyaming 已提交
3065
        tuple: The hidden value and cell value of lstm unit.
Y
yangyaming 已提交
3066 3067

    Raises:
3068 3069 3070 3071
        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
3072 3073 3074 3075 3076 3077

    Examples:

        .. code-block:: python

             x_t = fluid.layers.fc(input=x_t_data, size=10)
3078
             prev_hidden = fluid.layers.fc(input=prev_hidden_data, size=30)
Y
yangyaming 已提交
3079
             prev_cell = fluid.layers.fc(input=prev_cell_data, size=30)
Y
yangyaming 已提交
3080
             hidden_value, cell_value = fluid.layers.lstm_unit(x_t=x_t,
Y
yangyaming 已提交
3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096
                                                    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
3097
        raise ValueError("The 1st dimensions of x_t, hidden_t_prev and "
3098 3099 3100 3101
                         "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
3102 3103
                         "cell_t_prev must be the same.")

Y
yangyaming 已提交
3104 3105 3106
    if bias_attr is None:
        bias_attr = ParamAttr()

Y
yangyaming 已提交
3107
    size = cell_t_prev.shape[1]
3108
    concat_out = concat(input=[x_t, hidden_t_prev], axis=1)
Y
yangyaming 已提交
3109 3110
    fc_out = fc(input=concat_out,
                size=4 * size,
Y
yangyaming 已提交
3111
                param_attr=param_attr,
3112
                bias_attr=bias_attr)
Y
yangyaming 已提交
3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124
    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
3125
    return h, c
G
guosheng 已提交
3126 3127


C
caoying03 已提交
3128
def reduce_sum(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
3129
    """
Y
yangyaming 已提交
3130
    Computes the sum of tensor elements over the given dimension.
G
guosheng 已提交
3131 3132 3133

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
3134
        dim (list|int|None): The dimensions along which the sum is performed. If
Y
yangyaming 已提交
3135 3136
            :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
3137 3138
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
3139
        keep_dim (bool|False): Whether to reserve the reduced dimension in the
Y
yangyaming 已提交
3140
            output Tensor. The result tensor will have one fewer dimension
G
guosheng 已提交
3141
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
3142 3143
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
3144 3145 3146

    Returns:
        Variable: The reduced Tensor variable.
F
fengjiayi 已提交
3147

G
guosheng 已提交
3148 3149 3150 3151 3152 3153
    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
3154
            # Each example is followed by the corresponding output tensor.
G
guosheng 已提交
3155 3156 3157 3158
            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
3159 3160 3161 3162

            # 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
3163
            # Each example is followed by the corresponding output tensor.
W
whs 已提交
3164 3165 3166
            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
3167 3168 3169
    """
    helper = LayerHelper('reduce_sum', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3170 3171
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
G
guosheng 已提交
3172 3173 3174 3175 3176
    helper.append_op(
        type='reduce_sum',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3177
            'dim': dim if dim != None else [0],
G
guosheng 已提交
3178 3179 3180 3181
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
3182 3183


C
caoying03 已提交
3184
def reduce_mean(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
3185
    """
Y
Yibing Liu 已提交
3186
    Computes the mean of the input tensor's elements along the given dimension.
G
guosheng 已提交
3187 3188 3189

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
Yibing Liu 已提交
3190 3191 3192
        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
3193
            must be in the range :math:`[-rank(input), rank(input))`. If
3194
            :math:`dim[i] < 0`, the dimension to reduce is
Y
Yibing Liu 已提交
3195
            :math:`rank(input) + dim[i]`.
Y
yangyaming 已提交
3196 3197
        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
3198
            than the :attr:`input` unless :attr:`keep_dim` is true.
Y
Yibing Liu 已提交
3199
        name(str|None): A name for this layer(optional). If set `None`, the layer
C
caoying03 已提交
3200
                       will be named automatically.
G
guosheng 已提交
3201 3202

    Returns:
Y
Yibing Liu 已提交
3203
        Variable: The reduced mean Variable.
F
fengjiayi 已提交
3204

G
guosheng 已提交
3205 3206 3207 3208 3209 3210 3211 3212 3213 3214
    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
3215 3216
            fluid.layers.reduce_mean(
                x, dim=1, keep_dim=True)  # [[0.475], [0.4]]
W
whs 已提交
3217 3218 3219 3220 3221 3222 3223

            # 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
3224 3225 3226
    """
    helper = LayerHelper('reduce_mean', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3227 3228
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
G
guosheng 已提交
3229 3230 3231 3232 3233
    helper.append_op(
        type='reduce_mean',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3234
            'dim': dim if dim != None else [0],
G
guosheng 已提交
3235 3236 3237 3238
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
3239 3240


C
caoying03 已提交
3241
def reduce_max(input, dim=None, keep_dim=False, name=None):
3242
    """
Y
yangyaming 已提交
3243
    Computes the maximum of tensor elements over the given dimension.
3244 3245 3246

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
3247
        dim (list|int|None): The dimension along which the maximum is computed.
Y
yangyaming 已提交
3248 3249 3250
            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
3251
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
Y
yangyaming 已提交
3252 3253
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
3254
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
3255 3256
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
3257 3258 3259

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
3260

3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271
    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
3272 3273 3274 3275 3276 3277 3278

            # 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]
3279 3280 3281
    """
    helper = LayerHelper('reduce_max', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3282 3283
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3284 3285 3286 3287 3288
    helper.append_op(
        type='reduce_max',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3289
            'dim': dim if dim != None else [0],
3290 3291 3292 3293 3294 3295
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


C
caoying03 已提交
3296
def reduce_min(input, dim=None, keep_dim=False, name=None):
3297
    """
Y
yangyaming 已提交
3298
    Computes the minimum of tensor elements over the given dimension.
3299 3300 3301

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
3302
        dim (list|int|None): The dimensions along which the minimum is computed.
Y
yangyaming 已提交
3303 3304 3305
            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
3306
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
Y
yangyaming 已提交
3307 3308
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
3309
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
3310 3311
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
3312 3313 3314

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
3315

3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326
    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
3327 3328 3329 3330 3331 3332 3333

            # 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]
3334 3335 3336
    """
    helper = LayerHelper('reduce_min', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3337 3338
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3339 3340 3341 3342 3343
    helper.append_op(
        type='reduce_min',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3344
            'dim': dim if dim != None else [0],
3345 3346 3347 3348
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
3349 3350


3351 3352 3353 3354 3355 3356
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
3357
        dim (list|int|None): The dimensions along which the product is performed. If
3358 3359
            :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
3360 3361
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
3362 3363 3364
        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
3365
        name(str|None): A name for this layer(optional). If set None, the
Z
zhouhanqing 已提交
3366
            layer will be named automatically.
3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380

    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
3381
            fluid.layers.reduce_prod(x, dim=1,
Z
zhouhanqing 已提交
3382
                                     keep_dim=True)  # [[0.027], [0.0084]]
W
whs 已提交
3383 3384 3385 3386 3387 3388 3389

            # 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]
3390 3391 3392
    """
    helper = LayerHelper('reduce_prod', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3393 3394
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3395 3396 3397 3398 3399
    helper.append_op(
        type='reduce_prod',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3400
            'dim': dim if dim != None else [0],
3401 3402 3403 3404 3405 3406
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


C
caoying03 已提交
3407
def split(input, num_or_sections, dim=-1, name=None):
G
guosheng 已提交
3408
    """
C
caoying03 已提交
3409
    Split the input tensor into multiple sub-tensors.
G
guosheng 已提交
3410 3411 3412

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
C
caoying03 已提交
3413 3414 3415 3416 3417
        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
3418
            :attr:`dim` dimension orderly.
C
caoying03 已提交
3419
        dim (int): The dimension along which to split. If :math:`dim < 0`, the
G
guosheng 已提交
3420
            dimension to split along is :math:`rank(input) + dim`.
C
caoying03 已提交
3421 3422
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
3423 3424

    Returns:
D
dzhwinter 已提交
3425
        list(Variable): The list of segmented tensor variables.
G
guosheng 已提交
3426 3427 3428 3429 3430 3431 3432 3433 3434

    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
3435 3436
            x0, x1, x2 = fluid.layers.split(
                x, num_or_sections=[2, 3, 4], dim=1)
G
guosheng 已提交
3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465
            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
3466 3467 3468 3469 3470 3471 3472 3473 3474


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

3475
    .. math::
3476 3477

        y = \\frac{x}{ \sqrt{\sum {x^2} + epsion }}
C
caoying03 已提交
3478 3479 3480 3481 3482

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

    Args:
3483
        x(Variable|list): The input tensor to l2_normalize layer.
3484
        axis(int): The axis on which to apply normalization. If `axis < 0`, \
3485 3486
            the dimension to normalization is rank(X) + axis. -1 is the
            last dimension.
3487
        epsilon(float): The epsilon value is used to avoid division by zero, \
3488
            the defalut value is 1e-10.
3489
        name(str|None): A name for this layer(optional). If set None, the layer \
3490
            will be named automatically.
C
caoying03 已提交
3491 3492

    Returns:
3493
        Variable: The output tensor variable is the same shape with `x`.
C
caoying03 已提交
3494 3495

    Examples:
3496

C
caoying03 已提交
3497 3498
        .. code-block:: python

3499 3500 3501 3502
            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
3503 3504
    """

F
fengjiayi 已提交
3505 3506
    if len(x.shape) == 1:
        axis = 0
C
caoying03 已提交
3507 3508
    helper = LayerHelper("l2_normalize", **locals())

3509 3510
    out = helper.create_tmp_variable(dtype=x.dtype)
    norm = helper.create_tmp_variable(dtype=x.dtype)
C
caoying03 已提交
3511
    helper.append_op(
3512 3513 3514 3515
        type="norm",
        inputs={"X": x},
        outputs={"Out": out,
                 "Norm": norm},
C
caoying03 已提交
3516
        attrs={
3517 3518
            "axis": 1 if axis is None else axis,
            "epsilon": epsilon,
C
caoying03 已提交
3519 3520
        })
    return out
3521 3522


S
sneaxiy 已提交
3523
def matmul(x, y, transpose_x=False, transpose_y=False, alpha=1.0, name=None):
G
guosheng 已提交
3524
    """
Y
ying 已提交
3525 3526 3527 3528
    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
3529

C
chengduoZH 已提交
3530
    The actual behavior depends on the shapes of :math:`x`, :math:`y` and the
3531
    flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically:
G
guosheng 已提交
3532

3533 3534 3535 3536 3537
    - 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
3538
      :math:`[1, D]` in transposed form.
G
guosheng 已提交
3539

C
chengduoZH 已提交
3540
    - After transpose, the two tensors are 2-D or n-D and matrix multiplication
3541
      performs in the following way.
G
guosheng 已提交
3542

3543
      - If both are 2-D, they are multiplied like conventional matrices.
C
chengduoZH 已提交
3544
      - If either is n-D, it is treated as a stack of matrices residing in the
Y
ying 已提交
3545
        last two dimensions and a batched matrix multiply supporting broadcast
3546
        applies on the two tensors.
G
guosheng 已提交
3547

Y
ying 已提交
3548 3549
    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
3550
    removed after matrix multiplication.
G
guosheng 已提交
3551 3552 3553

    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
3554 3555 3556
        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
3557
        alpha (float): The scale of output. Default 1.0.
3558
        name(str|None): A name for this layer(optional). If set None, the layer
3559
            will be named automatically.
G
guosheng 已提交
3560 3561

    Returns:
3562
        Variable: The product Tensor variable.
G
guosheng 已提交
3563

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

3567
            # Examples to clarify shapes of the inputs and output
C
chengduoZH 已提交
3568 3569
            # x: [B, ..., M, K], y: [B, ..., K, N]
            fluid.layers.matmul(x, y)  # out: [B, ..., M, N]
Y
ying 已提交
3570

3571 3572
            # x: [B, M, K], y: [B, K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
3573

3574 3575
            # x: [B, M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
3576

3577 3578
            # x: [M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [M, N]
Y
ying 已提交
3579 3580 3581 3582

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

3583 3584
            # x: [K], y: [K]
            fluid.layers.matmul(x, y)  # out: [1]
3585

Y
ying 已提交
3586
            # x: [M], y: [N]
3587
            fluid.layers.matmul(x, y, True, True)  # out: [M, N]
G
guosheng 已提交
3588
    """
Y
ying 已提交
3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600

    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
3601
            y_shape = y_shape + [1]
Y
ying 已提交
3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617

        # 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)

3618
    helper = LayerHelper('matmul', **locals())
Y
ying 已提交
3619
    out = helper.create_tmp_variable(dtype=x.dtype)
G
guosheng 已提交
3620
    helper.append_op(
3621 3622 3623 3624
        type='matmul',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
S
sneaxiy 已提交
3625 3626 3627
        attrs={
            'transpose_X': transpose_x,
            'transpose_Y': transpose_y,
S
sneaxiy 已提交
3628
            'alpha': float(alpha),
S
sneaxiy 已提交
3629
        })
3630
    return out
3631 3632


3633
def topk(input, k, name=None):
Q
qingqing01 已提交
3634 3635 3636 3637
    """
    This operator is used to find values and indices of the k largest entries
    for the last dimension.

F
fengjiayi 已提交
3638
    If the input is a vector (1-D Tensor), finds the k largest entries in the vector
Q
qingqing01 已提交
3639 3640 3641 3642 3643 3644
    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
3645 3646 3647 3648 3649 3650 3651 3652 3653 3654 3655 3656 3657 3658 3659 3660 3661 3662 3663 3664 3665
    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
3666 3667 3668
    Args:
        input(Variable): The input variable which can be a vector or Tensor with
            higher rank.
3669
        k(int):  The number of top elements to look for along the last dimension
F
fengjiayi 已提交
3670
                 of input.
3671
        name(str|None): A name for this layer(optional). If set None, the layer
3672
                       will be named automatically.
F
fengjiayi 已提交
3673
                       Default: None
Q
qingqing01 已提交
3674 3675

    Returns:
3676 3677 3678
        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
3679
        within the last dimension of input.
Q
qingqing01 已提交
3680

F
fengjiayi 已提交
3681 3682
    Raises:
        ValueError: If k < 1 or k is not less than the last dimension of input
Q
qingqing01 已提交
3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698 3699 3700 3701 3702

    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


3703
def edit_distance(input, label, normalized=True, ignored_tokens=None):
3704
    """
Y
ying 已提交
3705 3706 3707 3708 3709 3710 3711 3712 3713
    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
3714

Y
ying 已提交
3715
    "kitten" -> "sitten" -> "sittin" -> "sitting"
W
wanghaoshuang 已提交
3716

3717
    The input is a LoDTensor consisting of all the hypothesis strings with
Y
ying 已提交
3718 3719
    the total number denoted by `batch_size`, and the separation is specified
    by the LoD information. And the `batch_size` reference strings are arranged
3720
    in order in the same way in the input LoDTensor.
W
wanghaoshuang 已提交
3721

3722
    The output contains the `batch_size` results and each stands for the edit
Y
ying 已提交
3723 3724
    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
3725

3726 3727 3728
    Args:
        input(Variable): The indices for hypothesis strings.
        label(Variable): The indices for reference strings.
3729
        normalized(bool, default True): Indicated whether to normalize the edit distance by
Y
ying 已提交
3730
                          the length of reference string.
3731
        ignored_tokens(list<int>, default None): Tokens that should be removed before
Y
ying 已提交
3732
                                     calculating edit distance.
3733
        name (str): The name of this layer. It is optional.
3734

W
wanghaoshuang 已提交
3735
    Returns:
W
wanghaoshuang 已提交
3736
        Variable: sequence-to-sequence edit distance in shape [batch_size, 1].
W
wanghaoshuang 已提交
3737 3738 3739 3740 3741

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
3742
            y = fluid.layers.data(name='y', shape=[7], dtype='float32')
3743
            cost = fluid.layers.edit_distance(input=x,label=y)
3744
    """
3745
    helper = LayerHelper("edit_distance", **locals())
3746

3747
    # remove some tokens from input and labels
W
wanghaoshuang 已提交
3748
    if ignored_tokens is not None and len(ignored_tokens) > 0:
3749 3750 3751 3752 3753 3754 3755
        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
3756
            attrs={"tokens": ignored_tokens})
3757 3758 3759 3760 3761
        input = erased_input

        helper.append_op(
            type="sequence_erase",
            inputs={"X": [label]},
W
whs 已提交
3762
            outputs={"Out": [erased_label]},
W
wanghaoshuang 已提交
3763
            attrs={"tokens": ignored_tokens})
3764 3765
        label = erased_label

3766 3767
    # edit distance op
    edit_distance_out = helper.create_tmp_variable(dtype="int64")
3768
    sequence_num = helper.create_tmp_variable(dtype="int64")
3769 3770 3771 3772
    helper.append_op(
        type="edit_distance",
        inputs={"Hyps": [input],
                "Refs": [label]},
3773 3774
        outputs={"Out": [edit_distance_out],
                 "SequenceNum": [sequence_num]},
3775 3776
        attrs={"normalized": normalized})

3777
    return edit_distance_out, sequence_num
3778 3779 3780 3781 3782


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
3783

Y
ying 已提交
3784 3785 3786 3787
    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.
3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798 3799 3800 3801 3802 3803 3804

    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]]

3805
        input.lod = [[4, 4]]
3806 3807 3808 3809 3810 3811 3812

        Then:

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

3813
        output.lod = [[2, 1]]
3814 3815 3816

    Args:

Y
ying 已提交
3817 3818 3819 3820 3821 3822 3823 3824 3825
        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).
3826
        name (str): The name of this layer. It is optional.
3827 3828

    Returns:
3829
        Variable: CTC greedy decode result. If all the sequences in result were
3830
        empty, the result LoDTensor will be [-1] with LoD [[]] and dims [1, 1].
3831 3832 3833 3834 3835

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
W
wanghaoshuang 已提交
3836

3837
            cost = fluid.layers.ctc_greedy_decoder(input=x, blank=0)
W
wanghaoshuang 已提交
3838
    """
3839
    helper = LayerHelper("ctc_greedy_decoder", **locals())
Q
qingqing01 已提交
3840
    _, topk_indices = topk(input, k=1)
3841 3842 3843 3844 3845 3846

    # 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
3847
        outputs={"Output": [ctc_out]},
3848 3849
        attrs={"merge_repeated": True,
               "blank": blank})
3850
    return ctc_out
3851 3852


F
fengjiayi 已提交
3853
def warpctc(input, label, blank=0, norm_by_times=False):
W
wanghaoshuang 已提交
3854
    """
3855 3856
    An operator integrating the open source Warp-CTC library
    (https://github.com/baidu-research/warp-ctc)
W
wanghaoshuang 已提交
3857
    to compute Connectionist Temporal Classification (CTC) loss.
3858 3859
    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
3860 3861 3862
    input tensor.

    Args:
3863
       input (Variable): The unscaled probabilities of variable-length sequences,
W
wanghaoshuang 已提交
3864 3865 3866 3867
         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).
3868
       label (Variable): The ground truth of variable-length sequence,
3869 3870 3871
         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
3872 3873
         Temporal Classification (CTC) loss, which is in the
         half-opened interval [0, num_classes + 1).
3874 3875 3876
       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
3877
         follewed by a mean_op.
W
wanghaoshuang 已提交
3878 3879

    Returns:
3880 3881
        Variable: The Connectionist Temporal Classification (CTC) loss,
        which is a 2-D Tensor of the shape [batch_size, 1].
W
wanghaoshuang 已提交
3882 3883

    Examples:
3884

W
wanghaoshuang 已提交
3885
        .. code-block:: python
3886

3887 3888 3889
            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
3890 3891

    """
F
fengjiayi 已提交
3892
    helper = LayerHelper('warpctc', **locals())
W
wanghaoshuang 已提交
3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903
    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
3904 3905 3906 3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918


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]]
3919 3920 3921
            x.data = [[1,  2], [3,  4],
                      [5,  6], [7,  8],
                      [9, 10], [11, 12]]
3922 3923 3924 3925 3926
            x.dims = [6, 2]

        set new_dim = 4

        then out is a LoDTensor:
3927

3928
            out.lod  = [[0, 1, 3]]
3929 3930 3931 3932

            out.data = [[1,  2,  3,  4],
                        [5,  6,  7,  8],
                        [9, 10, 11, 12]]
3933 3934 3935 3936 3937 3938 3939
            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:
3940 3941 3942

       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.
3943 3944

    Returns:
3945

3946 3947 3948 3949 3950
        Variable: Reshaped LoDTensor according to new dimension.

    Examples:
        .. code-block:: python

3951
            x = fluid.layers.data(shape=[5, 20], dtype='float32', lod_level=1)
3952
            x_reshaped = fluid.layers.sequence_reshape(input=x, new_dim=10)
3953 3954 3955 3956 3957 3958 3959 3960 3961
    """
    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
3962 3963


3964 3965 3966 3967
# 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
3968 3969 3970 3971 3972 3973 3974
def nce(input,
        label,
        num_total_classes,
        sample_weight=None,
        param_attr=None,
        bias_attr=None,
        num_neg_samples=None):
3975 3976 3977 3978 3979 3980 3981
    """
    ${comment}

    Args:
        input (Variable): input variable.
        label (Variable): label.
        num_total_classes (int):${num_total_classes_comment}
3982 3983
        sample_weight (Variable|None): A Variable of shape [batch_size, 1]
            storing a weight for each sample. The default weight for each
3984
            sample is 1.0.
3985 3986 3987
        param_attr (ParamAttr|None): attributes for parameter
        bias_attr (ParamAttr|None): attributes for bias
        num_neg_samples (int): ${num_neg_samples_comment}
F
fengjiayi 已提交
3988

3989
    Returns:
Y
Yibing Liu 已提交
3990 3991 3992 3993 3994 3995 3996 3997 3998 3999 4000 4001 4002 4003 4004 4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016
        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')
4017
    """
Y
Yang Yu 已提交
4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031 4032 4033 4034 4035 4036
    helper = LayerHelper('nce', **locals())
    assert isinstance(input, Variable)
    dim = input.shape[1]
    assert isinstance(label, Variable)
    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)
    b = helper.create_parameter(
        attr=helper.bias_attr,
        shape=[num_total_classes, 1],
        is_bias=True,
        dtype=input.dtype)
    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
4037 4038 4039 4040 4041 4042 4043 4044 4045
    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
4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057 4058 4059 4060 4061

    helper.append_op(
        type='nce',
        inputs={
            'Input': input,
            'Label': label,
            'Weight': w,
            'Bias': b,
            'SampleWeight': sample_weight if sample_weight is not None else []
        },
        outputs={
            'Cost': cost,
            'SampleLogits': sample_logits,
            'SampleLabels': sample_labels
        },
        attrs=attrs)
Y
Yang Yu 已提交
4062
    return cost / (num_neg_samples + 1)
4063 4064


G
guosheng 已提交
4065
def hsigmoid(input, label, num_classes, param_attr=None, bias_attr=None):
W
weixing02 已提交
4066 4067
    """
    The hierarchical sigmoid operator is used to accelerate the training
M
minqiyang 已提交
4068
    process of language model. This operator organizes the classes into a
G
guosheng 已提交
4069 4070 4071 4072 4073 4074 4075 4076 4077
    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
4078

W
weixing02 已提交
4079
    Args:
M
minqiyang 已提交
4080
        input (Variable): The input tensor variable with shape
G
guosheng 已提交
4081 4082 4083 4084 4085
            :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.
W
weixing02 已提交
4086 4087
        param_attr (ParamAttr|list of ParamAttr, default None): The parameter
             attribute for learnable parameters/weights of this layer.
M
minqiyang 已提交
4088
        bias_attr (ParamAttr|list of ParamAttr, default None):  The parameter
G
guosheng 已提交
4089 4090
             attribute for the bias of this layer. If it is set to False, no
             bias will be applied.
W
weixing02 已提交
4091 4092 4093 4094 4095 4096 4097 4098

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

    Examples:

        .. code-block:: python

G
guosheng 已提交
4099 4100 4101
            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
4102 4103 4104 4105 4106 4107 4108 4109
    """

    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
4110
        raise ValueError("num_classes must not be less than 2.")
W
weixing02 已提交
4111 4112 4113 4114 4115
    weights = helper.create_parameter(
        attr=helper.param_attr,
        shape=[num_classes - 1, dim],
        is_bias=False,
        dtype=input.dtype)
W
weixing02 已提交
4116 4117 4118 4119 4120 4121 4122 4123
    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
4124 4125
    helper.append_op(
        type="hierarchical_sigmoid",
W
weixing02 已提交
4126
        inputs=inputs,
W
weixing02 已提交
4127 4128 4129 4130 4131 4132
        outputs={"Out": out,
                 "PreOut": pre_out},
        attrs={"num_classes": num_classes})
    return out


Y
fix ci.  
ying 已提交
4133
def transpose(x, perm, name=None):
Y
ying 已提交
4134 4135 4136 4137 4138 4139 4140
    """
    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:
4141 4142 4143
        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
4144 4145 4146 4147 4148 4149 4150 4151

    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
4152
            x_transposed = layers.transpose(x, perm=[1, 0, 2])
Y
ying 已提交
4153 4154
    """

Y
fix ci.  
ying 已提交
4155
    if len(perm) != len(x.shape):
Y
ying 已提交
4156 4157 4158
        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
4159 4160 4161 4162 4163 4164
    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
4165 4166

    helper = LayerHelper('transpose', **locals())
Y
fix ci.  
ying 已提交
4167
    out = helper.create_tmp_variable(x.dtype)
4168
    x_shape = helper.create_tmp_variable(x.dtype)
Y
ying 已提交
4169
    helper.append_op(
4170
        type='transpose2',
Y
fix ci.  
ying 已提交
4171
        inputs={'X': [x]},
4172 4173
        outputs={'Out': [out],
                 'XShape': [x_shape]},
Y
ying 已提交
4174 4175
        attrs={'axis': perm})
    return out
4176 4177


4178 4179 4180 4181 4182 4183 4184
def im2sequence(input,
                filter_size=1,
                stride=1,
                padding=0,
                input_image_size=None,
                out_stride=1,
                name=None):
4185
    """
4186 4187 4188 4189 4190 4191 4192
    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:
4193 4194 4195 4196 4197 4198 4199 4200 4201 4202

    .. 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
4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220

        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.

4221 4222 4223 4224 4225 4226 4227 4228 4229
        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.

4230 4231 4232
        name (int): The name of this layer. It is optional.

    Returns:
W
wanghaoshuang 已提交
4233 4234 4235 4236 4237
        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.
4238 4239 4240 4241 4242 4243 4244 4245 4246 4247 4248 4249 4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261 4262 4263 4264

    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
4265 4266 4267
            filter = [2, 2]
            stride = [1, 1]
            padding = [0, 0]
4268 4269 4270 4271 4272 4273 4274 4275 4276 4277 4278 4279

            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.]]

4280
            output.dims = {8, 8}
4281

4282
            output.lod = [[4, 4]]
4283

D
dzhwinter 已提交
4284
     Examples:
4285 4286 4287

        .. code-block:: python

4288 4289
            output = fluid.layers.im2sequence(
                input=layer, stride=[1, 1], filter_size=[2, 2])
4290 4291

    """
W
wanghaoshuang 已提交
4292 4293 4294 4295 4296 4297 4298 4299 4300 4301

    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])
4302 4303 4304 4305 4306 4307 4308
    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
4309
    helper = LayerHelper('im2sequence', **locals())
4310 4311
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
4312
        type='im2sequence', inputs=inputs, outputs={'Out': out}, attrs=attrs)
4313
    return out
4314 4315


Y
yuyang18 已提交
4316
@templatedoc()
4317
def row_conv(input, future_context_size, param_attr=None, act=None):
Y
yuyang18 已提交
4318 4319
    """
    ${comment}
4320 4321

    Args:
Y
yuyang18 已提交
4322
        input (${x_type}): ${x_comment}.
Y
yangyaming 已提交
4323 4324
        future_context_size (int): Future context size. Please note, the shape
            of convolution kernel is [future_context_size + 1, D].
4325 4326 4327 4328 4329
        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
4330
        ${out_comment}.
4331 4332

    Examples:
Y
yuyang18 已提交
4333 4334 4335 4336
        >>> 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)
4337 4338 4339 4340 4341 4342 4343 4344 4345 4346 4347 4348
    """
    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
4349
    return helper.append_activation(out)
4350 4351


Y
yuyang18 已提交
4352
@templatedoc()
4353 4354
def multiplex(inputs, index):
    """
Y
yuyang18 已提交
4355 4356 4357 4358 4359 4360 4361
    ${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)
4362 4363

    Args:
Y
yuyang18 已提交
4364 4365
       inputs (list): ${x_comment}.
       index (${ids_type}): ${ids_comment}.
4366 4367

    Returns:
Y
yuyang18 已提交
4368
        ${out_comment}.
4369 4370
    """
    helper = LayerHelper('multiplex', **locals())
Y
yangyaming 已提交
4371 4372 4373 4374 4375 4376

    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)
4377 4378 4379 4380 4381 4382
    helper.append_op(
        type='multiplex',
        inputs={'X': inputs,
                'Ids': index},
        outputs={'Out': [out]})
    return out
4383 4384


4385 4386 4387 4388
def softmax_with_cross_entropy(logits,
                               label,
                               soft_label=False,
                               ignore_index=-100):
4389 4390
    """
    **Softmax With Cross Entropy Operator.**
4391

4392 4393 4394 4395
    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.
4396

4397 4398 4399
    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.
4400

4401 4402 4403
    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.
4404

4405
    The equation is as follows:
4406

4407
    1) Hard label (one-hot label, so every sample has exactly one class)
4408

4409 4410 4411 4412
    .. math::

        loss_j =  -\\text{logit}_{label_j} +
        \\log\\left(\\sum_{i=0}^{K}\\exp(\\text{logit}_i)\\right), j = 1,..., K
4413

4414 4415 4416
    2) Soft label (each sample can have a distribution over all classes)

    .. math::
4417

4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429
        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
4430 4431
        ignore_index (int): Specifies a target value that is ignored and does
                            not contribute to the input gradient. Only valid
4432 4433
                            if soft_label is set to False. Default: -100

4434 4435 4436 4437 4438 4439 4440 4441 4442
    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
4443 4444
            out = fluid.layers.softmax_with_cross_entropy(
                logits=fc, label=label)
4445 4446 4447 4448 4449 4450 4451 4452 4453 4454
    """
    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},
4455 4456
        attrs={'soft_label': soft_label,
               'ignore_index': ignore_index})
4457 4458 4459 4460 4461
    return loss


def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None):
    """
Y
Yibing Liu 已提交
4462 4463
    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
4464
    For each instance, it computes the smooth L1 loss element by element first
4465
    and then sums all the losses. So the shape of ouput Variable is
4466
    [batch_size, 1].
4467

4468 4469
    Args:
        x (Variable): A tensor with rank at least 2. The input value of smooth
Q
qingqing01 已提交
4470
            L1 loss op with shape [batch_size, dim1, ..., dimN].
4471
        y (Variable): A tensor with rank at least 2. The target value of smooth
Y
Yibing Liu 已提交
4472
            L1 loss op with same shape as :attr:`x`.
4473
        inside_weight (Variable|None):  A tensor with rank at least 2. This
4474 4475
            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
4476
            by this tensor element by element.
4477
        outside_weight (Variable|None): A tensor with rank at least 2. This
4478 4479
            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
4480
            element by element.
4481
        sigma (float|None): Hyper parameter of smooth L1 loss layer. A float
4482 4483
           scalar with default value 1.0.

4484
    Returns:
4485
        Variable: The output smooth L1 loss with shape [batch_size, 1].
4486 4487 4488 4489 4490

    Examples:
        .. code-block:: python

            data = fluid.layers.data(name='data', shape=[128], dtype='float32')
F
stash  
fengjiayi 已提交
4491 4492
            label = fluid.layers.data(
                name='label', shape=[100], dtype='float32')
4493
            fc = fluid.layers.fc(input=data, size=100)
F
fengjiayi 已提交
4494
            out = fluid.layers.smooth_l1(x=fc, y=label)
4495
    """
4496

4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511
    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
4512 4513 4514 4515


def one_hot(input, depth):
    """
Y
Yibing Liu 已提交
4516
    This layer creates the one-hot representations for input indices.
4517 4518

    Args:
Y
Yibing Liu 已提交
4519 4520
        input(Variable): Input indices, last dimension must be 1.
        depth(scalar): An interger defining the depth of the one-hot dimension.
4521 4522

    Returns:
Y
Yibing Liu 已提交
4523
        Variable: The one-hot representations of input.
4524 4525

    Examples:
C
caoying03 已提交
4526
        .. code-block:: python
4527

Y
Yibing Liu 已提交
4528 4529
            label = layers.data(name="label", shape=[1], dtype="float32")
            one_hot_label = layers.one_hot(input=label, depth=10)
4530 4531 4532 4533 4534 4535 4536 4537 4538
    """
    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
4539 4540


Y
Yu Yang 已提交
4541
def autoincreased_step_counter(counter_name=None, begin=1, step=1):
Y
Yu Yang 已提交
4542
    """
Y
yi.wu 已提交
4543 4544 4545
    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
4546 4547 4548 4549 4550 4551

    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.

4552 4553
    Returns:
        Variable: The global run counter.
Y
yi.wu 已提交
4554 4555 4556 4557 4558 4559

    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
4560 4561
    """
    helper = LayerHelper('global_step_counter')
Y
Yu Yang 已提交
4562 4563
    if counter_name is None:
        counter_name = '@STEP_COUNTER@'
Y
Yu Yang 已提交
4564 4565 4566 4567 4568
    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
4569
                value=begin - 1, force_cpu=True))
W
Wu Yi 已提交
4570
        helper.main_program.global_block()._prepend_op(
Y
Yu Yang 已提交
4571 4572
            type='increment',
            inputs={'X': [counter]},
Y
Yu Yang 已提交
4573 4574
            outputs={'Out': [counter]},
            attrs={'step': float(step)})
Y
Yu Yang 已提交
4575 4576 4577
        counter.stop_gradient = True

    return counter
Y
yangyaming 已提交
4578 4579


4580
def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None):
C
caoying03 已提交
4581
    """
C
caoying03 已提交
4582 4583
    Gives a new shape to the input Tensor without changing its data.

4584 4585 4586 4587 4588
    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
4589

4590
    Some tricks exist when specifying the target shape.
C
caoying03 已提交
4591

4592 4593 4594 4595
    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.

4596
    2. 0 means the actual dimension value is going to be copied from the
4597 4598 4599 4600
    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
4601 4602

    1. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
W
wanghaoshuang 已提交
4603
    is [6, 8], the reshape operator will transform x into a 2-D tensor with
4604
    shape [6, 8] and leaving x's data unchanged.
C
caoying03 已提交
4605

4606
    2. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
4607 4608
    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
4609 4610
    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
4611
    dimensions.
C
caoying03 已提交
4612

4613
    3. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
4614 4615 4616 4617
    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
4618 4619

    Args:
4620
        x(variable): The input tensor.
C
caoying03 已提交
4621 4622
        shape(list): The new shape. At most one dimension of the new shape can
                     be -1.
4623 4624 4625 4626 4627
        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
4628
        act (str): The non-linear activation to be applied to output variable.
X
Xin Pan 已提交
4629 4630 4631 4632
        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.
4633
        name (str): The name of this layer. It is optional.
C
caoying03 已提交
4634

4635 4636
    Returns:
        Variable: The output tensor.
C
caoying03 已提交
4637

X
Xin Pan 已提交
4638 4639 4640
    Raises:
        TypeError: if actual_shape is neither Variable nor None.

C
caoying03 已提交
4641 4642
    Examples:
        .. code-block:: python
G
guosheng 已提交
4643

4644
            data = fluid.layers.data(
4645
                name='data', shape=[2, 4, 6], dtype='float32')
C
caoying03 已提交
4646
            reshaped = fluid.layers.reshape(
4647
                x=data, shape=[-1, 0, 3, 2], act='tanh', inplace=True)
C
caoying03 已提交
4648 4649 4650
    """

    if not (isinstance(shape, list) or isinstance(shape, tuple)):
L
luotao1 已提交
4651
        raise ValueError("Input shape must be a python list or tuple.")
X
Xin Pan 已提交
4652 4653 4654 4655 4656
    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
4657

4658 4659 4660 4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672
    # 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.")

4673
    helper = LayerHelper("reshape2", **locals())
D
dzhwinter 已提交
4674
    out = helper.create_tmp_variable(dtype=x.dtype)
4675
    x_shape = helper.create_tmp_variable(dtype=x.dtype)
C
caoying03 已提交
4676
    helper.append_op(
4677
        type="reshape2",
X
Xin Pan 已提交
4678
        inputs=inputs,
D
dzhwinter 已提交
4679
        attrs={"shape": shape},
4680 4681
        outputs={"Out": out,
                 "XShape": x_shape})
C
caoying03 已提交
4682

D
dzhwinter 已提交
4683
    return helper.append_activation(out)
4684

4685

4686
def squeeze(input, axes, name=None):
Y
Yibing Liu 已提交
4687
    """
M
minqiyang 已提交
4688 4689 4690
    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
4691
    selected with shape entry not equal to one, an error is raised.
M
minqiyang 已提交
4692

Y
Yibing Liu 已提交
4693 4694
    Examples:
    Case 1:
M
minqiyang 已提交
4695
      Given
Y
Yibing Liu 已提交
4696 4697 4698 4699 4700 4701 4702 4703
        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
4704
        and
Y
Yibing Liu 已提交
4705 4706 4707
          axes = []
        we get:
          Out.shape = (3, 5)
M
minqiyang 已提交
4708

Y
Yibing Liu 已提交
4709
    Args:
4710
        input (Variable): The input variable to be squeezed.
Y
Yibing Liu 已提交
4711
        axes (list): List of integers, indicating the dimensions to be squeezed.
4712
        name (str|None): Name for this layer.
Y
Yibing Liu 已提交
4713 4714 4715 4716 4717 4718 4719 4720

    Returns:
        Variable: Output squeezed variable.

    Examples:
        .. code-block:: python

            x = layers.data(name='x', shape=[5, 1, 10])
4721
            y = layers.sequeeze(input=x, axes=[1])
Y
Yibing Liu 已提交
4722 4723
    """
    helper = LayerHelper("squeeze", **locals())
4724
    out = helper.create_tmp_variable(dtype=input.dtype)
4725
    x_shape = helper.create_tmp_variable(dtype=input.dtype)
Y
Yibing Liu 已提交
4726
    helper.append_op(
4727
        type="squeeze2",
4728
        inputs={"X": input},
Y
Yibing Liu 已提交
4729
        attrs={"axes": axes},
4730 4731
        outputs={"Out": out,
                 "XShape": x_shape})
Y
Yibing Liu 已提交
4732

4733 4734 4735
    return out


4736
def unsqueeze(input, axes, name=None):
Y
Yibing Liu 已提交
4737
    """
M
minqiyang 已提交
4738 4739 4740
    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
4741

M
minqiyang 已提交
4742 4743
    For example:
      Given a tensor such that tensor with shape [3, 4, 5],
Y
Yibing Liu 已提交
4744
      then Unsqueezed tensor with axes=[0, 4] has shape [1, 3, 4, 5, 1].
M
minqiyang 已提交
4745

Y
Yibing Liu 已提交
4746
    Args:
4747
        input (Variable): The input variable to be unsqueezed.
Y
Yibing Liu 已提交
4748
        axes (list): List of integers, indicating the dimensions to be inserted.
4749
        name (str|None): Name for this layer.
Y
Yibing Liu 已提交
4750 4751 4752 4753 4754 4755 4756 4757

    Returns:
        Variable: Output unsqueezed variable.

    Examples:
        .. code-block:: python

            x = layers.data(name='x', shape=[5, 10])
4758
            y = layers.unsequeeze(input=x, axes=[1])
Y
Yibing Liu 已提交
4759 4760
    """
    helper = LayerHelper("unsqueeze", **locals())
4761
    out = helper.create_tmp_variable(dtype=input.dtype)
4762
    x_shape = helper.create_tmp_variable(dtype=input.dtype)
Y
Yibing Liu 已提交
4763
    helper.append_op(
4764
        type="unsqueeze2",
4765
        inputs={"X": input},
Y
Yibing Liu 已提交
4766
        attrs={"axes": axes},
4767 4768
        outputs={"Out": out,
                 "XShape": x_shape})
Y
Yibing Liu 已提交
4769

4770 4771
    return out

4772

Y
yangyaming 已提交
4773
def lod_reset(x, y=None, target_lod=None):
Y
yangyaming 已提交
4774
    """
Y
Yibing Liu 已提交
4775
    Set LoD of :attr:`x` to a new one specified by :attr:`y` or
4776 4777 4778 4779
    :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
4780
    :attr:`Y.data` or :attr:`target_lod`, only one level LoD is supported.
Y
yangyaming 已提交
4781 4782 4783 4784 4785 4786

    .. code-block:: text

        * Example 1:

            Given a 1-level LoDTensor x:
4787
                x.lod =  [[ 2,           3,                   1 ]]
Y
yangyaming 已提交
4788 4789 4790
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

4791
            target_lod: [4, 2]
Y
yangyaming 已提交
4792 4793

            then we get a 1-level LoDTensor:
4794
                out.lod =  [[4,                          2]]
Y
yangyaming 已提交
4795 4796 4797 4798 4799 4800
                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:
4801
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
4802 4803 4804 4805
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a Tensor:
4806
                y.data = [[2, 4]]
Y
yangyaming 已提交
4807 4808 4809
                y.dims = [1, 3]

            then we get a 1-level LoDTensor:
4810
                out.lod =  [[2,            4]]
Y
yangyaming 已提交
4811 4812 4813 4814 4815 4816
                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:
4817
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
4818 4819 4820 4821
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a 2-level LoDTensor:
4822
                y.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
4823 4824 4825 4826
                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:
4827
                out.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
4828 4829 4830 4831 4832
                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.
4833
        y (Variable|None): If provided, output's LoD would be derived
Y
Yibing Liu 已提交
4834
                           from :attr:`y`.
Y
yangyaming 已提交
4835
        target_lod (list|tuple|None): One level LoD which should be considered
Y
Yibing Liu 已提交
4836
                                      as target LoD when :attr:`y` not provided.
Y
yangyaming 已提交
4837 4838

    Returns:
Y
Yibing Liu 已提交
4839
        Variable: Output variable with LoD specified by this layer.
Y
yangyaming 已提交
4840 4841

    Raises:
Y
Yibing Liu 已提交
4842
        ValueError: If :attr:`y` and :attr:`target_lod` are both None.
Y
yangyaming 已提交
4843 4844 4845 4846 4847 4848 4849 4850 4851 4852 4853 4854 4855 4856 4857 4858 4859 4860 4861 4862 4863 4864 4865 4866

    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
4867 4868 4869 4870 4871 4872 4873 4874 4875 4876 4877


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
4878
      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
4879 4880 4881 4882 4883 4884 4885 4886 4887 4888 4889 4890 4891 4892 4893 4894 4895 4896 4897 4898 4899 4900 4901 4902 4903 4904 4905 4906

    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
4907 4908
          data = fluid.layers.data(
              name="data", shape=[3, 112, 112], dtype="float32")
D
dragonwarrior 已提交
4909 4910 4911 4912 4913 4914 4915 4916 4917 4918 4919 4920 4921 4922 4923 4924 4925 4926 4927 4928 4929 4930 4931 4932 4933 4934 4935
          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
4936 4937 4938 4939


def pad(x, paddings, pad_value=0., name=None):
    """
G
guosheng 已提交
4940
    Pads a tensor with a constant value given by :attr:`pad_value`, and the
W
wanghaoshuang 已提交
4941
    padded width is specified by :attr:`paddings`.
G
guosheng 已提交
4942

G
guosheng 已提交
4943 4944 4945 4946
    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
4947 4948 4949 4950 4951 4952 4953 4954 4955 4956 4957 4958 4959 4960 4961 4962 4963 4964 4965 4966 4967 4968

    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
4969
                         The length of :attr:paddings must be
G
guosheng 已提交
4970 4971 4972 4973 4974 4975 4976 4977 4978 4979
                         :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
4980

G
guosheng 已提交
4981 4982 4983 4984 4985 4986 4987 4988 4989 4990 4991 4992 4993 4994
            # 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
4995 4996


C
chengduo 已提交
4997 4998 4999 5000 5001 5002 5003 5004 5005 5006 5007 5008 5009 5010 5011 5012 5013 5014 5015 5016 5017 5018 5019 5020 5021 5022 5023 5024 5025 5026 5027 5028 5029 5030 5031 5032 5033 5034 5035 5036 5037 5038 5039 5040 5041 5042 5043 5044 5045 5046 5047 5048 5049 5050 5051 5052 5053 5054 5055 5056 5057 5058 5059 5060 5061 5062 5063 5064 5065 5066 5067 5068 5069 5070 5071 5072 5073 5074 5075 5076
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


5077 5078 5079 5080 5081 5082 5083
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
5084 5085
    called label-smoothing regularization (LSR).

5086 5087 5088 5089 5090 5091 5092 5093 5094 5095 5096 5097 5098 5099 5100 5101 5102 5103 5104 5105 5106 5107 5108
    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
5109
                              be :math:`(1, class\_num)`.
5110 5111
        epsilon(float): The weight used to mix up the original ground-truth
                        distribution and the fixed distribution.
5112
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32,
5113 5114 5115 5116 5117 5118 5119 5120 5121 5122 5123 5124 5125 5126 5127 5128 5129 5130 5131 5132 5133 5134 5135 5136 5137 5138 5139
                                                  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
5140 5141


Y
yi.wu 已提交
5142
@templatedoc()
5143 5144
def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0):
    """
Y
yi.wu 已提交
5145
    ${comment}
5146 5147

    Args:
Y
yi.wu 已提交
5148 5149
        input (Variable): ${x_comment}
        rois (Variable): ROIs (Regions of Interest) to pool over.
Y
yi.wu 已提交
5150 5151 5152
        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
5153 5154

    Returns:
Y
update  
yi.wu 已提交
5155
        Variable: ${out_comment}.
5156 5157

    Examples:
5158 5159
        .. code-block:: python

5160
            pool_out = fluid.layers.roi_pool(input=x, rois=rois, 7, 7, 1.0)
5161 5162 5163 5164 5165 5166 5167 5168 5169 5170 5171 5172 5173 5174 5175 5176 5177
    """
    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
5178 5179 5180 5181 5182 5183 5184 5185 5186 5187 5188 5189 5190 5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201 5202 5203 5204 5205


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:
5206 5207
        .. code-block:: python

W
whs 已提交
5208 5209 5210 5211
            predictions = fluid.layers.softmax(x)
            loss = fluid.layers.dice_loss(input=predictions, label=label, 2)
    """
    label = one_hot(label, depth=input.shape[-1])
5212
    reduce_dim = list(range(1, len(input.shape)))
W
whs 已提交
5213 5214 5215 5216 5217 5218
    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)
5219 5220


5221 5222 5223 5224 5225
def image_resize(input,
                 out_shape=None,
                 scale=None,
                 name=None,
                 resample='BILINEAR'):
5226
    """
Q
qiaolongfei 已提交
5227
    **Resize a Batch of Images**
F
stash  
fengjiayi 已提交
5228

5229
    The input must be a tensor of the shape (num_batches, channels, in_h, in_w),
5230 5231 5232
    and the resizing only applies on the last two dimensions(hight and width).

    Supporting resample methods:
Q
update  
qiaolongfei 已提交
5233

5234
        'BILINEAR' : Bilinear interpolation
F
stash  
fengjiayi 已提交
5235

5236
    Args:
5237
        input (Variable): The input tensor of image resize layer,
5238 5239
                          This is a 4-D tensor of the shape
                          (num_batches, channels, in_h, in_w).
5240
        out_shape(list|tuple|Variable|None): Output shape of image resize
5241 5242
                                    layer, the shape is (out_h, out_w).
                                    Default: None
B
baiyf 已提交
5243
        scale(float|None): The multiplier for the input height or width.
5244 5245 5246
                         At least one of out_shape or scale must be set.
                         And out_shape has a higher priority than scale.
                         Default: None
5247 5248
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
5249 5250
        resample(str): The resample method. It can only be 'BILINEAR' currently.
                       Default: 'BILINEAR'
5251 5252

    Returns:
Q
update  
qiaolongfei 已提交
5253 5254
        Variable: The output is a 4-D tensor of the shape
        (num_batches, channls, out_h, out_w).
F
stash  
fengjiayi 已提交
5255

5256 5257 5258
    Examples:
        .. code-block:: python

5259
            out = fluid.layers.image_resize(input, out_shape=[12, 12])
5260
    """
5261 5262 5263 5264
    resample_methods = {'BILINEAR': 'bilinear_interp'}
    if resample not in resample_methods:
        raise ValueError(
            "The 'resample' of image_resize can only be 'BILINEAR' currently.")
5265 5266
    if out_shape is None and scale is None:
        raise ValueError("One of out_shape and scale must not be None")
5267 5268
    helper = LayerHelper('bilinear_interp', **locals())
    dtype = helper.input_dtype()
5269 5270 5271 5272

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

5273 5274 5275
    out_h = 0
    out_w = 0
    inputs = {"X": input}
5276
    if out_shape is not None:
B
baiyf 已提交
5277 5278 5279
        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')
5280 5281 5282 5283 5284 5285
        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
5286 5287 5288 5289
    else:
        out_h = int(input.shape[2] * scale)
        out_w = int(input.shape[3] * scale)

5290 5291
    out = helper.create_tmp_variable(dtype)
    helper.append_op(
5292
        type=resample_methods[resample],
5293
        inputs=inputs,
5294 5295 5296 5297
        outputs={"Out": out},
        attrs={"out_h": out_h,
               "out_w": out_w})
    return out
F
stash  
fengjiayi 已提交
5298 5299


Y
yuyang18 已提交
5300
@templatedoc(op_type="bilinear_interp")
5301 5302
def resize_bilinear(input, out_shape=None, scale=None, name=None):
    """
Y
yuyang18 已提交
5303 5304 5305 5306 5307 5308
    ${comment}

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

        out_shape(${out_size_type}): ${out_size_comment}.
5309

Y
yuyang18 已提交
5310 5311 5312 5313 5314 5315 5316 5317
        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}.
5318 5319 5320 5321 5322 5323 5324
    """

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


def image_resize_short(input, out_short_len, resample='BILINEAR'):
    """
5325 5326 5327
    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
5328 5329 5330 5331 5332 5333 5334
    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.
5335
        resample (str): resample method, default: BILINEAR.
F
fengjiayi 已提交
5336

5337
    Returns:
Q
update  
qiaolongfei 已提交
5338
        Variable: The output is a 4-D tensor of the shape
5339
        (num_batches, channls, out_h, out_w).
5340 5341 5342 5343 5344 5345 5346 5347 5348 5349
    """
    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
5350 5351 5352
    out_shape[long_idx] = int(
        float(out_shape[long_idx]) * (float(out_short_len) / float(hw[
            short_idx])) + 0.5)
5353 5354 5355
    return image_resize(input=input, out_shape=out_shape, resample=resample)


W
whs 已提交
5356 5357
def gather(input, index):
    """
Q
qiaolongfei 已提交
5358 5359
    **Gather Layer**

5360
    Output is obtained by gathering entries of the outer-most dimension
W
whs 已提交
5361 5362 5363 5364
    of X indexed by `index` and concatenate them together.

    .. math::

5365
        Out = X[Index]
W
whs 已提交
5366 5367 5368 5369 5370 5371 5372


    .. code-block:: text


                Given:

5373 5374
                X = [[1, 2],
                     [3, 4],
W
whs 已提交
5375 5376 5377 5378 5379 5380 5381 5382 5383 5384
                     [5, 6]]

                Index = [1, 2]

                Then:

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

    Args:
5385
        input (Variable): The source input with rank>=1.
W
whs 已提交
5386 5387 5388 5389 5390 5391
        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
5392

W
whs 已提交
5393 5394 5395 5396 5397 5398 5399 5400 5401 5402 5403 5404 5405 5406 5407
        .. 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


5408 5409 5410 5411 5412 5413 5414 5415 5416 5417 5418 5419 5420 5421 5422 5423 5424 5425 5426 5427 5428 5429 5430 5431 5432 5433 5434 5435 5436 5437 5438 5439 5440 5441 5442 5443 5444 5445 5446 5447 5448
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
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 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 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
5509 5510 5511 5512 5513 5514 5515 5516 5517 5518 5519 5520 5521
@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}
5522

5523 5524 5525
    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
5526
    """
F
stash  
fengjiayi 已提交
5527
    helper = LayerHelper("random_crop", **locals())
F
fengjiayi 已提交
5528
    dtype = x.dtype
F
stash  
fengjiayi 已提交
5529
    out = helper.create_tmp_variable(dtype)
Y
yuyang18 已提交
5530
    if seed is None:
5531
        seed = np.random.randint(-65536, 65536)
F
fengjiayi 已提交
5532
    op_attrs = {"shape": shape}
F
stash  
fengjiayi 已提交
5533
    if isinstance(seed, int):
F
fengjiayi 已提交
5534 5535 5536 5537 5538
        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
5539 5540 5541 5542
    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
5543
        inputs={"X": x,
F
stash  
fengjiayi 已提交
5544 5545
                "Seed": seed},
        outputs={"Out": out,
F
fengjiayi 已提交
5546 5547
                 "SeedOut": seed},
        attrs=op_attrs)
F
stash  
fengjiayi 已提交
5548
    return out
W
whs 已提交
5549 5550


5551
def log(x, name=None):
W
wanghaoshuang 已提交
5552 5553 5554 5555 5556
    """
    Calculates the natural log of the given input tensor, element-wise.

    .. math::

5557
        Out = \\ln(x)
W
wanghaoshuang 已提交
5558 5559

    Args:
5560
        x (Variable): Input tensor.
5561 5562
        name (str|None, default None): A name for this layer If set None,
            the layer will be named automatically.
W
wanghaoshuang 已提交
5563 5564 5565 5566 5567 5568 5569 5570

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

    Examples:

        .. code-block:: python

5571
            output = fluid.layers.log(x)
W
wanghaoshuang 已提交
5572 5573
    """
    helper = LayerHelper('log', **locals())
W
wanghaoshuang 已提交
5574
    dtype = helper.input_dtype(input_param_name='x')
W
wanghaoshuang 已提交
5575
    out = helper.create_tmp_variable(dtype)
W
wanghaoshuang 已提交
5576
    helper.append_op(type="log", inputs={"X": x}, outputs={"Out": out})
W
wanghaoshuang 已提交
5577 5578 5579
    return out


5580
def relu(x, name=None):
W
wanghaoshuang 已提交
5581 5582
    """
    Relu takes one input data (Tensor) and produces one output data (Tensor)
5583
    where the rectified linear function, y = max(0, x), is applied to
W
wanghaoshuang 已提交
5584 5585 5586 5587
    the tensor elementwise.

    .. math::

5588
        Out = \\max(0, x)
W
wanghaoshuang 已提交
5589 5590

    Args:
5591
        x (Variable): The input tensor.
5592 5593
        name (str|None, default None): A name for this layer If set None,
            the layer will be named automatically.
W
wanghaoshuang 已提交
5594 5595 5596 5597 5598 5599 5600 5601

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

    Examples:

        .. code-block:: python

5602
            output = fluid.layers.relu(x)
W
wanghaoshuang 已提交
5603 5604
    """
    helper = LayerHelper('relu', **locals())
W
wanghaoshuang 已提交
5605
    dtype = helper.input_dtype(input_param_name='x')
W
wanghaoshuang 已提交
5606
    out = helper.create_tmp_variable(dtype)
W
wanghaoshuang 已提交
5607
    helper.append_op(type="relu", inputs={"X": x}, outputs={"Out": out})
W
wanghaoshuang 已提交
5608
    return out
5609 5610


W
whs 已提交
5611 5612 5613
def mean_iou(input, label, num_classes):
    """
    Mean Intersection-Over-Union is a common evaluation metric for
5614 5615 5616 5617
    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
5618
    .. math::
5619 5620

        IOU = \\frac{true\_positiv}{(true\_positive + false\_positive + false\_negative)}.
W
whs 已提交
5621

5622
    The predictions are accumulated in a confusion matrix and mean-IOU
W
whs 已提交
5623 5624 5625 5626 5627
    is then calculated from it.


    Args:
        input (Variable): A Tensor of prediction results for semantic labels with type int32 or int64.
5628
        label (Variable): A Tensor of ground truth labels with type int32 or int64.
W
whs 已提交
5629
                           Its shape should be the same as input.
5630
        num_classes (int): The possible number of labels.
W
whs 已提交
5631 5632 5633 5634

    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.
5635
        out_correct(Variable): A Tensor with shape [num_classes]. The correct numbers of each class.
W
whs 已提交
5636 5637 5638 5639

    Examples:

        .. code-block:: python
5640

W
whs 已提交
5641 5642 5643 5644 5645 5646 5647 5648 5649
            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
5650 5651
        inputs={"Predictions": input,
                "Labels": label},
W
whs 已提交
5652
        outputs={
W
whs 已提交
5653 5654 5655
            "OutMeanIou": out_mean_iou,
            "OutWrong": out_wrong,
            "OutCorrect": out_correct
W
whs 已提交
5656 5657 5658
        },
        attrs={"num_classes": num_classes})
    return out_mean_iou, out_wrong, out_correct
5659 5660 5661 5662 5663 5664 5665 5666 5667 5668 5669 5670 5671 5672 5673 5674 5675 5676 5677 5678 5679 5680 5681 5682 5683 5684 5685 5686 5687 5688 5689 5690 5691 5692 5693 5694 5695 5696 5697 5698 5699 5700 5701 5702 5703 5704 5705 5706 5707 5708 5709 5710 5711 5712 5713 5714 5715 5716 5717 5718 5719 5720 5721 5722 5723 5724 5725 5726 5727 5728 5729 5730 5731 5732


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
5733
                    isinstance(shape, Variable)):
5734 5735 5736 5737 5738 5739 5740 5741 5742 5743 5744 5745 5746 5747 5748 5749 5750 5751 5752 5753 5754 5755 5756
        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
5757 5758 5759 5760 5761 5762 5763 5764 5765 5766


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
5767

5768 5769
    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
5770

5771 5772 5773 5774
    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
5775

5776 5777 5778 5779 5780
    $$
      C_{i,j} = -\tilde{P_{ij}} * o_{i,j} + \log(1 + e^{o_{i,j}}) \\
      o_{i,j} =  o_i - o_j  \\
      \tilde{P_{i,j}} = \left \{0, 0.5, 1 \right \} \ or \ \left \{0, 1 \right \}
    $$
M
minqiyang 已提交
5781 5782 5783

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

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 5812 5813 5814 5815 5816 5817 5818 5819 5820 5821 5822 5823 5824 5825 5826 5827
    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
5828 5829


W
whs 已提交
5830 5831 5832 5833 5834 5835 5836 5837 5838 5839 5840 5841 5842 5843
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
5844

W
whs 已提交
5845 5846
      X = [[1, 2, 3],
           [4, 5, 6]]
M
minqiyang 已提交
5847

W
whs 已提交
5848
      Case 0:
M
minqiyang 已提交
5849

W
whs 已提交
5850 5851 5852
        paddings = [0, 1, 2, 3],
        mode = 'constant'
        pad_value = 0
M
minqiyang 已提交
5853

W
whs 已提交
5854 5855 5856
        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
5857

W
whs 已提交
5858
      Case 1:
M
minqiyang 已提交
5859

W
whs 已提交
5860 5861
        paddings = [0, 1, 2, 1],
        mode = 'reflect'
M
minqiyang 已提交
5862

W
whs 已提交
5863 5864 5865
        Out = [[3, 2, 1, 2, 3, 2]
               [6, 5, 4, 5, 6, 5]
               [3, 2, 1, 2, 3, 2]]
M
minqiyang 已提交
5866

W
whs 已提交
5867
      Case 2:
M
minqiyang 已提交
5868

W
whs 已提交
5869 5870
        paddings = [0, 1, 2, 1],
        mode = 'edge'
M
minqiyang 已提交
5871

W
whs 已提交
5872 5873 5874
        Out = [[1, 1, 1, 2, 3, 3]
               [4, 4, 4, 5, 6, 6]
               [4, 4, 4, 5, 6, 6]]
M
minqiyang 已提交
5875 5876


W
whs 已提交
5877 5878 5879 5880 5881 5882 5883 5884 5885 5886 5887 5888 5889 5890 5891 5892 5893 5894 5895 5896 5897 5898 5899 5900 5901 5902 5903 5904 5905 5906 5907 5908 5909 5910 5911 5912 5913 5914 5915 5916 5917
    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


5918 5919 5920 5921 5922 5923 5924 5925 5926 5927 5928 5929 5930 5931 5932 5933 5934 5935 5936 5937 5938 5939 5940 5941 5942 5943 5944 5945 5946 5947 5948 5949 5950 5951 5952 5953 5954 5955 5956 5957 5958 5959 5960 5961 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 6036 6037 6038 6039 6040 6041 6042 6043 6044 6045 6046 6047 6048 6049 6050 6051 6052 6053 6054 6055 6056 6057 6058 6059
@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
6060 6061 6062 6063 6064 6065 6066 6067 6068 6069 6070 6071 6072 6073
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
6074
	name(str|None): A name for this layer(optional). If set None, the layer
M
minqiyang 已提交
6075
                        will be named automatically.
J
jerrywgz 已提交
6076 6077 6078 6079 6080 6081 6082 6083 6084 6085 6086 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

    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


6113 6114 6115 6116 6117 6118 6119 6120 6121 6122 6123 6124 6125 6126 6127 6128 6129 6130 6131 6132 6133 6134 6135 6136 6137 6138 6139 6140 6141 6142 6143 6144 6145 6146 6147 6148 6149 6150 6151 6152 6153 6154 6155 6156 6157 6158 6159 6160 6161 6162 6163 6164 6165 6166 6167 6168 6169 6170 6171 6172 6173 6174 6175 6176 6177 6178 6179 6180
@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


6181 6182 6183 6184 6185 6186 6187 6188 6189 6190 6191 6192 6193
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)
6194

6195 6196 6197 6198 6199 6200 6201 6202 6203 6204
    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.
6205 6206
        axis (int): Indicate up to which input dimensions (exclusive) should
                    be flattened to the outer dimension of the output.
6207 6208 6209 6210 6211 6212 6213 6214 6215 6216 6217 6218 6219 6220 6221
                    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.
6222
        ValueError: If axis is not in range [0, rank(x)].
6223 6224 6225 6226 6227 6228 6229 6230 6231 6232 6233 6234 6235 6236 6237 6238 6239

    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)
6240
    x_shape = helper.create_tmp_variable(x.dtype)
6241
    helper.append_op(
6242
        type='flatten2',
6243
        inputs={"X": x},
6244 6245
        outputs={'Out': out,
                 'XShape': x_shape},
6246 6247
        attrs={"axis": axis})
    return out
X
Xin Pan 已提交
6248 6249


C
chenweihang 已提交
6250
def sequence_enumerate(input, win_size, pad_value=0, name=None):
C
chenweihang 已提交
6251
    """
C
chenweihang 已提交
6252
    Generate a new sequence for the input index sequence, which enumerates all the
M
minqiyang 已提交
6253
    sub-sequences with length `win_size` of the input.
C
chenweihang 已提交
6254 6255
    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
6256

C
chenweihang 已提交
6257 6258 6259 6260
    Examples:
    Case 1:
      Input:
        X.lod = [[0, 3, 5]]
6261
        X.data = [[1], [2], [3], [4], [5]]
C
chenweihang 已提交
6262 6263 6264 6265 6266 6267
        X.dims = [5, 1]
      Attrs:
        win_size = 2
        pad_value = 0
      Output:
        Out.lod = [[0, 3, 5]]
6268
        Out.data = [[1, 2], [2, 3], [3, 0], [4, 5], [5, 0]]
C
chenweihang 已提交
6269 6270 6271
        Out.dims = [5, 2]

    Args:
C
chenweihang 已提交
6272 6273 6274
        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
6275 6276 6277 6278 6279 6280 6281 6282 6283 6284 6285

    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
6286
    out = helper.create_tmp_variable(helper.input_dtype(), stop_gradient=True)
C
chenweihang 已提交
6287 6288 6289 6290 6291 6292
    helper.append_op(
        type='sequence_enumerate',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={'win_size': win_size,
               'pad_value': pad_value})
6293

6294

S
sneaxiy 已提交
6295 6296 6297 6298 6299 6300 6301 6302 6303
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:
6304

S
sneaxiy 已提交
6305
    .. math::
6306

S
sneaxiy 已提交
6307 6308 6309
        y(i_1, i_2,..., i_n, j) = (j < x(i_1, i_2,..., i_n))

    Args:
6310
        x (Variable): Input tensor of sequence_mask layer,
S
sneaxiy 已提交
6311 6312 6313 6314
                      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.
6315 6316 6317
        name (str|None): A name for this layer(optional). If set None, the
                         layer will be named automatically.

S
sneaxiy 已提交
6318 6319
    Returns:
        Variable: The output sequence mask.
6320

S
sneaxiy 已提交
6321 6322
    """

Q
qingqing01 已提交
6323
    helper = LayerHelper('sequence_mask', **locals())
S
sneaxiy 已提交
6324 6325 6326 6327 6328
    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
6329 6330 6331
    helper.append_op(
        type='sequence_mask',
        inputs={'X': [x]},
S
sneaxiy 已提交
6332 6333
        outputs={'Y': out},
        attrs={
6334
            'maxlen': maxlen if maxlen is not None else -1,
S
sneaxiy 已提交
6335 6336 6337
            'out_dtype': out.dtype
        })
    return out
S
sneaxiy 已提交
6338 6339


X
Xin Pan 已提交
6340
def stack(x, axis=0):
S
sneaxiy 已提交
6341 6342 6343 6344
    """
    **Stack Layer**

    This layer stacks all of the input :code:`x` along axis.
6345 6346 6347 6348 6349 6350 6351

    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
6352
    If :code:`axis` < 0, it would be replaced with :code:`axis+rank(x[0])+1`.
6353
    If :code:`axis` is None, it would be replaced with 0.
S
sneaxiy 已提交
6354 6355

    Args:
6356
        x (Variable|list(Variable)|tuple(Variable)): Input variables.
S
sneaxiy 已提交
6357
        axis (int|None): The axis along which all inputs are stacked.
6358

S
sneaxiy 已提交
6359 6360
    Returns:
        Variable: The stacked variable.
6361

S
sneaxiy 已提交
6362 6363
    """

X
Xin Pan 已提交
6364 6365 6366 6367 6368 6369 6370 6371
    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
6372 6373
        type='stack', inputs={'X': x}, outputs={'Y': out},
        attrs={'axis': axis})
6374

X
Xin Pan 已提交
6375
    return out
D
dzhwinter 已提交
6376 6377 6378 6379 6380 6381 6382


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
6383

D
dzhwinter 已提交
6384 6385 6386
    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
6387
    raised.
D
dzhwinter 已提交
6388 6389

    Args:
M
minqiyang 已提交
6390
        x (Variable): Input variable.
D
dzhwinter 已提交
6391 6392
        axis (int): The axis along which the input is unstacked.
        num (int|None): The number of output variables.
M
minqiyang 已提交
6393

D
dzhwinter 已提交
6394 6395
    Returns:
        list(Variable): The unstacked variables.
M
minqiyang 已提交
6396

D
dzhwinter 已提交
6397 6398 6399 6400 6401 6402 6403 6404 6405 6406 6407 6408 6409 6410 6411 6412 6413 6414 6415 6416
    """

    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
6417 6418 6419 6420 6421 6422 6423 6424 6425 6426 6427 6428


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
6429

W
whs 已提交
6430 6431 6432 6433
                [
                   [[1], [2], [3]],
                   [[4], [5], [6]]
                ]
M
minqiyang 已提交
6434

W
whs 已提交
6435
        Attr(expand_times):  [1, 2, 2]
M
minqiyang 已提交
6436

W
whs 已提交
6437
        Output(Out) is a 3-D tensor with shape [2, 6, 2]:
M
minqiyang 已提交
6438

W
whs 已提交
6439 6440 6441 6442
                [
                    [[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
6443

W
whs 已提交
6444 6445 6446 6447 6448 6449 6450 6451 6452 6453 6454 6455 6456 6457 6458 6459 6460 6461 6462 6463 6464 6465 6466
    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
6467 6468


G
fix  
gongweibao 已提交
6469 6470 6471
from paddle.fluid.framework import convert_np_dtype_to_dtype_


G
gongweibao 已提交
6472
@templatedoc()
G
fix  
gongweibao 已提交
6473 6474 6475 6476 6477 6478 6479 6480 6481
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
6482
    ${comment}
G
fix  
gongweibao 已提交
6483 6484

    Args:
G
gongweibao 已提交
6485 6486 6487
        input (Variable): ${input_comment}
        shape (tuple|list): ${shape_comment}
        input_dim_idx (Int): ${input_dim_idx_comment}
G
gongweibao 已提交
6488
        output_dim_idx (Int): ${output_dim_idx_comment}
G
gongweibao 已提交
6489 6490 6491
        min (Float): ${min_comment}
        max (Float): ${max_comment}
        seed (Int): ${seed_comment}
G
fix  
gongweibao 已提交
6492 6493
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc
    Returns:
G
gongweibao 已提交
6494
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
6495 6496 6497 6498 6499 6500 6501 6502 6503 6504 6505 6506 6507 6508 6509 6510 6511 6512 6513 6514 6515

    """

    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
6516 6517


G
gongweibao 已提交
6518
@templatedoc()
X
Xin Pan 已提交
6519
def gaussian_random(shape, mean=0.0, std=1.0, seed=0, dtype='float32'):
G
fix  
gongweibao 已提交
6520
    """
G
gongweibao 已提交
6521
    ${comment}
G
fix  
gongweibao 已提交
6522 6523

    Args:
G
gongweibao 已提交
6524 6525 6526 6527
        shape (tuple|list): ${shape_comment}
        mean (Float): ${mean_comment}
        std (Float): ${std_comment}
        seed (Int): ${seed_comment}
G
fix  
gongweibao 已提交
6528 6529 6530
        dtype(np.dtype|core.VarDesc.VarType|str): Output data type.

    Returns:
G
gongweibao 已提交
6531
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
6532 6533 6534 6535 6536 6537 6538 6539 6540 6541 6542 6543 6544 6545 6546

    """

    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
6547
            'use_mkldnn': False
G
fix  
gongweibao 已提交
6548 6549 6550 6551 6552
        })

    return out


G
gongweibao 已提交
6553
@templatedoc()
G
fix  
gongweibao 已提交
6554
def sampling_id(x, min=0.0, max=1.0, seed=0, dtype='float32'):
G
fix  
gongweibao 已提交
6555
    """
G
gongweibao 已提交
6556
    ${comment}
G
fix  
gongweibao 已提交
6557 6558

    Args:
G
gongweibao 已提交
6559 6560 6561 6562
        x (Variable): ${x_comment}
        min (Float): ${min_comment}
        max (Float): ${max_comment}
        seed (Float): ${seed_comment}
G
fix  
gongweibao 已提交
6563
        dtype(np.dtype|core.VarDesc.VarType|str): The type of output data : float32, float_16, int etc
G
fix  
gongweibao 已提交
6564 6565

    Returns:
G
gongweibao 已提交
6566
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
6567 6568 6569 6570

    """

    helper = LayerHelper('sampling_id', **locals())
G
fix  
gongweibao 已提交
6571
    out = helper.create_tmp_variable(dtype)
G
fix  
gongweibao 已提交
6572 6573 6574 6575 6576 6577 6578 6579 6580 6581 6582
    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
6583
@templatedoc()
G
fix  
gongweibao 已提交
6584 6585 6586 6587 6588 6589 6590 6591 6592
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
6593
    ${comment}
G
fix  
gongweibao 已提交
6594 6595

    Args:
G
gongweibao 已提交
6596 6597
        input (Variable): ${input_comment}
        shape (tuple|list): ${shape_comment}
G
gongweibao 已提交
6598
        input_dim_idx (Int): ${input_dim_idx_comment}
G
gongweibao 已提交
6599 6600 6601 6602
        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
6603
        dtype(np.dtype|core.VarDesc.VarType|str): The type of output data : float32, float_16, int etc
G
fix  
gongweibao 已提交
6604 6605

    Returns:
G
gongweibao 已提交
6606
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
6607 6608 6609 6610 6611 6612 6613 6614 6615 6616 6617 6618 6619 6620 6621 6622 6623 6624 6625 6626 6627 6628
    """

    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
6629
@templatedoc()
X
Xin Pan 已提交
6630
def sum(x):
G
fix  
gongweibao 已提交
6631
    """
G
gongweibao 已提交
6632
    ${comment}
G
fix  
gongweibao 已提交
6633 6634

    Args:
G
gongweibao 已提交
6635
        x (Variable): ${x_comment}
G
fix  
gongweibao 已提交
6636 6637

    Returns:
G
gongweibao 已提交
6638
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
6639 6640 6641
    """

    helper = LayerHelper('sum', **locals())
G
fix  
gongweibao 已提交
6642
    out = helper.create_tmp_variable(dtype=helper.input_dtype('x'))
G
fix  
gongweibao 已提交
6643 6644 6645 6646
    helper.append_op(
        type='sum',
        inputs={'X': x},
        outputs={'Out': out},
X
Xin Pan 已提交
6647
        attrs={'use_mkldnn': False})
G
fix  
gongweibao 已提交
6648 6649 6650 6651

    return out


G
gongweibao 已提交
6652
@templatedoc()
G
fix  
gongweibao 已提交
6653 6654
def slice(input, axes, starts, ends):
    """
G
gongweibao 已提交
6655
    ${comment}
G
fix  
gongweibao 已提交
6656 6657

    Args:
G
gongweibao 已提交
6658 6659 6660 6661
        input (Variable): ${input_comment}.
        axes (List): ${axes_comment}
        starts (List): ${starts_comment}
        ends (List): ${ends_comment}
G
fix  
gongweibao 已提交
6662 6663

    Returns:
G
gongweibao 已提交
6664
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
6665 6666 6667 6668

    """

    helper = LayerHelper('slice', **locals())
G
fix  
gongweibao 已提交
6669
    out = helper.create_tmp_variable(dtype=helper.input_dtype('input'))
G
fix  
gongweibao 已提交
6670 6671 6672 6673 6674 6675 6676 6677 6678 6679 6680
    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
6681
@templatedoc()
G
fix  
gongweibao 已提交
6682 6683
def shape(input):
    """
G
gongweibao 已提交
6684
    ${comment}
G
fix  
gongweibao 已提交
6685 6686

    Args:
G
gongweibao 已提交
6687
        input (Variable): ${input_comment}
G
fix  
gongweibao 已提交
6688 6689

    Returns:
G
gongweibao 已提交
6690
        out (Variable): ${out_comment}
G
fix  
gongweibao 已提交
6691 6692 6693 6694

    """

    helper = LayerHelper('shape', **locals())
G
fix  
gongweibao 已提交
6695
    out = helper.create_tmp_variable(dtype=helper.input_dtype('input'))
G
fix  
gongweibao 已提交
6696
    helper.append_op(
G
fix  
gongweibao 已提交
6697
        type='shape', inputs={'Input': input}, outputs={'Out': out})
G
fix  
gongweibao 已提交
6698 6699

    return out
G
merge  
gongweibao 已提交
6700 6701


S
sneaxiy 已提交
6702 6703 6704 6705 6706 6707 6708 6709
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
6710 6711 6712 6713 6714 6715
    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
6716

S
sneaxiy 已提交
6717 6718 6719 6720 6721 6722 6723 6724 6725 6726 6727
    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
6728
def scale(x, scale=1.0, bias=0.0, bias_after_scale=True, act=None, name=None):
S
sneaxiy 已提交
6729 6730 6731 6732 6733 6734 6735 6736
    """
    ${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
6737
        act(basestring|None): Activation applied to the output.
M
minqiyang 已提交
6738
        name(basestring|None): Name of the output.
S
sneaxiy 已提交
6739 6740 6741 6742 6743 6744

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

    helper = LayerHelper('scale', **locals())
S
sneaxiy 已提交
6745 6746 6747 6748 6749
    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
6750 6751 6752 6753 6754 6755 6756 6757 6758 6759

    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
6760
    return helper.append_activation(out)
S
sneaxiy 已提交
6761 6762


X
Xin Pan 已提交
6763
def elementwise_add(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
6764 6765 6766
    return _elementwise_op(LayerHelper('elementwise_add', **locals()))


X
Xin Pan 已提交
6767
def elementwise_div(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
6768 6769 6770
    return _elementwise_op(LayerHelper('elementwise_div', **locals()))


X
Xin Pan 已提交
6771
def elementwise_sub(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
6772 6773 6774
    return _elementwise_op(LayerHelper('elementwise_sub', **locals()))


X
Xin Pan 已提交
6775
def elementwise_mul(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
6776 6777 6778
    return _elementwise_op(LayerHelper('elementwise_mul', **locals()))


X
Xin Pan 已提交
6779
def elementwise_max(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
6780 6781 6782
    return _elementwise_op(LayerHelper('elementwise_max', **locals()))


X
Xin Pan 已提交
6783
def elementwise_min(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
6784 6785 6786
    return _elementwise_op(LayerHelper('elementwise_min', **locals()))


X
Xin Pan 已提交
6787
def elementwise_pow(x, y, axis=-1, act=None, name=None):
S
sneaxiy 已提交
6788 6789 6790 6791 6792 6793 6794 6795 6796 6797 6798
    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
6799 6800
            "act (basestring|None): Activation applied to the output.",
            "name (basestring|None): Name of the output."
S
sneaxiy 已提交
6801
        ])
M
minqiyang 已提交
6802 6803


6804
def _logical_op(op_name, x, y, out=None, name=None, binary_op=True):
M
minqiyang 已提交
6805 6806
    helper = LayerHelper(op_name, **locals())

M
minqiyang 已提交
6807 6808
    if binary_op:
        assert x.dtype == y.dtype
M
minqiyang 已提交
6809 6810 6811 6812 6813 6814 6815 6816 6817 6818 6819 6820 6821 6822 6823 6824 6825 6826 6827

    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()
6828
def logical_and(x, y, out=None, name=None):
M
minqiyang 已提交
6829 6830 6831 6832 6833 6834 6835 6836 6837 6838 6839 6840 6841 6842 6843 6844 6845 6846
    """
    ${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()
6847
def logical_or(x, y, out=None, name=None):
M
minqiyang 已提交
6848 6849 6850 6851 6852 6853 6854 6855 6856 6857 6858 6859 6860 6861 6862 6863 6864 6865
    """
    ${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()
6866
def logical_xor(x, y, out=None, name=None):
M
minqiyang 已提交
6867 6868 6869 6870 6871 6872 6873 6874 6875 6876 6877 6878 6879 6880 6881 6882 6883 6884
    """
    ${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()
6885
def logical_not(x, out=None, name=None):
M
minqiyang 已提交
6886 6887 6888 6889 6890 6891 6892 6893 6894 6895 6896 6897 6898 6899
    """
    ${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)
6900 6901 6902 6903 6904 6905 6906 6907 6908 6909 6910 6911 6912 6913 6914 6915 6916 6917 6918 6919 6920 6921 6922 6923 6924 6925 6926 6927 6928 6929 6930 6931 6932 6933 6934 6935 6936 6937 6938 6939 6940 6941 6942 6943 6944 6945 6946 6947 6948 6949 6950 6951 6952 6953 6954 6955 6956 6957 6958 6959 6960 6961 6962 6963


@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
6964 6965 6966 6967 6968 6969 6970 6971 6972 6973 6974 6975 6976 6977 6978 6979 6980 6981 6982 6983 6984 6985 6986 6987 6988 6989 6990 6991 6992 6993 6994 6995 6996 6997 6998 6999 7000 7001 7002 7003 7004 7005 7006 7007 7008 7009 7010 7011 7012 7013 7014 7015 7016 7017 7018 7019 7020 7021


@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
7022 7023
            "x_num_col_dims": x_num_col_dims,
            "y_num_col_dims": y_num_col_dims
X
Xin Pan 已提交
7024 7025 7026 7027 7028 7029 7030 7031 7032 7033 7034 7035 7036 7037 7038 7039 7040 7041 7042 7043 7044 7045 7046 7047 7048 7049 7050 7051 7052 7053 7054 7055 7056 7057 7058 7059 7060 7061 7062 7063 7064 7065 7066 7067 7068 7069 7070 7071 7072 7073 7074 7075 7076 7077 7078 7079 7080 7081 7082 7083 7084 7085 7086
        },
        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