nn.py 195.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
# 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.

#   Copyright (c ) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
16
#
D
dzhwinter 已提交
17 18 19
# 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
D
dzhwinter 已提交
20
#
D
dzhwinter 已提交
21
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
22
#
D
dzhwinter 已提交
23 24 25 26 27
# 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
28
"""
29
All layers just related to the neural network.
Y
Yu Yang 已提交
30 31 32 33 34
"""

from ..layer_helper import LayerHelper
from ..initializer import Normal, Constant
from ..framework import Variable
Y
yangyaming 已提交
35
from ..param_attr import ParamAttr
36 37 38
from .layer_function_generator import autodoc, templatedoc
from .tensor import concat
from . import utils
Y
yuyang18 已提交
39
import random
F
fengjiayi 已提交
40
from .. import unique_name
41
from functools import reduce
Y
Yu Yang 已提交
42 43

__all__ = [
Y
ying 已提交
44 45 46
    'fc',
    'embedding',
    'dynamic_lstm',
Y
Yibing Liu 已提交
47
    'dynamic_lstmp',
G
guosheng 已提交
48
    'dynamic_gru',
Y
ying 已提交
49 50 51 52 53 54 55 56 57
    'gru_unit',
    'linear_chain_crf',
    'crf_decoding',
    'cos_sim',
    'cross_entropy',
    'square_error_cost',
    'chunk_eval',
    'sequence_conv',
    'conv2d',
Y
yuyang18 已提交
58
    'conv3d',
Y
ying 已提交
59
    'sequence_pool',
60 61
    'sequence_softmax',
    'softmax',
Y
ying 已提交
62
    'pool2d',
Y
yuyang18 已提交
63
    'pool3d',
Y
ying 已提交
64 65 66
    'batch_norm',
    'beam_search_decode',
    'conv2d_transpose',
Y
yuyang18 已提交
67
    'conv3d_transpose',
Y
ying 已提交
68 69 70 71 72 73
    'sequence_expand',
    'lstm_unit',
    'reduce_sum',
    'reduce_mean',
    'reduce_max',
    'reduce_min',
74
    'reduce_prod',
Y
ying 已提交
75 76 77 78
    'sequence_first_step',
    'sequence_last_step',
    'dropout',
    'split',
79 80
    'ctc_greedy_decoder',
    'edit_distance',
Y
ying 已提交
81 82
    'l2_normalize',
    'matmul',
Q
qingqing01 已提交
83
    'topk',
Y
ying 已提交
84 85
    'warpctc',
    'sequence_reshape',
86
    'transpose',
87
    'im2sequence',
88
    'nce',
W
weixing02 已提交
89
    'hsigmoid',
Q
Qiao Longfei 已提交
90
    'beam_search',
91
    'row_conv',
92
    'multiplex',
G
guosheng 已提交
93
    'layer_norm',
94 95
    'softmax_with_cross_entropy',
    'smooth_l1',
96
    'one_hot',
Y
Yu Yang 已提交
97
    'autoincreased_step_counter',
C
caoying03 已提交
98
    'reshape',
Y
yangyaming 已提交
99
    'lod_reset',
D
dragonwarrior 已提交
100
    'lrn',
G
guosheng 已提交
101
    'pad',
102
    'label_smooth',
103
    'roi_pool',
W
whs 已提交
104
    'dice_loss',
F
fengjiayi 已提交
105 106
    'image_resize',
    'image_resize_short',
B
baiyf 已提交
107
    'resize_bilinear',
W
whs 已提交
108
    'gather',
109
    'random_crop',
Y
yuyang18 已提交
110 111 112
    'mean_iou',
    'relu',
    'log',
113
    'crop',
114
    'rank_loss',
Y
Yu Yang 已提交
115 116 117 118 119 120 121 122
]


def fc(input,
       size,
       num_flatten_dims=1,
       param_attr=None,
       bias_attr=None,
123
       use_mkldnn=False,
Y
Yu Yang 已提交
124
       act=None,
J
Jacek Czaja 已提交
125
       is_test=False,
126
       name=None):
Y
Yu Yang 已提交
127
    """
128
    **Fully Connected Layer**
Y
Yu Yang 已提交
129

130 131 132 133 134 135 136 137
    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
138
    to the output as well.
C
caoying03 已提交
139

C
caoying03 已提交
140
    This process can be formulated as follows:
141 142 143

    .. math::

144
        Out = Act({\sum_{i=0}^{N-1}X_iW_i + b})
145 146 147

    In the above equation:

C
caoying03 已提交
148 149 150 151
    * :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).
152
    * :math:`Act`: The activation function.
C
caoying03 已提交
153
    * :math:`Out`: The output tensor.
Y
Yu Yang 已提交
154 155

    Args:
R
ranqiu 已提交
156 157 158 159 160 161 162 163 164 165 166 167 168 169 170
        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
171 172
            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
173
        act (str, default None): Activation to be applied to the output of this layer.
J
Jacek Czaja 已提交
174
        is_test(bool): A flag indicating whether execution is in test phase.
M
mozga-intel 已提交
175 176
        use_mkldnn(bool): Use mkldnn kernel or not, it is valid only when the mkldnn
            library is installed. Default: False
R
ranqiu 已提交
177
        name (str, default None): The name of this layer.
Y
Yu Yang 已提交
178

179
    Returns:
F
fengjiayi 已提交
180
        Variable: The transformation result.
181 182

    Raises:
C
caoying03 已提交
183
        ValueError: If rank of the input tensor is less than 2.
184 185 186 187

    Examples:
        .. code-block:: python

F
fengjiayi 已提交
188
          data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
189
          fc = fluid.layers.fc(input=data, size=1000, act="tanh")
Y
Yu Yang 已提交
190
    """
C
caoying03 已提交
191

C
caoying03 已提交
192
    helper = LayerHelper("fc", **locals())
Y
Yu Yang 已提交
193 194 195 196

    dtype = helper.input_dtype()

    mul_results = []
197 198
    for input_var, param_attr in helper.iter_inputs_and_params():
        input_shape = input_var.shape
Y
Yu Yang 已提交
199 200 201
        param_shape = [
            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
        ] + [size]
Y
ying 已提交
202

Y
Yu Yang 已提交
203
        w = helper.create_parameter(
204 205
            attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False)
        tmp = helper.create_tmp_variable(dtype)
206
        helper.append_op(
207 208 209
            type="mul",
            inputs={"X": input_var,
                    "Y": w},
210
            outputs={"Out": tmp},
M
mozga-intel 已提交
211 212
            attrs={"x_num_col_dims": num_flatten_dims,
                   "y_num_col_dims": 1})
213 214 215 216
        mul_results.append(tmp)

    if len(mul_results) == 1:
        pre_bias = mul_results[0]
217
    else:
218 219
        pre_bias = helper.create_tmp_variable(dtype)
        helper.append_op(
220 221 222 223
            type="sum",
            inputs={"X": mul_results},
            outputs={"Out": pre_bias},
            attrs={"use_mkldnn": use_mkldnn})
224 225 226 227
    # 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
228 229


230 231 232
def embedding(input,
              size,
              is_sparse=False,
233
              is_distributed=False,
234 235 236
              padding_idx=None,
              param_attr=None,
              dtype='float32'):
Y
Yu Yang 已提交
237
    """
238 239
    **Embedding Layer**

240
    This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in
241 242
    a lookup table. The result of this lookup is the embedding of each ID in the
    :attr:`input`.
243 244 245

    All the input variables are passed in as local variables to the LayerHelper
    constructor.
Y
Yu Yang 已提交
246 247

    Args:
248 249 250 251 252
        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.
253
        is_distributed(bool): Whether to run lookup table from remote parameter server.
254 255
        padding_idx(int|long|None): If :attr:`None`, it makes no effect to lookup.
            Otherwise the given :attr:`padding_idx` indicates padding the output
256
            with zeros whenever lookup encounters it in :attr:`input`. If
257
            :math:`padding_idx < 0`, the :attr:`padding_idx` to use in lookup is
258 259
            :math:`size[0] + dim`.
        param_attr(ParamAttr): Parameters for this layer
260
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc
Y
Yu Yang 已提交
261

262 263 264
    Returns:
        Variable: The tensor variable storing the embeddings of the \
                  supplied inputs.
Y
Yu Yang 已提交
265

266 267
    Examples:
        .. code-block:: python
Y
Yu Yang 已提交
268

C
chengduoZH 已提交
269
          dict_size = len(dataset.ids)
270
          data = fluid.layers.data(name='ids', shape=[32, 32], dtype='float32')
C
chengduoZH 已提交
271
          fc = fluid.layers.embedding(input=data, size=[dict_size, 16])
Y
Yu Yang 已提交
272 273 274 275 276 277
    """

    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)
278 279
    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
280 281 282 283 284
    helper.append_op(
        type='lookup_table',
        inputs={'Ids': input,
                'W': w},
        outputs={'Out': tmp},
285 286 287 288 289
        attrs={
            'is_sparse': is_sparse,
            'is_distributed': is_distributed,
            'padding_idx': padding_idx
        })
Y
Yu Yang 已提交
290 291 292
    return tmp


Y
yi.wu 已提交
293
@templatedoc(op_type="lstm")
Y
Yu Yang 已提交
294 295
def dynamic_lstm(input,
                 size,
Y
Yancey 已提交
296 297
                 h_0=None,
                 c_0=None,
Y
Yu Yang 已提交
298 299 300 301 302 303 304
                 param_attr=None,
                 bias_attr=None,
                 use_peepholes=True,
                 is_reverse=False,
                 gate_activation='sigmoid',
                 cell_activation='tanh',
                 candidate_activation='tanh',
305 306
                 dtype='float32',
                 name=None):
Y
Yibing Liu 已提交
307
    """
Y
yi.wu 已提交
308
    ${comment}
Y
Yibing Liu 已提交
309 310

    Args:
Y
yi.wu 已提交
311 312
        input (Variable): ${input_comment}
        size (int): 4 * hidden size.
Y
Yancey 已提交
313 314 315 316 317 318 319
        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.

320
        param_attr(ParamAttr|None): The parameter attribute for the learnable
321
                               hidden-hidden weights.
Y
Yibing Liu 已提交
322 323 324

                               - Weights = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}
325 326
                               - The shape is (D x 4D), where D is the hidden
                                 size.
Y
yi.wu 已提交
327
        bias_attr (ParamAttr|None): The bias attribute for the learnable bias
328 329 330
                              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
331

332
                              1. `use_peepholes = False`
Y
yi.wu 已提交
333 334
                                 - Biases = {:math:`b_c, b_i, b_f, b_o`}.
                                 - The shape is (1 x 4D).
335
                              2. `use_peepholes = True`
Y
yi.wu 已提交
336
                                 - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
Y
Yibing Liu 已提交
337
                                                 W_{fc}, W_{oc}`}.
Y
yi.wu 已提交
338
                                 - The shape is (1 x 7D).
Y
yi.wu 已提交
339 340 341 342 343 344 345 346
        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
347 348

    Returns:
Y
Yibing Liu 已提交
349 350
        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
351

Y
Yibing Liu 已提交
352
    Examples:
Y
Yibing Liu 已提交
353 354
        .. code-block:: python

Y
Yibing Liu 已提交
355 356
            hidden_dim = 512
            forward_proj = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
357
                                           act=None, bias_attr=None)
Y
Yibing Liu 已提交
358 359
            forward, _ = fluid.layers.dynamic_lstm(
                input=forward_proj, size=hidden_dim * 4, use_peepholes=False)
Y
Yibing Liu 已提交
360
    """
361

Y
Yu Yang 已提交
362
    helper = LayerHelper('lstm', **locals())
M
minqiyang 已提交
363
    size = size // 4
Y
Yu Yang 已提交
364 365 366 367 368 369 370 371 372 373 374 375
    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
376 377 378 379 380 381 382 383 384 385
    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
386 387 388

    helper.append_op(
        type='lstm',
Y
Yancey 已提交
389
        inputs=inputs,
Y
Yu Yang 已提交
390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405
        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
406 407 408 409 410 411 412 413 414 415 416
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',
417 418
                  dtype='float32',
                  name=None):
Y
Yibing Liu 已提交
419 420 421
    """
    **Dynamic LSTMP Layer**

422 423 424 425 426 427
    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
428 429 430 431 432

    The formula is as follows:

    .. math::

433
        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
434

435
        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
436

437
        \\tilde{c_t} & = act_g(W_{cx}x_t + W_{cr}r_{t-1} + b_c)
Y
Yibing Liu 已提交
438

439
        o_t & = \sigma(W_{ox}x_{t} + W_{or}r_{t-1} + W_{oc}c_t + b_o)
Y
Yibing Liu 已提交
440

441
        c_t & = f_t \odot c_{t-1} + i_t \odot \\tilde{c_t}
Y
Yibing Liu 已提交
442

443
        h_t & = o_t \odot act_h(c_t)
Y
Yibing Liu 已提交
444

445
        r_t & = \overline{act_h}(W_{rh}h_t)
Y
Yibing Liu 已提交
446

Y
Yibing Liu 已提交
447 448 449 450 451 452
    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, \
453
          we use vectors to reprenset these diagonal weight matrices.
Y
Yibing Liu 已提交
454
    * :math:`b`: Denotes bias vectors (e.g. :math:`b_i` is the input gate \
455
          bias vector).
Y
Yibing Liu 已提交
456 457 458
    * :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 \
459
          the same size as the cell output activation vector :math:`h`.
Y
Yibing Liu 已提交
460
    * :math:`h`: The hidden state.
461
    * :math:`r`: The recurrent projection of the hidden state.
Y
Yibing Liu 已提交
462 463
    * :math:`\\tilde{c_t}`: The candidate hidden state, whose \
          computation is based on the current input and previous hidden state.
464
    * :math:`\odot`: The element-wise product of the vectors.
Y
Yibing Liu 已提交
465
    * :math:`act_g` and :math:`act_h`: The cell input and cell output \
466
          activation functions and `tanh` is usually used for them.
Y
Yibing Liu 已提交
467 468
    * :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
469 470 471 472

    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.
473

Y
Yibing Liu 已提交
474 475 476 477 478 479 480 481 482 483 484 485
    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.
486
        param_attr(ParamAttr|None): The parameter attribute for the learnable
Y
Yibing Liu 已提交
487 488
                               hidden-hidden weight and projection weight.

489 490
                               - Hidden-hidden weight = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}.
491 492
                               - 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
493 494
                                 size.
                               - Projection weight = {:math:`W_{rh}`}.
495 496
                               - 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
497 498 499 500 501 502
                              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`}.
503
                                - The shape is (1 x 4D).
Y
Yibing Liu 已提交
504 505 506
                              2. `use_peepholes = True`
                                - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
                                                 W_{fc}, W_{oc}`}.
507
                                - The shape is (1 x 7D).
Y
Yibing Liu 已提交
508 509 510 511 512 513 514 515 516
        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.
517
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
Y
Yibing Liu 已提交
518 519
                              default "tanh".
        proj_activation(str): The activation for projection output.
520
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
Y
Yibing Liu 已提交
521 522
                              default "tanh".
        dtype(str): Data type. Choices = ["float32", "float64"], default "float32".
523 524
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
Y
Yibing Liu 已提交
525 526

    Returns:
527 528 529 530
        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
531 532

    Examples:
533

Y
Yibing Liu 已提交
534 535
        .. code-block:: python

536 537 538 539
            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
540
            hidden_dim, proj_dim = 512, 256
541
            fc_out = fluid.layers.fc(input=emb, size=hidden_dim * 4,
Y
Yibing Liu 已提交
542
                                     act=None, bias_attr=None)
543 544 545
            proj_out, _ = fluid.layers.dynamic_lstmp(input=fc_out,
                                                     size=hidden_dim * 4,
                                                     proj_size=proj_dim,
Y
Yibing Liu 已提交
546 547 548 549
                                                     use_peepholes=False,
                                                     is_reverse=True,
                                                     cell_activation="tanh",
                                                     proj_activation="tanh")
Y
Yibing Liu 已提交
550
    """
551

Y
Yibing Liu 已提交
552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597
    helper = LayerHelper('lstmp', **locals())
    size = size / 4
    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
598 599 600 601 602 603 604 605 606
def dynamic_gru(input,
                size,
                param_attr=None,
                bias_attr=None,
                is_reverse=False,
                gate_activation='sigmoid',
                candidate_activation='tanh',
                h_0=None):
    """
607
    **Gated Recurrent Unit (GRU) Layer**
G
guosheng 已提交
608

609
    Refer to `Empirical Evaluation of Gated Recurrent Neural Networks on
610
    Sequence Modeling <https://arxiv.org/abs/1412.3555>`_ .
611

G
guosheng 已提交
612 613 614 615 616 617 618 619 620
    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)
621

G
guosheng 已提交
622
        h_t & = (1-u_t) \odot h_{t-1} + u_t \odot \\tilde{h_t}
623

G
guosheng 已提交
624
    The :math:`\odot` is the element-wise product of the vectors. :math:`act_g`
625 626
    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
627 628 629 630
    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
631
    to use fully-connect layer before GRU layer.
G
guosheng 已提交
632 633

    Args:
634 635
        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
636
            Variable is a matrix with shape :math:`(T \\times 3D)`, where
637
            :math:`T` is the total time steps in this mini-batch, :math:`D`
G
guosheng 已提交
638 639
            is the hidden size.
        size(int): The dimension of the gru cell.
640
        param_attr(ParamAttr|None): The parameter attribute for the learnable
G
guosheng 已提交
641 642
            hidden-hidden weight matrix. Note:

643
            - The shape of the weight matrix is :math:`(T \\times 3D)`, where
G
guosheng 已提交
644
              :math:`D` is the hidden size.
645
            - All elements in the weight matrix can be divided into two parts.
G
guosheng 已提交
646
              The first part are weights of the update gate and reset gate with
647
              shape :math:`(D \\times 2D)`, and the second part are weights for
G
guosheng 已提交
648
              candidate hidden state with shape :math:`(D \\times D)`.
649
        bias_attr(ParamAttr): The parameter attribute for learnable the
G
guosheng 已提交
650
            hidden-hidden bias.
651
        is_reverse(bool): Whether to compute reversed GRU, default
G
guosheng 已提交
652 653 654
            :attr:`False`.
        gate_activation(str): The activation for update gate and reset gate.
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "sigmoid".
655
        candidate_activation(str): The activation for candidate hidden state.
G
guosheng 已提交
656
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh".
657 658 659 660
        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
661 662

    Returns:
G
guosheng 已提交
663
        Variable: The hidden state of GRU. The shape is :math:`(T \\times D)`, \
664
            and sequence length is the same with the input.
665

G
guosheng 已提交
666
    Examples:
667

G
guosheng 已提交
668 669
        .. code-block:: python

670 671 672 673
            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
674
            hidden_dim = 512
675
            x = fluid.layers.fc(input=emb, size=hidden_dim * 3)
G
guosheng 已提交
676 677 678 679 680 681 682 683 684 685
            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
686
    batch_size = input.shape[0]
G
guosheng 已提交
687 688 689
    inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
    if h_0 != None:
        assert h_0.shape == (
Y
Yancey 已提交
690 691 692
            batch_size, size
        ), 'The shape of h0 should be(batch_size, %d)' % size
        inputs['H0'] = h_0
G
guosheng 已提交
693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715

    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
716 717 718
def gru_unit(input,
             hidden,
             size,
719 720
             param_attr=None,
             bias_attr=None,
Y
Yu Yang 已提交
721
             activation='tanh',
722
             gate_activation='sigmoid'):
Y
Yu Yang 已提交
723
    """
724
    GRU unit layer. The equation of a gru step is:
Y
Yu Yang 已提交
725

726 727
        .. math::
            u_t & = actGate(xu_{t} + W_u h_{t-1} + b_u)
Y
Yu Yang 已提交
728

729
            r_t & = actGate(xr_{t} + W_r h_{t-1} + b_r)
Y
Yu Yang 已提交
730

731
            m_t & = actNode(xm_t + W_c dot(r_t, h_{t-1}) + b_m)
732

733
            h_t & = dot((1-u_t), m_t) + dot(u_t, h_{t-1})
734 735

    The inputs of gru unit includes :math:`z_t`, :math:`h_{t-1}`. In terms
736 737 738
    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
739 740
    connected layer has to be applied, such that :math:`z_t = W_{fc}x_t`.

741 742
    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
743 744 745
    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`.
746 747 748 749 750

    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.
751 752
        param_attr (ParamAttr): The weight parameters for gru unit. Default: None
        bias_attr (ParamAttr): The bias parameters for gru unit. Default: None
753 754 755 756
        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
757

758 759 760 761 762 763
    Returns:
        tuple: The hidden value, reset-hidden value and gate values.

    Examples:

        .. code-block:: python
Y
Yu Yang 已提交
764

765
             # assuming we have x_t_data and prev_hidden of size=10
766
             x_t = fluid.layers.fc(input=x_t_data, size=30)
767 768
             hidden_val, r_h_val, gate_val = fluid.layers.gru_unit(input=x_t,
                                                    hidden = prev_hidden)
Y
Yu Yang 已提交
769 770 771 772 773 774 775 776 777 778 779 780 781 782 783

    """
    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()
    size = size / 3

    # create weight
784 785
    weight = helper.create_parameter(
        attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype)
Y
Yu Yang 已提交
786

787 788 789 790
    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
791
    # create bias
792
    if helper.bias_attr:
Y
Yu Yang 已提交
793 794 795
        bias_size = [1, 3 * size]
        bias = helper.create_parameter(
            attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
796
        inputs['Bias'] = bias
Y
Yu Yang 已提交
797 798 799

    helper.append_op(
        type='gru_unit',
800
        inputs=inputs,
Y
Yu Yang 已提交
801 802 803 804 805 806
        outputs={
            'Gate': gate,
            'ResetHiddenPrev': reset_hidden_pre,
            'Hidden': updated_hidden,
        },
        attrs={
807 808
            'activation': 2,  # tanh
            'gate_activation': 1,  # sigmoid
Y
Yu Yang 已提交
809 810 811 812 813
        })

    return updated_hidden, reset_hidden_pre, gate


Y
yuyang18 已提交
814
@templatedoc()
815
def linear_chain_crf(input, label, param_attr=None):
Y
yuyang18 已提交
816 817 818 819 820 821 822
    """
    Linear Chain CRF.

    ${comment}

    Args:
        input(${emission_type}): ${emission_comment}
D
dzhwinter 已提交
823
        input(${transition_type}): ${transition_comment}
Y
yuyang18 已提交
824 825 826 827
        label(${label_type}): ${label_comment}
        param_attr(ParamAttr): The attribute of the learnable parameter.

    Returns:
D
dzhwinter 已提交
828 829 830
        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
831 832

    """
Y
Yu Yang 已提交
833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857
    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
858
@templatedoc()
859
def crf_decoding(input, param_attr, label=None):
Y
yuyang18 已提交
860 861 862 863 864
    """
    ${comment}

    Args:
        input(${emission_type}): ${emission_comment}
Y
yi.wu 已提交
865

Y
yuyang18 已提交
866
        param_attr(ParamAttr): The parameter attribute for training.
Y
yi.wu 已提交
867

Y
yuyang18 已提交
868 869 870
        label(${label_type}): ${label_comment}

    Returns:
Y
update  
yi.wu 已提交
871
        Variable: ${viterbi_path_comment}
872

Y
yi.wu 已提交
873 874 875 876 877
    Examples:
        .. code-block:: python

           crf_decode = layers.crf_decoding(
                input=hidden, param_attr=ParamAttr(name="crfw"))
Y
yuyang18 已提交
878
    """
Y
Yu Yang 已提交
879 880 881 882 883 884 885 886 887 888 889 890 891
    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
892
@templatedoc()
F
fengjiayi 已提交
893
def cos_sim(X, Y):
Y
Yu Yang 已提交
894
    """
Y
yi.wu 已提交
895 896 897
    ${comment}

    Args:
898 899
        X (Variable): ${x_comment}.
        Y (Variable): ${y_comment}.
F
fengjiayi 已提交
900

Y
yi.wu 已提交
901
    Returns:
902
        Variable: the output of cosine(X, Y).
Y
Yu Yang 已提交
903
    """
F
fengjiayi 已提交
904
    helper = LayerHelper('cos_sim', **locals())
Y
Yu Yang 已提交
905 906 907 908 909 910 911 912 913 914 915 916 917
    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


918
def dropout(x, dropout_prob, is_test=False, seed=None, name=None):
919 920 921 922 923
    """
    Computes dropout.

    Drop or keep each element of `x` independently. Dropout is a regularization
    technique for reducing overfitting by preventing neuron co-adaption during
924
    training. The dropout operator randomly sets (according to the given dropout
925 926 927 928
    probability) the outputs of some units to zero, while others are remain
    unchanged.

    Args:
929 930
        x (Variable): The input tensor variable.
        dropout_prob (float): Probability of setting units to zero.
931 932 933 934 935 936 937
        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.
938 939

    Returns:
940
        Variable: A tensor variable is the shape with `x`.
941 942

    Examples:
943

944 945
        .. code-block:: python

946 947
            x = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
            droped = fluid.layers.dropout(x, dropout_prob=0.5)
948 949
    """

F
fengjiayi 已提交
950
    helper = LayerHelper('dropout', **locals())
951 952 953 954 955 956 957
    out = helper.create_tmp_variable(dtype=x.dtype)
    mask = helper.create_tmp_variable(dtype=x.dtype, stop_gradient=True)
    helper.append_op(
        type='dropout',
        inputs={'X': [x]},
        outputs={'Out': [out],
                 'Mask': [mask]},
958 959 960 961 962 963
        attrs={
            'dropout_prob': dropout_prob,
            'is_test': is_test,
            'fix_seed': seed is not None,
            'seed': seed if seed is not None else 0
        })
964 965 966
    return out


F
fengjiayi 已提交
967
def cross_entropy(input, label, soft_label=False):
Y
Yu Yang 已提交
968
    """
Y
Yibing Liu 已提交
969 970
    **Cross Entropy Layer**

971 972 973
    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
974 975

    1) One-hot cross-entropy:
F
fengjiayi 已提交
976
        `soft_label = False`, `Label[i, 0]` indicates the class index for sample i:
Y
yangyaming 已提交
977

Y
Yibing Liu 已提交
978
        .. math::
Y
yangyaming 已提交
979

Y
Yibing Liu 已提交
980 981 982
            Y[i] = -\log(X[i, Label[i]])

    2) Soft-label cross-entropy:
F
fengjiayi 已提交
983 984
        `soft_label = True`, `Label[i, j]` indicates the soft label of class j
        for sample i:
Y
Yibing Liu 已提交
985 986 987 988 989

        .. math::

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

Y
Yibing Liu 已提交
990
       Please make sure that in this case the summation of each row of `label`
Y
Yibing Liu 已提交
991 992 993
       equals one.

    3) One-hot cross-entropy with vecterized `label`:
F
fengjiayi 已提交
994 995
         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
996
         to a one-hot cross-entropy with one-hot label representation.
Y
yangyaming 已提交
997

Y
Yibing Liu 已提交
998
    Args:
Y
yangyaming 已提交
999
        input (Variable|list):  a 2-D tensor with shape [N x D], where N is the
1000 1001 1002 1003
                                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
1004
        label (Variable|list): the ground truth which is a 2-D tensor. When
1005 1006 1007 1008
                               `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
1009
        soft_label (bool): a flag indicating whether to
1010 1011
                                           interpretate the given labels as soft
                                           labels, default `False`.
Y
Yibing Liu 已提交
1012 1013 1014 1015 1016

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

    Raises:
1017 1018 1019 1020 1021
        `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
1022 1023 1024 1025 1026 1027

    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
1028
    """
F
fengjiayi 已提交
1029
    helper = LayerHelper('cross_entropy', **locals())
Y
Yu Yang 已提交
1030 1031 1032 1033 1034 1035
    out = helper.create_tmp_variable(dtype=input.dtype)
    helper.append_op(
        type='cross_entropy',
        inputs={'X': [input],
                'Label': [label]},
        outputs={'Y': [out]},
F
fengjiayi 已提交
1036
        attrs={"soft_label": soft_label})
Y
Yu Yang 已提交
1037 1038 1039
    return out


F
fengjiayi 已提交
1040
def square_error_cost(input, label):
Y
Yu Yang 已提交
1041
    """
1042 1043
    **Square error cost layer**

1044 1045
    This layer accepts input predictions and target label and returns the
    squared error cost.
Y
ying 已提交
1046

1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059
    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:
1060 1061
        input (Variable): Input tensor, has predictions.
        label (Variable): Label tensor, has target labels.
1062 1063

    Returns:
G
guosheng 已提交
1064
        Variable: The tensor variable storing the element-wise squared error \
1065
                  difference of input and label.
1066 1067 1068 1069 1070 1071 1072 1073

    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
1074
    """
F
fengjiayi 已提交
1075
    helper = LayerHelper('square_error_cost', **locals())
Y
Yu Yang 已提交
1076 1077 1078 1079 1080 1081 1082 1083 1084
    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
1085 1086
        type='square', inputs={'X': [minus_out]},
        outputs={'Out': [square_out]})
Y
Yu Yang 已提交
1087 1088 1089
    return square_out


Y
yi.wu 已提交
1090
@templatedoc()
Y
Yu Yang 已提交
1091 1092 1093 1094
def chunk_eval(input,
               label,
               chunk_scheme,
               num_chunk_types,
F
fengjiayi 已提交
1095
               excluded_chunk_types=None):
Y
Yu Yang 已提交
1096
    """
Y
yi.wu 已提交
1097
    **Chunk Evaluator**
Y
yi.wu 已提交
1098

Y
yangyaming 已提交
1099
    This function computes and outputs the precision, recall and
1100
    F1-score of chunk detection.
Y
yi.wu 已提交
1101

Y
yi.wu 已提交
1102 1103 1104 1105 1106 1107 1108 1109
    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
1110

Y
yi.wu 已提交
1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135
       ====== ====== ======  =====  ==  ============   =====  ===== =====  ==  =========
              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
1136

Y
yi.wu 已提交
1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160
       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
1161
    Args:
1162 1163 1164 1165 1166
        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
1167

Y
yi.wu 已提交
1168
    Returns:
Y
update  
yi.wu 已提交
1169 1170 1171
        tuple: tuple containing: precision, recall, f1_score,
        num_infer_chunks, num_label_chunks,
        num_correct_chunks
1172

Y
yi.wu 已提交
1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184
    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
1185
    """
F
fengjiayi 已提交
1186
    helper = LayerHelper("chunk_eval", **locals())
Y
Yu Yang 已提交
1187 1188 1189 1190 1191

    # prepare output
    precision = helper.create_tmp_variable(dtype="float32")
    recall = helper.create_tmp_variable(dtype="float32")
    f1_score = helper.create_tmp_variable(dtype="float32")
1192 1193 1194
    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
1195 1196 1197 1198 1199 1200 1201 1202

    helper.append_op(
        type="chunk_eval",
        inputs={"Inference": [input],
                "Label": [label]},
        outputs={
            "Precision": [precision],
            "Recall": [recall],
1203 1204 1205 1206
            "F1-Score": [f1_score],
            "NumInferChunks": [num_infer_chunks],
            "NumLabelChunks": [num_label_chunks],
            "NumCorrectChunks": [num_correct_chunks]
Y
Yu Yang 已提交
1207 1208 1209
        },
        attrs={
            "num_chunk_types": num_chunk_types,
G
guosheng 已提交
1210 1211
            "chunk_scheme": chunk_scheme,
            "excluded_chunk_types": excluded_chunk_types or []
Y
Yu Yang 已提交
1212
        })
1213 1214
    return (precision, recall, f1_score, num_infer_chunks, num_label_chunks,
            num_correct_chunks)
Y
Yu Yang 已提交
1215 1216


1217
@templatedoc()
Y
Yu Yang 已提交
1218 1219 1220 1221 1222 1223 1224
def sequence_conv(input,
                  num_filters,
                  filter_size=3,
                  filter_stride=1,
                  padding=None,
                  bias_attr=None,
                  param_attr=None,
1225
                  act=None):
Y
Yu Yang 已提交
1226 1227 1228 1229
    """
    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.
1230 1231 1232 1233 1234 1235 1236 1237 1238 1239

    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
1240

1241 1242
    Returns:
        Variable: output of sequence_conv
Y
Yu Yang 已提交
1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267
    """

    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,
            'contextStart': -int(filter_size / 2),
            'contextLength': filter_size
        })
    pre_act = helper.append_bias_op(pre_bias)
    return helper.append_activation(pre_act)


1268
def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=True):
1269 1270 1271
    """
    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
1272
    input Tensor can be either :math:`[N, 1]` or :math:`[N]`, where :math:`N`
1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291
    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 \
        library is installed. Default: True
1292

1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303
    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)
    """
1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314
    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


1315
def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True, name=None):
Q
qiaolongfei 已提交
1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354
    """
    The input of the softmax layer is a 2-D tensor with shape N x K (N is the
    batch_size, K is the dimension of input feature). The output tensor has the
    same shape as the input tensor.

    For each row of the input tensor, the softmax operator squashes the
    K-dimensional vector of arbitrary real values to a K-dimensional vector of real
    values in the range [0, 1] that add up to 1.

    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.

    For each row :math:`i` and each column :math:`j` in Input(X), we have:

    .. 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)

    """
1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365
    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
1366 1367 1368
def conv2d(input,
           num_filters,
           filter_size,
C
chengduoZH 已提交
1369 1370
           stride=1,
           padding=0,
1371
           dilation=1,
Y
Yu Yang 已提交
1372 1373 1374
           groups=None,
           param_attr=None,
           bias_attr=None,
C
chengduoZH 已提交
1375
           use_cudnn=True,
1376
           use_mkldnn=False,
1377 1378
           act=None,
           name=None):
Y
Yu Yang 已提交
1379
    """
C
chengduoZH 已提交
1380
    The convolution2D layer calculates the output based on the input, filter
T
tensor-tang 已提交
1381 1382
    and strides, paddings, dilations, groups parameters. Input and
    Output are in NCHW format, where N is batch size, C is the number of
1383
    channels, H is the height of the feature, and W is the width of the feature.
T
tensor-tang 已提交
1384 1385 1386 1387 1388 1389 1390
    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.
1391 1392 1393
    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
1394

1395
    For each input :math:`X`, the equation is:
C
refine  
chengduoZH 已提交
1396

C
chengduoZH 已提交
1397 1398
    .. math::

C
refine  
chengduoZH 已提交
1399
        Out = \sigma (W \\ast X + b)
C
chengduoZH 已提交
1400

T
tensor-tang 已提交
1401
    Where:
C
chengduoZH 已提交
1402

1403 1404 1405 1406 1407
    * :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
1408
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
1409 1410 1411

    Example:

1412 1413
        - Input:

W
weixing02 已提交
1414
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
C
refine  
chengduoZH 已提交
1415

W
weixing02 已提交
1416
          Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)`
C
refine  
chengduoZH 已提交
1417

1418
        - Output:
T
tensor-tang 已提交
1419

W
weixing02 已提交
1420
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
C
refine  
chengduoZH 已提交
1421

C
chengduoZH 已提交
1422
        Where
1423 1424

        .. math::
C
chengduoZH 已提交
1425

W
weixing02 已提交
1426 1427
            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
1428 1429

    Args:
1430
        input (Variable): The input image with [N, C, H, W] format.
T
tensor-tang 已提交
1431
        num_filters(int): The number of filter. It is as same as the output
1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453
            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
T
tensor-tang 已提交
1454 1455
        use_mkldnn (bool): Use mkldnn kernels or not, it is valid only when compiled
            with mkldnn library. Default: False
1456 1457 1458
        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
1459 1460

    Returns:
G
guosheng 已提交
1461
        Variable: The tensor variable storing the convolution and \
C
chengduoZH 已提交
1462 1463
                  non-linearity activation result.

C
refine  
chengduoZH 已提交
1464
    Raises:
1465 1466
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
C
refine  
chengduoZH 已提交
1467

C
chengduoZH 已提交
1468 1469 1470
    Examples:
        .. code-block:: python

1471 1472
          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
1473 1474 1475
    """

    num_channels = input.shape[1]
1476 1477

    l_type = 'conv2d'
X
xzl 已提交
1478 1479
    if (num_channels == groups and num_filters % num_channels == 0 and
            not use_cudnn):
1480
        l_type = 'depthwise_conv2d'
1481 1482 1483 1484

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

Y
Yu Yang 已提交
1485 1486 1487 1488 1489 1490 1491
    if groups is None:
        num_filter_channels = num_channels
    else:
        if num_channels % groups != 0:
            raise ValueError("num_channels must be divisible by groups.")
        num_filter_channels = num_channels / groups

C
chengduoZH 已提交
1492 1493 1494
    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')
1495
    dilation = utils.convert_to_list(dilation, 2, 'dilation')
C
chengduoZH 已提交
1496

C
chengduoZH 已提交
1497 1498
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1499 1500

    input_shape = input.shape
M
minqiyang 已提交
1501
    filter_shape = [num_filters, int(num_filter_channels)] + filter_size
Y
Yu Yang 已提交
1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515

    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(
1516
        type=l_type,
Y
Yu Yang 已提交
1517 1518 1519 1520 1521
        inputs={
            'Input': input,
            'Filter': filter_param,
        },
        outputs={"Output": pre_bias},
C
chengduoZH 已提交
1522 1523 1524
        attrs={
            'strides': stride,
            'paddings': padding,
1525
            'dilations': dilation,
C
chengduoZH 已提交
1526
            'groups': groups,
1527 1528
            'use_cudnn': use_cudnn,
            'use_mkldnn': use_mkldnn
C
chengduoZH 已提交
1529
        })
Y
Yu Yang 已提交
1530 1531 1532 1533 1534 1535

    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
1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553
def conv3d(input,
           num_filters,
           filter_size,
           stride=1,
           padding=0,
           dilation=1,
           groups=None,
           param_attr=None,
           bias_attr=None,
           use_cudnn=True,
           use_mkldnn=False,
           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
1554 1555 1556 1557 1558 1559
    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
1560 1561 1562 1563 1564 1565 1566 1567 1568

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

    .. math::

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

    In the above equation:

1569 1570
    * :math:`X`: Input value, a tensor with NCDHW format.
    * :math:`W`: Filter value, a tensor with MCDHW format.
C
chengduoZH 已提交
1571 1572 1573
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
1574
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599

    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,
1600
            it must contain three integers, (filter_size_D, filter_size_H, filter_size_W).
C
chengduoZH 已提交
1601 1602
            Otherwise, the filter will be a square.
        stride (int|tuple): The stride size. If stride is a tuple, it must
1603
            contain three integers, (stride_D, stride_H, stride_W). Otherwise, the
C
chengduoZH 已提交
1604 1605
            stride_D = stride_H = stride_W = stride. Default: stride = 1.
        padding (int|tuple): The padding size. If padding is a tuple, it must
1606
            contain three integers, (padding_D, padding_H, padding_W). Otherwise, the
C
chengduoZH 已提交
1607 1608
            padding_D = padding_H = padding_W = padding. Default: padding = 0.
        dilation (int|tuple): The dilation size. If dilation is a tuple, it must
1609
            contain three integers, (dilation_D, dilation_H, dilation_W). Otherwise, the
C
chengduoZH 已提交
1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635
            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
        use_mkldnn (bool): Use mkldnn kernels or not.
        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

1636 1637
          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
1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 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 1677 1678 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.")
        num_filter_channels = num_channels / groups

    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,
            'use_mkldnn': use_mkldnn
        })

1693
    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
C
chengduoZH 已提交
1694 1695 1696 1697

    return helper.append_activation(pre_act)


F
fengjiayi 已提交
1698
def sequence_pool(input, pool_type):
Y
Yu Yang 已提交
1699
    """
Y
yangyaming 已提交
1700 1701 1702
    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
1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713

    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:
1714
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
1715 1716 1717 1718 1719
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
1720
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
1721 1722 1723 1724 1725 1726 1727

       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)
1728 1729
         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
1730

L
Luo Tao 已提交
1731 1732
    Args:
        input(variable): The input variable which is a LoDTensor.
Y
yangyaming 已提交
1733
        pool_type (string): The pooling type of sequence_pool.
L
Luo Tao 已提交
1734 1735 1736 1737 1738 1739 1740 1741
            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
1742

Y
yangyaming 已提交
1743
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1744 1745 1746 1747 1748
                              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')
1749 1750
             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
1751
    """
F
fengjiayi 已提交
1752
    helper = LayerHelper('sequence_pool', **locals())
Y
Yu Yang 已提交
1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763
    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
1764 1765 1766 1767 1768
    # 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
1769 1770 1771
    return pool_out


F
fengjiayi 已提交
1772
def sequence_first_step(input):
L
Luo Tao 已提交
1773
    """
L
Luo Tao 已提交
1774
    This function gets the first step of sequence.
L
Luo Tao 已提交
1775 1776 1777 1778

    .. code-block:: text

       x is a 1-level LoDTensor:
1779
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
1780 1781 1782 1783 1784
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
1785
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
1786
         out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1)
F
fengjiayi 已提交
1787

L
Luo Tao 已提交
1788 1789 1790 1791 1792 1793 1794 1795 1796
    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
1797

Y
yangyaming 已提交
1798
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1799 1800 1801
                              dtype='float32', lod_level=1)
             x_first_step = fluid.layers.sequence_first_step(input=x)
    """
1802 1803 1804
    return sequence_pool(input=input, pool_type="first")


F
fengjiayi 已提交
1805
def sequence_last_step(input):
L
Luo Tao 已提交
1806
    """
L
Luo Tao 已提交
1807
    This function gets the last step of sequence.
L
Luo Tao 已提交
1808 1809 1810 1811

    .. code-block:: text

       x is a 1-level LoDTensor:
1812
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
1813 1814 1815 1816 1817
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
1818
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
1819
         out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1)
F
fengjiayi 已提交
1820

L
Luo Tao 已提交
1821 1822 1823 1824 1825 1826 1827 1828 1829
    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
1830

Y
yangyaming 已提交
1831
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1832 1833 1834
                              dtype='float32', lod_level=1)
             x_last_step = fluid.layers.sequence_last_step(input=x)
    """
1835 1836 1837
    return sequence_pool(input=input, pool_type="last")


F
fengjiayi 已提交
1838
@templatedoc()
Y
Yu Yang 已提交
1839
def pool2d(input,
C
chengduoZH 已提交
1840 1841
           pool_size=-1,
           pool_type="max",
C
chengduoZH 已提交
1842 1843
           pool_stride=1,
           pool_padding=0,
C
caoying03 已提交
1844
           global_pooling=False,
C
chengduoZH 已提交
1845
           use_cudnn=True,
1846
           ceil_mode=False,
1847
           use_mkldnn=False,
C
caoying03 已提交
1848
           name=None):
Y
Yu Yang 已提交
1849
    """
F
fengjiayi 已提交
1850
    ${comment}
1851 1852

    Args:
1853 1854 1855
        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
1856
                          feature, and W is the width of the feature.
1857
        pool_size (int): The side length of pooling windows. All pooling
F
fengjiayi 已提交
1858
                         windows are squares with pool_size on a side.
F
fengjiayi 已提交
1859
        pool_type: ${pooling_type_comment}
1860 1861
        pool_stride (int): stride of the pooling layer.
        pool_padding (int): padding size.
F
fengjiayi 已提交
1862 1863 1864 1865
        global_pooling: ${global_pooling_comment}
        use_cudnn: ${use_cudnn_comment}
        ceil_mode: ${ceil_mode_comment}
        use_mkldnn: ${use_mkldnn_comment}
1866
        name (str|None): A name for this layer(optional). If set None, the
F
fengjiayi 已提交
1867 1868
                        layer will be named automatically.

1869
    Returns:
F
fengjiayi 已提交
1870
        Variable: The pooling result.
F
fengjiayi 已提交
1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883

    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(
1884 1885 1886 1887
                            input=data,
                            pool_size=2,
                            pool_type='max',
                            pool_stride=1,
F
fengjiayi 已提交
1888
                            global_pooling=False)
Y
Yu Yang 已提交
1889 1890 1891 1892 1893
    """
    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
1894

C
chengduoZH 已提交
1895 1896 1897 1898 1899
    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
1900 1901 1902 1903
    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
1904 1905
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1906

C
Add doc  
chengduoZH 已提交
1907
    l_type = 'pool2d'
1908 1909

    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
1910 1911 1912 1913
    dtype = helper.input_dtype()
    pool_out = helper.create_tmp_variable(dtype)

    helper.append_op(
1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942
        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,
            "use_mkldnn": use_mkldnn
        })

    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,
           use_mkldnn=False,
           name=None):
    """
    This function adds the operator for pooling in 3-dimensions, using the
Y
Yu Yang 已提交
1943
    pooling configurations mentioned in input parameters.
1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956

    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}
        use_mkldnn (bool): ${use_mkldnn_comment}
        name (str): A name for this layer(optional). If set None, the layer
            will be named automatically.
1957

1958
    Returns:
1959
        Variable: output of pool3d layer.
Y
Yu Yang 已提交
1960 1961 1962 1963 1964
    """
    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
1965

C
chengduoZH 已提交
1966 1967 1968 1969 1970
    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))

1971 1972 1973
    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
1974

C
chengduoZH 已提交
1975 1976
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1977

1978 1979
    l_type = "pool3d"
    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
1980 1981 1982 1983
    dtype = helper.input_dtype()
    pool_out = helper.create_tmp_variable(dtype)

    helper.append_op(
1984
        type=l_type,
Y
Yu Yang 已提交
1985 1986 1987 1988 1989 1990 1991
        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
1992
            "paddings": pool_padding,
1993
            "use_cudnn": use_cudnn,
1994 1995
            "ceil_mode": ceil_mode,
            "use_mkldnn": use_mkldnn
Y
Yu Yang 已提交
1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007
        })

    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
2008
               data_layout='NCHW',
Y
Yang Yang 已提交
2009
               in_place=False,
2010
               use_mkldnn=False,
2011 2012
               name=None,
               moving_mean_name=None,
W
wanghaoshuang 已提交
2013
               moving_variance_name=None,
2014 2015
               do_model_average_for_mean_and_var=False,
               fuse_with_relu=False):
Y
Yu Yang 已提交
2016
    """
Q
qiaolongfei 已提交
2017 2018 2019 2020
    **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
2021

Q
qiaolongfei 已提交
2022
    1. NHWC `[batch, in_height, in_width, in_channels]`
Q
qiaolongfei 已提交
2023

Q
qiaolongfei 已提交
2024 2025
    2. NCHW `[batch, in_channels, in_height, in_width]`

Q
qiaolongfei 已提交
2026 2027 2028
    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
2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040

    :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
2041 2042

    Args:
Q
qiaolongfei 已提交
2043
        input(variable): The input variable which is a LoDTensor.
Q
qiaolongfei 已提交
2044 2045 2046 2047
        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
2048 2049 2050
        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
Q
qiaolongfei 已提交
2051
        in_place(bool, Default False): Make the input and output of batch norm reuse memory.
Q
qiaolongfei 已提交
2052 2053 2054 2055 2056
        use_mkldnn(bool, Default false): ${use_mkldnn_comment}
        name(string, Default None): A name for this layer(optional). If set None, the layer
            will be named automatically.
        moving_mean_name(string, Default None): The name of moving_mean which store the global Mean.
        moving_variance_name(string, Default None): The name of the moving_variance which store the global Variance.
Q
qiaolongfei 已提交
2057
        do_model_average_for_mean_and_var(bool, Default False): Do model average for mean and variance or not.
2058
        fuse_with_relu (bool): if True, this OP performs relu after batch norm.
2059 2060

    Returns:
Q
qiaolongfei 已提交
2061
        Variable: A tensor variable which is the result after applying batch normalization on the input.
Q
qiaolongfei 已提交
2062 2063 2064 2065 2066 2067 2068

    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
2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091
    """
    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(
2092
        attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
Y
Yu Yang 已提交
2093

2094 2095
    mean = helper.create_parameter(
        attr=ParamAttr(
W
wanghaoshuang 已提交
2096 2097 2098
            name=moving_mean_name,
            initializer=Constant(0.0),
            trainable=False,
W
wanghaoshuang 已提交
2099
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
2100
        shape=param_shape,
2101 2102 2103 2104 2105 2106 2107
        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
2108
            trainable=False,
W
wanghaoshuang 已提交
2109
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
2110
        shape=param_shape,
2111 2112
        dtype=input.dtype)
    variance.stop_gradient = True
Y
Yu Yang 已提交
2113 2114 2115 2116 2117 2118

    # 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
2119 2120
    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
2121

Y
Yang Yang 已提交
2122
    batch_norm_out = input if in_place else helper.create_tmp_variable(dtype)
Y
Yu Yang 已提交
2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139

    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
        },
2140 2141 2142 2143
        attrs={
            "momentum": momentum,
            "epsilon": epsilon,
            "is_test": is_test,
2144 2145
            "use_mkldnn": use_mkldnn,
            "fuse_with_relu": fuse_with_relu
2146
        })
Y
Yu Yang 已提交
2147 2148 2149 2150

    return helper.append_activation(batch_norm_out)


Y
yuyang18 已提交
2151
@templatedoc()
G
guosheng 已提交
2152 2153 2154 2155 2156 2157 2158 2159 2160 2161
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
2162
    ${comment}
G
guosheng 已提交
2163 2164 2165

    The formula is as follows:

Y
yuyang18 已提交
2166
    ..  math::
G
guosheng 已提交
2167 2168 2169 2170 2171 2172 2173

        \\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
2174 2175 2176 2177 2178 2179 2180 2181
    * :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
2182

G
guosheng 已提交
2183 2184
    Args:
        input(Variable): The input tensor variable.
2185
        scale(bool): Whether to learn the adaptive gain :math:`g` after
G
guosheng 已提交
2186
            normalization.
2187
        shift(bool): Whether to learn the adaptive bias :math:`b` after
G
guosheng 已提交
2188
            normalization.
2189
        begin_norm_axis(bool): The normalization will be performed along
G
guosheng 已提交
2190
            dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`.
2191
        epsilon(float): The small value added to the variance to prevent
G
guosheng 已提交
2192 2193 2194 2195 2196 2197
            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.
2198
        name (str): The name of this layer. It is optional.
G
guosheng 已提交
2199 2200

    Returns:
Y
yuyang18 已提交
2201
        ${y_comment}
G
guosheng 已提交
2202 2203 2204

    Examples:

Y
yuyang18 已提交
2205 2206 2207
        >>> 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
2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222
    """
    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
2223
    if shift:
G
guosheng 已提交
2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247
        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
2248 2249 2250 2251
def conv2d_transpose(input,
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
2252 2253 2254
                     padding=0,
                     stride=1,
                     dilation=1,
2255
                     groups=None,
C
caoying03 已提交
2256
                     param_attr=None,
2257
                     bias_attr=None,
C
chengduoZH 已提交
2258
                     use_cudnn=True,
2259
                     act=None,
C
caoying03 已提交
2260
                     name=None):
Y
Yu Yang 已提交
2261
    """
2262 2263 2264 2265 2266 2267 2268 2269
    **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
2270 2271
    layer, please refer to the following explanation and references
    `therein <http://www.matthewzeiler.com/wp-content/uploads/2017/07/cvpr2010.pdf>`_.
2272 2273 2274
    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.
2275 2276 2277 2278 2279

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

    .. math::

2280
        Out = \sigma (W \\ast X + b)
2281

2282
    Where:
2283 2284 2285

    * :math:`X`: Input value, a tensor with NCHW format.
    * :math:`W`: Filter value, a tensor with MCHW format.
2286 2287 2288 2289
    * :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
2290

2291 2292 2293 2294
    Example:

        - Input:

2295
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
2296

2297
          Filter shape: :math:`(C_{in}, C_{out}, H_f, W_f)`
2298 2299 2300

        - Output:

2301
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
2302 2303

        Where
Y
Yu Yang 已提交
2304

2305 2306 2307 2308
        .. math::

           H_{out} &= (H_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (H_f - 1) + 1 \\\\
           W_{out} &= (W_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (W_f - 1) + 1
Y
Yu Yang 已提交
2309 2310

    Args:
2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343
        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
            tuple, it must contain two integers, (image_H, image_W). This
            parameter only works when filter_size is None.
        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
2344 2345

    Returns:
2346
        Variable: The tensor variable storing the convolution transpose result.
2347 2348

    Raises:
2349 2350
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
2351 2352 2353 2354

    Examples:
       .. code-block:: python

2355 2356
          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
2357
    """
2358 2359 2360 2361 2362 2363 2364 2365 2366

    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
2367 2368 2369
    if not isinstance(input, Variable):
        raise TypeError("Input of conv2d_transpose must be Variable")

C
chengduoZH 已提交
2370 2371 2372
    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
2373

C
chengduoZH 已提交
2374 2375
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
G
guosheng 已提交
2376

Y
Yu Yang 已提交
2377 2378 2379 2380 2381
    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
2382

Y
Yu Yang 已提交
2383 2384
        h_in = input.shape[2]
        w_in = input.shape[3]
G
guosheng 已提交
2385

C
chengduoZH 已提交
2386 2387 2388 2389
        filter_size_h = (output_size[0] - (h_in - 1) * stride[0] + 2 *
                         padding[0] - 1) / dilation[0] + 1
        filter_size_w = (output_size[1] - (w_in - 1) * stride[1] + 2 *
                         padding[1] - 1) / dilation[1] + 1
Y
Yu Yang 已提交
2390
        filter_size = [filter_size_h, filter_size_w]
C
chengduoZH 已提交
2391 2392 2393
    else:
        filter_size = utils.convert_to_list(filter_size, 2,
                                            'conv2d_transpose.filter_size')
2394

2395 2396
    groups = 1 if groups is None else groups
    filter_shape = [input_channel, num_filters / groups] + filter_size
Y
Yu Yang 已提交
2397 2398 2399
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

2400
    pre_bias = helper.create_tmp_variable(dtype=input.dtype)
Y
Yu Yang 已提交
2401
    helper.append_op(
2402
        type=op_type,
Y
Yu Yang 已提交
2403 2404
        inputs={'Input': [input],
                'Filter': [img_filter]},
2405
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
2406
        attrs={
2407 2408 2409 2410 2411
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn
Y
Yu Yang 已提交
2412 2413
        })

2414 2415 2416
    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
2417 2418


2419
def conv3d_transpose(input,
Y
Yu Yang 已提交
2420 2421 2422
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
2423 2424 2425
                     padding=0,
                     stride=1,
                     dilation=1,
2426
                     groups=None,
C
caoying03 已提交
2427
                     param_attr=None,
2428
                     bias_attr=None,
C
chengduoZH 已提交
2429
                     use_cudnn=True,
2430
                     act=None,
C
caoying03 已提交
2431
                     name=None):
Y
Yu Yang 已提交
2432
    """
2433
    **Convlution3D transpose layer**
2434

2435
    The convolution3D transpose layer calculates the output based on the input,
2436
    filter, and dilations, strides, paddings. Input(Input) and output(Output)
2437 2438 2439 2440 2441 2442
    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>`_.
2443 2444 2445
    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.
2446 2447 2448 2449 2450

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

    .. math::

2451
        Out = \sigma (W \\ast X + b)
2452 2453 2454

    In the above equation:

2455 2456
    * :math:`X`: Input value, a tensor with NCDHW format.
    * :math:`W`: Filter value, a tensor with MCDHW format.
2457 2458 2459 2460
    * :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
2461

2462 2463 2464 2465
    Example:

        - Input:

2466
          Input shape: :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})`
2467

2468
          Filter shape: :math:`(C_{in}, C_{out}, D_f, H_f, W_f)`
2469 2470 2471

        - Output:

2472
          Output shape: :math:`(N, C_{out}, D_{out}, H_{out}, W_{out})`
2473 2474

        Where
Y
Yu Yang 已提交
2475

2476 2477
        .. math::

2478 2479 2480
           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
2481 2482

    Args:
2483
        input(Variable): The input image with [N, C, D, H, W] format.
2484 2485 2486
        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
2487
            tuple, it must contain three integers, (image_D, image_H, image_W). This
2488 2489
            parameter only works when filter_size is None.
        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
2490
            it must contain three integers, (filter_size_D, filter_size_H, filter_size_W).
2491 2492 2493
            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
2494 2495
            contain three integers, (padding_D, padding_H, padding_W). Otherwise, the
            padding_D = padding_H = padding_W = padding. Default: padding = 0.
2496
        stride(int|tuple): The stride size. If stride is a tuple, it must
2497 2498
            contain three integers, (stride_D, stride_H, stride_W). Otherwise, the
            stride_D = stride_H = stride_W = stride. Default: stride = 1.
2499
        dilation(int|tuple): The dilation size. If dilation is a tuple, it must
2500 2501 2502
            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
2503 2504 2505 2506 2507
            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
2508 2509 2510
        param_attr(ParamAttr): The parameters to the Conv3d_transpose Layer.
            Default: None
        bias_attr(ParamAttr): Bias parameter for the Conv3d layer. Default: None
2511 2512 2513 2514 2515
        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
2516 2517

    Returns:
2518
        Variable: The tensor variable storing the convolution transpose result.
2519 2520

    Raises:
2521 2522
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
2523 2524 2525 2526

    Examples:
       .. code-block:: python

2527 2528
          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
2529
    """
2530 2531
    l_type = "conv3d_transpose"
    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
2532
    if not isinstance(input, Variable):
2533
        raise TypeError("Input of conv3d_transpose must be Variable")
Y
Yu Yang 已提交
2534 2535
    input_channel = input.shape[1]

2536 2537 2538
    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
2539

C
chengduoZH 已提交
2540 2541 2542
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")

Y
Yu Yang 已提交
2543 2544 2545 2546 2547 2548
    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]

2549 2550 2551
        d_in = input.shape[2]
        h_in = input.shape[3]
        w_in = input.shape[4]
C
chengduoZH 已提交
2552

2553
        filter_size_d = (output_size[0] - (d_in - 1) * stride[0] + 2 *
C
chengduoZH 已提交
2554
                         padding[0] - 1) / dilation[0] + 1
2555
        filter_size_h = (output_size[1] - (h_in - 1) * stride[1] + 2 *
C
chengduoZH 已提交
2556
                         padding[1] - 1) / dilation[1] + 1
2557 2558 2559
        filter_size_w = (output_size[2] - (w_in - 1) * stride[2] + 2 *
                         padding[2] - 1) / dilation[2] + 1
        filter_size = [filter_size_d, filter_size_h, filter_size_w]
C
chengduoZH 已提交
2560
    else:
2561 2562
        filter_size = utils.convert_to_list(filter_size, 3,
                                            'conv3d_transpose.filter_size')
Y
Yu Yang 已提交
2563

2564 2565
    groups = 1 if groups is None else groups
    filter_shape = [input_channel, num_filters / groups] + filter_size
Y
Yu Yang 已提交
2566 2567 2568
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

2569
    pre_bias = helper.create_tmp_variable(dtype=input.dtype)
Y
Yu Yang 已提交
2570
    helper.append_op(
2571
        type=l_type,
Y
Yu Yang 已提交
2572 2573
        inputs={'Input': [input],
                'Filter': [img_filter]},
2574
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
2575 2576 2577 2578
        attrs={
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
2579
            'groups': groups,
C
chengduoZH 已提交
2580 2581
            'use_cudnn': use_cudnn
        })
Y
Yu Yang 已提交
2582

2583 2584
    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
2585
    return out
Y
yangyaming 已提交
2586 2587


Y
yangyaming 已提交
2588
def sequence_expand(x, y, ref_level=-1, name=None):
2589
    """Sequence Expand Layer. This layer will expand the input variable **x**
Y
yangyaming 已提交
2590 2591 2592 2593
    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:
2594 2595 2596 2597 2598

    .. code-block:: text

        * Case 1
            x is a LoDTensor:
2599
                x.lod  = [[2,        2]]
Y
yangyaming 已提交
2600
                x.data = [[a], [b], [c], [d]]
2601 2602 2603
                x.dims = [4, 1]

            y is a LoDTensor:
2604 2605
                y.lod = [[2,    2],
                         [3, 3, 1, 1]]
2606

Y
yangyaming 已提交
2607
            ref_level: 0
2608

Y
yangyaming 已提交
2609
            then output is a 1-level LoDTensor:
2610
                out.lod =  [[2,        2,        2,        2]]
Y
yangyaming 已提交
2611
                out.data = [[a], [b], [a], [b], [c], [d], [c], [d]]
2612 2613 2614 2615
                out.dims = [8, 1]

        * Case 2
            x is a Tensor:
Y
yangyaming 已提交
2616
                x.data = [[a], [b], [c]]
2617 2618 2619
                x.dims = [3, 1]

            y is a LoDTensor:
2620
                y.lod = [[2, 0, 3]]
2621

Y
yangyaming 已提交
2622
            ref_level: -1
2623

Y
yangyaming 已提交
2624 2625 2626
            then output is a Tensor:
                out.data = [[a], [a], [c], [c], [c]]
                out.dims = [5, 1]
2627 2628 2629
    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
2630 2631
        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
2632
        name(str|None): A name for this layer(optional). If set None, the layer
Y
yangyaming 已提交
2633
                        will be named automatically.
2634 2635 2636 2637 2638 2639 2640 2641 2642 2643

    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
2644
            out = layers.sequence_expand(x=x, y=y, ref_level=0)
2645
    """
Y
yangyaming 已提交
2646
    helper = LayerHelper('sequence_expand', input=x, **locals())
2647 2648 2649
    dtype = helper.input_dtype()
    tmp = helper.create_tmp_variable(dtype)
    helper.append_op(
Y
yangyaming 已提交
2650 2651 2652 2653 2654
        type='sequence_expand',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': tmp},
        attrs={'ref_level': ref_level})
2655
    return tmp
2656 2657


2658 2659 2660 2661 2662 2663 2664 2665 2666
def beam_search(pre_ids,
                pre_scores,
                ids,
                scores,
                beam_size,
                end_id,
                level=0,
                name=None):
    """
2667 2668
    Beam search is a classical algorithm for selecting candidate words in a
    machine translation task.
Y
Yan Chunwei 已提交
2669 2670 2671

    Refer to `Beam search <https://en.wikipedia.org/wiki/Beam_search>`_
    for more details.
M
minqiyang 已提交
2672 2673

    This layer does the search in beams for one time step. Specifically, it
2674 2675 2676 2677 2678 2679
    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
2680

2681 2682 2683 2684 2685 2686 2687 2688
    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
2689

2690
    Args:
2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715
        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
2716

2717
    Returns:
2718 2719
        Variable: The LodTensor pair containing the selected ids and the \
            corresponding scores.
Y
Yan Chunwei 已提交
2720 2721 2722 2723

    Examples:
        .. code-block:: python

2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740
            # 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
2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751
    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,
2752
            'pre_scores': pre_scores,
Q
Qiao Longfei 已提交
2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769
            '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


2770 2771 2772 2773 2774 2775 2776
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
2777

2778 2779 2780 2781 2782 2783 2784 2785 2786
    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
2787

2788 2789 2790 2791 2792 2793
    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
2794

2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819
    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
2820 2821 2822 2823
def lstm_unit(x_t,
              hidden_t_prev,
              cell_t_prev,
              forget_bias=0.0,
Y
yangyaming 已提交
2824
              param_attr=None,
C
caoying03 已提交
2825 2826
              bias_attr=None,
              name=None):
Y
yangyaming 已提交
2827 2828 2829 2830
    """Lstm unit layer. The equation of a lstm step is:

        .. math::

2831
            i_t & = \sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i)
Y
yangyaming 已提交
2832

2833
            f_t & = \sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + b_f)
Y
yangyaming 已提交
2834

2835
            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
2836

2837
            o_t & = \sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + b_o)
Y
yangyaming 已提交
2838 2839 2840

            h_t & = o_t tanh(c_t)

2841 2842 2843 2844 2845 2846
    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
2847 2848 2849

        .. math::

2850
            L_{i_t} = W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i
Y
yangyaming 已提交
2851 2852 2853 2854 2855 2856 2857 2858

    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
2859
    This layer has two outputs including :math:`h_t` and :math:`o_t`.
Y
yangyaming 已提交
2860 2861

    Args:
Y
yangyaming 已提交
2862 2863 2864 2865 2866 2867
        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
2868
        forget_bias (float): The forget bias of lstm unit.
Y
yangyaming 已提交
2869 2870
        param_attr (ParamAttr): The attributes of parameter weights, used to set
            initializer, name etc.
Y
yangyaming 已提交
2871 2872
        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
2873 2874
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
Y
yangyaming 已提交
2875 2876

    Returns:
Y
yangyaming 已提交
2877
        tuple: The hidden value and cell value of lstm unit.
Y
yangyaming 已提交
2878 2879

    Raises:
2880 2881 2882 2883
        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
2884 2885 2886 2887 2888 2889

    Examples:

        .. code-block:: python

             x_t = fluid.layers.fc(input=x_t_data, size=10)
2890
             prev_hidden = fluid.layers.fc(input=prev_hidden_data, size=30)
Y
yangyaming 已提交
2891
             prev_cell = fluid.layers.fc(input=prev_cell_data, size=30)
Y
yangyaming 已提交
2892
             hidden_value, cell_value = fluid.layers.lstm_unit(x_t=x_t,
Y
yangyaming 已提交
2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908
                                                    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
2909
        raise ValueError("The 1st dimensions of x_t, hidden_t_prev and "
2910 2911 2912 2913
                         "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
2914 2915
                         "cell_t_prev must be the same.")

Y
yangyaming 已提交
2916 2917 2918
    if bias_attr is None:
        bias_attr = ParamAttr()

Y
yangyaming 已提交
2919
    size = cell_t_prev.shape[1]
2920
    concat_out = concat(input=[x_t, hidden_t_prev], axis=1)
Y
yangyaming 已提交
2921 2922
    fc_out = fc(input=concat_out,
                size=4 * size,
Y
yangyaming 已提交
2923
                param_attr=param_attr,
2924
                bias_attr=bias_attr)
Y
yangyaming 已提交
2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936
    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
2937
    return h, c
G
guosheng 已提交
2938 2939


C
caoying03 已提交
2940
def reduce_sum(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
2941
    """
Y
yangyaming 已提交
2942
    Computes the sum of tensor elements over the given dimension.
G
guosheng 已提交
2943 2944 2945

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
2946
        dim (list|int|None): The dimensions along which the sum is performed. If
Y
yangyaming 已提交
2947 2948
            :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
2949 2950
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
2951
        keep_dim (bool|False): Whether to reserve the reduced dimension in the
Y
yangyaming 已提交
2952
            output Tensor. The result tensor will have one fewer dimension
G
guosheng 已提交
2953
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
2954 2955
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
2956 2957 2958

    Returns:
        Variable: The reduced Tensor variable.
F
fengjiayi 已提交
2959

G
guosheng 已提交
2960 2961 2962 2963 2964 2965
    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
2966
            # Each example is followed by the corresponding output tensor.
G
guosheng 已提交
2967 2968 2969 2970
            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
2971 2972 2973 2974

            # 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
2975
            # Each example is followed by the corresponding output tensor.
W
whs 已提交
2976 2977 2978
            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
2979 2980 2981
    """
    helper = LayerHelper('reduce_sum', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
2982 2983
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
G
guosheng 已提交
2984 2985 2986 2987 2988
    helper.append_op(
        type='reduce_sum',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
2989
            'dim': dim if dim != None else [0],
G
guosheng 已提交
2990 2991 2992 2993
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
2994 2995


C
caoying03 已提交
2996
def reduce_mean(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
2997
    """
Y
Yibing Liu 已提交
2998
    Computes the mean of the input tensor's elements along the given dimension.
G
guosheng 已提交
2999 3000 3001

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
Yibing Liu 已提交
3002 3003 3004
        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
3005
            must be in the range :math:`[-rank(input), rank(input))`. If
3006
            :math:`dim[i] < 0`, the dimension to reduce is
Y
Yibing Liu 已提交
3007
            :math:`rank(input) + dim[i]`.
Y
yangyaming 已提交
3008 3009
        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
3010
            than the :attr:`input` unless :attr:`keep_dim` is true.
Y
Yibing Liu 已提交
3011
        name(str|None): A name for this layer(optional). If set `None`, the layer
C
caoying03 已提交
3012
                       will be named automatically.
G
guosheng 已提交
3013 3014

    Returns:
Y
Yibing Liu 已提交
3015
        Variable: The reduced mean Variable.
F
fengjiayi 已提交
3016

G
guosheng 已提交
3017 3018 3019 3020 3021 3022 3023 3024 3025 3026
    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
3027 3028
            fluid.layers.reduce_mean(
                x, dim=1, keep_dim=True)  # [[0.475], [0.4]]
W
whs 已提交
3029 3030 3031 3032 3033 3034 3035

            # 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
3036 3037 3038
    """
    helper = LayerHelper('reduce_mean', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3039 3040
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
G
guosheng 已提交
3041 3042 3043 3044 3045
    helper.append_op(
        type='reduce_mean',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3046
            'dim': dim if dim != None else [0],
G
guosheng 已提交
3047 3048 3049 3050
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
3051 3052


C
caoying03 已提交
3053
def reduce_max(input, dim=None, keep_dim=False, name=None):
3054
    """
Y
yangyaming 已提交
3055
    Computes the maximum of tensor elements over the given dimension.
3056 3057 3058

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
3059
        dim (list|int|None): The dimension along which the maximum is computed.
Y
yangyaming 已提交
3060 3061 3062
            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
3063
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
Y
yangyaming 已提交
3064 3065
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
3066
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
3067 3068
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
3069 3070 3071

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
3072

3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083
    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
3084 3085 3086 3087 3088 3089 3090

            # 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]
3091 3092 3093
    """
    helper = LayerHelper('reduce_max', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3094 3095
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3096 3097 3098 3099 3100
    helper.append_op(
        type='reduce_max',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3101
            'dim': dim if dim != None else [0],
3102 3103 3104 3105 3106 3107
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


C
caoying03 已提交
3108
def reduce_min(input, dim=None, keep_dim=False, name=None):
3109
    """
Y
yangyaming 已提交
3110
    Computes the minimum of tensor elements over the given dimension.
3111 3112 3113

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
3114
        dim (list|int|None): The dimensions along which the minimum is computed.
Y
yangyaming 已提交
3115 3116 3117
            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
3118
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
Y
yangyaming 已提交
3119 3120
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
3121
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
3122 3123
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
3124 3125 3126

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
3127

3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138
    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
3139 3140 3141 3142 3143 3144 3145

            # 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]
3146 3147 3148
    """
    helper = LayerHelper('reduce_min', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3149 3150
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3151 3152 3153 3154 3155
    helper.append_op(
        type='reduce_min',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3156
            'dim': dim if dim != None else [0],
3157 3158 3159 3160
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
3161 3162


3163 3164 3165 3166 3167 3168
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
3169
        dim (list|int|None): The dimensions along which the product is performed. If
3170 3171
            :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
3172 3173
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
3174 3175 3176
        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
3177
        name(str|None): A name for this layer(optional). If set None, the
Z
zhouhanqing 已提交
3178
            layer will be named automatically.
3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192

    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
3193
            fluid.layers.reduce_prod(x, dim=1,
Z
zhouhanqing 已提交
3194
                                     keep_dim=True)  # [[0.027], [0.0084]]
W
whs 已提交
3195 3196 3197 3198 3199 3200 3201

            # 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]
3202 3203 3204
    """
    helper = LayerHelper('reduce_prod', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3205 3206
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3207 3208 3209 3210 3211
    helper.append_op(
        type='reduce_prod',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3212
            'dim': dim if dim != None else [0],
3213 3214 3215 3216 3217 3218
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


C
caoying03 已提交
3219
def split(input, num_or_sections, dim=-1, name=None):
G
guosheng 已提交
3220
    """
C
caoying03 已提交
3221
    Split the input tensor into multiple sub-tensors.
G
guosheng 已提交
3222 3223 3224

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
C
caoying03 已提交
3225 3226 3227 3228 3229
        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
3230
            :attr:`dim` dimension orderly.
C
caoying03 已提交
3231
        dim (int): The dimension along which to split. If :math:`dim < 0`, the
G
guosheng 已提交
3232
            dimension to split along is :math:`rank(input) + dim`.
C
caoying03 已提交
3233 3234
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
3235 3236

    Returns:
D
dzhwinter 已提交
3237
        list(Variable): The list of segmented tensor variables.
G
guosheng 已提交
3238 3239 3240 3241 3242 3243 3244 3245 3246

    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
3247 3248
            x0, x1, x2 = fluid.layers.split(
                x, num_or_sections=[2, 3, 4], dim=1)
G
guosheng 已提交
3249 3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277
            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
3278 3279 3280 3281 3282 3283 3284 3285 3286


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

3287
    .. math::
3288 3289

        y = \\frac{x}{ \sqrt{\sum {x^2} + epsion }}
C
caoying03 已提交
3290 3291 3292 3293 3294

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

    Args:
3295
        x(Variable|list): The input tensor to l2_normalize layer.
3296
        axis(int): The axis on which to apply normalization. If `axis < 0`, \
3297 3298
            the dimension to normalization is rank(X) + axis. -1 is the
            last dimension.
3299
        epsilon(float): The epsilon value is used to avoid division by zero, \
3300
            the defalut value is 1e-10.
3301
        name(str|None): A name for this layer(optional). If set None, the layer \
3302
            will be named automatically.
C
caoying03 已提交
3303 3304

    Returns:
3305
        Variable: The output tensor variable is the same shape with `x`.
C
caoying03 已提交
3306 3307

    Examples:
3308

C
caoying03 已提交
3309 3310
        .. code-block:: python

3311 3312 3313 3314
            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
3315 3316
    """

F
fengjiayi 已提交
3317 3318
    if len(x.shape) == 1:
        axis = 0
C
caoying03 已提交
3319 3320
    helper = LayerHelper("l2_normalize", **locals())

3321 3322
    out = helper.create_tmp_variable(dtype=x.dtype)
    norm = helper.create_tmp_variable(dtype=x.dtype)
C
caoying03 已提交
3323
    helper.append_op(
3324 3325 3326 3327
        type="norm",
        inputs={"X": x},
        outputs={"Out": out,
                 "Norm": norm},
C
caoying03 已提交
3328
        attrs={
3329 3330
            "axis": 1 if axis is None else axis,
            "epsilon": epsilon,
C
caoying03 已提交
3331 3332
        })
    return out
3333 3334


3335
def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
G
guosheng 已提交
3336
    """
Y
ying 已提交
3337 3338 3339 3340
    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
3341

C
chengduoZH 已提交
3342
    The actual behavior depends on the shapes of :math:`x`, :math:`y` and the
3343
    flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically:
G
guosheng 已提交
3344

3345 3346 3347 3348 3349
    - 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
3350
      :math:`[1, D]` in transposed form.
G
guosheng 已提交
3351

C
chengduoZH 已提交
3352
    - After transpose, the two tensors are 2-D or n-D and matrix multiplication
3353
      performs in the following way.
G
guosheng 已提交
3354

3355
      - If both are 2-D, they are multiplied like conventional matrices.
C
chengduoZH 已提交
3356
      - If either is n-D, it is treated as a stack of matrices residing in the
Y
ying 已提交
3357
        last two dimensions and a batched matrix multiply supporting broadcast
3358
        applies on the two tensors.
G
guosheng 已提交
3359

Y
ying 已提交
3360 3361
    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
3362
    removed after matrix multiplication.
G
guosheng 已提交
3363 3364 3365

    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
3366 3367 3368
        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.
3369
        name(str|None): A name for this layer(optional). If set None, the layer
3370
            will be named automatically.
G
guosheng 已提交
3371 3372

    Returns:
3373
        Variable: The product Tensor variable.
G
guosheng 已提交
3374

G
guosheng 已提交
3375 3376 3377
    Examples:
        .. code-block:: python

3378
            # Examples to clarify shapes of the inputs and output
C
chengduoZH 已提交
3379 3380
            # x: [B, ..., M, K], y: [B, ..., K, N]
            fluid.layers.matmul(x, y)  # out: [B, ..., M, N]
Y
ying 已提交
3381

3382 3383
            # x: [B, M, K], y: [B, K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
3384

3385 3386
            # x: [B, M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
3387

3388 3389
            # x: [M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [M, N]
Y
ying 已提交
3390 3391 3392 3393

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

3394 3395
            # x: [K], y: [K]
            fluid.layers.matmul(x, y)  # out: [1]
3396

Y
ying 已提交
3397
            # x: [M], y: [N]
3398
            fluid.layers.matmul(x, y, True, True)  # out: [M, N]
G
guosheng 已提交
3399
    """
Y
ying 已提交
3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411

    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
3412
            y_shape = y_shape + [1]
Y
ying 已提交
3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428

        # 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)

3429
    helper = LayerHelper('matmul', **locals())
Y
ying 已提交
3430
    out = helper.create_tmp_variable(dtype=x.dtype)
G
guosheng 已提交
3431
    helper.append_op(
3432 3433 3434 3435 3436 3437 3438
        type='matmul',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
        attrs={'transpose_X': transpose_x,
               'transpose_Y': transpose_y})
    return out
3439 3440


3441
def topk(input, k, name=None):
Q
qingqing01 已提交
3442 3443 3444 3445
    """
    This operator is used to find values and indices of the k largest entries
    for the last dimension.

F
fengjiayi 已提交
3446
    If the input is a vector (1-D Tensor), finds the k largest entries in the vector
Q
qingqing01 已提交
3447 3448 3449 3450 3451 3452
    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
3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473
    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
3474 3475 3476
    Args:
        input(Variable): The input variable which can be a vector or Tensor with
            higher rank.
3477
        k(int):  The number of top elements to look for along the last dimension
F
fengjiayi 已提交
3478
                 of input.
3479
        name(str|None): A name for this layer(optional). If set None, the layer
3480
                       will be named automatically.
F
fengjiayi 已提交
3481
                       Default: None
Q
qingqing01 已提交
3482 3483

    Returns:
3484 3485 3486
        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
3487
        within the last dimension of input.
Q
qingqing01 已提交
3488

F
fengjiayi 已提交
3489 3490
    Raises:
        ValueError: If k < 1 or k is not less than the last dimension of input
Q
qingqing01 已提交
3491 3492 3493 3494 3495 3496 3497

    Examples:
        .. code-block:: python

            top5_values, top5_indices = layers.topk(input, k=5)
    """
    shape = input.shape
F
fengjiayi 已提交
3498
    if k < 1 or k >= shape[-1]:
Q
qingqing01 已提交
3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515
        raise ValueError("k must be greater than 0 and less than %d." %
                         (shape[-1]))

    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


3516
def edit_distance(input, label, normalized=True, ignored_tokens=None):
3517
    """
Y
ying 已提交
3518 3519 3520 3521 3522 3523 3524 3525 3526
    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
3527

Y
ying 已提交
3528
    "kitten" -> "sitten" -> "sittin" -> "sitting"
W
wanghaoshuang 已提交
3529

3530
    The input is a LoDTensor consisting of all the hypothesis strings with
Y
ying 已提交
3531 3532
    the total number denoted by `batch_size`, and the separation is specified
    by the LoD information. And the `batch_size` reference strings are arranged
3533
    in order in the same way in the input LoDTensor.
W
wanghaoshuang 已提交
3534

3535
    The output contains the `batch_size` results and each stands for the edit
Y
ying 已提交
3536 3537
    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
3538

3539 3540 3541
    Args:
        input(Variable): The indices for hypothesis strings.
        label(Variable): The indices for reference strings.
3542
        normalized(bool, default True): Indicated whether to normalize the edit distance by
Y
ying 已提交
3543
                          the length of reference string.
3544
        ignored_tokens(list<int>, default None): Tokens that should be removed before
Y
ying 已提交
3545
                                     calculating edit distance.
3546
        name (str): The name of this layer. It is optional.
3547

W
wanghaoshuang 已提交
3548
    Returns:
W
wanghaoshuang 已提交
3549
        Variable: sequence-to-sequence edit distance in shape [batch_size, 1].
W
wanghaoshuang 已提交
3550 3551 3552 3553 3554

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
3555
            y = fluid.layers.data(name='y', shape=[7], dtype='float32')
3556
            cost = fluid.layers.edit_distance(input=x,label=y)
3557
    """
3558
    helper = LayerHelper("edit_distance", **locals())
3559

3560
    # remove some tokens from input and labels
W
wanghaoshuang 已提交
3561
    if ignored_tokens is not None and len(ignored_tokens) > 0:
3562 3563 3564 3565 3566 3567 3568
        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
3569
            attrs={"tokens": ignored_tokens})
3570 3571 3572 3573 3574
        input = erased_input

        helper.append_op(
            type="sequence_erase",
            inputs={"X": [label]},
W
whs 已提交
3575
            outputs={"Out": [erased_label]},
W
wanghaoshuang 已提交
3576
            attrs={"tokens": ignored_tokens})
3577 3578
        label = erased_label

3579 3580
    # edit distance op
    edit_distance_out = helper.create_tmp_variable(dtype="int64")
3581
    sequence_num = helper.create_tmp_variable(dtype="int64")
3582 3583 3584 3585
    helper.append_op(
        type="edit_distance",
        inputs={"Hyps": [input],
                "Refs": [label]},
3586 3587
        outputs={"Out": [edit_distance_out],
                 "SequenceNum": [sequence_num]},
3588 3589
        attrs={"normalized": normalized})

3590
    return edit_distance_out, sequence_num
3591 3592 3593 3594 3595


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
3596

Y
ying 已提交
3597 3598 3599 3600
    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.
3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617

    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]]

3618
        input.lod = [[4, 4]]
3619 3620 3621 3622 3623 3624 3625

        Then:

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

3626
        output.lod = [[2, 1]]
3627 3628 3629

    Args:

Y
ying 已提交
3630 3631 3632 3633 3634 3635 3636 3637 3638
        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).
3639
        name (str): The name of this layer. It is optional.
3640 3641

    Returns:
3642
        Variable: CTC greedy decode result. If all the sequences in result were
3643
        empty, the result LoDTensor will be [-1] with LoD [[]] and dims [1, 1].
3644 3645 3646 3647 3648

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
W
wanghaoshuang 已提交
3649

3650
            cost = fluid.layers.ctc_greedy_decoder(input=x, blank=0)
W
wanghaoshuang 已提交
3651
    """
3652
    helper = LayerHelper("ctc_greedy_decoder", **locals())
Q
qingqing01 已提交
3653
    _, topk_indices = topk(input, k=1)
3654 3655 3656 3657 3658 3659

    # 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
3660
        outputs={"Output": [ctc_out]},
3661 3662
        attrs={"merge_repeated": True,
               "blank": blank})
3663
    return ctc_out
3664 3665


F
fengjiayi 已提交
3666
def warpctc(input, label, blank=0, norm_by_times=False):
W
wanghaoshuang 已提交
3667
    """
3668 3669
    An operator integrating the open source Warp-CTC library
    (https://github.com/baidu-research/warp-ctc)
W
wanghaoshuang 已提交
3670
    to compute Connectionist Temporal Classification (CTC) loss.
3671 3672
    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
3673 3674 3675
    input tensor.

    Args:
3676
       input (Variable): The unscaled probabilities of variable-length sequences,
W
wanghaoshuang 已提交
3677 3678 3679 3680
         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).
3681
       label (Variable): The ground truth of variable-length sequence,
3682 3683 3684
         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
3685 3686
         Temporal Classification (CTC) loss, which is in the
         half-opened interval [0, num_classes + 1).
3687 3688 3689
       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
3690
         follewed by a mean_op.
W
wanghaoshuang 已提交
3691 3692

    Returns:
3693 3694
        Variable: The Connectionist Temporal Classification (CTC) loss,
        which is a 2-D Tensor of the shape [batch_size, 1].
W
wanghaoshuang 已提交
3695 3696

    Examples:
3697

W
wanghaoshuang 已提交
3698
        .. code-block:: python
3699

3700 3701 3702
            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
3703 3704

    """
F
fengjiayi 已提交
3705
    helper = LayerHelper('warpctc', **locals())
W
wanghaoshuang 已提交
3706 3707 3708 3709 3710 3711 3712 3713 3714 3715 3716
    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
3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730 3731


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]]
3732 3733 3734
            x.data = [[1,  2], [3,  4],
                      [5,  6], [7,  8],
                      [9, 10], [11, 12]]
3735 3736 3737 3738 3739
            x.dims = [6, 2]

        set new_dim = 4

        then out is a LoDTensor:
3740

3741
            out.lod  = [[0, 1, 3]]
3742 3743 3744 3745

            out.data = [[1,  2,  3,  4],
                        [5,  6,  7,  8],
                        [9, 10, 11, 12]]
3746 3747 3748 3749 3750 3751 3752
            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:
3753 3754 3755

       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.
3756 3757

    Returns:
3758

3759 3760 3761 3762 3763
        Variable: Reshaped LoDTensor according to new dimension.

    Examples:
        .. code-block:: python

3764
            x = fluid.layers.data(shape=[5, 20], dtype='float32', lod_level=1)
3765
            x_reshaped = fluid.layers.sequence_reshape(input=x, new_dim=10)
3766 3767 3768 3769 3770 3771 3772 3773 3774
    """
    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
3775 3776


3777 3778 3779 3780
# 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
3781 3782 3783 3784 3785 3786 3787
def nce(input,
        label,
        num_total_classes,
        sample_weight=None,
        param_attr=None,
        bias_attr=None,
        num_neg_samples=None):
3788 3789 3790 3791 3792 3793 3794
    """
    ${comment}

    Args:
        input (Variable): input variable.
        label (Variable): label.
        num_total_classes (int):${num_total_classes_comment}
3795 3796
        sample_weight (Variable|None): A Variable of shape [batch_size, 1]
            storing a weight for each sample. The default weight for each
3797
            sample is 1.0.
3798 3799 3800
        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
3801

3802
    Returns:
Y
Yibing Liu 已提交
3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826 3827 3828 3829
        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')
3830
    """
Y
Yang Yu 已提交
3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848 3849
    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
3850 3851 3852 3853 3854 3855 3856 3857 3858
    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
3859 3860 3861 3862 3863 3864 3865 3866 3867 3868 3869 3870 3871 3872 3873 3874

    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
3875
    return cost / (num_neg_samples + 1)
3876 3877


G
guosheng 已提交
3878
def hsigmoid(input, label, num_classes, param_attr=None, bias_attr=None):
W
weixing02 已提交
3879 3880
    """
    The hierarchical sigmoid operator is used to accelerate the training
M
minqiyang 已提交
3881
    process of language model. This operator organizes the classes into a
G
guosheng 已提交
3882 3883 3884 3885 3886 3887 3888 3889 3890
    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
3891

W
weixing02 已提交
3892
    Args:
M
minqiyang 已提交
3893
        input (Variable): The input tensor variable with shape
G
guosheng 已提交
3894 3895 3896 3897 3898
            :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
3899 3900
        param_attr (ParamAttr|list of ParamAttr, default None): The parameter
             attribute for learnable parameters/weights of this layer.
M
minqiyang 已提交
3901
        bias_attr (ParamAttr|list of ParamAttr, default None):  The parameter
G
guosheng 已提交
3902 3903
             attribute for the bias of this layer. If it is set to False, no
             bias will be applied.
W
weixing02 已提交
3904 3905 3906 3907 3908 3909 3910 3911

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

    Examples:

        .. code-block:: python

G
guosheng 已提交
3912 3913 3914
            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
3915 3916 3917 3918 3919 3920 3921 3922
    """

    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
3923
        raise ValueError("num_classes must not be less than 2.")
W
weixing02 已提交
3924 3925 3926 3927 3928
    weights = helper.create_parameter(
        attr=helper.param_attr,
        shape=[num_classes - 1, dim],
        is_bias=False,
        dtype=input.dtype)
W
weixing02 已提交
3929 3930 3931 3932 3933 3934 3935 3936
    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
3937 3938
    helper.append_op(
        type="hierarchical_sigmoid",
W
weixing02 已提交
3939
        inputs=inputs,
W
weixing02 已提交
3940 3941 3942 3943 3944 3945
        outputs={"Out": out,
                 "PreOut": pre_out},
        attrs={"num_classes": num_classes})
    return out


Y
fix ci.  
ying 已提交
3946
def transpose(x, perm, name=None):
Y
ying 已提交
3947 3948 3949 3950 3951 3952 3953
    """
    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:
3954 3955 3956
        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
3957 3958 3959 3960 3961 3962 3963 3964

    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
3965
            x_transposed = layers.transpose(x, perm=[1, 0, 2])
Y
ying 已提交
3966 3967
    """

Y
fix ci.  
ying 已提交
3968
    if len(perm) != len(x.shape):
Y
ying 已提交
3969 3970 3971
        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
3972 3973 3974 3975 3976 3977
    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
3978 3979

    helper = LayerHelper('transpose', **locals())
Y
fix ci.  
ying 已提交
3980
    out = helper.create_tmp_variable(x.dtype)
Y
ying 已提交
3981 3982
    helper.append_op(
        type='transpose',
Y
fix ci.  
ying 已提交
3983
        inputs={'X': [x]},
Y
ying 已提交
3984 3985 3986
        outputs={'Out': [out]},
        attrs={'axis': perm})
    return out
3987 3988


3989 3990 3991 3992 3993 3994 3995
def im2sequence(input,
                filter_size=1,
                stride=1,
                padding=0,
                input_image_size=None,
                out_stride=1,
                name=None):
3996
    """
3997 3998 3999 4000 4001 4002 4003
    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:
4004 4005 4006 4007 4008 4009 4010 4011 4012 4013

    .. 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
4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031

        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.

4032 4033 4034 4035 4036 4037 4038 4039 4040
        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.

4041 4042 4043
        name (int): The name of this layer. It is optional.

    Returns:
W
wanghaoshuang 已提交
4044 4045 4046 4047 4048
        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.
4049 4050 4051 4052 4053 4054 4055 4056 4057 4058 4059 4060 4061 4062 4063 4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075

    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
4076 4077 4078
            filter = [2, 2]
            stride = [1, 1]
            padding = [0, 0]
4079 4080 4081 4082 4083 4084 4085 4086 4087 4088 4089 4090

            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.]]

4091
            output.dims = {8, 8}
4092

4093
            output.lod = [[4, 4]]
4094

D
dzhwinter 已提交
4095
     Examples:
4096 4097 4098

        .. code-block:: python

4099 4100
            output = fluid.layers.im2sequence(
                input=layer, stride=[1, 1], filter_size=[2, 2])
4101 4102

    """
W
wanghaoshuang 已提交
4103 4104 4105 4106 4107 4108 4109 4110 4111 4112

    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])
4113 4114 4115 4116 4117 4118 4119
    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
4120
    helper = LayerHelper('im2sequence', **locals())
4121 4122
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
4123
        type='im2sequence', inputs=inputs, outputs={'Out': out}, attrs=attrs)
4124
    return out
4125 4126


Y
yuyang18 已提交
4127
@templatedoc()
4128
def row_conv(input, future_context_size, param_attr=None, act=None):
Y
yuyang18 已提交
4129 4130
    """
    ${comment}
4131 4132

    Args:
Y
yuyang18 已提交
4133
        input (${x_type}): ${x_comment}.
Y
yangyaming 已提交
4134 4135
        future_context_size (int): Future context size. Please note, the shape
            of convolution kernel is [future_context_size + 1, D].
4136 4137 4138 4139 4140
        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
4141
        ${out_comment}.
4142 4143

    Examples:
Y
yuyang18 已提交
4144 4145 4146 4147
        >>> 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)
4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159
    """
    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
4160
    return helper.append_activation(out)
4161 4162


Y
yuyang18 已提交
4163
@templatedoc()
4164 4165
def multiplex(inputs, index):
    """
Y
yuyang18 已提交
4166 4167 4168 4169 4170 4171 4172
    ${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)
4173 4174

    Args:
Y
yuyang18 已提交
4175 4176
       inputs (list): ${x_comment}.
       index (${ids_type}): ${ids_comment}.
4177 4178

    Returns:
Y
yuyang18 已提交
4179
        ${out_comment}.
4180 4181
    """
    helper = LayerHelper('multiplex', **locals())
Y
yangyaming 已提交
4182 4183 4184 4185 4186 4187

    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)
4188 4189 4190 4191 4192 4193
    helper.append_op(
        type='multiplex',
        inputs={'X': inputs,
                'Ids': index},
        outputs={'Out': [out]})
    return out
4194 4195 4196 4197 4198


def softmax_with_cross_entropy(logits, label, soft_label=False):
    """
    **Softmax With Cross Entropy Operator.**
4199

4200 4201 4202 4203
    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.
4204

4205 4206 4207
    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.
4208

4209 4210 4211
    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.
4212

4213
    The equation is as follows:
4214

4215
    1) Hard label (one-hot label, so every sample has exactly one class)
4216

4217 4218 4219 4220
    .. math::

        loss_j =  -\\text{logit}_{label_j} +
        \\log\\left(\\sum_{i=0}^{K}\\exp(\\text{logit}_i)\\right), j = 1,..., K
4221

4222 4223 4224
    2) Soft label (each sample can have a distribution over all classes)

    .. math::
4225

4226 4227 4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238 4239 4240 4241 4242 4243 4244 4245 4246
        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.
    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
4247 4248
            out = fluid.layers.softmax_with_cross_entropy(
                logits=fc, label=label)
4249 4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261 4262 4263 4264
    """
    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},
        attrs={'soft_label': soft_label})
    return loss


def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None):
    """
Y
Yibing Liu 已提交
4265 4266
    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
4267
    For each instance, it computes the smooth L1 loss element by element first
4268
    and then sums all the losses. So the shape of ouput Variable is
4269
    [batch_size, 1].
4270

4271 4272
    Args:
        x (Variable): A tensor with rank at least 2. The input value of smooth
Q
qingqing01 已提交
4273
            L1 loss op with shape [batch_size, dim1, ..., dimN].
4274
        y (Variable): A tensor with rank at least 2. The target value of smooth
Y
Yibing Liu 已提交
4275
            L1 loss op with same shape as :attr:`x`.
4276
        inside_weight (Variable|None):  A tensor with rank at least 2. This
4277 4278
            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
4279
            by this tensor element by element.
4280
        outside_weight (Variable|None): A tensor with rank at least 2. This
4281 4282
            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
4283
            element by element.
4284
        sigma (float|None): Hyper parameter of smooth L1 loss layer. A float
4285 4286
           scalar with default value 1.0.

4287
    Returns:
4288
        Variable: The output smooth L1 loss with shape [batch_size, 1].
4289 4290 4291 4292 4293

    Examples:
        .. code-block:: python

            data = fluid.layers.data(name='data', shape=[128], dtype='float32')
F
stash  
fengjiayi 已提交
4294 4295
            label = fluid.layers.data(
                name='label', shape=[100], dtype='float32')
4296
            fc = fluid.layers.fc(input=data, size=100)
F
fengjiayi 已提交
4297
            out = fluid.layers.smooth_l1(x=fc, y=label)
4298
    """
4299

4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312 4313 4314
    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
4315 4316 4317 4318


def one_hot(input, depth):
    """
Y
Yibing Liu 已提交
4319
    This layer creates the one-hot representations for input indices.
4320 4321

    Args:
Y
Yibing Liu 已提交
4322 4323
        input(Variable): Input indices, last dimension must be 1.
        depth(scalar): An interger defining the depth of the one-hot dimension.
4324 4325

    Returns:
Y
Yibing Liu 已提交
4326
        Variable: The one-hot representations of input.
4327 4328

    Examples:
C
caoying03 已提交
4329
        .. code-block:: python
4330

Y
Yibing Liu 已提交
4331 4332
            label = layers.data(name="label", shape=[1], dtype="float32")
            one_hot_label = layers.one_hot(input=label, depth=10)
4333 4334 4335 4336 4337 4338 4339 4340 4341
    """
    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
4342 4343


Y
Yu Yang 已提交
4344
def autoincreased_step_counter(counter_name=None, begin=1, step=1):
Y
Yu Yang 已提交
4345
    """
Y
yi.wu 已提交
4346 4347 4348
    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
4349 4350 4351 4352 4353 4354

    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.

4355 4356
    Returns:
        Variable: The global run counter.
Y
yi.wu 已提交
4357 4358 4359 4360 4361 4362

    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
4363 4364
    """
    helper = LayerHelper('global_step_counter')
Y
Yu Yang 已提交
4365 4366
    if counter_name is None:
        counter_name = '@STEP_COUNTER@'
Y
Yu Yang 已提交
4367 4368 4369 4370 4371
    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
4372
                value=begin - 1, force_cpu=True))
W
Wu Yi 已提交
4373
        helper.main_program.global_block()._prepend_op(
Y
Yu Yang 已提交
4374 4375
            type='increment',
            inputs={'X': [counter]},
Y
Yu Yang 已提交
4376 4377
            outputs={'Out': [counter]},
            attrs={'step': float(step)})
Y
Yu Yang 已提交
4378 4379 4380
        counter.stop_gradient = True

    return counter
Y
yangyaming 已提交
4381 4382


4383
def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None):
C
caoying03 已提交
4384
    """
C
caoying03 已提交
4385 4386
    Gives a new shape to the input Tensor without changing its data.

4387 4388 4389 4390 4391
    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
4392

4393
    Some tricks exist when specifying the target shape.
C
caoying03 已提交
4394

4395 4396 4397 4398
    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.

4399
    2. 0 means the actual dimension value is going to be copied from the
4400 4401 4402 4403
    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
4404 4405

    1. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
W
wanghaoshuang 已提交
4406
    is [6, 8], the reshape operator will transform x into a 2-D tensor with
4407
    shape [6, 8] and leaving x's data unchanged.
C
caoying03 已提交
4408

4409
    2. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
4410 4411
    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
4412 4413
    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
4414
    dimensions.
C
caoying03 已提交
4415

4416
    3. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
4417 4418 4419 4420
    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
4421 4422

    Args:
4423
        x(variable): The input tensor.
C
caoying03 已提交
4424 4425
        shape(list): The new shape. At most one dimension of the new shape can
                     be -1.
4426 4427 4428 4429 4430
        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
4431
        act (str): The non-linear activation to be applied to output variable.
X
Xin Pan 已提交
4432 4433 4434 4435
        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.
4436
        name (str): The name of this layer. It is optional.
C
caoying03 已提交
4437

4438 4439
    Returns:
        Variable: The output tensor.
C
caoying03 已提交
4440

X
Xin Pan 已提交
4441 4442 4443
    Raises:
        TypeError: if actual_shape is neither Variable nor None.

C
caoying03 已提交
4444 4445
    Examples:
        .. code-block:: python
G
guosheng 已提交
4446

4447
            data = fluid.layers.data(
4448
                name='data', shape=[2, 4, 6], dtype='float32')
C
caoying03 已提交
4449
            reshaped = fluid.layers.reshape(
4450
                x=data, shape=[-1, 0, 3, 2], act='tanh', inplace=True)
C
caoying03 已提交
4451 4452 4453 4454
    """

    if not (isinstance(shape, list) or isinstance(shape, tuple)):
        raise ValueError("Input shape must be a python lsit or tuple.")
X
Xin Pan 已提交
4455 4456 4457 4458 4459
    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
4460

4461 4462 4463 4464 4465 4466 4467 4468 4469 4470 4471 4472 4473 4474 4475
    # 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.")

C
caoying03 已提交
4476
    helper = LayerHelper("reshape", **locals())
D
dzhwinter 已提交
4477
    out = helper.create_tmp_variable(dtype=x.dtype)
C
caoying03 已提交
4478 4479
    helper.append_op(
        type="reshape",
X
Xin Pan 已提交
4480
        inputs=inputs,
D
dzhwinter 已提交
4481 4482
        attrs={"shape": shape},
        outputs={"Out": out})
C
caoying03 已提交
4483

D
dzhwinter 已提交
4484
    return helper.append_activation(out)
4485 4486


Y
yangyaming 已提交
4487
def lod_reset(x, y=None, target_lod=None):
Y
yangyaming 已提交
4488
    """
Y
Yibing Liu 已提交
4489
    Set LoD of :attr:`x` to a new one specified by :attr:`y` or
4490 4491 4492 4493
    :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
4494
    :attr:`Y.data` or :attr:`target_lod`, only one level LoD is supported.
Y
yangyaming 已提交
4495 4496 4497 4498 4499 4500

    .. code-block:: text

        * Example 1:

            Given a 1-level LoDTensor x:
4501
                x.lod =  [[ 2,           3,                   1 ]]
Y
yangyaming 已提交
4502 4503 4504
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

4505
            target_lod: [4, 2]
Y
yangyaming 已提交
4506 4507

            then we get a 1-level LoDTensor:
4508
                out.lod =  [[4,                          2]]
Y
yangyaming 已提交
4509 4510 4511 4512 4513 4514
                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:
4515
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
4516 4517 4518 4519
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a Tensor:
4520
                y.data = [[2, 4]]
Y
yangyaming 已提交
4521 4522 4523
                y.dims = [1, 3]

            then we get a 1-level LoDTensor:
4524
                out.lod =  [[2,            4]]
Y
yangyaming 已提交
4525 4526 4527 4528 4529 4530
                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:
4531
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
4532 4533 4534 4535
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a 2-level LoDTensor:
4536
                y.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
4537 4538 4539 4540
                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:
4541
                out.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
4542 4543 4544 4545 4546
                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.
4547
        y (Variable|None): If provided, output's LoD would be derived
Y
Yibing Liu 已提交
4548
                           from :attr:`y`.
Y
yangyaming 已提交
4549
        target_lod (list|tuple|None): One level LoD which should be considered
Y
Yibing Liu 已提交
4550
                                      as target LoD when :attr:`y` not provided.
Y
yangyaming 已提交
4551 4552

    Returns:
Y
Yibing Liu 已提交
4553
        Variable: Output variable with LoD specified by this layer.
Y
yangyaming 已提交
4554 4555

    Raises:
Y
Yibing Liu 已提交
4556
        ValueError: If :attr:`y` and :attr:`target_lod` are both None.
Y
yangyaming 已提交
4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576 4577 4578 4579 4580

    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
4581 4582 4583 4584 4585 4586 4587 4588 4589 4590 4591


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
4592
      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
4593 4594 4595 4596 4597 4598 4599 4600 4601 4602 4603 4604 4605 4606 4607 4608 4609 4610 4611 4612 4613 4614 4615 4616 4617 4618 4619 4620

    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
4621 4622
          data = fluid.layers.data(
              name="data", shape=[3, 112, 112], dtype="float32")
D
dragonwarrior 已提交
4623 4624 4625 4626 4627 4628 4629 4630 4631 4632 4633 4634 4635 4636 4637 4638 4639 4640 4641 4642 4643 4644 4645 4646 4647 4648 4649
          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
4650 4651 4652 4653


def pad(x, paddings, pad_value=0., name=None):
    """
G
guosheng 已提交
4654
    Pads a tensor with a constant value given by :attr:`pad_value`, and the
W
wanghaoshuang 已提交
4655
    padded width is specified by :attr:`paddings`.
G
guosheng 已提交
4656

G
guosheng 已提交
4657 4658 4659 4660
    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
4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675 4676 4677 4678 4679 4680 4681 4682

    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
4683
                         The length of :attr:paddings must be
G
guosheng 已提交
4684 4685 4686 4687 4688 4689 4690 4691 4692 4693
                         :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
4694

G
guosheng 已提交
4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706 4707 4708
            # 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
4709 4710 4711 4712 4713 4714 4715 4716 4717


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
4718 4719
    called label-smoothing regularization (LSR).

4720 4721 4722 4723 4724 4725 4726 4727 4728 4729 4730 4731 4732 4733 4734 4735 4736 4737 4738 4739 4740 4741 4742
    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
4743
                              be :math:`(1, class\_num)`.
4744 4745
        epsilon(float): The weight used to mix up the original ground-truth
                        distribution and the fixed distribution.
4746
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32,
4747 4748 4749 4750 4751 4752 4753 4754 4755 4756 4757 4758 4759 4760 4761 4762 4763 4764 4765 4766 4767 4768 4769 4770 4771 4772 4773
                                                  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
4774 4775


Y
yi.wu 已提交
4776
@templatedoc()
4777 4778
def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0):
    """
Y
yi.wu 已提交
4779
    ${comment}
4780 4781

    Args:
Y
yi.wu 已提交
4782 4783
        input (Variable): ${x_comment}
        rois (Variable): ROIs (Regions of Interest) to pool over.
Y
yi.wu 已提交
4784 4785 4786
        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
4787 4788

    Returns:
Y
update  
yi.wu 已提交
4789
        Variable: ${out_comment}.
4790 4791

    Examples:
4792 4793
        .. code-block:: python

4794
            pool_out = fluid.layers.roi_pool(input=x, rois=rois, 7, 7, 1.0)
4795 4796 4797 4798 4799 4800 4801 4802 4803 4804 4805 4806 4807 4808 4809 4810 4811
    """
    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
4812 4813 4814 4815 4816 4817 4818 4819 4820 4821 4822 4823 4824 4825 4826 4827 4828 4829 4830 4831 4832 4833 4834 4835 4836 4837 4838 4839


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:
4840 4841
        .. code-block:: python

W
whs 已提交
4842 4843 4844 4845
            predictions = fluid.layers.softmax(x)
            loss = fluid.layers.dice_loss(input=predictions, label=label, 2)
    """
    label = one_hot(label, depth=input.shape[-1])
4846
    reduce_dim = list(range(1, len(input.shape)))
W
whs 已提交
4847 4848 4849 4850 4851 4852
    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)
4853 4854


4855 4856 4857 4858 4859
def image_resize(input,
                 out_shape=None,
                 scale=None,
                 name=None,
                 resample='BILINEAR'):
4860
    """
Q
qiaolongfei 已提交
4861
    **Resize a Batch of Images**
F
stash  
fengjiayi 已提交
4862

4863
    The input must be a tensor of the shape (num_batches, channels, in_h, in_w),
4864 4865 4866
    and the resizing only applies on the last two dimensions(hight and width).

    Supporting resample methods:
Q
update  
qiaolongfei 已提交
4867

4868
        'BILINEAR' : Bilinear interpolation
F
stash  
fengjiayi 已提交
4869

4870
    Args:
4871
        input (Variable): The input tensor of image resize layer,
4872 4873
                          This is a 4-D tensor of the shape
                          (num_batches, channels, in_h, in_w).
4874
        out_shape(list|tuple|Variable|None): Output shape of image resize
4875 4876
                                    layer, the shape is (out_h, out_w).
                                    Default: None
B
baiyf 已提交
4877
        scale(float|None): The multiplier for the input height or width.
4878 4879 4880
                         At least one of out_shape or scale must be set.
                         And out_shape has a higher priority than scale.
                         Default: None
4881 4882
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
4883 4884
        resample(str): The resample method. It can only be 'BILINEAR' currently.
                       Default: 'BILINEAR'
4885 4886

    Returns:
Q
update  
qiaolongfei 已提交
4887 4888
        Variable: The output is a 4-D tensor of the shape
        (num_batches, channls, out_h, out_w).
F
stash  
fengjiayi 已提交
4889

4890 4891 4892
    Examples:
        .. code-block:: python

4893
            out = fluid.layers.image_resize(input, out_shape=[12, 12])
4894
    """
4895 4896 4897 4898
    resample_methods = {'BILINEAR': 'bilinear_interp'}
    if resample not in resample_methods:
        raise ValueError(
            "The 'resample' of image_resize can only be 'BILINEAR' currently.")
4899 4900
    if out_shape is None and scale is None:
        raise ValueError("One of out_shape and scale must not be None")
4901 4902
    helper = LayerHelper('bilinear_interp', **locals())
    dtype = helper.input_dtype()
4903 4904 4905 4906

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

4907 4908 4909
    out_h = 0
    out_w = 0
    inputs = {"X": input}
4910
    if out_shape is not None:
B
baiyf 已提交
4911 4912 4913
        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')
4914 4915 4916 4917 4918 4919
        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
4920 4921 4922 4923
    else:
        out_h = int(input.shape[2] * scale)
        out_w = int(input.shape[3] * scale)

4924 4925
    out = helper.create_tmp_variable(dtype)
    helper.append_op(
4926
        type=resample_methods[resample],
4927
        inputs=inputs,
4928 4929 4930 4931
        outputs={"Out": out},
        attrs={"out_h": out_h,
               "out_w": out_w})
    return out
F
stash  
fengjiayi 已提交
4932 4933


Y
yuyang18 已提交
4934
@templatedoc(op_type="bilinear_interp")
4935 4936
def resize_bilinear(input, out_shape=None, scale=None, name=None):
    """
Y
yuyang18 已提交
4937 4938 4939 4940 4941 4942
    ${comment}

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

        out_shape(${out_size_type}): ${out_size_comment}.
4943

Y
yuyang18 已提交
4944 4945 4946 4947 4948 4949 4950 4951
        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}.
4952 4953 4954 4955 4956 4957 4958
    """

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


def image_resize_short(input, out_short_len, resample='BILINEAR'):
    """
4959 4960 4961
    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
4962 4963 4964 4965 4966 4967 4968
    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.
4969
        resample (str): resample method, default: BILINEAR.
F
fengjiayi 已提交
4970

4971
    Returns:
Q
update  
qiaolongfei 已提交
4972
        Variable: The output is a 4-D tensor of the shape
4973
        (num_batches, channls, out_h, out_w).
4974 4975 4976 4977 4978 4979 4980 4981 4982 4983
    """
    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
4984 4985 4986
    out_shape[long_idx] = int(
        float(out_shape[long_idx]) * (float(out_short_len) / float(hw[
            short_idx])) + 0.5)
4987 4988 4989
    return image_resize(input=input, out_shape=out_shape, resample=resample)


W
whs 已提交
4990 4991
def gather(input, index):
    """
Q
qiaolongfei 已提交
4992 4993
    **Gather Layer**

4994
    Output is obtained by gathering entries of the outer-most dimension
W
whs 已提交
4995 4996 4997 4998
    of X indexed by `index` and concatenate them together.

    .. math::

4999
        Out = X[Index]
W
whs 已提交
5000 5001 5002 5003 5004 5005 5006


    .. code-block:: text


                Given:

5007 5008
                X = [[1, 2],
                     [3, 4],
W
whs 已提交
5009 5010 5011 5012 5013 5014 5015 5016 5017 5018
                     [5, 6]]

                Index = [1, 2]

                Then:

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

    Args:
5019
        input (Variable): The source input with rank>=1.
W
whs 已提交
5020 5021 5022 5023 5024 5025
        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
5026

W
whs 已提交
5027 5028 5029 5030 5031 5032 5033 5034 5035 5036 5037 5038 5039 5040 5041
        .. 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


Y
yuyang18 已提交
5042 5043 5044 5045 5046 5047 5048 5049 5050 5051 5052 5053 5054
@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}
5055

5056 5057 5058
    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
5059
    """
F
stash  
fengjiayi 已提交
5060
    helper = LayerHelper("random_crop", **locals())
F
fengjiayi 已提交
5061
    dtype = x.dtype
F
stash  
fengjiayi 已提交
5062
    out = helper.create_tmp_variable(dtype)
Y
yuyang18 已提交
5063 5064
    if seed is None:
        seed = random.randint(-65536, 65535)
F
fengjiayi 已提交
5065
    op_attrs = {"shape": shape}
F
stash  
fengjiayi 已提交
5066
    if isinstance(seed, int):
F
fengjiayi 已提交
5067 5068 5069 5070 5071
        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
5072 5073 5074 5075
    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
5076
        inputs={"X": x,
F
stash  
fengjiayi 已提交
5077 5078
                "Seed": seed},
        outputs={"Out": out,
F
fengjiayi 已提交
5079 5080
                 "SeedOut": seed},
        attrs=op_attrs)
F
stash  
fengjiayi 已提交
5081
    return out
W
whs 已提交
5082 5083


5084
def log(x):
W
wanghaoshuang 已提交
5085 5086 5087 5088 5089
    """
    Calculates the natural log of the given input tensor, element-wise.

    .. math::

5090
        Out = \\ln(x)
W
wanghaoshuang 已提交
5091 5092

    Args:
5093
        x (Variable): Input tensor.
W
wanghaoshuang 已提交
5094 5095 5096 5097 5098 5099 5100 5101

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

    Examples:

        .. code-block:: python

5102
            output = fluid.layers.log(x)
W
wanghaoshuang 已提交
5103 5104
    """
    helper = LayerHelper('log', **locals())
W
wanghaoshuang 已提交
5105
    dtype = helper.input_dtype(input_param_name='x')
W
wanghaoshuang 已提交
5106
    out = helper.create_tmp_variable(dtype)
W
wanghaoshuang 已提交
5107
    helper.append_op(type="log", inputs={"X": x}, outputs={"Out": out})
W
wanghaoshuang 已提交
5108 5109 5110
    return out


5111
def relu(x):
W
wanghaoshuang 已提交
5112 5113
    """
    Relu takes one input data (Tensor) and produces one output data (Tensor)
5114
    where the rectified linear function, y = max(0, x), is applied to
W
wanghaoshuang 已提交
5115 5116 5117 5118
    the tensor elementwise.

    .. math::

5119
        Out = \\max(0, x)
W
wanghaoshuang 已提交
5120 5121

    Args:
5122
        x (Variable): The input tensor.
W
wanghaoshuang 已提交
5123 5124 5125 5126 5127 5128 5129 5130

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

    Examples:

        .. code-block:: python

5131
            output = fluid.layers.relu(x)
W
wanghaoshuang 已提交
5132 5133
    """
    helper = LayerHelper('relu', **locals())
W
wanghaoshuang 已提交
5134
    dtype = helper.input_dtype(input_param_name='x')
W
wanghaoshuang 已提交
5135
    out = helper.create_tmp_variable(dtype)
W
wanghaoshuang 已提交
5136
    helper.append_op(type="relu", inputs={"X": x}, outputs={"Out": out})
W
wanghaoshuang 已提交
5137
    return out
5138 5139


W
whs 已提交
5140 5141 5142
def mean_iou(input, label, num_classes):
    """
    Mean Intersection-Over-Union is a common evaluation metric for
5143 5144 5145 5146
    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
5147
    .. math::
5148 5149

        IOU = \\frac{true\_positiv}{(true\_positive + false\_positive + false\_negative)}.
W
whs 已提交
5150

5151
    The predictions are accumulated in a confusion matrix and mean-IOU
W
whs 已提交
5152 5153 5154 5155 5156
    is then calculated from it.


    Args:
        input (Variable): A Tensor of prediction results for semantic labels with type int32 or int64.
5157
        label (Variable): A Tensor of ground truth labels with type int32 or int64.
W
whs 已提交
5158
                           Its shape should be the same as input.
5159
        num_classes (int): The possible number of labels.
W
whs 已提交
5160 5161 5162 5163

    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.
5164
        out_correct(Variable): A Tensor with shape [num_classes]. The correct numbers of each class.
W
whs 已提交
5165 5166 5167 5168

    Examples:

        .. code-block:: python
5169

W
whs 已提交
5170 5171 5172 5173 5174 5175 5176 5177 5178
            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
5179 5180
        inputs={"Predictions": input,
                "Labels": label},
W
whs 已提交
5181
        outputs={
W
whs 已提交
5182 5183 5184
            "OutMeanIou": out_mean_iou,
            "OutWrong": out_wrong,
            "OutCorrect": out_correct
W
whs 已提交
5185 5186 5187
        },
        attrs={"num_classes": num_classes})
    return out_mean_iou, out_wrong, out_correct
5188 5189 5190 5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201 5202 5203 5204 5205 5206 5207 5208 5209 5210 5211 5212 5213 5214 5215 5216 5217 5218 5219 5220 5221 5222 5223 5224 5225 5226 5227 5228 5229 5230 5231 5232 5233 5234 5235 5236 5237 5238 5239 5240 5241 5242 5243 5244 5245 5246 5247 5248 5249 5250 5251 5252 5253 5254 5255 5256 5257 5258 5259 5260 5261 5262 5263 5264 5265 5266 5267 5268 5269 5270 5271 5272 5273 5274 5275 5276 5277 5278 5279 5280 5281 5282 5283 5284 5285


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 \
        isinstance(shape, Variable)):
        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
5286 5287 5288 5289 5290 5291 5292 5293 5294 5295


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
5296

5297 5298
    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
5299

5300 5301 5302 5303
    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
5304

5305 5306 5307 5308 5309
    $$
      C_{i,j} = -\tilde{P_{ij}} * o_{i,j} + \log(1 + e^{o_{i,j}}) \\
      o_{i,j} =  o_i - o_j  \\
      \tilde{P_{i,j}} = \left \{0, 0.5, 1 \right \} \ or \ \left \{0, 1 \right \}
    $$
M
minqiyang 已提交
5310 5311 5312

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

5313 5314 5315 5316 5317 5318 5319 5320 5321 5322 5323 5324 5325 5326 5327 5328 5329 5330 5331 5332 5333 5334 5335 5336 5337 5338 5339 5340 5341 5342 5343 5344 5345 5346 5347 5348 5349 5350 5351 5352 5353 5354 5355 5356
    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