nn.py 192.6 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
Y
yuyang18 已提交
36
from layer_function_generator import autodoc, templatedoc
Y
yangyaming 已提交
37
from tensor import concat
C
chengduoZH 已提交
38
import utils
Y
yuyang18 已提交
39
import random
F
fengjiayi 已提交
40
from .. import unique_name
Y
Yu Yang 已提交
41 42

__all__ = [
Y
ying 已提交
43 44 45
    'fc',
    'embedding',
    'dynamic_lstm',
Y
Yibing Liu 已提交
46
    'dynamic_lstmp',
G
guosheng 已提交
47
    'dynamic_gru',
Y
ying 已提交
48 49 50 51 52 53 54 55 56
    'gru_unit',
    'linear_chain_crf',
    'crf_decoding',
    'cos_sim',
    'cross_entropy',
    'square_error_cost',
    'chunk_eval',
    'sequence_conv',
    'conv2d',
Y
yuyang18 已提交
57
    'conv3d',
Y
ying 已提交
58
    'sequence_pool',
59 60
    'sequence_softmax',
    'softmax',
Y
ying 已提交
61
    'pool2d',
Y
yuyang18 已提交
62
    'pool3d',
Y
ying 已提交
63 64 65
    'batch_norm',
    'beam_search_decode',
    'conv2d_transpose',
Y
yuyang18 已提交
66
    'conv3d_transpose',
Y
ying 已提交
67 68 69 70 71 72
    'sequence_expand',
    'lstm_unit',
    'reduce_sum',
    'reduce_mean',
    'reduce_max',
    'reduce_min',
73
    'reduce_prod',
Y
ying 已提交
74 75 76 77
    'sequence_first_step',
    'sequence_last_step',
    'dropout',
    'split',
78 79
    'ctc_greedy_decoder',
    'edit_distance',
Y
ying 已提交
80 81
    'l2_normalize',
    'matmul',
Q
qingqing01 已提交
82
    'topk',
Y
ying 已提交
83 84
    'warpctc',
    'sequence_reshape',
85
    'transpose',
86
    'im2sequence',
87
    'nce',
W
weixing02 已提交
88
    'hsigmoid',
Q
Qiao Longfei 已提交
89
    'beam_search',
90
    'row_conv',
91
    'multiplex',
G
guosheng 已提交
92
    'layer_norm',
93 94
    'softmax_with_cross_entropy',
    'smooth_l1',
95
    'one_hot',
Y
Yu Yang 已提交
96
    'autoincreased_step_counter',
C
caoying03 已提交
97
    'reshape',
Y
yangyaming 已提交
98
    'lod_reset',
D
dragonwarrior 已提交
99
    'lrn',
G
guosheng 已提交
100
    'pad',
101
    'label_smooth',
102
    'roi_pool',
W
whs 已提交
103
    'dice_loss',
F
fengjiayi 已提交
104 105
    'image_resize',
    'image_resize_short',
B
baiyf 已提交
106
    'resize_bilinear',
W
whs 已提交
107
    'gather',
108
    'random_crop',
Y
yuyang18 已提交
109 110 111
    'mean_iou',
    'relu',
    'log',
112
    'crop',
Y
Yu Yang 已提交
113 114 115 116 117 118 119 120
]


def fc(input,
       size,
       num_flatten_dims=1,
       param_attr=None,
       bias_attr=None,
121
       use_mkldnn=False,
Y
Yu Yang 已提交
122
       act=None,
J
Jacek Czaja 已提交
123
       is_test=False,
124
       name=None):
Y
Yu Yang 已提交
125
    """
126
    **Fully Connected Layer**
Y
Yu Yang 已提交
127

128 129 130 131 132 133 134 135
    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
136
    to the output as well.
C
caoying03 已提交
137

C
caoying03 已提交
138
    This process can be formulated as follows:
139 140 141

    .. math::

142
        Out = Act({\sum_{i=0}^{N-1}X_iW_i + b})
143 144 145

    In the above equation:

C
caoying03 已提交
146 147 148 149
    * :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).
150
    * :math:`Act`: The activation function.
C
caoying03 已提交
151
    * :math:`Out`: The output tensor.
Y
Yu Yang 已提交
152 153

    Args:
R
ranqiu 已提交
154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170
        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
            of this layer. If it is set to None, no bias will be added to the output units.
        act (str, default None): Activation to be applied to the output of this layer.
J
Jacek Czaja 已提交
171
        is_test(bool): A flag indicating whether execution is in test phase.
M
mozga-intel 已提交
172 173
        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
174
        name (str, default None): The name of this layer.
Y
Yu Yang 已提交
175

176
    Returns:
F
fengjiayi 已提交
177
        Variable: The transformation result.
178 179

    Raises:
C
caoying03 已提交
180
        ValueError: If rank of the input tensor is less than 2.
181 182 183 184

    Examples:
        .. code-block:: python

F
fengjiayi 已提交
185
          data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
186
          fc = fluid.layers.fc(input=data, size=1000, act="tanh")
Y
Yu Yang 已提交
187
    """
C
caoying03 已提交
188

C
caoying03 已提交
189
    helper = LayerHelper("fc", **locals())
Y
Yu Yang 已提交
190 191 192 193

    dtype = helper.input_dtype()

    mul_results = []
194 195
    for input_var, param_attr in helper.iter_inputs_and_params():
        input_shape = input_var.shape
Y
Yu Yang 已提交
196 197 198
        param_shape = [
            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
        ] + [size]
Y
ying 已提交
199

Y
Yu Yang 已提交
200
        w = helper.create_parameter(
201 202
            attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False)
        tmp = helper.create_tmp_variable(dtype)
203
        helper.append_op(
204 205 206
            type="mul",
            inputs={"X": input_var,
                    "Y": w},
207
            outputs={"Out": tmp},
M
mozga-intel 已提交
208 209
            attrs={"x_num_col_dims": num_flatten_dims,
                   "y_num_col_dims": 1})
210 211 212 213
        mul_results.append(tmp)

    if len(mul_results) == 1:
        pre_bias = mul_results[0]
214
    else:
215 216
        pre_bias = helper.create_tmp_variable(dtype)
        helper.append_op(
217 218 219 220
            type="sum",
            inputs={"X": mul_results},
            outputs={"Out": pre_bias},
            attrs={"use_mkldnn": use_mkldnn})
221 222 223 224
    # 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
225 226


227 228 229
def embedding(input,
              size,
              is_sparse=False,
230
              is_distributed=False,
231 232 233
              padding_idx=None,
              param_attr=None,
              dtype='float32'):
Y
Yu Yang 已提交
234
    """
235 236
    **Embedding Layer**

237
    This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in
238 239
    a lookup table. The result of this lookup is the embedding of each ID in the
    :attr:`input`.
240 241 242

    All the input variables are passed in as local variables to the LayerHelper
    constructor.
Y
Yu Yang 已提交
243 244

    Args:
245 246 247 248 249
        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.
250
        is_distributed(bool): Whether to run lookup table from remote parameter server.
251 252
        padding_idx(int|long|None): If :attr:`None`, it makes no effect to lookup.
            Otherwise the given :attr:`padding_idx` indicates padding the output
253
            with zeros whenever lookup encounters it in :attr:`input`. If
254
            :math:`padding_idx < 0`, the :attr:`padding_idx` to use in lookup is
255 256
            :math:`size[0] + dim`.
        param_attr(ParamAttr): Parameters for this layer
257
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc
Y
Yu Yang 已提交
258

259 260 261
    Returns:
        Variable: The tensor variable storing the embeddings of the \
                  supplied inputs.
Y
Yu Yang 已提交
262

263 264
    Examples:
        .. code-block:: python
Y
Yu Yang 已提交
265

C
chengduoZH 已提交
266
          dict_size = len(dataset.ids)
267
          data = fluid.layers.data(name='ids', shape=[32, 32], dtype='float32')
C
chengduoZH 已提交
268
          fc = fluid.layers.embedding(input=data, size=[dict_size, 16])
Y
Yu Yang 已提交
269 270 271 272 273 274
    """

    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)
275 276
    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
277 278 279 280 281
    helper.append_op(
        type='lookup_table',
        inputs={'Ids': input,
                'W': w},
        outputs={'Out': tmp},
282 283 284 285 286
        attrs={
            'is_sparse': is_sparse,
            'is_distributed': is_distributed,
            'padding_idx': padding_idx
        })
Y
Yu Yang 已提交
287 288 289
    return tmp


Y
yi.wu 已提交
290
@templatedoc(op_type="lstm")
Y
Yu Yang 已提交
291 292
def dynamic_lstm(input,
                 size,
Y
Yancey 已提交
293 294
                 h_0=None,
                 c_0=None,
Y
Yu Yang 已提交
295 296 297 298 299 300 301
                 param_attr=None,
                 bias_attr=None,
                 use_peepholes=True,
                 is_reverse=False,
                 gate_activation='sigmoid',
                 cell_activation='tanh',
                 candidate_activation='tanh',
302 303
                 dtype='float32',
                 name=None):
Y
Yibing Liu 已提交
304
    """
Y
yi.wu 已提交
305
    ${comment}
Y
Yibing Liu 已提交
306 307

    Args:
Y
yi.wu 已提交
308 309
        input (Variable): ${input_comment}
        size (int): 4 * hidden size.
Y
Yancey 已提交
310 311 312 313 314 315 316
        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.

317
        param_attr(ParamAttr|None): The parameter attribute for the learnable
318
                               hidden-hidden weights.
Y
Yibing Liu 已提交
319 320 321

                               - Weights = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}
322 323
                               - The shape is (D x 4D), where D is the hidden
                                 size.
Y
yi.wu 已提交
324
        bias_attr (ParamAttr|None): The bias attribute for the learnable bias
325 326 327
                              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
328

329
                              1. `use_peepholes = False`
Y
yi.wu 已提交
330 331
                                 - Biases = {:math:`b_c, b_i, b_f, b_o`}.
                                 - The shape is (1 x 4D).
332
                              2. `use_peepholes = True`
Y
yi.wu 已提交
333
                                 - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
Y
Yibing Liu 已提交
334
                                                 W_{fc}, W_{oc}`}.
Y
yi.wu 已提交
335
                                 - The shape is (1 x 7D).
Y
yi.wu 已提交
336 337 338 339 340 341 342 343
        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
344 345

    Returns:
Y
Yibing Liu 已提交
346 347
        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
348

Y
Yibing Liu 已提交
349
    Examples:
Y
Yibing Liu 已提交
350 351
        .. code-block:: python

Y
Yibing Liu 已提交
352 353
            hidden_dim = 512
            forward_proj = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
354
                                           act=None, bias_attr=None)
Y
Yibing Liu 已提交
355 356
            forward, _ = fluid.layers.dynamic_lstm(
                input=forward_proj, size=hidden_dim * 4, use_peepholes=False)
Y
Yibing Liu 已提交
357
    """
358

Y
Yu Yang 已提交
359 360 361 362 363 364 365 366 367 368 369 370 371 372
    helper = LayerHelper('lstm', **locals())
    size = size / 4
    weight = helper.create_parameter(
        attr=helper.param_attr, shape=[size, 4 * size], dtype=dtype)
    bias_size = [1, 7 * size]
    if not use_peepholes:
        bias_size[1] = 4 * size
    bias = helper.create_parameter(
        attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)

    hidden = helper.create_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
373 374 375 376 377 378 379 380 381 382
    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
383 384 385

    helper.append_op(
        type='lstm',
Y
Yancey 已提交
386
        inputs=inputs,
Y
Yu Yang 已提交
387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402
        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
403 404 405 406 407 408 409 410 411 412 413
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',
414 415
                  dtype='float32',
                  name=None):
Y
Yibing Liu 已提交
416 417 418
    """
    **Dynamic LSTMP Layer**

419 420 421 422 423 424
    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
425 426 427 428 429

    The formula is as follows:

    .. math::

430
        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
431

432
        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
433

434
        \\tilde{c_t} & = act_g(W_{cx}x_t + W_{cr}r_{t-1} + b_c)
Y
Yibing Liu 已提交
435

436
        o_t & = \sigma(W_{ox}x_{t} + W_{or}r_{t-1} + W_{oc}c_t + b_o)
Y
Yibing Liu 已提交
437

438
        c_t & = f_t \odot c_{t-1} + i_t \odot \\tilde{c_t}
Y
Yibing Liu 已提交
439

440
        h_t & = o_t \odot act_h(c_t)
Y
Yibing Liu 已提交
441

442
        r_t & = \overline{act_h}(W_{rh}h_t)
Y
Yibing Liu 已提交
443

Y
Yibing Liu 已提交
444 445 446 447 448 449
    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, \
450
          we use vectors to reprenset these diagonal weight matrices.
Y
Yibing Liu 已提交
451
    * :math:`b`: Denotes bias vectors (e.g. :math:`b_i` is the input gate \
452
          bias vector).
Y
Yibing Liu 已提交
453 454 455
    * :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 \
456
          the same size as the cell output activation vector :math:`h`.
Y
Yibing Liu 已提交
457
    * :math:`h`: The hidden state.
458
    * :math:`r`: The recurrent projection of the hidden state.
Y
Yibing Liu 已提交
459 460
    * :math:`\\tilde{c_t}`: The candidate hidden state, whose \
          computation is based on the current input and previous hidden state.
461
    * :math:`\odot`: The element-wise product of the vectors.
Y
Yibing Liu 已提交
462
    * :math:`act_g` and :math:`act_h`: The cell input and cell output \
463
          activation functions and `tanh` is usually used for them.
Y
Yibing Liu 已提交
464 465
    * :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
466 467 468 469

    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.
470

Y
Yibing Liu 已提交
471 472 473 474 475 476 477 478 479 480 481 482
    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.
483
        param_attr(ParamAttr|None): The parameter attribute for the learnable
Y
Yibing Liu 已提交
484 485
                               hidden-hidden weight and projection weight.

486 487
                               - Hidden-hidden weight = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}.
488 489
                               - 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
490 491
                                 size.
                               - Projection weight = {:math:`W_{rh}`}.
492 493
                               - 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
494 495 496 497 498 499
                              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`}.
500
                                - The shape is (1 x 4D).
Y
Yibing Liu 已提交
501 502 503
                              2. `use_peepholes = True`
                                - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
                                                 W_{fc}, W_{oc}`}.
504
                                - The shape is (1 x 7D).
Y
Yibing Liu 已提交
505 506 507 508 509 510 511 512 513
        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.
514
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
Y
Yibing Liu 已提交
515 516
                              default "tanh".
        proj_activation(str): The activation for projection output.
517
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
Y
Yibing Liu 已提交
518 519
                              default "tanh".
        dtype(str): Data type. Choices = ["float32", "float64"], default "float32".
520 521
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
Y
Yibing Liu 已提交
522 523

    Returns:
524 525 526 527
        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
528 529

    Examples:
530

Y
Yibing Liu 已提交
531 532
        .. code-block:: python

533 534 535 536
            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
537
            hidden_dim, proj_dim = 512, 256
538
            fc_out = fluid.layers.fc(input=emb, size=hidden_dim * 4,
Y
Yibing Liu 已提交
539
                                     act=None, bias_attr=None)
540 541 542
            proj_out, _ = fluid.layers.dynamic_lstmp(input=fc_out,
                                                     size=hidden_dim * 4,
                                                     proj_size=proj_dim,
Y
Yibing Liu 已提交
543 544 545 546
                                                     use_peepholes=False,
                                                     is_reverse=True,
                                                     cell_activation="tanh",
                                                     proj_activation="tanh")
Y
Yibing Liu 已提交
547
    """
548

Y
Yibing Liu 已提交
549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594
    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
595 596 597 598 599 600 601 602 603
def dynamic_gru(input,
                size,
                param_attr=None,
                bias_attr=None,
                is_reverse=False,
                gate_activation='sigmoid',
                candidate_activation='tanh',
                h_0=None):
    """
604
    **Gated Recurrent Unit (GRU) Layer**
G
guosheng 已提交
605

606
    Refer to `Empirical Evaluation of Gated Recurrent Neural Networks on
607
    Sequence Modeling <https://arxiv.org/abs/1412.3555>`_ .
608

G
guosheng 已提交
609 610 611 612 613 614 615 616 617
    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)
618

G
guosheng 已提交
619
        h_t & = (1-u_t) \odot h_{t-1} + u_t \odot \\tilde{h_t}
620

G
guosheng 已提交
621
    The :math:`\odot` is the element-wise product of the vectors. :math:`act_g`
622 623
    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
624 625 626 627
    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
628
    to use fully-connect layer before GRU layer.
G
guosheng 已提交
629 630

    Args:
631 632
        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
633
            Variable is a matrix with shape :math:`(T \\times 3D)`, where
634
            :math:`T` is the total time steps in this mini-batch, :math:`D`
G
guosheng 已提交
635 636
            is the hidden size.
        size(int): The dimension of the gru cell.
637
        param_attr(ParamAttr|None): The parameter attribute for the learnable
G
guosheng 已提交
638 639
            hidden-hidden weight matrix. Note:

640
            - The shape of the weight matrix is :math:`(T \\times 3D)`, where
G
guosheng 已提交
641
              :math:`D` is the hidden size.
642
            - All elements in the weight matrix can be divided into two parts.
G
guosheng 已提交
643
              The first part are weights of the update gate and reset gate with
644
              shape :math:`(D \\times 2D)`, and the second part are weights for
G
guosheng 已提交
645
              candidate hidden state with shape :math:`(D \\times D)`.
646
        bias_attr(ParamAttr): The parameter attribute for learnable the
G
guosheng 已提交
647
            hidden-hidden bias.
648
        is_reverse(bool): Whether to compute reversed GRU, default
G
guosheng 已提交
649 650 651
            :attr:`False`.
        gate_activation(str): The activation for update gate and reset gate.
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "sigmoid".
652
        candidate_activation(str): The activation for candidate hidden state.
G
guosheng 已提交
653
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh".
654 655 656 657
        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
658 659

    Returns:
G
guosheng 已提交
660
        Variable: The hidden state of GRU. The shape is :math:`(T \\times D)`, \
661
            and sequence length is the same with the input.
662

G
guosheng 已提交
663
    Examples:
664

G
guosheng 已提交
665 666
        .. code-block:: python

667 668 669 670
            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
671
            hidden_dim = 512
672
            x = fluid.layers.fc(input=emb, size=hidden_dim * 3)
G
guosheng 已提交
673 674 675 676 677 678 679 680 681 682
            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
683
    batch_size = input.shape[0]
G
guosheng 已提交
684 685 686
    inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
    if h_0 != None:
        assert h_0.shape == (
Y
Yancey 已提交
687 688 689
            batch_size, size
        ), 'The shape of h0 should be(batch_size, %d)' % size
        inputs['H0'] = h_0
G
guosheng 已提交
690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712

    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
713 714 715
def gru_unit(input,
             hidden,
             size,
716 717
             param_attr=None,
             bias_attr=None,
Y
Yu Yang 已提交
718
             activation='tanh',
719
             gate_activation='sigmoid'):
Y
Yu Yang 已提交
720
    """
721
    GRU unit layer. The equation of a gru step is:
Y
Yu Yang 已提交
722

723 724
        .. math::
            u_t & = actGate(xu_{t} + W_u h_{t-1} + b_u)
Y
Yu Yang 已提交
725

726
            r_t & = actGate(xr_{t} + W_r h_{t-1} + b_r)
Y
Yu Yang 已提交
727

728
            m_t & = actNode(xm_t + W_c dot(r_t, h_{t-1}) + b_m)
729

730
            h_t & = dot((1-u_t), m_t) + dot(u_t, h_{t-1})
731 732

    The inputs of gru unit includes :math:`z_t`, :math:`h_{t-1}`. In terms
733 734 735
    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
736 737
    connected layer has to be applied, such that :math:`z_t = W_{fc}x_t`.

738 739
    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
740 741 742
    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`.
743 744 745 746 747

    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.
748 749
        param_attr (ParamAttr): The weight parameters for gru unit. Default: None
        bias_attr (ParamAttr): The bias parameters for gru unit. Default: None
750 751 752 753
        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
754

755 756 757 758 759 760
    Returns:
        tuple: The hidden value, reset-hidden value and gate values.

    Examples:

        .. code-block:: python
Y
Yu Yang 已提交
761

762
             # assuming we have x_t_data and prev_hidden of size=10
763
             x_t = fluid.layers.fc(input=x_t_data, size=30)
764 765
             hidden_val, r_h_val, gate_val = fluid.layers.gru_unit(input=x_t,
                                                    hidden = prev_hidden)
Y
Yu Yang 已提交
766 767 768 769 770 771 772 773 774 775 776 777 778 779 780

    """
    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
781 782
    weight = helper.create_parameter(
        attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype)
Y
Yu Yang 已提交
783

784 785 786 787
    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
788
    # create bias
789
    if helper.bias_attr:
Y
Yu Yang 已提交
790 791 792
        bias_size = [1, 3 * size]
        bias = helper.create_parameter(
            attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
793
        inputs['Bias'] = bias
Y
Yu Yang 已提交
794 795 796

    helper.append_op(
        type='gru_unit',
797
        inputs=inputs,
Y
Yu Yang 已提交
798 799 800 801 802 803
        outputs={
            'Gate': gate,
            'ResetHiddenPrev': reset_hidden_pre,
            'Hidden': updated_hidden,
        },
        attrs={
804 805
            'activation': 2,  # tanh
            'gate_activation': 1,  # sigmoid
Y
Yu Yang 已提交
806 807 808 809 810
        })

    return updated_hidden, reset_hidden_pre, gate


Y
yuyang18 已提交
811
@templatedoc()
812
def linear_chain_crf(input, label, param_attr=None):
Y
yuyang18 已提交
813 814 815 816 817 818 819
    """
    Linear Chain CRF.

    ${comment}

    Args:
        input(${emission_type}): ${emission_comment}
D
dzhwinter 已提交
820
        input(${transition_type}): ${transition_comment}
Y
yuyang18 已提交
821 822 823 824
        label(${label_type}): ${label_comment}
        param_attr(ParamAttr): The attribute of the learnable parameter.

    Returns:
D
dzhwinter 已提交
825 826 827
        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
828 829

    """
Y
Yu Yang 已提交
830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854
    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
855
@templatedoc()
856
def crf_decoding(input, param_attr, label=None):
Y
yuyang18 已提交
857 858 859 860 861
    """
    ${comment}

    Args:
        input(${emission_type}): ${emission_comment}
Y
yi.wu 已提交
862

Y
yuyang18 已提交
863
        param_attr(ParamAttr): The parameter attribute for training.
Y
yi.wu 已提交
864

Y
yuyang18 已提交
865 866 867
        label(${label_type}): ${label_comment}

    Returns:
Y
update  
yi.wu 已提交
868
        Variable: ${viterbi_path_comment}
869

Y
yi.wu 已提交
870 871 872 873 874
    Examples:
        .. code-block:: python

           crf_decode = layers.crf_decoding(
                input=hidden, param_attr=ParamAttr(name="crfw"))
Y
yuyang18 已提交
875
    """
Y
Yu Yang 已提交
876 877 878 879 880 881 882 883 884 885 886 887 888
    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
889
@templatedoc()
F
fengjiayi 已提交
890
def cos_sim(X, Y):
Y
Yu Yang 已提交
891
    """
Y
yi.wu 已提交
892 893 894
    ${comment}

    Args:
895 896
        X (Variable): ${x_comment}.
        Y (Variable): ${y_comment}.
F
fengjiayi 已提交
897

Y
yi.wu 已提交
898
    Returns:
899
        Variable: the output of cosine(X, Y).
Y
Yu Yang 已提交
900
    """
F
fengjiayi 已提交
901
    helper = LayerHelper('cos_sim', **locals())
Y
Yu Yang 已提交
902 903 904 905 906 907 908 909 910 911 912 913 914
    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


915
def dropout(x, dropout_prob, is_test=False, seed=None, name=None):
916 917 918 919 920
    """
    Computes dropout.

    Drop or keep each element of `x` independently. Dropout is a regularization
    technique for reducing overfitting by preventing neuron co-adaption during
921
    training. The dropout operator randomly sets (according to the given dropout
922 923 924 925
    probability) the outputs of some units to zero, while others are remain
    unchanged.

    Args:
926 927
        x (Variable): The input tensor variable.
        dropout_prob (float): Probability of setting units to zero.
928 929 930 931 932 933 934
        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.
935 936

    Returns:
937
        Variable: A tensor variable is the shape with `x`.
938 939

    Examples:
940

941 942
        .. code-block:: python

943 944
            x = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
            droped = fluid.layers.dropout(x, dropout_prob=0.5)
945 946
    """

F
fengjiayi 已提交
947
    helper = LayerHelper('dropout', **locals())
948 949 950 951 952 953 954
    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]},
955 956 957 958 959 960
        attrs={
            'dropout_prob': dropout_prob,
            'is_test': is_test,
            'fix_seed': seed is not None,
            'seed': seed if seed is not None else 0
        })
961 962 963
    return out


F
fengjiayi 已提交
964
def cross_entropy(input, label, soft_label=False):
Y
Yu Yang 已提交
965
    """
Y
Yibing Liu 已提交
966 967
    **Cross Entropy Layer**

968 969 970
    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
971 972

    1) One-hot cross-entropy:
F
fengjiayi 已提交
973
        `soft_label = False`, `Label[i, 0]` indicates the class index for sample i:
Y
yangyaming 已提交
974

Y
Yibing Liu 已提交
975
        .. math::
Y
yangyaming 已提交
976

Y
Yibing Liu 已提交
977 978 979
            Y[i] = -\log(X[i, Label[i]])

    2) Soft-label cross-entropy:
F
fengjiayi 已提交
980 981
        `soft_label = True`, `Label[i, j]` indicates the soft label of class j
        for sample i:
Y
Yibing Liu 已提交
982 983 984 985 986

        .. math::

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

Y
Yibing Liu 已提交
987
       Please make sure that in this case the summation of each row of `label`
Y
Yibing Liu 已提交
988 989 990
       equals one.

    3) One-hot cross-entropy with vecterized `label`:
F
fengjiayi 已提交
991 992
         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
993
         to a one-hot cross-entropy with one-hot label representation.
Y
yangyaming 已提交
994

Y
Yibing Liu 已提交
995
    Args:
Y
yangyaming 已提交
996
        input (Variable|list):  a 2-D tensor with shape [N x D], where N is the
997 998 999 1000
                                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
1001
        label (Variable|list): the ground truth which is a 2-D tensor. When
1002 1003 1004 1005
                               `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
1006
        soft_label (bool): a flag indicating whether to
1007 1008
                                           interpretate the given labels as soft
                                           labels, default `False`.
Y
Yibing Liu 已提交
1009 1010 1011 1012 1013

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

    Raises:
1014 1015 1016 1017 1018
        `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
1019 1020 1021 1022 1023 1024

    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
1025
    """
F
fengjiayi 已提交
1026
    helper = LayerHelper('cross_entropy', **locals())
Y
Yu Yang 已提交
1027 1028 1029 1030 1031 1032
    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
1033
        attrs={"soft_label": soft_label})
Y
Yu Yang 已提交
1034 1035 1036
    return out


F
fengjiayi 已提交
1037
def square_error_cost(input, label):
Y
Yu Yang 已提交
1038
    """
1039 1040
    **Square error cost layer**

1041 1042
    This layer accepts input predictions and target label and returns the
    squared error cost.
Y
ying 已提交
1043

1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056
    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:
1057 1058
        input (Variable): Input tensor, has predictions.
        label (Variable): Label tensor, has target labels.
1059 1060

    Returns:
G
guosheng 已提交
1061
        Variable: The tensor variable storing the element-wise squared error \
1062
                  difference of input and label.
1063 1064 1065 1066 1067 1068 1069 1070

    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
1071
    """
F
fengjiayi 已提交
1072
    helper = LayerHelper('square_error_cost', **locals())
Y
Yu Yang 已提交
1073 1074 1075 1076 1077 1078 1079 1080 1081
    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
1082 1083
        type='square', inputs={'X': [minus_out]},
        outputs={'Out': [square_out]})
Y
Yu Yang 已提交
1084 1085 1086
    return square_out


Y
yi.wu 已提交
1087
@templatedoc()
Y
Yu Yang 已提交
1088 1089 1090 1091
def chunk_eval(input,
               label,
               chunk_scheme,
               num_chunk_types,
F
fengjiayi 已提交
1092
               excluded_chunk_types=None):
Y
Yu Yang 已提交
1093
    """
Y
yi.wu 已提交
1094
    **Chunk Evaluator**
Y
yi.wu 已提交
1095

Y
yangyaming 已提交
1096
    This function computes and outputs the precision, recall and
1097
    F1-score of chunk detection.
Y
yi.wu 已提交
1098

Y
yi.wu 已提交
1099 1100 1101 1102 1103 1104 1105 1106
    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
1107

Y
yi.wu 已提交
1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132
       ====== ====== ======  =====  ==  ============   =====  ===== =====  ==  =========
              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
1133

Y
yi.wu 已提交
1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157
       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
1158
    Args:
1159 1160 1161 1162 1163
        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
1164

Y
yi.wu 已提交
1165
    Returns:
Y
update  
yi.wu 已提交
1166 1167 1168
        tuple: tuple containing: precision, recall, f1_score,
        num_infer_chunks, num_label_chunks,
        num_correct_chunks
1169

Y
yi.wu 已提交
1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181
    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
1182
    """
F
fengjiayi 已提交
1183
    helper = LayerHelper("chunk_eval", **locals())
Y
Yu Yang 已提交
1184 1185 1186 1187 1188

    # prepare output
    precision = helper.create_tmp_variable(dtype="float32")
    recall = helper.create_tmp_variable(dtype="float32")
    f1_score = helper.create_tmp_variable(dtype="float32")
1189 1190 1191
    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
1192 1193 1194 1195 1196 1197 1198 1199

    helper.append_op(
        type="chunk_eval",
        inputs={"Inference": [input],
                "Label": [label]},
        outputs={
            "Precision": [precision],
            "Recall": [recall],
1200 1201 1202 1203
            "F1-Score": [f1_score],
            "NumInferChunks": [num_infer_chunks],
            "NumLabelChunks": [num_label_chunks],
            "NumCorrectChunks": [num_correct_chunks]
Y
Yu Yang 已提交
1204 1205 1206
        },
        attrs={
            "num_chunk_types": num_chunk_types,
G
guosheng 已提交
1207 1208
            "chunk_scheme": chunk_scheme,
            "excluded_chunk_types": excluded_chunk_types or []
Y
Yu Yang 已提交
1209
        })
1210 1211
    return (precision, recall, f1_score, num_infer_chunks, num_label_chunks,
            num_correct_chunks)
Y
Yu Yang 已提交
1212 1213


1214
@templatedoc()
Y
Yu Yang 已提交
1215 1216 1217 1218 1219 1220 1221
def sequence_conv(input,
                  num_filters,
                  filter_size=3,
                  filter_stride=1,
                  padding=None,
                  bias_attr=None,
                  param_attr=None,
1222
                  act=None):
Y
Yu Yang 已提交
1223 1224 1225 1226
    """
    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.
1227 1228 1229 1230 1231 1232 1233 1234 1235 1236

    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
1237

1238 1239
    Returns:
        Variable: output of sequence_conv
Y
Yu Yang 已提交
1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264
    """

    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)


1265
def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=True):
1266 1267 1268
    """
    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
1269
    input Tensor can be either :math:`[N, 1]` or :math:`[N]`, where :math:`N`
1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288
    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
1289

1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300
    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)
    """
1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311
    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


1312
def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True, name=None):
Q
qiaolongfei 已提交
1313 1314 1315 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
    """
    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)

    """
1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362
    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
1363 1364 1365
def conv2d(input,
           num_filters,
           filter_size,
C
chengduoZH 已提交
1366 1367
           stride=1,
           padding=0,
1368
           dilation=1,
Y
Yu Yang 已提交
1369 1370 1371
           groups=None,
           param_attr=None,
           bias_attr=None,
C
chengduoZH 已提交
1372
           use_cudnn=True,
1373
           use_mkldnn=False,
1374 1375
           act=None,
           name=None):
Y
Yu Yang 已提交
1376
    """
C
chengduoZH 已提交
1377
    The convolution2D layer calculates the output based on the input, filter
T
tensor-tang 已提交
1378 1379
    and strides, paddings, dilations, groups parameters. Input and
    Output are in NCHW format, where N is batch size, C is the number of
1380
    channels, H is the height of the feature, and W is the width of the feature.
T
tensor-tang 已提交
1381 1382 1383 1384 1385 1386 1387
    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.
1388 1389 1390
    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
1391

1392
    For each input :math:`X`, the equation is:
C
refine  
chengduoZH 已提交
1393

C
chengduoZH 已提交
1394 1395
    .. math::

C
refine  
chengduoZH 已提交
1396
        Out = \sigma (W \\ast X + b)
C
chengduoZH 已提交
1397

T
tensor-tang 已提交
1398
    Where:
C
chengduoZH 已提交
1399

1400 1401 1402 1403 1404
    * :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
1405
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
1406 1407 1408

    Example:

1409 1410
        - Input:

W
weixing02 已提交
1411
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
C
refine  
chengduoZH 已提交
1412

W
weixing02 已提交
1413
          Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)`
C
refine  
chengduoZH 已提交
1414

1415
        - Output:
T
tensor-tang 已提交
1416

W
weixing02 已提交
1417
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
C
refine  
chengduoZH 已提交
1418

C
chengduoZH 已提交
1419
        Where
1420 1421

        .. math::
C
chengduoZH 已提交
1422

W
weixing02 已提交
1423 1424
            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
1425 1426

    Args:
1427
        input (Variable): The input image with [N, C, H, W] format.
T
tensor-tang 已提交
1428
        num_filters(int): The number of filter. It is as same as the output
1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450
            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
1451 1452
        use_mkldnn (bool): Use mkldnn kernels or not, it is valid only when compiled
            with mkldnn library. Default: False
1453 1454 1455
        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
1456 1457

    Returns:
G
guosheng 已提交
1458
        Variable: The tensor variable storing the convolution and \
C
chengduoZH 已提交
1459 1460
                  non-linearity activation result.

C
refine  
chengduoZH 已提交
1461
    Raises:
1462 1463
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
C
refine  
chengduoZH 已提交
1464

C
chengduoZH 已提交
1465 1466 1467
    Examples:
        .. code-block:: python

1468 1469
          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
1470 1471 1472
    """

    num_channels = input.shape[1]
1473 1474

    l_type = 'conv2d'
X
xzl 已提交
1475 1476
    if (num_channels == groups and num_filters % num_channels == 0 and
            not use_cudnn):
1477
        l_type = 'depthwise_conv2d'
1478 1479 1480 1481

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

Y
Yu Yang 已提交
1482 1483 1484 1485 1486 1487 1488
    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
1489 1490 1491
    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')
1492
    dilation = utils.convert_to_list(dilation, 2, 'dilation')
C
chengduoZH 已提交
1493

C
chengduoZH 已提交
1494 1495
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512

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

    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(
1513
        type=l_type,
Y
Yu Yang 已提交
1514 1515 1516 1517 1518
        inputs={
            'Input': input,
            'Filter': filter_param,
        },
        outputs={"Output": pre_bias},
C
chengduoZH 已提交
1519 1520 1521
        attrs={
            'strides': stride,
            'paddings': padding,
1522
            'dilations': dilation,
C
chengduoZH 已提交
1523
            'groups': groups,
1524 1525
            'use_cudnn': use_cudnn,
            'use_mkldnn': use_mkldnn
C
chengduoZH 已提交
1526
        })
Y
Yu Yang 已提交
1527 1528 1529 1530 1531 1532

    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
1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550
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
1551 1552 1553 1554 1555 1556
    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
1557 1558 1559 1560 1561 1562 1563 1564 1565

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

    .. math::

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

    In the above equation:

1566 1567
    * :math:`X`: Input value, a tensor with NCDHW format.
    * :math:`W`: Filter value, a tensor with MCDHW format.
C
chengduoZH 已提交
1568 1569 1570
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
1571
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
C
chengduoZH 已提交
1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596

    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,
1597
            it must contain three integers, (filter_size_D, filter_size_H, filter_size_W).
C
chengduoZH 已提交
1598 1599
            Otherwise, the filter will be a square.
        stride (int|tuple): The stride size. If stride is a tuple, it must
1600
            contain three integers, (stride_D, stride_H, stride_W). Otherwise, the
C
chengduoZH 已提交
1601 1602
            stride_D = stride_H = stride_W = stride. Default: stride = 1.
        padding (int|tuple): The padding size. If padding is a tuple, it must
1603
            contain three integers, (padding_D, padding_H, padding_W). Otherwise, the
C
chengduoZH 已提交
1604 1605
            padding_D = padding_H = padding_W = padding. Default: padding = 0.
        dilation (int|tuple): The dilation size. If dilation is a tuple, it must
1606
            contain three integers, (dilation_D, dilation_H, dilation_W). Otherwise, the
C
chengduoZH 已提交
1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632
            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

1633 1634
          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
1635 1636 1637 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
    """

    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
        })

1690
    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
C
chengduoZH 已提交
1691 1692 1693 1694

    return helper.append_activation(pre_act)


F
fengjiayi 已提交
1695
def sequence_pool(input, pool_type):
Y
Yu Yang 已提交
1696
    """
Y
yangyaming 已提交
1697 1698 1699
    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
1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710

    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:
1711
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
1712 1713 1714 1715 1716
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
1717
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
1718 1719 1720 1721 1722 1723 1724

       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)
1725 1726
         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
1727

L
Luo Tao 已提交
1728 1729
    Args:
        input(variable): The input variable which is a LoDTensor.
Y
yangyaming 已提交
1730
        pool_type (string): The pooling type of sequence_pool.
L
Luo Tao 已提交
1731 1732 1733 1734 1735 1736 1737 1738
            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
1739

Y
yangyaming 已提交
1740
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1741 1742 1743 1744 1745
                              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')
1746 1747
             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
1748
    """
F
fengjiayi 已提交
1749
    helper = LayerHelper('sequence_pool', **locals())
Y
Yu Yang 已提交
1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760
    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
1761 1762 1763 1764 1765
    # 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
1766 1767 1768
    return pool_out


F
fengjiayi 已提交
1769
def sequence_first_step(input):
L
Luo Tao 已提交
1770
    """
L
Luo Tao 已提交
1771
    This function gets the first step of sequence.
L
Luo Tao 已提交
1772 1773 1774 1775

    .. code-block:: text

       x is a 1-level LoDTensor:
1776
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
1777 1778 1779 1780 1781
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
1782
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
1783
         out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1)
F
fengjiayi 已提交
1784

L
Luo Tao 已提交
1785 1786 1787 1788 1789 1790 1791 1792 1793
    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
1794

Y
yangyaming 已提交
1795
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1796 1797 1798
                              dtype='float32', lod_level=1)
             x_first_step = fluid.layers.sequence_first_step(input=x)
    """
1799 1800 1801
    return sequence_pool(input=input, pool_type="first")


F
fengjiayi 已提交
1802
def sequence_last_step(input):
L
Luo Tao 已提交
1803
    """
L
Luo Tao 已提交
1804
    This function gets the last step of sequence.
L
Luo Tao 已提交
1805 1806 1807 1808

    .. code-block:: text

       x is a 1-level LoDTensor:
1809
         x.lod = [[2, 3, 2]]
L
Luo Tao 已提交
1810 1811 1812 1813 1814
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
1815
         with condition len(x.lod[-1]) == out.dims[0]
L
Luo Tao 已提交
1816
         out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1)
F
fengjiayi 已提交
1817

L
Luo Tao 已提交
1818 1819 1820 1821 1822 1823 1824 1825 1826
    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
1827

Y
yangyaming 已提交
1828
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1829 1830 1831
                              dtype='float32', lod_level=1)
             x_last_step = fluid.layers.sequence_last_step(input=x)
    """
1832 1833 1834
    return sequence_pool(input=input, pool_type="last")


F
fengjiayi 已提交
1835
@templatedoc()
Y
Yu Yang 已提交
1836
def pool2d(input,
C
chengduoZH 已提交
1837 1838
           pool_size=-1,
           pool_type="max",
C
chengduoZH 已提交
1839 1840
           pool_stride=1,
           pool_padding=0,
C
caoying03 已提交
1841
           global_pooling=False,
C
chengduoZH 已提交
1842
           use_cudnn=True,
1843
           ceil_mode=False,
1844
           use_mkldnn=False,
C
caoying03 已提交
1845
           name=None):
Y
Yu Yang 已提交
1846
    """
F
fengjiayi 已提交
1847
    ${comment}
1848 1849

    Args:
1850 1851 1852
        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
1853
                          feature, and W is the width of the feature.
1854
        pool_size (int): The side length of pooling windows. All pooling
F
fengjiayi 已提交
1855
                         windows are squares with pool_size on a side.
F
fengjiayi 已提交
1856
        pool_type: ${pooling_type_comment}
1857 1858
        pool_stride (int): stride of the pooling layer.
        pool_padding (int): padding size.
F
fengjiayi 已提交
1859 1860 1861 1862
        global_pooling: ${global_pooling_comment}
        use_cudnn: ${use_cudnn_comment}
        ceil_mode: ${ceil_mode_comment}
        use_mkldnn: ${use_mkldnn_comment}
1863
        name (str|None): A name for this layer(optional). If set None, the
F
fengjiayi 已提交
1864 1865
                        layer will be named automatically.

1866
    Returns:
F
fengjiayi 已提交
1867
        Variable: The pooling result.
F
fengjiayi 已提交
1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880

    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(
1881 1882 1883 1884
                            input=data,
                            pool_size=2,
                            pool_type='max',
                            pool_stride=1,
F
fengjiayi 已提交
1885
                            global_pooling=False)
Y
Yu Yang 已提交
1886 1887 1888 1889 1890
    """
    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
1891

C
chengduoZH 已提交
1892 1893 1894 1895 1896
    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
1897 1898 1899 1900
    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
1901 1902
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1903

C
Add doc  
chengduoZH 已提交
1904
    l_type = 'pool2d'
1905 1906

    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
1907 1908 1909 1910
    dtype = helper.input_dtype()
    pool_out = helper.create_tmp_variable(dtype)

    helper.append_op(
1911 1912 1913 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
        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
1940
    pooling configurations mentioned in input parameters.
1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953

    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.
1954

1955
    Returns:
1956
        Variable: output of pool3d layer.
Y
Yu Yang 已提交
1957 1958 1959 1960 1961
    """
    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
1962

C
chengduoZH 已提交
1963 1964 1965 1966 1967
    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))

1968 1969 1970
    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
1971

C
chengduoZH 已提交
1972 1973
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1974

1975 1976
    l_type = "pool3d"
    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
1977 1978 1979 1980
    dtype = helper.input_dtype()
    pool_out = helper.create_tmp_variable(dtype)

    helper.append_op(
1981
        type=l_type,
Y
Yu Yang 已提交
1982 1983 1984 1985 1986 1987 1988
        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
1989
            "paddings": pool_padding,
1990
            "use_cudnn": use_cudnn,
1991 1992
            "ceil_mode": ceil_mode,
            "use_mkldnn": use_mkldnn
Y
Yu Yang 已提交
1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004
        })

    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
2005
               data_layout='NCHW',
Y
Yang Yang 已提交
2006
               in_place=False,
2007
               use_mkldnn=False,
2008 2009
               name=None,
               moving_mean_name=None,
W
wanghaoshuang 已提交
2010
               moving_variance_name=None,
2011 2012
               do_model_average_for_mean_and_var=False,
               fuse_with_relu=False):
Y
Yu Yang 已提交
2013
    """
Q
qiaolongfei 已提交
2014 2015 2016 2017
    **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
2018

Q
qiaolongfei 已提交
2019
    1. NHWC `[batch, in_height, in_width, in_channels]`
Q
qiaolongfei 已提交
2020

Q
qiaolongfei 已提交
2021 2022
    2. NCHW `[batch, in_channels, in_height, in_width]`

Q
qiaolongfei 已提交
2023 2024 2025
    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
2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037

    :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
2038 2039

    Args:
Q
qiaolongfei 已提交
2040
        input(variable): The input variable which is a LoDTensor.
Q
qiaolongfei 已提交
2041 2042 2043 2044
        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
2045 2046 2047
        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
2048
        in_place(bool, Default False): Make the input and output of batch norm reuse memory.
Q
qiaolongfei 已提交
2049 2050 2051 2052 2053
        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
2054
        do_model_average_for_mean_and_var(bool, Default False): Do model average for mean and variance or not.
2055
        fuse_with_relu (bool): if True, this OP performs relu after batch norm.
2056 2057

    Returns:
Q
qiaolongfei 已提交
2058
        Variable: A tensor variable which is the result after applying batch normalization on the input.
Q
qiaolongfei 已提交
2059 2060 2061 2062 2063 2064 2065

    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
2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088
    """
    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(
2089
        attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
Y
Yu Yang 已提交
2090

2091 2092
    mean = helper.create_parameter(
        attr=ParamAttr(
W
wanghaoshuang 已提交
2093 2094 2095
            name=moving_mean_name,
            initializer=Constant(0.0),
            trainable=False,
W
wanghaoshuang 已提交
2096
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
2097
        shape=param_shape,
2098 2099 2100 2101 2102 2103 2104
        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
2105
            trainable=False,
W
wanghaoshuang 已提交
2106
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
2107
        shape=param_shape,
2108 2109
        dtype=input.dtype)
    variance.stop_gradient = True
Y
Yu Yang 已提交
2110 2111 2112 2113 2114 2115

    # 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
2116 2117
    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
2118

Y
Yang Yang 已提交
2119
    batch_norm_out = input if in_place else helper.create_tmp_variable(dtype)
Y
Yu Yang 已提交
2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136

    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
        },
2137 2138 2139 2140
        attrs={
            "momentum": momentum,
            "epsilon": epsilon,
            "is_test": is_test,
2141 2142
            "use_mkldnn": use_mkldnn,
            "fuse_with_relu": fuse_with_relu
2143
        })
Y
Yu Yang 已提交
2144 2145 2146 2147

    return helper.append_activation(batch_norm_out)


Y
yuyang18 已提交
2148
@templatedoc()
G
guosheng 已提交
2149 2150 2151 2152 2153 2154 2155 2156 2157 2158
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
2159
    ${comment}
G
guosheng 已提交
2160 2161 2162

    The formula is as follows:

Y
yuyang18 已提交
2163
    ..  math::
G
guosheng 已提交
2164 2165 2166 2167 2168 2169 2170

        \\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
2171 2172 2173 2174 2175 2176 2177 2178
    * :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
2179

G
guosheng 已提交
2180 2181
    Args:
        input(Variable): The input tensor variable.
2182
        scale(bool): Whether to learn the adaptive gain :math:`g` after
G
guosheng 已提交
2183
            normalization.
2184
        shift(bool): Whether to learn the adaptive bias :math:`b` after
G
guosheng 已提交
2185
            normalization.
2186
        begin_norm_axis(bool): The normalization will be performed along
G
guosheng 已提交
2187
            dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`.
2188
        epsilon(float): The small value added to the variance to prevent
G
guosheng 已提交
2189 2190 2191 2192 2193 2194
            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.
2195
        name (str): The name of this layer. It is optional.
G
guosheng 已提交
2196 2197

    Returns:
Y
yuyang18 已提交
2198
        ${y_comment}
G
guosheng 已提交
2199 2200 2201

    Examples:

Y
yuyang18 已提交
2202 2203 2204
        >>> 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
2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219
    """
    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
2220
    if shift:
G
guosheng 已提交
2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244
        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
2245 2246 2247 2248
def conv2d_transpose(input,
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
2249 2250 2251
                     padding=0,
                     stride=1,
                     dilation=1,
2252
                     groups=None,
C
caoying03 已提交
2253
                     param_attr=None,
2254
                     bias_attr=None,
C
chengduoZH 已提交
2255
                     use_cudnn=True,
2256
                     act=None,
C
caoying03 已提交
2257
                     name=None):
Y
Yu Yang 已提交
2258
    """
2259 2260 2261 2262 2263 2264 2265 2266
    **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
2267 2268
    layer, please refer to the following explanation and references
    `therein <http://www.matthewzeiler.com/wp-content/uploads/2017/07/cvpr2010.pdf>`_.
2269 2270 2271
    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.
2272 2273 2274 2275 2276

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

    .. math::

2277
        Out = \sigma (W \\ast X + b)
2278

2279
    Where:
2280 2281 2282

    * :math:`X`: Input value, a tensor with NCHW format.
    * :math:`W`: Filter value, a tensor with MCHW format.
2283 2284 2285 2286
    * :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
2287

2288 2289 2290 2291
    Example:

        - Input:

2292
          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
2293

2294
          Filter shape: :math:`(C_{in}, C_{out}, H_f, W_f)`
2295 2296 2297

        - Output:

2298
          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
2299 2300

        Where
Y
Yu Yang 已提交
2301

2302 2303 2304 2305
        .. 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
2306 2307

    Args:
2308 2309 2310 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
        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
2341 2342

    Returns:
2343
        Variable: The tensor variable storing the convolution transpose result.
2344 2345

    Raises:
2346 2347
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
2348 2349 2350 2351

    Examples:
       .. code-block:: python

2352 2353
          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
2354
    """
2355 2356 2357 2358 2359 2360 2361 2362 2363

    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
2364 2365 2366
    if not isinstance(input, Variable):
        raise TypeError("Input of conv2d_transpose must be Variable")

C
chengduoZH 已提交
2367 2368 2369
    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
2370

C
chengduoZH 已提交
2371 2372
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
G
guosheng 已提交
2373

Y
Yu Yang 已提交
2374 2375 2376 2377 2378
    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
2379

Y
Yu Yang 已提交
2380 2381
        h_in = input.shape[2]
        w_in = input.shape[3]
G
guosheng 已提交
2382

C
chengduoZH 已提交
2383 2384 2385 2386
        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
2387
        filter_size = [filter_size_h, filter_size_w]
C
chengduoZH 已提交
2388 2389 2390
    else:
        filter_size = utils.convert_to_list(filter_size, 2,
                                            'conv2d_transpose.filter_size')
2391

2392 2393
    groups = 1 if groups is None else groups
    filter_shape = [input_channel, num_filters / groups] + filter_size
Y
Yu Yang 已提交
2394 2395 2396
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

2397
    pre_bias = helper.create_tmp_variable(dtype=input.dtype)
Y
Yu Yang 已提交
2398
    helper.append_op(
2399
        type=op_type,
Y
Yu Yang 已提交
2400 2401
        inputs={'Input': [input],
                'Filter': [img_filter]},
2402
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
2403
        attrs={
2404 2405 2406 2407 2408
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn
Y
Yu Yang 已提交
2409 2410
        })

2411 2412 2413
    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
2414 2415


2416
def conv3d_transpose(input,
Y
Yu Yang 已提交
2417 2418 2419
                     num_filters,
                     output_size=None,
                     filter_size=None,
C
chengduoZH 已提交
2420 2421 2422
                     padding=0,
                     stride=1,
                     dilation=1,
2423
                     groups=None,
C
caoying03 已提交
2424
                     param_attr=None,
2425
                     bias_attr=None,
C
chengduoZH 已提交
2426
                     use_cudnn=True,
2427
                     act=None,
C
caoying03 已提交
2428
                     name=None):
Y
Yu Yang 已提交
2429
    """
2430
    **Convlution3D transpose layer**
2431

2432
    The convolution3D transpose layer calculates the output based on the input,
2433
    filter, and dilations, strides, paddings. Input(Input) and output(Output)
2434 2435 2436 2437 2438 2439
    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>`_.
2440 2441 2442
    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.
2443 2444 2445 2446 2447

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

    .. math::

2448
        Out = \sigma (W \\ast X + b)
2449 2450 2451

    In the above equation:

2452 2453
    * :math:`X`: Input value, a tensor with NCDHW format.
    * :math:`W`: Filter value, a tensor with MCDHW format.
2454 2455 2456 2457
    * :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
2458

2459 2460 2461 2462
    Example:

        - Input:

2463
          Input shape: :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})`
2464

2465
          Filter shape: :math:`(C_{in}, C_{out}, D_f, H_f, W_f)`
2466 2467 2468

        - Output:

2469
          Output shape: :math:`(N, C_{out}, D_{out}, H_{out}, W_{out})`
2470 2471

        Where
Y
Yu Yang 已提交
2472

2473 2474
        .. math::

2475 2476 2477
           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
2478 2479

    Args:
2480
        input(Variable): The input image with [N, C, D, H, W] format.
2481 2482 2483
        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
2484
            tuple, it must contain three integers, (image_D, image_H, image_W). This
2485 2486
            parameter only works when filter_size is None.
        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
2487
            it must contain three integers, (filter_size_D, filter_size_H, filter_size_W).
2488 2489 2490
            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
2491 2492
            contain three integers, (padding_D, padding_H, padding_W). Otherwise, the
            padding_D = padding_H = padding_W = padding. Default: padding = 0.
2493
        stride(int|tuple): The stride size. If stride is a tuple, it must
2494 2495
            contain three integers, (stride_D, stride_H, stride_W). Otherwise, the
            stride_D = stride_H = stride_W = stride. Default: stride = 1.
2496
        dilation(int|tuple): The dilation size. If dilation is a tuple, it must
2497 2498 2499
            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
2500 2501 2502 2503 2504
            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
2505 2506 2507
        param_attr(ParamAttr): The parameters to the Conv3d_transpose Layer.
            Default: None
        bias_attr(ParamAttr): Bias parameter for the Conv3d layer. Default: None
2508 2509 2510 2511 2512
        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
2513 2514

    Returns:
2515
        Variable: The tensor variable storing the convolution transpose result.
2516 2517

    Raises:
2518 2519
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
2520 2521 2522 2523

    Examples:
       .. code-block:: python

2524 2525
          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
2526
    """
2527 2528
    l_type = "conv3d_transpose"
    helper = LayerHelper(l_type, **locals())
Y
Yu Yang 已提交
2529
    if not isinstance(input, Variable):
2530
        raise TypeError("Input of conv3d_transpose must be Variable")
Y
Yu Yang 已提交
2531 2532
    input_channel = input.shape[1]

2533 2534 2535
    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
2536

C
chengduoZH 已提交
2537 2538 2539
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")

Y
Yu Yang 已提交
2540 2541 2542 2543 2544 2545
    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]

2546 2547 2548
        d_in = input.shape[2]
        h_in = input.shape[3]
        w_in = input.shape[4]
C
chengduoZH 已提交
2549

2550
        filter_size_d = (output_size[0] - (d_in - 1) * stride[0] + 2 *
C
chengduoZH 已提交
2551
                         padding[0] - 1) / dilation[0] + 1
2552
        filter_size_h = (output_size[1] - (h_in - 1) * stride[1] + 2 *
C
chengduoZH 已提交
2553
                         padding[1] - 1) / dilation[1] + 1
2554 2555 2556
        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
2557
    else:
2558 2559
        filter_size = utils.convert_to_list(filter_size, 3,
                                            'conv3d_transpose.filter_size')
Y
Yu Yang 已提交
2560

2561 2562
    groups = 1 if groups is None else groups
    filter_shape = [input_channel, num_filters / groups] + filter_size
Y
Yu Yang 已提交
2563 2564 2565
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

2566
    pre_bias = helper.create_tmp_variable(dtype=input.dtype)
Y
Yu Yang 已提交
2567
    helper.append_op(
2568
        type=l_type,
Y
Yu Yang 已提交
2569 2570
        inputs={'Input': [input],
                'Filter': [img_filter]},
2571
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
2572 2573 2574 2575
        attrs={
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
2576
            'groups': groups,
C
chengduoZH 已提交
2577 2578
            'use_cudnn': use_cudnn
        })
Y
Yu Yang 已提交
2579

2580 2581
    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
2582
    return out
Y
yangyaming 已提交
2583 2584


Y
yangyaming 已提交
2585
def sequence_expand(x, y, ref_level=-1, name=None):
2586
    """Sequence Expand Layer. This layer will expand the input variable **x**
Y
yangyaming 已提交
2587 2588 2589 2590
    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:
2591 2592 2593 2594 2595

    .. code-block:: text

        * Case 1
            x is a LoDTensor:
2596
                x.lod  = [[2,        2]]
Y
yangyaming 已提交
2597
                x.data = [[a], [b], [c], [d]]
2598 2599 2600
                x.dims = [4, 1]

            y is a LoDTensor:
2601 2602
                y.lod = [[2,    2],
                         [3, 3, 1, 1]]
2603

Y
yangyaming 已提交
2604
            ref_level: 0
2605

Y
yangyaming 已提交
2606
            then output is a 1-level LoDTensor:
2607
                out.lod =  [[2,        2,        2,        2]]
Y
yangyaming 已提交
2608
                out.data = [[a], [b], [a], [b], [c], [d], [c], [d]]
2609 2610 2611 2612
                out.dims = [8, 1]

        * Case 2
            x is a Tensor:
Y
yangyaming 已提交
2613
                x.data = [[a], [b], [c]]
2614 2615 2616
                x.dims = [3, 1]

            y is a LoDTensor:
2617
                y.lod = [[2, 0, 3]]
2618

Y
yangyaming 已提交
2619
            ref_level: -1
2620

Y
yangyaming 已提交
2621 2622 2623
            then output is a Tensor:
                out.data = [[a], [a], [c], [c], [c]]
                out.dims = [5, 1]
2624 2625 2626
    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
2627 2628
        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
2629
        name(str|None): A name for this layer(optional). If set None, the layer
Y
yangyaming 已提交
2630
                        will be named automatically.
2631 2632 2633 2634 2635 2636 2637 2638 2639 2640

    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
2641
            out = layers.sequence_expand(x=x, y=y, ref_level=0)
2642
    """
Y
yangyaming 已提交
2643
    helper = LayerHelper('sequence_expand', input=x, **locals())
2644 2645 2646
    dtype = helper.input_dtype()
    tmp = helper.create_tmp_variable(dtype)
    helper.append_op(
Y
yangyaming 已提交
2647 2648 2649 2650 2651
        type='sequence_expand',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': tmp},
        attrs={'ref_level': ref_level})
2652
    return tmp
2653 2654


2655 2656 2657 2658 2659 2660 2661 2662 2663
def beam_search(pre_ids,
                pre_scores,
                ids,
                scores,
                beam_size,
                end_id,
                level=0,
                name=None):
    """
2664 2665
    Beam search is a classical algorithm for selecting candidate words in a
    machine translation task.
Y
Yan Chunwei 已提交
2666 2667 2668

    Refer to `Beam search <https://en.wikipedia.org/wiki/Beam_search>`_
    for more details.
2669 2670 2671 2672 2673 2674 2675 2676
    
    This layer does the search in beams for one time step. Specifically, it 
    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.
2677 2678 2679 2680 2681 2682 2683 2684 2685
 
    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
2686

2687
    Args:
2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712
        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
2713

2714
    Returns:
2715 2716
        Variable: The LodTensor pair containing the selected ids and the \
            corresponding scores.
Y
Yan Chunwei 已提交
2717 2718 2719 2720

    Examples:
        .. code-block:: python

2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737
            # 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
2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748
    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,
2749
            'pre_scores': pre_scores,
Q
Qiao Longfei 已提交
2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766
            '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


2767 2768 2769 2770 2771 2772 2773
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
2774

2775 2776 2777 2778 2779 2780 2781 2782 2783
    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
2784

2785 2786 2787 2788 2789 2790
    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
2791

2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816
    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
2817 2818 2819 2820
def lstm_unit(x_t,
              hidden_t_prev,
              cell_t_prev,
              forget_bias=0.0,
Y
yangyaming 已提交
2821
              param_attr=None,
C
caoying03 已提交
2822 2823
              bias_attr=None,
              name=None):
Y
yangyaming 已提交
2824 2825 2826 2827
    """Lstm unit layer. The equation of a lstm step is:

        .. math::

2828
            i_t & = \sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i)
Y
yangyaming 已提交
2829

2830
            f_t & = \sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + b_f)
Y
yangyaming 已提交
2831

2832
            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
2833

2834
            o_t & = \sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + b_o)
Y
yangyaming 已提交
2835 2836 2837

            h_t & = o_t tanh(c_t)

2838 2839 2840 2841 2842 2843
    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
2844 2845 2846

        .. math::

2847
            L_{i_t} = W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i
Y
yangyaming 已提交
2848 2849 2850 2851 2852 2853 2854 2855

    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
2856
    This layer has two outputs including :math:`h_t` and :math:`o_t`.
Y
yangyaming 已提交
2857 2858

    Args:
Y
yangyaming 已提交
2859 2860 2861 2862 2863 2864
        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
2865
        forget_bias (float): The forget bias of lstm unit.
Y
yangyaming 已提交
2866 2867
        param_attr (ParamAttr): The attributes of parameter weights, used to set
            initializer, name etc.
Y
yangyaming 已提交
2868 2869
        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
2870 2871
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
Y
yangyaming 已提交
2872 2873

    Returns:
Y
yangyaming 已提交
2874
        tuple: The hidden value and cell value of lstm unit.
Y
yangyaming 已提交
2875 2876

    Raises:
2877 2878 2879 2880
        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
2881 2882 2883 2884 2885 2886

    Examples:

        .. code-block:: python

             x_t = fluid.layers.fc(input=x_t_data, size=10)
2887
             prev_hidden = fluid.layers.fc(input=prev_hidden_data, size=30)
Y
yangyaming 已提交
2888
             prev_cell = fluid.layers.fc(input=prev_cell_data, size=30)
Y
yangyaming 已提交
2889
             hidden_value, cell_value = fluid.layers.lstm_unit(x_t=x_t,
Y
yangyaming 已提交
2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905
                                                    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
2906
        raise ValueError("The 1st dimensions of x_t, hidden_t_prev and "
2907 2908 2909 2910
                         "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
2911 2912
                         "cell_t_prev must be the same.")

Y
yangyaming 已提交
2913 2914 2915
    if bias_attr is None:
        bias_attr = ParamAttr()

Y
yangyaming 已提交
2916
    size = cell_t_prev.shape[1]
2917
    concat_out = concat(input=[x_t, hidden_t_prev], axis=1)
Y
yangyaming 已提交
2918 2919
    fc_out = fc(input=concat_out,
                size=4 * size,
Y
yangyaming 已提交
2920
                param_attr=param_attr,
2921
                bias_attr=bias_attr)
Y
yangyaming 已提交
2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933
    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
2934
    return h, c
G
guosheng 已提交
2935 2936


C
caoying03 已提交
2937
def reduce_sum(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
2938
    """
Y
yangyaming 已提交
2939
    Computes the sum of tensor elements over the given dimension.
G
guosheng 已提交
2940 2941 2942

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
2943
        dim (list|int|None): The dimensions along which the sum is performed. If
Y
yangyaming 已提交
2944 2945
            :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
2946 2947
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
2948
        keep_dim (bool|False): Whether to reserve the reduced dimension in the
Y
yangyaming 已提交
2949
            output Tensor. The result tensor will have one fewer dimension
G
guosheng 已提交
2950
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
2951 2952
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
2953 2954 2955

    Returns:
        Variable: The reduced Tensor variable.
F
fengjiayi 已提交
2956

G
guosheng 已提交
2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967
    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_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
2968 2969 2970 2971 2972 2973 2974 2975

            # x is a Tensor variable with shape [2, 2, 2] and elements as below:
            #      [[[1, 2], [3, 4]],
            #      [[5, 6], [7, 8]]]
            # Each example is followed by the correspending output tensor.
            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
2976 2977 2978
    """
    helper = LayerHelper('reduce_sum', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
2979 2980
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
G
guosheng 已提交
2981 2982 2983 2984 2985
    helper.append_op(
        type='reduce_sum',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
2986
            'dim': dim if dim != None else [0],
G
guosheng 已提交
2987 2988 2989 2990
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
2991 2992


C
caoying03 已提交
2993
def reduce_mean(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
2994
    """
Y
Yibing Liu 已提交
2995
    Computes the mean of the input tensor's elements along the given dimension.
G
guosheng 已提交
2996 2997 2998

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
Yibing Liu 已提交
2999 3000 3001
        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
3002
            must be in the range :math:`[-rank(input), rank(input))`. If
3003
            :math:`dim[i] < 0`, the dimension to reduce is
Y
Yibing Liu 已提交
3004
            :math:`rank(input) + dim[i]`.
Y
yangyaming 已提交
3005 3006
        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
3007
            than the :attr:`input` unless :attr:`keep_dim` is true.
Y
Yibing Liu 已提交
3008
        name(str|None): A name for this layer(optional). If set `None`, the layer
C
caoying03 已提交
3009
                       will be named automatically.
G
guosheng 已提交
3010 3011

    Returns:
Y
Yibing Liu 已提交
3012
        Variable: The reduced mean Variable.
F
fengjiayi 已提交
3013

G
guosheng 已提交
3014 3015 3016 3017 3018 3019 3020 3021 3022 3023
    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
3024 3025
            fluid.layers.reduce_mean(
                x, dim=1, keep_dim=True)  # [[0.475], [0.4]]
W
whs 已提交
3026 3027 3028 3029 3030 3031 3032

            # 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
3033 3034 3035
    """
    helper = LayerHelper('reduce_mean', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3036 3037
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
G
guosheng 已提交
3038 3039 3040 3041 3042
    helper.append_op(
        type='reduce_mean',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3043
            'dim': dim if dim != None else [0],
G
guosheng 已提交
3044 3045 3046 3047
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
3048 3049


C
caoying03 已提交
3050
def reduce_max(input, dim=None, keep_dim=False, name=None):
3051
    """
Y
yangyaming 已提交
3052
    Computes the maximum of tensor elements over the given dimension.
3053 3054 3055

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
3056
        dim (list|int|None): The dimension along which the maximum is computed.
Y
yangyaming 已提交
3057 3058 3059
            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
3060
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
Y
yangyaming 已提交
3061 3062
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
3063
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
3064 3065
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
3066 3067 3068

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
3069

3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080
    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
3081 3082 3083 3084 3085 3086 3087

            # 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]
3088 3089 3090
    """
    helper = LayerHelper('reduce_max', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3091 3092
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3093 3094 3095 3096 3097
    helper.append_op(
        type='reduce_max',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3098
            'dim': dim if dim != None else [0],
3099 3100 3101 3102 3103 3104
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


C
caoying03 已提交
3105
def reduce_min(input, dim=None, keep_dim=False, name=None):
3106
    """
Y
yangyaming 已提交
3107
    Computes the minimum of tensor elements over the given dimension.
3108 3109 3110

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
W
whs 已提交
3111
        dim (list|int|None): The dimensions along which the minimum is computed.
Y
yangyaming 已提交
3112 3113 3114
            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
3115
            If :math:`dim[i] < 0`, the dimension to reduce is :math:`rank + dim[i]`.
Y
yangyaming 已提交
3116 3117
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
3118
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
3119 3120
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
3121 3122 3123

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
3124

3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135
    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
3136 3137 3138 3139 3140 3141 3142

            # 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]
3143 3144 3145
    """
    helper = LayerHelper('reduce_min', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3146 3147
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3148 3149 3150 3151 3152
    helper.append_op(
        type='reduce_min',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3153
            'dim': dim if dim != None else [0],
3154 3155 3156 3157
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
3158 3159


3160 3161 3162 3163 3164 3165
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
3166
        dim (list|int|None): The dimensions along which the product is performed. If
3167 3168
            :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
3169 3170
            range :math:`[-rank(input), rank(input))`. If :math:`dim[i] < 0`,
            the dimension to reduce is :math:`rank + dim[i]`.
3171 3172 3173
        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
3174
        name(str|None): A name for this layer(optional). If set None, the
Z
zhouhanqing 已提交
3175
            layer will be named automatically.
3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189

    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
3190
            fluid.layers.reduce_prod(x, dim=1,
Z
zhouhanqing 已提交
3191
                                     keep_dim=True)  # [[0.027], [0.0084]]
W
whs 已提交
3192 3193 3194 3195 3196 3197 3198

            # 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]
3199 3200 3201
    """
    helper = LayerHelper('reduce_prod', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
W
whs 已提交
3202 3203
    if dim is not None and not isinstance(dim, list):
        dim = [dim]
3204 3205 3206 3207 3208
    helper.append_op(
        type='reduce_prod',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
whs 已提交
3209
            'dim': dim if dim != None else [0],
3210 3211 3212 3213 3214 3215
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


C
caoying03 已提交
3216
def split(input, num_or_sections, dim=-1, name=None):
G
guosheng 已提交
3217
    """
C
caoying03 已提交
3218
    Split the input tensor into multiple sub-tensors.
G
guosheng 已提交
3219 3220 3221

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
C
caoying03 已提交
3222 3223 3224 3225 3226
        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
3227
            :attr:`dim` dimension orderly.
C
caoying03 已提交
3228
        dim (int): The dimension along which to split. If :math:`dim < 0`, the
G
guosheng 已提交
3229
            dimension to split along is :math:`rank(input) + dim`.
C
caoying03 已提交
3230 3231
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
3232 3233

    Returns:
D
dzhwinter 已提交
3234
        list(Variable): The list of segmented tensor variables.
G
guosheng 已提交
3235 3236 3237 3238 3239 3240 3241 3242 3243

    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
3244 3245
            x0, x1, x2 = fluid.layers.split(
                x, num_or_sections=[2, 3, 4], dim=1)
G
guosheng 已提交
3246 3247 3248 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
            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
3275 3276 3277 3278 3279 3280 3281 3282 3283


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

3284
    .. math::
3285 3286

        y = \\frac{x}{ \sqrt{\sum {x^2} + epsion }}
C
caoying03 已提交
3287 3288 3289 3290 3291

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

    Args:
3292
        x(Variable|list): The input tensor to l2_normalize layer.
3293
        axis(int): The axis on which to apply normalization. If `axis < 0`, \
3294 3295
            the dimension to normalization is rank(X) + axis. -1 is the
            last dimension.
3296
        epsilon(float): The epsilon value is used to avoid division by zero, \
3297
            the defalut value is 1e-10.
3298
        name(str|None): A name for this layer(optional). If set None, the layer \
3299
            will be named automatically.
C
caoying03 已提交
3300 3301

    Returns:
3302
        Variable: The output tensor variable is the same shape with `x`.
C
caoying03 已提交
3303 3304

    Examples:
3305

C
caoying03 已提交
3306 3307
        .. code-block:: python

3308 3309 3310 3311
            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
3312 3313
    """

F
fengjiayi 已提交
3314 3315
    if len(x.shape) == 1:
        axis = 0
C
caoying03 已提交
3316 3317
    helper = LayerHelper("l2_normalize", **locals())

3318 3319
    out = helper.create_tmp_variable(dtype=x.dtype)
    norm = helper.create_tmp_variable(dtype=x.dtype)
C
caoying03 已提交
3320
    helper.append_op(
3321 3322 3323 3324
        type="norm",
        inputs={"X": x},
        outputs={"Out": out,
                 "Norm": norm},
C
caoying03 已提交
3325
        attrs={
3326 3327
            "axis": 1 if axis is None else axis,
            "epsilon": epsilon,
C
caoying03 已提交
3328 3329
        })
    return out
3330 3331


3332
def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
G
guosheng 已提交
3333
    """
Y
ying 已提交
3334 3335 3336 3337
    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
3338

C
chengduoZH 已提交
3339
    The actual behavior depends on the shapes of :math:`x`, :math:`y` and the
3340
    flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically:
G
guosheng 已提交
3341

3342 3343 3344 3345 3346
    - 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
3347
      :math:`[1, D]` in transposed form.
G
guosheng 已提交
3348

C
chengduoZH 已提交
3349
    - After transpose, the two tensors are 2-D or n-D and matrix multiplication
3350
      performs in the following way.
G
guosheng 已提交
3351

3352
      - If both are 2-D, they are multiplied like conventional matrices.
C
chengduoZH 已提交
3353
      - If either is n-D, it is treated as a stack of matrices residing in the
Y
ying 已提交
3354
        last two dimensions and a batched matrix multiply supporting broadcast
3355
        applies on the two tensors.
G
guosheng 已提交
3356

Y
ying 已提交
3357 3358
    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
3359
    removed after matrix multiplication.
G
guosheng 已提交
3360 3361 3362

    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
3363 3364 3365
        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.
3366
        name(str|None): A name for this layer(optional). If set None, the layer
3367
            will be named automatically.
G
guosheng 已提交
3368 3369

    Returns:
3370
        Variable: The product Tensor variable.
G
guosheng 已提交
3371

G
guosheng 已提交
3372 3373 3374
    Examples:
        .. code-block:: python

3375
            # Examples to clarify shapes of the inputs and output
C
chengduoZH 已提交
3376 3377
            # x: [B, ..., M, K], y: [B, ..., K, N]
            fluid.layers.matmul(x, y)  # out: [B, ..., M, N]
Y
ying 已提交
3378

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

3385 3386
            # x: [M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [M, N]
Y
ying 已提交
3387 3388 3389 3390

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

3391 3392
            # x: [K], y: [K]
            fluid.layers.matmul(x, y)  # out: [1]
3393

Y
ying 已提交
3394
            # x: [M], y: [N]
3395
            fluid.layers.matmul(x, y, True, True)  # out: [M, N]
G
guosheng 已提交
3396
    """
Y
ying 已提交
3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408

    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
3409
            y_shape = y_shape + [1]
Y
ying 已提交
3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425

        # 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)

3426
    helper = LayerHelper('matmul', **locals())
Y
ying 已提交
3427
    out = helper.create_tmp_variable(dtype=x.dtype)
G
guosheng 已提交
3428
    helper.append_op(
3429 3430 3431 3432 3433 3434 3435
        type='matmul',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
        attrs={'transpose_X': transpose_x,
               'transpose_Y': transpose_y})
    return out
3436 3437


3438
def topk(input, k, name=None):
Q
qingqing01 已提交
3439 3440 3441 3442
    """
    This operator is used to find values and indices of the k largest entries
    for the last dimension.

F
fengjiayi 已提交
3443
    If the input is a vector (1-D Tensor), finds the k largest entries in the vector
Q
qingqing01 已提交
3444 3445 3446 3447 3448 3449
    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
3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470
    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
3471 3472 3473
    Args:
        input(Variable): The input variable which can be a vector or Tensor with
            higher rank.
3474
        k(int):  The number of top elements to look for along the last dimension
F
fengjiayi 已提交
3475
                 of input.
3476
        name(str|None): A name for this layer(optional). If set None, the layer
3477
                       will be named automatically.
F
fengjiayi 已提交
3478
                       Default: None
Q
qingqing01 已提交
3479 3480

    Returns:
3481 3482 3483
        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
3484
        within the last dimension of input.
Q
qingqing01 已提交
3485

F
fengjiayi 已提交
3486 3487
    Raises:
        ValueError: If k < 1 or k is not less than the last dimension of input
Q
qingqing01 已提交
3488 3489 3490 3491 3492 3493 3494

    Examples:
        .. code-block:: python

            top5_values, top5_indices = layers.topk(input, k=5)
    """
    shape = input.shape
F
fengjiayi 已提交
3495
    if k < 1 or k >= shape[-1]:
Q
qingqing01 已提交
3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512
        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


3513
def edit_distance(input, label, normalized=True, ignored_tokens=None):
3514
    """
Y
ying 已提交
3515 3516 3517 3518 3519 3520 3521 3522 3523
    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
3524

Y
ying 已提交
3525
    "kitten" -> "sitten" -> "sittin" -> "sitting"
W
wanghaoshuang 已提交
3526

3527
    The input is a LoDTensor consisting of all the hypothesis strings with
Y
ying 已提交
3528 3529
    the total number denoted by `batch_size`, and the separation is specified
    by the LoD information. And the `batch_size` reference strings are arranged
3530
    in order in the same way in the input LoDTensor.
W
wanghaoshuang 已提交
3531

3532
    The output contains the `batch_size` results and each stands for the edit
Y
ying 已提交
3533 3534
    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
3535

3536 3537 3538
    Args:
        input(Variable): The indices for hypothesis strings.
        label(Variable): The indices for reference strings.
3539
        normalized(bool, default True): Indicated whether to normalize the edit distance by
Y
ying 已提交
3540
                          the length of reference string.
3541
        ignored_tokens(list<int>, default None): Tokens that should be removed before
Y
ying 已提交
3542
                                     calculating edit distance.
3543
        name (str): The name of this layer. It is optional.
3544

W
wanghaoshuang 已提交
3545
    Returns:
W
wanghaoshuang 已提交
3546
        Variable: sequence-to-sequence edit distance in shape [batch_size, 1].
W
wanghaoshuang 已提交
3547 3548 3549 3550 3551

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
3552
            y = fluid.layers.data(name='y', shape=[7], dtype='float32')
3553
            cost = fluid.layers.edit_distance(input=x,label=y)
3554
    """
3555
    helper = LayerHelper("edit_distance", **locals())
3556

3557
    # remove some tokens from input and labels
W
wanghaoshuang 已提交
3558
    if ignored_tokens is not None and len(ignored_tokens) > 0:
3559 3560 3561 3562 3563 3564 3565
        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
3566
            attrs={"tokens": ignored_tokens})
3567 3568 3569 3570 3571
        input = erased_input

        helper.append_op(
            type="sequence_erase",
            inputs={"X": [label]},
W
whs 已提交
3572
            outputs={"Out": [erased_label]},
W
wanghaoshuang 已提交
3573
            attrs={"tokens": ignored_tokens})
3574 3575
        label = erased_label

3576 3577
    # edit distance op
    edit_distance_out = helper.create_tmp_variable(dtype="int64")
3578
    sequence_num = helper.create_tmp_variable(dtype="int64")
3579 3580 3581 3582
    helper.append_op(
        type="edit_distance",
        inputs={"Hyps": [input],
                "Refs": [label]},
3583 3584
        outputs={"Out": [edit_distance_out],
                 "SequenceNum": [sequence_num]},
3585 3586
        attrs={"normalized": normalized})

3587
    return edit_distance_out, sequence_num
3588 3589 3590 3591 3592


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
3593

Y
ying 已提交
3594 3595 3596 3597
    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.
3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614

    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]]

3615
        input.lod = [[4, 4]]
3616 3617 3618 3619 3620 3621 3622

        Then:

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

3623
        output.lod = [[2, 1]]
3624 3625 3626

    Args:

Y
ying 已提交
3627 3628 3629 3630 3631 3632 3633 3634 3635
        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).
3636
        name (str): The name of this layer. It is optional.
3637 3638

    Returns:
3639
        Variable: CTC greedy decode result. If all the sequences in result were
3640
        empty, the result LoDTensor will be [-1] with LoD [[]] and dims [1, 1].
3641 3642 3643 3644 3645

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
W
wanghaoshuang 已提交
3646

3647
            cost = fluid.layers.ctc_greedy_decoder(input=x, blank=0)
W
wanghaoshuang 已提交
3648
    """
3649
    helper = LayerHelper("ctc_greedy_decoder", **locals())
Q
qingqing01 已提交
3650
    _, topk_indices = topk(input, k=1)
3651 3652 3653 3654 3655 3656

    # 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
3657
        outputs={"Output": [ctc_out]},
3658 3659
        attrs={"merge_repeated": True,
               "blank": blank})
3660
    return ctc_out
3661 3662


F
fengjiayi 已提交
3663
def warpctc(input, label, blank=0, norm_by_times=False):
W
wanghaoshuang 已提交
3664
    """
3665 3666
    An operator integrating the open source Warp-CTC library
    (https://github.com/baidu-research/warp-ctc)
W
wanghaoshuang 已提交
3667
    to compute Connectionist Temporal Classification (CTC) loss.
3668 3669
    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
3670 3671 3672
    input tensor.

    Args:
3673
       input (Variable): The unscaled probabilities of variable-length sequences,
W
wanghaoshuang 已提交
3674 3675 3676 3677
         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).
3678
       label (Variable): The ground truth of variable-length sequence,
3679 3680 3681
         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
3682 3683
         Temporal Classification (CTC) loss, which is in the
         half-opened interval [0, num_classes + 1).
3684 3685 3686
       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
3687
         follewed by a mean_op.
W
wanghaoshuang 已提交
3688 3689

    Returns:
3690 3691
        Variable: The Connectionist Temporal Classification (CTC) loss,
        which is a 2-D Tensor of the shape [batch_size, 1].
W
wanghaoshuang 已提交
3692 3693

    Examples:
3694

W
wanghaoshuang 已提交
3695
        .. code-block:: python
3696

3697 3698 3699
            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
3700 3701

    """
F
fengjiayi 已提交
3702
    helper = LayerHelper('warpctc', **locals())
W
wanghaoshuang 已提交
3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713
    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
3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728


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]]
3729 3730 3731
            x.data = [[1,  2], [3,  4],
                      [5,  6], [7,  8],
                      [9, 10], [11, 12]]
3732 3733 3734 3735 3736
            x.dims = [6, 2]

        set new_dim = 4

        then out is a LoDTensor:
3737

3738
            out.lod  = [[0, 1, 3]]
3739 3740 3741 3742

            out.data = [[1,  2,  3,  4],
                        [5,  6,  7,  8],
                        [9, 10, 11, 12]]
3743 3744 3745 3746 3747 3748 3749
            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:
3750 3751 3752

       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.
3753 3754

    Returns:
3755

3756 3757 3758 3759 3760
        Variable: Reshaped LoDTensor according to new dimension.

    Examples:
        .. code-block:: python

3761
            x = fluid.layers.data(shape=[5, 20], dtype='float32', lod_level=1)
3762
            x_reshaped = fluid.layers.sequence_reshape(input=x, new_dim=10)
3763 3764 3765 3766 3767 3768 3769 3770 3771
    """
    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
3772 3773


3774 3775 3776 3777
# 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
3778 3779 3780 3781 3782 3783 3784
def nce(input,
        label,
        num_total_classes,
        sample_weight=None,
        param_attr=None,
        bias_attr=None,
        num_neg_samples=None):
3785 3786 3787 3788 3789 3790 3791
    """
    ${comment}

    Args:
        input (Variable): input variable.
        label (Variable): label.
        num_total_classes (int):${num_total_classes_comment}
3792 3793
        sample_weight (Variable|None): A Variable of shape [batch_size, 1]
            storing a weight for each sample. The default weight for each
3794
            sample is 1.0.
3795 3796 3797
        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
3798

3799
    Returns:
Y
Yibing Liu 已提交
3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826
        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')
3827
    """
Y
Yang Yu 已提交
3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846
    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
3847 3848 3849 3850 3851 3852 3853 3854 3855
    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
3856 3857 3858 3859 3860 3861 3862 3863 3864 3865 3866 3867 3868 3869 3870 3871

    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
3872
    return cost / (num_neg_samples + 1)
3873 3874


G
guosheng 已提交
3875
def hsigmoid(input, label, num_classes, param_attr=None, bias_attr=None):
W
weixing02 已提交
3876 3877 3878
    """
    The hierarchical sigmoid operator is used to accelerate the training
    process of language model. This operator organizes the classes into a 
G
guosheng 已提交
3879 3880 3881 3882 3883 3884 3885 3886 3887 3888
    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>`_
    
W
weixing02 已提交
3889
    Args:
G
guosheng 已提交
3890 3891 3892 3893 3894 3895
        input (Variable): The input tensor variable with shape 
            :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
3896 3897 3898
        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 
G
guosheng 已提交
3899 3900
             attribute for the bias of this layer. If it is set to False, no
             bias will be applied.
W
weixing02 已提交
3901 3902 3903 3904 3905 3906 3907 3908

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

    Examples:

        .. code-block:: python

G
guosheng 已提交
3909 3910 3911
            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
3912 3913 3914 3915 3916 3917 3918 3919
    """

    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
3920
        raise ValueError("num_classes must not be less than 2.")
W
weixing02 已提交
3921 3922 3923 3924 3925
    weights = helper.create_parameter(
        attr=helper.param_attr,
        shape=[num_classes - 1, dim],
        is_bias=False,
        dtype=input.dtype)
W
weixing02 已提交
3926 3927 3928 3929 3930 3931 3932 3933
    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
3934 3935
    helper.append_op(
        type="hierarchical_sigmoid",
W
weixing02 已提交
3936
        inputs=inputs,
W
weixing02 已提交
3937 3938 3939 3940 3941 3942
        outputs={"Out": out,
                 "PreOut": pre_out},
        attrs={"num_classes": num_classes})
    return out


Y
fix ci.  
ying 已提交
3943
def transpose(x, perm, name=None):
Y
ying 已提交
3944 3945 3946 3947 3948 3949 3950
    """
    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:
3951 3952 3953
        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
3954 3955 3956 3957 3958 3959 3960 3961

    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
3962
            x_transposed = layers.transpose(x, perm=[1, 0, 2])
Y
ying 已提交
3963 3964
    """

Y
fix ci.  
ying 已提交
3965
    if len(perm) != len(x.shape):
Y
ying 已提交
3966 3967 3968
        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
3969 3970 3971 3972 3973 3974
    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
3975 3976

    helper = LayerHelper('transpose', **locals())
Y
fix ci.  
ying 已提交
3977
    out = helper.create_tmp_variable(x.dtype)
Y
ying 已提交
3978 3979
    helper.append_op(
        type='transpose',
Y
fix ci.  
ying 已提交
3980
        inputs={'X': [x]},
Y
ying 已提交
3981 3982 3983
        outputs={'Out': [out]},
        attrs={'axis': perm})
    return out
3984 3985


3986 3987 3988 3989 3990 3991 3992
def im2sequence(input,
                filter_size=1,
                stride=1,
                padding=0,
                input_image_size=None,
                out_stride=1,
                name=None):
3993
    """
3994 3995 3996 3997 3998 3999 4000
    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:
4001 4002 4003 4004 4005 4006 4007 4008 4009 4010

    .. 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
4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028

        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.

4029 4030 4031 4032 4033 4034 4035 4036 4037
        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.

4038 4039 4040
        name (int): The name of this layer. It is optional.

    Returns:
W
wanghaoshuang 已提交
4041 4042 4043 4044 4045
        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.
4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057 4058 4059 4060 4061 4062 4063 4064 4065 4066 4067 4068 4069 4070 4071 4072

    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
4073 4074 4075
            filter = [2, 2]
            stride = [1, 1]
            padding = [0, 0]
4076 4077 4078 4079 4080 4081 4082 4083 4084 4085 4086 4087

            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.]]

4088
            output.dims = {8, 8}
4089

4090
            output.lod = [[4, 4]]
4091

D
dzhwinter 已提交
4092
     Examples:
4093 4094 4095

        .. code-block:: python

4096 4097
            output = fluid.layers.im2sequence(
                input=layer, stride=[1, 1], filter_size=[2, 2])
4098 4099

    """
W
wanghaoshuang 已提交
4100 4101 4102 4103 4104 4105 4106 4107 4108 4109

    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])
4110 4111 4112 4113 4114 4115 4116
    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
4117
    helper = LayerHelper('im2sequence', **locals())
4118 4119
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
4120
        type='im2sequence', inputs=inputs, outputs={'Out': out}, attrs=attrs)
4121
    return out
4122 4123


Y
yuyang18 已提交
4124
@templatedoc()
4125
def row_conv(input, future_context_size, param_attr=None, act=None):
Y
yuyang18 已提交
4126 4127
    """
    ${comment}
4128 4129

    Args:
Y
yuyang18 已提交
4130
        input (${x_type}): ${x_comment}.
Y
yangyaming 已提交
4131 4132
        future_context_size (int): Future context size. Please note, the shape
            of convolution kernel is [future_context_size + 1, D].
4133 4134 4135 4136 4137
        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
4138
        ${out_comment}.
4139 4140

    Examples:
Y
yuyang18 已提交
4141 4142 4143 4144
        >>> 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)
4145 4146 4147 4148 4149 4150 4151 4152 4153 4154 4155 4156
    """
    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
4157
    return helper.append_activation(out)
4158 4159


Y
yuyang18 已提交
4160
@templatedoc()
4161 4162
def multiplex(inputs, index):
    """
Y
yuyang18 已提交
4163 4164 4165 4166 4167 4168 4169
    ${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)
4170 4171

    Args:
Y
yuyang18 已提交
4172 4173
       inputs (list): ${x_comment}.
       index (${ids_type}): ${ids_comment}.
4174 4175

    Returns:
Y
yuyang18 已提交
4176
        ${out_comment}.
4177 4178
    """
    helper = LayerHelper('multiplex', **locals())
Y
yangyaming 已提交
4179 4180 4181 4182 4183 4184

    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)
4185 4186 4187 4188 4189 4190
    helper.append_op(
        type='multiplex',
        inputs={'X': inputs,
                'Ids': index},
        outputs={'Out': [out]})
    return out
4191 4192 4193 4194 4195


def softmax_with_cross_entropy(logits, label, soft_label=False):
    """
    **Softmax With Cross Entropy Operator.**
4196

4197 4198 4199 4200
    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.
4201

4202 4203 4204
    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.
4205

4206 4207 4208
    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.
4209

4210
    The equation is as follows:
4211

4212
    1) Hard label (one-hot label, so every sample has exactly one class)
4213

4214 4215 4216 4217
    .. math::

        loss_j =  -\\text{logit}_{label_j} +
        \\log\\left(\\sum_{i=0}^{K}\\exp(\\text{logit}_i)\\right), j = 1,..., K
4218

4219 4220 4221
    2) Soft label (each sample can have a distribution over all classes)

    .. math::
4222

4223 4224 4225 4226 4227 4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238 4239 4240 4241 4242 4243
        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
4244 4245
            out = fluid.layers.softmax_with_cross_entropy(
                logits=fc, label=label)
4246 4247 4248 4249 4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261
    """
    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
4262 4263
    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
4264
    For each instance, it computes the smooth L1 loss element by element first
4265
    and then sums all the losses. So the shape of ouput Variable is
4266
    [batch_size, 1].
4267

4268 4269
    Args:
        x (Variable): A tensor with rank at least 2. The input value of smooth
Q
qingqing01 已提交
4270
            L1 loss op with shape [batch_size, dim1, ..., dimN].
4271
        y (Variable): A tensor with rank at least 2. The target value of smooth
Y
Yibing Liu 已提交
4272
            L1 loss op with same shape as :attr:`x`.
4273
        inside_weight (Variable|None):  A tensor with rank at least 2. This
4274 4275
            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
4276
            by this tensor element by element.
4277
        outside_weight (Variable|None): A tensor with rank at least 2. This
4278 4279
            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
4280
            element by element.
4281
        sigma (float|None): Hyper parameter of smooth L1 loss layer. A float
4282 4283
           scalar with default value 1.0.

4284
    Returns:
4285
        Variable: The output smooth L1 loss with shape [batch_size, 1].
4286 4287 4288 4289 4290

    Examples:
        .. code-block:: python

            data = fluid.layers.data(name='data', shape=[128], dtype='float32')
F
stash  
fengjiayi 已提交
4291 4292
            label = fluid.layers.data(
                name='label', shape=[100], dtype='float32')
4293
            fc = fluid.layers.fc(input=data, size=100)
F
fengjiayi 已提交
4294
            out = fluid.layers.smooth_l1(x=fc, y=label)
4295
    """
4296

4297 4298 4299 4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311
    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
4312 4313 4314 4315


def one_hot(input, depth):
    """
Y
Yibing Liu 已提交
4316
    This layer creates the one-hot representations for input indices.
4317 4318

    Args:
Y
Yibing Liu 已提交
4319 4320
        input(Variable): Input indices, last dimension must be 1.
        depth(scalar): An interger defining the depth of the one-hot dimension.
4321 4322

    Returns:
Y
Yibing Liu 已提交
4323
        Variable: The one-hot representations of input.
4324 4325

    Examples:
C
caoying03 已提交
4326
        .. code-block:: python
4327

Y
Yibing Liu 已提交
4328 4329
            label = layers.data(name="label", shape=[1], dtype="float32")
            one_hot_label = layers.one_hot(input=label, depth=10)
4330 4331 4332 4333 4334 4335 4336 4337 4338
    """
    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
4339 4340


Y
Yu Yang 已提交
4341
def autoincreased_step_counter(counter_name=None, begin=1, step=1):
Y
Yu Yang 已提交
4342
    """
Y
yi.wu 已提交
4343 4344 4345
    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
4346 4347 4348 4349 4350 4351

    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.

4352 4353
    Returns:
        Variable: The global run counter.
Y
yi.wu 已提交
4354 4355 4356 4357 4358 4359

    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
4360 4361
    """
    helper = LayerHelper('global_step_counter')
Y
Yu Yang 已提交
4362 4363
    if counter_name is None:
        counter_name = '@STEP_COUNTER@'
Y
Yu Yang 已提交
4364 4365 4366 4367 4368
    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
4369
                value=begin - 1, force_cpu=True))
W
Wu Yi 已提交
4370
        helper.main_program.global_block()._prepend_op(
Y
Yu Yang 已提交
4371 4372
            type='increment',
            inputs={'X': [counter]},
Y
Yu Yang 已提交
4373 4374
            outputs={'Out': [counter]},
            attrs={'step': float(step)})
Y
Yu Yang 已提交
4375 4376 4377
        counter.stop_gradient = True

    return counter
Y
yangyaming 已提交
4378 4379


4380
def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None):
C
caoying03 已提交
4381
    """
C
caoying03 已提交
4382 4383
    Gives a new shape to the input Tensor without changing its data.

4384 4385 4386 4387 4388
    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
4389

4390
    Some tricks exist when specifying the target shape.
C
caoying03 已提交
4391

4392 4393 4394 4395
    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.

4396
    2. 0 means the actual dimension value is going to be copied from the
4397 4398 4399 4400
    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
4401 4402

    1. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
W
wanghaoshuang 已提交
4403
    is [6, 8], the reshape operator will transform x into a 2-D tensor with
4404
    shape [6, 8] and leaving x's data unchanged.
C
caoying03 已提交
4405

4406
    2. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
4407 4408
    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
4409 4410
    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
4411
    dimensions.
C
caoying03 已提交
4412

4413
    3. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
4414 4415 4416 4417
    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
4418 4419

    Args:
4420
        x(variable): The input tensor.
C
caoying03 已提交
4421 4422
        shape(list): The new shape. At most one dimension of the new shape can
                     be -1.
4423 4424 4425 4426 4427
        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
4428
        act (str): The non-linear activation to be applied to output variable.
X
Xin Pan 已提交
4429 4430 4431 4432
        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.
4433
        name (str): The name of this layer. It is optional.
C
caoying03 已提交
4434

4435 4436
    Returns:
        Variable: The output tensor.
C
caoying03 已提交
4437

X
Xin Pan 已提交
4438 4439 4440
    Raises:
        TypeError: if actual_shape is neither Variable nor None.

C
caoying03 已提交
4441 4442
    Examples:
        .. code-block:: python
G
guosheng 已提交
4443

4444
            data = fluid.layers.data(
4445
                name='data', shape=[2, 4, 6], dtype='float32')
C
caoying03 已提交
4446
            reshaped = fluid.layers.reshape(
4447
                x=data, shape=[-1, 0, 3, 2], act='tanh', inplace=True)
C
caoying03 已提交
4448 4449 4450 4451
    """

    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
4452 4453 4454 4455 4456
    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
4457

4458 4459 4460 4461 4462 4463 4464 4465 4466 4467 4468 4469 4470 4471 4472
    # 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
4473 4474 4475 4476
    helper = LayerHelper("reshape", **locals())
    reshaped = helper.create_tmp_variable(dtype=x.dtype)
    helper.append_op(
        type="reshape",
X
Xin Pan 已提交
4477
        inputs=inputs,
C
caoying03 已提交
4478 4479 4480 4481 4482
        attrs={"shape": shape,
               "inplace": inplace},
        outputs={"Out": reshaped})

    return helper.append_activation(reshaped)
4483 4484


Y
yangyaming 已提交
4485
def lod_reset(x, y=None, target_lod=None):
Y
yangyaming 已提交
4486
    """
Y
Yibing Liu 已提交
4487
    Set LoD of :attr:`x` to a new one specified by :attr:`y` or
4488 4489 4490 4491
    :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
4492
    :attr:`Y.data` or :attr:`target_lod`, only one level LoD is supported.
Y
yangyaming 已提交
4493 4494 4495 4496 4497 4498

    .. code-block:: text

        * Example 1:

            Given a 1-level LoDTensor x:
4499
                x.lod =  [[ 2,           3,                   1 ]]
Y
yangyaming 已提交
4500 4501 4502
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

4503
            target_lod: [4, 2]
Y
yangyaming 已提交
4504 4505

            then we get a 1-level LoDTensor:
4506
                out.lod =  [[4,                          2]]
Y
yangyaming 已提交
4507 4508 4509 4510 4511 4512
                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:
4513
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
4514 4515 4516 4517
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a Tensor:
4518
                y.data = [[2, 4]]
Y
yangyaming 已提交
4519 4520 4521
                y.dims = [1, 3]

            then we get a 1-level LoDTensor:
4522
                out.lod =  [[2,            4]]
Y
yangyaming 已提交
4523 4524 4525 4526 4527 4528
                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:
4529
                x.lod =  [[2,            3,                   1]]
Y
yangyaming 已提交
4530 4531 4532 4533
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a 2-level LoDTensor:
4534
                y.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
4535 4536 4537 4538
                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:
4539
                out.lod =  [[2, 2], [2, 2, 1, 1]]
Y
yangyaming 已提交
4540 4541 4542 4543 4544
                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.
4545
        y (Variable|None): If provided, output's LoD would be derived
Y
Yibing Liu 已提交
4546
                           from :attr:`y`.
Y
yangyaming 已提交
4547
        target_lod (list|tuple|None): One level LoD which should be considered
Y
Yibing Liu 已提交
4548
                                      as target LoD when :attr:`y` not provided.
Y
yangyaming 已提交
4549 4550

    Returns:
Y
Yibing Liu 已提交
4551
        Variable: Output variable with LoD specified by this layer.
Y
yangyaming 已提交
4552 4553

    Raises:
Y
Yibing Liu 已提交
4554
        ValueError: If :attr:`y` and :attr:`target_lod` are both None.
Y
yangyaming 已提交
4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576 4577 4578

    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
4579 4580 4581 4582 4583 4584 4585 4586 4587 4588 4589


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
4590
      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
4591 4592 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

    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
4619 4620
          data = fluid.layers.data(
              name="data", shape=[3, 112, 112], dtype="float32")
D
dragonwarrior 已提交
4621 4622 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
          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
4648 4649 4650 4651


def pad(x, paddings, pad_value=0., name=None):
    """
G
guosheng 已提交
4652
    Pads a tensor with a constant value given by :attr:`pad_value`, and the
W
wanghaoshuang 已提交
4653
    padded width is specified by :attr:`paddings`.
G
guosheng 已提交
4654

G
guosheng 已提交
4655 4656 4657 4658
    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
4659 4660 4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675 4676 4677 4678 4679 4680

    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
4681
                         The length of :attr:paddings must be
G
guosheng 已提交
4682 4683 4684 4685 4686 4687 4688 4689 4690 4691
                         :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
4692

G
guosheng 已提交
4693 4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706
            # 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
4707 4708 4709 4710 4711 4712 4713 4714 4715


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
4716 4717
    called label-smoothing regularization (LSR).

4718 4719 4720 4721 4722 4723 4724 4725 4726 4727 4728 4729 4730 4731 4732 4733 4734 4735 4736 4737 4738 4739 4740
    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
4741
                              be :math:`(1, class\_num)`.
4742 4743
        epsilon(float): The weight used to mix up the original ground-truth
                        distribution and the fixed distribution.
4744
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32,
4745 4746 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
                                                  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
4772 4773


Y
yi.wu 已提交
4774
@templatedoc()
4775 4776
def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0):
    """
Y
yi.wu 已提交
4777
    ${comment}
4778 4779

    Args:
Y
yi.wu 已提交
4780 4781
        input (Variable): ${x_comment}
        rois (Variable): ROIs (Regions of Interest) to pool over.
Y
yi.wu 已提交
4782 4783 4784
        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
4785 4786

    Returns:
Y
update  
yi.wu 已提交
4787
        Variable: ${out_comment}.
4788 4789

    Examples:
4790 4791
        .. code-block:: python

4792
            pool_out = fluid.layers.roi_pool(input=x, rois=rois, 7, 7, 1.0)
4793 4794 4795 4796 4797 4798 4799 4800 4801 4802 4803 4804 4805 4806 4807 4808 4809
    """
    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
4810 4811 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


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:
4838 4839
        .. code-block:: python

W
whs 已提交
4840 4841 4842 4843 4844 4845 4846 4847 4848 4849 4850
            predictions = fluid.layers.softmax(x)
            loss = fluid.layers.dice_loss(input=predictions, label=label, 2)
    """
    label = one_hot(label, depth=input.shape[-1])
    reduce_dim = range(1, len(input.shape))
    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)
4851 4852


4853 4854 4855 4856 4857
def image_resize(input,
                 out_shape=None,
                 scale=None,
                 name=None,
                 resample='BILINEAR'):
4858
    """
Q
qiaolongfei 已提交
4859
    **Resize a Batch of Images**
F
stash  
fengjiayi 已提交
4860

4861
    The input must be a tensor of the shape (num_batches, channels, in_h, in_w),
4862 4863 4864
    and the resizing only applies on the last two dimensions(hight and width).

    Supporting resample methods:
Q
update  
qiaolongfei 已提交
4865

4866
        'BILINEAR' : Bilinear interpolation
F
stash  
fengjiayi 已提交
4867

4868
    Args:
4869
        input (Variable): The input tensor of image resize layer,
4870 4871
                          This is a 4-D tensor of the shape
                          (num_batches, channels, in_h, in_w).
4872
        out_shape(list|tuple|Variable|None): Output shape of image resize
4873 4874
                                    layer, the shape is (out_h, out_w).
                                    Default: None
B
baiyf 已提交
4875
        scale(float|None): The multiplier for the input height or width.
4876 4877 4878
                         At least one of out_shape or scale must be set.
                         And out_shape has a higher priority than scale.
                         Default: None
4879 4880
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
4881 4882
        resample(str): The resample method. It can only be 'BILINEAR' currently.
                       Default: 'BILINEAR'
4883 4884

    Returns:
Q
update  
qiaolongfei 已提交
4885 4886
        Variable: The output is a 4-D tensor of the shape
        (num_batches, channls, out_h, out_w).
F
stash  
fengjiayi 已提交
4887

4888 4889 4890
    Examples:
        .. code-block:: python

4891
            out = fluid.layers.image_resize(input, out_shape=[12, 12])
4892
    """
4893 4894 4895 4896
    resample_methods = {'BILINEAR': 'bilinear_interp'}
    if resample not in resample_methods:
        raise ValueError(
            "The 'resample' of image_resize can only be 'BILINEAR' currently.")
4897 4898
    if out_shape is None and scale is None:
        raise ValueError("One of out_shape and scale must not be None")
4899 4900
    helper = LayerHelper('bilinear_interp', **locals())
    dtype = helper.input_dtype()
4901 4902 4903 4904

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

4905 4906 4907
    out_h = 0
    out_w = 0
    inputs = {"X": input}
4908
    if out_shape is not None:
B
baiyf 已提交
4909 4910 4911
        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')
4912 4913 4914 4915 4916 4917
        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
4918 4919 4920 4921
    else:
        out_h = int(input.shape[2] * scale)
        out_w = int(input.shape[3] * scale)

4922 4923
    out = helper.create_tmp_variable(dtype)
    helper.append_op(
4924
        type=resample_methods[resample],
4925
        inputs=inputs,
4926 4927 4928 4929
        outputs={"Out": out},
        attrs={"out_h": out_h,
               "out_w": out_w})
    return out
F
stash  
fengjiayi 已提交
4930 4931


Y
yuyang18 已提交
4932
@templatedoc(op_type="bilinear_interp")
4933 4934
def resize_bilinear(input, out_shape=None, scale=None, name=None):
    """
Y
yuyang18 已提交
4935 4936 4937 4938 4939 4940
    ${comment}

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

        out_shape(${out_size_type}): ${out_size_comment}.
4941

Y
yuyang18 已提交
4942 4943 4944 4945 4946 4947 4948 4949
        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}.
4950 4951 4952 4953 4954 4955 4956
    """

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


def image_resize_short(input, out_short_len, resample='BILINEAR'):
    """
4957 4958 4959
    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
4960 4961 4962 4963 4964 4965 4966
    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.
4967
        resample (str): resample method, default: BILINEAR.
F
fengjiayi 已提交
4968

4969
    Returns:
Q
update  
qiaolongfei 已提交
4970
        Variable: The output is a 4-D tensor of the shape
4971
        (num_batches, channls, out_h, out_w).
4972 4973 4974 4975 4976 4977 4978 4979 4980 4981
    """
    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
4982 4983 4984
    out_shape[long_idx] = int(
        float(out_shape[long_idx]) * (float(out_short_len) / float(hw[
            short_idx])) + 0.5)
4985 4986 4987
    return image_resize(input=input, out_shape=out_shape, resample=resample)


W
whs 已提交
4988 4989
def gather(input, index):
    """
Q
qiaolongfei 已提交
4990 4991
    **Gather Layer**

4992
    Output is obtained by gathering entries of the outer-most dimension
W
whs 已提交
4993 4994 4995 4996
    of X indexed by `index` and concatenate them together.

    .. math::

4997
        Out = X[Index]
W
whs 已提交
4998 4999 5000 5001 5002 5003 5004


    .. code-block:: text


                Given:

5005 5006
                X = [[1, 2],
                     [3, 4],
W
whs 已提交
5007 5008 5009 5010 5011 5012 5013 5014 5015 5016
                     [5, 6]]

                Index = [1, 2]

                Then:

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

    Args:
5017
        input (Variable): The source input with rank>=1.
W
whs 已提交
5018 5019 5020 5021 5022 5023
        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
5024

W
whs 已提交
5025 5026 5027 5028 5029 5030 5031 5032 5033 5034 5035 5036 5037 5038 5039
        .. 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
5040 5041 5042 5043 5044 5045 5046 5047 5048 5049 5050 5051 5052
@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}
5053

5054 5055 5056
    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
5057
    """
F
stash  
fengjiayi 已提交
5058
    helper = LayerHelper("random_crop", **locals())
F
fengjiayi 已提交
5059
    dtype = x.dtype
F
stash  
fengjiayi 已提交
5060
    out = helper.create_tmp_variable(dtype)
Y
yuyang18 已提交
5061 5062
    if seed is None:
        seed = random.randint(-65536, 65535)
F
fengjiayi 已提交
5063
    op_attrs = {"shape": shape}
F
stash  
fengjiayi 已提交
5064
    if isinstance(seed, int):
F
fengjiayi 已提交
5065 5066 5067 5068 5069
        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
5070 5071 5072 5073
    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
5074
        inputs={"X": x,
F
stash  
fengjiayi 已提交
5075 5076
                "Seed": seed},
        outputs={"Out": out,
F
fengjiayi 已提交
5077 5078
                 "SeedOut": seed},
        attrs=op_attrs)
F
stash  
fengjiayi 已提交
5079
    return out
W
whs 已提交
5080 5081


5082
def log(x):
W
wanghaoshuang 已提交
5083 5084 5085 5086 5087
    """
    Calculates the natural log of the given input tensor, element-wise.

    .. math::

5088
        Out = \\ln(x)
W
wanghaoshuang 已提交
5089 5090

    Args:
5091
        x (Variable): Input tensor.
W
wanghaoshuang 已提交
5092 5093 5094 5095 5096 5097 5098 5099

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

    Examples:

        .. code-block:: python

5100
            output = fluid.layers.log(x)
W
wanghaoshuang 已提交
5101 5102
    """
    helper = LayerHelper('log', **locals())
W
wanghaoshuang 已提交
5103
    dtype = helper.input_dtype(input_param_name='x')
W
wanghaoshuang 已提交
5104
    out = helper.create_tmp_variable(dtype)
W
wanghaoshuang 已提交
5105
    helper.append_op(type="log", inputs={"X": x}, outputs={"Out": out})
W
wanghaoshuang 已提交
5106 5107 5108
    return out


5109
def relu(x):
W
wanghaoshuang 已提交
5110 5111
    """
    Relu takes one input data (Tensor) and produces one output data (Tensor)
5112
    where the rectified linear function, y = max(0, x), is applied to
W
wanghaoshuang 已提交
5113 5114 5115 5116
    the tensor elementwise.

    .. math::

5117
        Out = \\max(0, x)
W
wanghaoshuang 已提交
5118 5119

    Args:
5120
        x (Variable): The input tensor.
W
wanghaoshuang 已提交
5121 5122 5123 5124 5125 5126 5127 5128

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

    Examples:

        .. code-block:: python

5129
            output = fluid.layers.relu(x)
W
wanghaoshuang 已提交
5130 5131
    """
    helper = LayerHelper('relu', **locals())
W
wanghaoshuang 已提交
5132
    dtype = helper.input_dtype(input_param_name='x')
W
wanghaoshuang 已提交
5133
    out = helper.create_tmp_variable(dtype)
W
wanghaoshuang 已提交
5134
    helper.append_op(type="relu", inputs={"X": x}, outputs={"Out": out})
W
wanghaoshuang 已提交
5135
    return out
5136 5137


W
whs 已提交
5138 5139 5140
def mean_iou(input, label, num_classes):
    """
    Mean Intersection-Over-Union is a common evaluation metric for
5141 5142 5143 5144
    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
5145
    .. math::
5146 5147

        IOU = \\frac{true\_positiv}{(true\_positive + false\_positive + false\_negative)}.
W
whs 已提交
5148

5149
    The predictions are accumulated in a confusion matrix and mean-IOU
W
whs 已提交
5150 5151 5152 5153 5154
    is then calculated from it.


    Args:
        input (Variable): A Tensor of prediction results for semantic labels with type int32 or int64.
5155
        label (Variable): A Tensor of ground truth labels with type int32 or int64.
W
whs 已提交
5156
                           Its shape should be the same as input.
5157
        num_classes (int): The possible number of labels.
W
whs 已提交
5158 5159 5160 5161

    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.
5162
        out_correct(Variable): A Tensor with shape [num_classes]. The correct numbers of each class.
W
whs 已提交
5163 5164 5165 5166

    Examples:

        .. code-block:: python
5167

W
whs 已提交
5168 5169 5170 5171 5172 5173 5174 5175 5176
            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
5177 5178
        inputs={"Predictions": input,
                "Labels": label},
W
whs 已提交
5179
        outputs={
W
whs 已提交
5180 5181 5182
            "OutMeanIou": out_mean_iou,
            "OutWrong": out_wrong,
            "OutCorrect": out_correct
W
whs 已提交
5183 5184 5185
        },
        attrs={"num_classes": num_classes})
    return out_mean_iou, out_wrong, out_correct
5186 5187 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


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