nn.py 115.1 KB
Newer Older
D
dzhwinter 已提交
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# 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
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
Y
Yu Yang 已提交
14 15 16
"""
All layers just related to the neural network.
"""
C
chengduoZH 已提交
17

Y
Yu Yang 已提交
18 19 20
from ..layer_helper import LayerHelper
from ..initializer import Normal, Constant
from ..framework import Variable
Y
yangyaming 已提交
21
from ..param_attr import ParamAttr
22
from layer_function_generator import autodoc
Y
yangyaming 已提交
23
from tensor import concat
C
chengduoZH 已提交
24 25
import math
from operator import mul
Y
Yu Yang 已提交
26 27

__all__ = [
Y
ying 已提交
28 29 30
    'fc',
    'embedding',
    'dynamic_lstm',
Y
Yibing Liu 已提交
31
    'dynamic_lstmp',
G
guosheng 已提交
32
    'dynamic_gru',
Y
ying 已提交
33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57
    'gru_unit',
    'linear_chain_crf',
    'crf_decoding',
    'cos_sim',
    'cross_entropy',
    'square_error_cost',
    'accuracy',
    'chunk_eval',
    'sequence_conv',
    'conv2d',
    'sequence_pool',
    'pool2d',
    'batch_norm',
    'beam_search_decode',
    'conv2d_transpose',
    'sequence_expand',
    'lstm_unit',
    'reduce_sum',
    'reduce_mean',
    'reduce_max',
    'reduce_min',
    'sequence_first_step',
    'sequence_last_step',
    'dropout',
    'split',
58 59
    'ctc_greedy_decoder',
    'edit_distance',
Y
ying 已提交
60 61 62 63
    'l2_normalize',
    'matmul',
    'warpctc',
    'sequence_reshape',
64
    'transpose',
65
    'im2sequence',
66
    'nce',
Q
Qiao Longfei 已提交
67
    'beam_search',
68
    'row_conv',
C
chengduoZH 已提交
69
    'reshape_with_axis',
C
chengduoZH 已提交
70
    'flatten',
71
    'multiplex',
G
guosheng 已提交
72
    'layer_norm',
Y
Yu Yang 已提交
73 74 75 76 77 78 79 80 81
]


def fc(input,
       size,
       num_flatten_dims=1,
       param_attr=None,
       bias_attr=None,
       act=None,
82
       name=None):
Y
Yu Yang 已提交
83
    """
84
    **Fully Connected Layer**
Y
Yu Yang 已提交
85

C
caoying03 已提交
86
    The fully connected layer can take multiple tensors as its inputs. It
Y
ying 已提交
87 88 89 90 91 92 93 94
    creates a variable (one for each input tensor) 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 biases variable will be created and added to the output. Finally,
    if activation is not None, it will be applied to the output as well.
C
caoying03 已提交
95

C
caoying03 已提交
96
    This process can be formulated as follows:
97 98 99

    .. math::

100
        Out = Act({\sum_{i=0}^{N-1}X_iW_i + b})
101 102 103

    In the above equation:

C
caoying03 已提交
104 105 106 107
    * :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).
C
caoying03 已提交
108 109
    * :math:`Act`: The activation funtion.
    * :math:`Out`: The output tensor.
Y
Yu Yang 已提交
110 111

    Args:
C
caoying03 已提交
112 113 114 115 116 117 118 119
       input(Variable|list): The input tensor(s) to the fully connected layer.
       size(int): The number of output units in the fully connected layer.
       num_flatten_dims(int): 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`
Y
ying 已提交
120 121 122 123 124 125 126 127 128 129 130
                              (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]. By default,
                              `num_flatten_dims` is set to 1.
C
caoying03 已提交
131 132 133 134 135 136 137 138 139 140 141 142 143 144 145
       param_attr(ParamAttr|list): The parameter attribute for learnable
                                   parameters/weights of the fully connected
                                   layer.
       param_initializer(ParamAttr|list): The initializer used for the
                                          weight/parameter. If set None,
                                          XavierInitializer() will be used.
       bias_attr(ParamAttr|list): The parameter attribute for the bias parameter
                                  for this layer. If set None, no bias will be
                                  added to the output units.
       bias_initializer(ParamAttr|list): The initializer used for the bias.
                                        If set None, then ConstantInitializer()
                                        will be used.
       act(str): Activation to be applied to the output of the fully connected
                 layer.
       name(str): Name/alias of the fully connected layer.
Y
Yu Yang 已提交
146 147


148
    Returns:
C
caoying03 已提交
149
        Variable: The output tensor variable.
150 151

    Raises:
C
caoying03 已提交
152
        ValueError: If rank of the input tensor is less than 2.
153 154 155 156

    Examples:
        .. code-block:: python

C
caoying03 已提交
157
          data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
158
          fc = fluid.layers.fc(input=data, size=1000, act="tanh")
Y
Yu Yang 已提交
159
    """
C
caoying03 已提交
160

C
caoying03 已提交
161
    helper = LayerHelper("fc", **locals())
Y
Yu Yang 已提交
162 163 164 165 166 167 168 169 170

    dtype = helper.input_dtype()

    mul_results = []
    for input_var, param_attr in helper.iter_inputs_and_params():
        input_shape = input_var.shape
        param_shape = [
            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
        ] + [size]
Y
ying 已提交
171

Y
Yu Yang 已提交
172 173 174 175 176
        w = helper.create_parameter(
            attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False)
        tmp = helper.create_tmp_variable(dtype)
        helper.append_op(
            type="mul",
Q
Qiao Longfei 已提交
177 178
            inputs={"X": input_var,
                    "Y": w},
Y
Yu Yang 已提交
179
            outputs={"Out": tmp},
C
caoying03 已提交
180 181
            attrs={"x_num_col_dims": num_flatten_dims,
                   "y_num_col_dims": 1})
Y
Yu Yang 已提交
182 183 184 185 186 187 188 189 190 191
        mul_results.append(tmp)

    # sum
    if len(mul_results) == 1:
        pre_bias = mul_results[0]
    else:
        pre_bias = helper.create_tmp_variable(dtype)
        helper.append_op(
            type="sum", inputs={"X": mul_results}, outputs={"Out": pre_bias})
    # add bias
192
    pre_activation = helper.append_bias_op(pre_bias, dim_start=num_flatten_dims)
Y
Yu Yang 已提交
193 194 195 196
    # add activation
    return helper.append_activation(pre_activation)


197 198 199 200 201 202
def embedding(input,
              size,
              is_sparse=False,
              padding_idx=None,
              param_attr=None,
              dtype='float32'):
Y
Yu Yang 已提交
203
    """
204 205
    **Embedding Layer**

206
    This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in
207 208
    a lookup table. The result of this lookup is the embedding of each ID in the
    :attr:`input`.
209 210 211

    All the input variables are passed in as local variables to the LayerHelper
    constructor.
Y
Yu Yang 已提交
212 213

    Args:
214 215 216 217 218 219 220
        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.
        padding_idx(int|long|None): If :attr:`None`, it makes no effect to lookup.
            Otherwise the given :attr:`padding_idx` indicates padding the output
221 222
            with zeros whenever lookup encounters it in :attr:`input`. If
            :math:`padding_idx < 0`, the padding_idx to use in lookup is
223 224 225
            :math:`size[0] + dim`.
        param_attr(ParamAttr): Parameters for this layer
        dtype(np.dtype|core.DataType|str): The type of data : float32, float_16, int etc
Y
Yu Yang 已提交
226

227 228 229
    Returns:
        Variable: The tensor variable storing the embeddings of the \
                  supplied inputs.
Y
Yu Yang 已提交
230

231 232
    Examples:
        .. code-block:: python
Y
Yu Yang 已提交
233

C
chengduoZH 已提交
234
          dict_size = len(dataset.ids)
235
          data = fluid.layers.data(name='ids', shape=[32, 32], dtype='float32')
C
chengduoZH 已提交
236
          fc = fluid.layers.embedding(input=data, size=[dict_size, 16])
Y
Yu Yang 已提交
237 238 239 240 241 242
    """

    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)
243 244
    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
245 246 247 248 249
    helper.append_op(
        type='lookup_table',
        inputs={'Ids': input,
                'W': w},
        outputs={'Out': tmp},
250 251
        attrs={'is_sparse': is_sparse,
               'padding_idx': padding_idx})
Y
Yu Yang 已提交
252 253 254 255 256 257 258 259 260 261 262 263 264
    return tmp


# TODO(qijun): expose H0 and C0
def dynamic_lstm(input,
                 size,
                 param_attr=None,
                 bias_attr=None,
                 use_peepholes=True,
                 is_reverse=False,
                 gate_activation='sigmoid',
                 cell_activation='tanh',
                 candidate_activation='tanh',
265 266
                 dtype='float32',
                 name=None):
Y
Yibing Liu 已提交
267 268 269 270 271 272
    """
    **Dynamic LSTM Layer**

    The defalut implementation is diagonal/peephole connection
    (https://arxiv.org/pdf/1402.1128.pdf), the formula is as follows:

Y
Yibing Liu 已提交
273
    .. math::
Y
Yibing Liu 已提交
274

275
        i_t & = \sigma(W_{ix}x_{t} + W_{ih}h_{t-1} + W_{ic}c_{t-1} + b_i)
Y
Yibing Liu 已提交
276

277
        f_t & = \sigma(W_{fx}x_{t} + W_{fh}h_{t-1} + W_{fc}c_{t-1} + b_f)
Y
Yibing Liu 已提交
278

279
        \\tilde{c_t} & = act_g(W_{cx}x_t + W_{ch}h_{t-1} + b_c)
Y
Yibing Liu 已提交
280

281 282 283
        o_t & = \sigma(W_{ox}x_{t} + W_{oh}h_{t-1} + W_{oc}c_t + b_o)

        c_t & = f_t \odot c_{t-1} + i_t \odot \\tilde{c_t}
Y
Yibing Liu 已提交
284

Y
Yibing Liu 已提交
285
        h_t & = o_t \odot act_h(c_t)
Y
Yibing Liu 已提交
286

287
    where the :math:`W` terms denote weight matrices (e.g. :math:`W_{xi}` is
288
    the matrix of weights from the input gate to the input), :math:`W_{ic}, \
289 290 291
    W_{fc}, W_{oc}` are diagonal weight matrices for peephole connections. In
    our implementation, we use vectors to reprenset these diagonal weight
    matrices. The :math:`b` terms denote bias vectors (:math:`b_i` is the input
Y
Yibing Liu 已提交
292
    gate bias vector), :math:`\sigma` is the non-linear activations, such as
293 294
    logistic sigmoid function, and :math:`i, f, o` and :math:`c` are the input
    gate, forget gate, output gate, and cell activation vectors, respectively,
295 296
    all of which have the same size as the cell output activation vector :math:`h`.

297 298 299 300
    The :math:`\odot` is the element-wise product of the vectors. :math:`act_g`
    and :math:`act_h` are the cell input and cell output activation functions
    and `tanh` is usually used for them. :math:`\\tilde{c_t}` is also called
    candidate hidden state, which is computed based on the current input and
301 302 303
    the previous hidden state.

    Set `use_peepholes` to `False` to disable peephole connection. The formula
Y
Yibing Liu 已提交
304 305 306
    is omitted here, please refer to the paper
    http://www.bioinf.jku.at/publications/older/2604.pdf for details.

Y
Yibing Liu 已提交
307 308 309
    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-connect layer before LSTM layer.
Y
Yibing Liu 已提交
310 311

    Args:
312 313 314 315
        input(Variable): The input of dynamic_lstm 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
Y
Yibing Liu 已提交
316 317
                         mini-batch, D is the hidden size.
        size(int): 4 * hidden size.
318
        param_attr(ParamAttr|None): The parameter attribute for the learnable
319
                               hidden-hidden weights.
Y
Yibing Liu 已提交
320 321 322

                               - Weights = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}
323 324 325
                               - The shape is (D x 4D), where D is the hidden
                                 size.
        bias_attr(ParamAttr|None): The bias attribute for the learnable bias
326 327 328
                              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
329

330
                              1. `use_peepholes = False`
Y
Yibing Liu 已提交
331
                                - Biases = {:math:`b_c, b_i, b_f, b_o`}.
332
                                - The shape is (1 x 4D).
333
                              2. `use_peepholes = True`
Y
Yibing Liu 已提交
334 335
                                - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
                                                 W_{fc}, W_{oc}`}.
336
                                - The shape is (1 x 7D).
337
        use_peepholes(bool): Whether to enable diagonal/peephole connections,
Y
Yibing Liu 已提交
338 339
                             default `True`.
        is_reverse(bool): Whether to compute reversed LSTM, default `False`.
340 341
        gate_activation(str): The activation for input gate, forget gate and
                              output gate. Choices = ["sigmoid", "tanh", "relu",
Y
Yibing Liu 已提交
342
                              "identity"], default "sigmoid".
343
        cell_activation(str): The activation for cell output. Choices = ["sigmoid",
Y
Yibing Liu 已提交
344 345 346 347 348
                              "tanh", "relu", "identity"], default "tanh".
        candidate_activation(str): The activation for candidate hidden state.
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
                              default "tanh".
        dtype(str): Data type. Choices = ["float32", "float64"], default "float32".
349 350
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
Y
Yibing Liu 已提交
351 352

    Returns:
Y
Yibing Liu 已提交
353 354
        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
355

Y
Yibing Liu 已提交
356
    Examples:
Y
Yibing Liu 已提交
357 358
        .. code-block:: python

Y
Yibing Liu 已提交
359 360
            hidden_dim = 512
            forward_proj = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
361
                                           act=None, bias_attr=None)
Y
Yibing Liu 已提交
362 363
            forward, _ = fluid.layers.dynamic_lstm(
                input=forward_proj, size=hidden_dim * 4, use_peepholes=False)
Y
Yibing Liu 已提交
364
    """
365

Y
Yu Yang 已提交
366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401
    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)

    helper.append_op(
        type='lstm',
        inputs={'Input': input,
                'Weight': weight,
                'Bias': bias},
        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
402 403 404 405 406 407 408 409 410 411 412
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',
413 414
                  dtype='float32',
                  name=None):
Y
Yibing Liu 已提交
415 416 417
    """
    **Dynamic LSTMP Layer**

418 419 420 421 422 423
    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
424 425 426 427 428

    The formula is as follows:

    .. math::

429
        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
430

431
        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
432

433
        \\tilde{c_t} & = act_g(W_{cx}x_t + W_{cr}r_{t-1} + b_c)
Y
Yibing Liu 已提交
434

435
        o_t & = \sigma(W_{ox}x_{t} + W_{or}r_{t-1} + W_{oc}c_t + b_o)
Y
Yibing Liu 已提交
436

437
        c_t & = f_t \odot c_{t-1} + i_t \odot \\tilde{c_t}
Y
Yibing Liu 已提交
438

439
        h_t & = o_t \odot act_h(c_t)
Y
Yibing Liu 已提交
440

441
        r_t & = \overline{act_h}(W_{rh}h_t)
Y
Yibing Liu 已提交
442

Y
Yibing Liu 已提交
443 444 445 446 447 448
    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, \
449
          we use vectors to reprenset these diagonal weight matrices.
Y
Yibing Liu 已提交
450
    * :math:`b`: Denotes bias vectors (e.g. :math:`b_i` is the input gate \
451
          bias vector).
Y
Yibing Liu 已提交
452 453 454
    * :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 \
455
          the same size as the cell output activation vector :math:`h`.
Y
Yibing Liu 已提交
456
    * :math:`h`: The hidden state.
457
    * :math:`r`: The recurrent projection of the hidden state.
Y
Yibing Liu 已提交
458 459
    * :math:`\\tilde{c_t}`: The candidate hidden state, whose \
          computation is based on the current input and previous hidden state.
460
    * :math:`\odot`: The element-wise product of the vectors.
Y
Yibing Liu 已提交
461
    * :math:`act_g` and :math:`act_h`: The cell input and cell output \
462
          activation functions and `tanh` is usually used for them.
Y
Yibing Liu 已提交
463 464
    * :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
465 466 467 468

    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.
469

Y
Yibing Liu 已提交
470 471 472 473 474 475 476 477 478 479 480 481
    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.
482
        param_attr(ParamAttr|None): The parameter attribute for the learnable
Y
Yibing Liu 已提交
483 484
                               hidden-hidden weight and projection weight.

485 486
                               - Hidden-hidden weight = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}.
487 488
                               - 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
489 490
                                 size.
                               - Projection weight = {:math:`W_{rh}`}.
491 492
                               - 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
493 494 495 496 497 498
                              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`}.
499
                                - The shape is (1 x 4D).
Y
Yibing Liu 已提交
500 501 502
                              2. `use_peepholes = True`
                                - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
                                                 W_{fc}, W_{oc}`}.
503
                                - The shape is (1 x 7D).
Y
Yibing Liu 已提交
504 505 506 507 508 509 510 511 512 513 514 515 516 517 518
        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.
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
                              default "tanh".
        proj_activation(str): The activation for projection output.
                              Choices = ["sigmoid", "tanh", "relu", "identity"],
                              default "tanh".
        dtype(str): Data type. Choices = ["float32", "float64"], default "float32".
519 520
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
Y
Yibing Liu 已提交
521 522

    Returns:
523 524
        tuple: The projection of hidden state, and cell state of LSTMP. The \
               shape of projection is (T x P), for the cell state which is \
Y
Yibing Liu 已提交
525 526 527 528 529
               (T x D), and both LoD is the same with the `input`.

    Examples:
        .. code-block:: python

Y
Yibing Liu 已提交
530
            hidden_dim, proj_dim = 512, 256
Y
Yibing Liu 已提交
531 532
            fc_out = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
                                     act=None, bias_attr=None)
533 534 535
            proj_out, _ = fluid.layers.dynamic_lstmp(input=fc_out,
                                                     size=hidden_dim * 4,
                                                     proj_size=proj_dim,
Y
Yibing Liu 已提交
536 537 538 539
                                                     use_peepholes=False,
                                                     is_reverse=True,
                                                     cell_activation="tanh",
                                                     proj_activation="tanh")
Y
Yibing Liu 已提交
540
    """
541

Y
Yibing Liu 已提交
542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587
    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
588 589 590 591 592 593 594 595 596 597 598
def dynamic_gru(input,
                size,
                param_attr=None,
                bias_attr=None,
                is_reverse=False,
                gate_activation='sigmoid',
                candidate_activation='tanh',
                h_0=None):
    """
    **Dynamic GRU Layer**

599
    Refer to `Empirical Evaluation of Gated Recurrent Neural Networks on
G
guosheng 已提交
600
    Sequence Modeling <https://arxiv.org/abs/1412.3555>`_
601

G
guosheng 已提交
602 603 604 605 606 607 608 609 610
    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)
611

G
guosheng 已提交
612
        h_t & = (1-u_t) \odot h_{t-1} + u_t \odot \\tilde{h_t}
613

G
guosheng 已提交
614
    The :math:`\odot` is the element-wise product of the vectors. :math:`act_g`
615 616
    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
617 618 619 620
    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
621
    to use fully-connect layer before GRU layer.
G
guosheng 已提交
622 623

    Args:
624 625
        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
626
            Variable is a matrix with shape :math:`(T \\times 3D)`, where
627
            :math:`T` is the total time steps in this mini-batch, :math:`D`
G
guosheng 已提交
628 629
            is the hidden size.
        size(int): The dimension of the gru cell.
630
        param_attr(ParamAttr|None): The parameter attribute for the learnable
G
guosheng 已提交
631 632
            hidden-hidden weight matrix. Note:

633
            - The shape of the weight matrix is :math:`(T \\times 3D)`, where
G
guosheng 已提交
634
              :math:`D` is the hidden size.
635
            - All elements in the weight matrix can be divided into two parts.
G
guosheng 已提交
636
              The first part are weights of the update gate and reset gate with
637
              shape :math:`(D \\times 2D)`, and the second part are weights for
G
guosheng 已提交
638
              candidate hidden state with shape :math:`(D \\times D)`.
639
        bias_attr(ParamAttr): The parameter attribute for learnable the
G
guosheng 已提交
640
            hidden-hidden bias.
641
        is_reverse(bool): Whether to compute reversed GRU, default
G
guosheng 已提交
642 643 644
            :attr:`False`.
        gate_activation(str): The activation for update gate and reset gate.
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "sigmoid".
645
        activation(str): The activation for candidate hidden state.
G
guosheng 已提交
646 647 648
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh".

    Returns:
G
guosheng 已提交
649 650
        Variable: The hidden state of GRU. The shape is :math:`(T \\times D)`, \
            and lod is the same with the input.
651

G
guosheng 已提交
652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694
    Examples:
        .. code-block:: python

            hidden_dim = 512
            x = fluid.layers.fc(input=data, size=hidden_dim * 3)
            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)
    inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
    if h_0 != None:
        assert h_0.shape == (
            size, size), 'The shape of h0 should be(%d, %d)' % (size, size)
        inputs['h0'] = h_0

    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
695 696 697 698 699 700
def gru_unit(input,
             hidden,
             size,
             weight=None,
             bias=None,
             activation='tanh',
701
             gate_activation='sigmoid'):
Y
Yu Yang 已提交
702
    """
703
    GRU unit layer. The equation of a gru step is:
Y
Yu Yang 已提交
704

705 706
        .. math::
            u_t & = actGate(xu_{t} + W_u h_{t-1} + b_u)
Y
Yu Yang 已提交
707

708
            r_t & = actGate(xr_{t} + W_r h_{t-1} + b_r)
Y
Yu Yang 已提交
709

710
            m_t & = actNode(xm_t + W_c dot(r_t, h_{t-1}) + b_m)
711

712
            h_t & = dot((1-u_t), m_t) + dot(u_t, h_{t-1})
713 714

    The inputs of gru unit includes :math:`z_t`, :math:`h_{t-1}`. In terms
715 716 717
    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
718 719
    connected layer has to be applied, such that :math:`z_t = W_{fc}x_t`.

720 721
    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
722 723 724
    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`.
725 726 727 728 729 730 731

    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.
        weight (ParamAttr): The weight parameters for gru unit. Default: None
        bias (ParamAttr): The bias parameters for gru unit. Default: None
732 733 734 735
        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
736

737 738 739 740 741 742
    Returns:
        tuple: The hidden value, reset-hidden value and gate values.

    Examples:

        .. code-block:: python
Y
Yu Yang 已提交
743

744
             # assuming we have x_t_data and prev_hidden of size=10
745
             x_t = fluid.layers.fc(input=x_t_data, size=30)
746 747
             hidden_val, r_h_val, gate_val = fluid.layers.gru_unit(input=x_t,
                                                    hidden = prev_hidden)
Y
Yu Yang 已提交
748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767

    """
    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
    if weight is None:
        weight = helper.create_parameter(
            attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype)

    # create bias
Y
Yibing Liu 已提交
768

Y
Yu Yang 已提交
769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795
    if bias is None:
        bias_size = [1, 3 * size]
        bias = helper.create_parameter(
            attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)

    gate = helper.create_tmp_variable(dtype)
    reset_hidden_pre = helper.create_tmp_variable(dtype)
    updated_hidden = helper.create_tmp_variable(dtype)

    helper.append_op(
        type='gru_unit',
        inputs={'Input': input,
                'HiddenPrev': hidden,
                'Weight': weight},
        outputs={
            'Gate': gate,
            'ResetHiddenPrev': reset_hidden_pre,
            'Hidden': updated_hidden,
        },
        attrs={
            'activation': 0,
            'gate_activation': 1,
        })

    return updated_hidden, reset_hidden_pre, gate


796
def linear_chain_crf(input, label, param_attr=None):
Y
Yu Yang 已提交
797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821
    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


822
def crf_decoding(input, param_attr, label=None):
Y
Yu Yang 已提交
823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854
    helper = LayerHelper('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


def cos_sim(X, Y, **kwargs):
    """
    This function performs the cosine similarity between two tensors
    X and Y and returns that as the output.
    """
    helper = LayerHelper('cos_sim', **kwargs)
    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


855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883
def dropout(x, dropout_prob, is_test=False, seed=None, **kwargs):
    """
    Computes dropout.

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

    Args:
       x(variable): The input tensor.
       dropout_prob(float): Probability of setting units to zero.
       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.

    Returns:
        Variable: A tensor variable.

    Examples:
        .. code-block:: python

          x = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
          droped = fluid.layers.dropout(input=x, dropout_rate=0.5)
    """

884 885 886 887 888 889 890 891
    helper = LayerHelper('dropout', **kwargs)
    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]},
892 893 894 895 896 897
        attrs={
            'dropout_prob': dropout_prob,
            'is_test': is_test,
            'fix_seed': seed is not None,
            'seed': seed if seed is not None else 0
        })
898 899 900
    return out


Y
Yu Yang 已提交
901 902
def cross_entropy(input, label, **kwargs):
    """
Y
Yibing Liu 已提交
903 904
    **Cross Entropy Layer**

905 906 907
    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
908 909

    1) One-hot cross-entropy:
Y
Yibing Liu 已提交
910
	`soft_label = False`, `Label[i, 0]` indicates the class index for sample i:
Y
yangyaming 已提交
911

Y
Yibing Liu 已提交
912
        .. math::
Y
yangyaming 已提交
913

Y
Yibing Liu 已提交
914 915 916
            Y[i] = -\log(X[i, Label[i]])

    2) Soft-label cross-entropy:
Y
Yibing Liu 已提交
917
	`soft_label = True`, `Label[i, j]` indicates the soft label of class j
Y
Yibing Liu 已提交
918 919 920 921 922 923
	for sample i:

        .. math::

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

Y
Yibing Liu 已提交
924
       Please make sure that in this case the summation of each row of `label`
Y
Yibing Liu 已提交
925 926 927 928
       equals one.

    3) One-hot cross-entropy with vecterized `label`:
	 As a special case of 2), when each row of 'label' has only one
Y
Yibing Liu 已提交
929 930
	 non-zero element which is equal to 1, soft-label cross-entropy degenerates
         to a one-hot cross-entropy with one-hot label representation.
Y
yangyaming 已提交
931

Y
Yibing Liu 已提交
932
    Args:
Y
yangyaming 已提交
933
        input (Variable|list):  a 2-D tensor with shape [N x D], where N is the
934 935 936 937
                                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
938
        label (Variable|list): the ground truth which is a 2-D tensor. When
939 940 941 942 943 944 945
                               `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].
        soft_label (bool, via `**kwargs`): a flag indicating whether to
                                           interpretate the given labels as soft
                                           labels, default `False`.
Y
Yibing Liu 已提交
946 947 948 949 950

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

    Raises:
951 952 953 954 955
        `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
956 957 958 959 960 961

    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
962 963 964 965 966 967 968 969 970 971 972 973 974 975
    """
    helper = LayerHelper('cross_entropy', **kwargs)
    out = helper.create_tmp_variable(dtype=input.dtype)
    helper.append_op(
        type='cross_entropy',
        inputs={'X': [input],
                'Label': [label]},
        outputs={'Y': [out]},
        attrs=kwargs)
    return out


def square_error_cost(input, label, **kwargs):
    """
976 977
    **Square error cost layer**

978 979
    This layer accepts input predictions and target label and returns the
    squared error cost.
Y
ying 已提交
980

981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997
    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:
       input(Variable): Input tensor, has predictions.
       label(Variable): Label tensor, has target labels.

    Returns:
G
guosheng 已提交
998
        Variable: The tensor variable storing the element-wise squared error \
999
                  difference of input and label.
1000 1001 1002 1003 1004 1005 1006 1007

    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
1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018
    """
    helper = LayerHelper('square_error_cost', **kwargs)
    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
1019 1020
        type='square', inputs={'X': [minus_out]},
        outputs={'Out': [square_out]})
Y
Yu Yang 已提交
1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064
    return square_out


def accuracy(input, label, k=1, correct=None, total=None, **kwargs):
    """
    This function computes the accuracy using the input and label.
    The output is the top_k inputs and their indices.
    """
    helper = LayerHelper("accuracy", **kwargs)
    topk_out = helper.create_tmp_variable(dtype=input.dtype)
    topk_indices = helper.create_tmp_variable(dtype="int64")
    helper.append_op(
        type="top_k",
        inputs={"X": [input]},
        outputs={"Out": [topk_out],
                 "Indices": [topk_indices]},
        attrs={"k": k})
    acc_out = helper.create_tmp_variable(dtype="float32")
    if correct is None:
        correct = helper.create_tmp_variable(dtype="int64")
    if total is None:
        total = helper.create_tmp_variable(dtype="int64")
    helper.append_op(
        type="accuracy",
        inputs={
            "Out": [topk_out],
            "Indices": [topk_indices],
            "Label": [label]
        },
        outputs={
            "Accuracy": [acc_out],
            "Correct": [correct],
            "Total": [total],
        })
    return acc_out


def chunk_eval(input,
               label,
               chunk_scheme,
               num_chunk_types,
               excluded_chunk_types=None,
               **kwargs):
    """
Y
yangyaming 已提交
1065
    This function computes and outputs the precision, recall and
1066
    F1-score of chunk detection.
Y
Yu Yang 已提交
1067 1068 1069 1070 1071 1072 1073
    """
    helper = LayerHelper("chunk_eval", **kwargs)

    # prepare output
    precision = helper.create_tmp_variable(dtype="float32")
    recall = helper.create_tmp_variable(dtype="float32")
    f1_score = helper.create_tmp_variable(dtype="float32")
1074 1075 1076
    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
1077 1078 1079 1080 1081 1082 1083 1084

    helper.append_op(
        type="chunk_eval",
        inputs={"Inference": [input],
                "Label": [label]},
        outputs={
            "Precision": [precision],
            "Recall": [recall],
1085 1086 1087 1088
            "F1-Score": [f1_score],
            "NumInferChunks": [num_infer_chunks],
            "NumLabelChunks": [num_label_chunks],
            "NumCorrectChunks": [num_correct_chunks]
Y
Yu Yang 已提交
1089 1090 1091
        },
        attrs={
            "num_chunk_types": num_chunk_types,
G
guosheng 已提交
1092 1093
            "chunk_scheme": chunk_scheme,
            "excluded_chunk_types": excluded_chunk_types or []
Y
Yu Yang 已提交
1094
        })
1095 1096
    return (precision, recall, f1_score, num_infer_chunks, num_label_chunks,
            num_correct_chunks)
Y
Yu Yang 已提交
1097 1098 1099 1100 1101 1102 1103 1104 1105


def sequence_conv(input,
                  num_filters,
                  filter_size=3,
                  filter_stride=1,
                  padding=None,
                  bias_attr=None,
                  param_attr=None,
1106
                  act=None):
Y
Yu Yang 已提交
1107 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 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147
    """
    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.
    """

    # FIXME(dzh) : want to unify the argument of python layer
    # function. So we ignore some unecessary attributes.
    # such as, padding_trainable, context_start.

    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)


def conv2d(input,
           num_filters,
           filter_size,
           stride=None,
           padding=None,
           groups=None,
           param_attr=None,
           bias_attr=None,
C
chengduoZH 已提交
1148
           use_cudnn=True,
C
chengduoZH 已提交
1149
           act=None):
Y
Yu Yang 已提交
1150
    """
C
chengduoZH 已提交
1151 1152 1153
    **Convlution2D Layer**

    The convolution2D layer calculates the output based on the input, filter
1154 1155 1156
    and strides, paddings, dilations, groups parameters. 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.
C
chengduoZH 已提交
1157 1158
    The details of convolution layer, please refer UFLDL's `convolution,
    <http://ufldl.stanford.edu/tutorial/supervised/FeatureExtractionUsingConvolution/>`_ .
1159 1160 1161
    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
1162

1163
    For each input :math:`X`, the equation is:
C
refine  
chengduoZH 已提交
1164

C
chengduoZH 已提交
1165 1166
    .. math::

C
refine  
chengduoZH 已提交
1167
        Out = \sigma (W \\ast X + b)
C
chengduoZH 已提交
1168

C
chengduoZH 已提交
1169
    In the above equation:
C
chengduoZH 已提交
1170

1171 1172 1173 1174 1175
    * :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.
1176 1177
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be
                   different.
C
chengduoZH 已提交
1178 1179 1180

    Example:

1181 1182 1183
        - Input:

          Input shape: $(N, C_{in}, H_{in}, W_{in})$
C
refine  
chengduoZH 已提交
1184

1185
          Filter shape: $(C_{out}, C_{in}, H_f, W_f)$
C
refine  
chengduoZH 已提交
1186

1187 1188
        - Output:
          Output shape: $(N, C_{out}, H_{out}, W_{out})$
C
refine  
chengduoZH 已提交
1189

C
chengduoZH 已提交
1190
        Where
1191 1192

        .. math::
C
chengduoZH 已提交
1193

C
chengduoZH 已提交
1194 1195
        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
1196 1197

    Args:
1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219
       input(Variable): The input image with [N, C, 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,
           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.
       groups(int): The groups number of the Conv2d Layer. According to grouped
           convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
           the first half of the filters is only connected to the first half
           of the input channels, while the second half of the filters is only
           connected to the second half of the input channels. Default: groups=1
       param_attr(ParamAttr): The parameters to the Conv2d Layer. Default: None
       bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None
       use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
           library is installed. Default: True
       act(str): Activation type. Default: None
C
chengduoZH 已提交
1220 1221

    Returns:
G
guosheng 已提交
1222
        Variable: The tensor variable storing the convolution and \
C
chengduoZH 已提交
1223 1224
                  non-linearity activation result.

C
refine  
chengduoZH 已提交
1225
    Raises:
1226 1227
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
C
refine  
chengduoZH 已提交
1228

C
chengduoZH 已提交
1229 1230 1231
    Examples:
        .. code-block:: python

1232 1233 1234 1235
          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
1236 1237 1238 1239 1240
    """
    if stride is None:
        stride = [1, 1]

    num_channels = input.shape[1]
1241 1242

    l_type = 'conv2d'
X
xzl 已提交
1243 1244
    if (num_channels == groups and num_filters % num_channels == 0 and
            not use_cudnn):
1245
        l_type = 'depthwise_conv2d'
1246 1247 1248 1249

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

Y
Yu Yang 已提交
1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262
    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

    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]
C
chengduoZH 已提交
1263 1264
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281

    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(
1282
        type=l_type,
Y
Yu Yang 已提交
1283 1284 1285 1286 1287
        inputs={
            'Input': input,
            'Filter': filter_param,
        },
        outputs={"Output": pre_bias},
C
chengduoZH 已提交
1288 1289 1290 1291 1292 1293
        attrs={
            'strides': stride,
            'paddings': padding,
            'groups': groups,
            'use_cudnn': use_cudnn
        })
Y
Yu Yang 已提交
1294 1295 1296 1297 1298 1299 1300 1301

    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)

    return helper.append_activation(pre_act)


def sequence_pool(input, pool_type, **kwargs):
    """
Y
yangyaming 已提交
1302 1303 1304
    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
1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329

    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:
         x.lod = [[0, 2, 5, 7]]
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
         with condition len(x.lod[-1]) - 1 == out.dims[0]

       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)
F
fengjiayi 已提交
1330

L
Luo Tao 已提交
1331 1332
    Args:
        input(variable): The input variable which is a LoDTensor.
Y
yangyaming 已提交
1333
        pool_type (string): The pooling type of sequence_pool.
L
Luo Tao 已提交
1334 1335 1336 1337 1338 1339 1340 1341
            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
1342

Y
yangyaming 已提交
1343
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1344 1345 1346 1347 1348
                              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')
Y
Yu Yang 已提交
1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361
    """
    helper = LayerHelper('sequence_pool', input=input, **kwargs)
    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
1362 1363 1364 1365 1366
    # 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
1367 1368 1369
    return pool_out


1370
def sequence_first_step(input, **kwargs):
L
Luo Tao 已提交
1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384
    """
    This funciton get the first step of sequence.

    .. code-block:: text

       x is a 1-level LoDTensor:
         x.lod = [[0, 2, 5, 7]]
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
         with condition len(x.lod[-1]) - 1 == out.dims[0]
         out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1)
F
fengjiayi 已提交
1385

L
Luo Tao 已提交
1386 1387 1388 1389 1390 1391 1392 1393 1394
    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
1395

Y
yangyaming 已提交
1396
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1397 1398 1399
                              dtype='float32', lod_level=1)
             x_first_step = fluid.layers.sequence_first_step(input=x)
    """
1400 1401 1402 1403
    return sequence_pool(input=input, pool_type="first")


def sequence_last_step(input, **kwargs):
L
Luo Tao 已提交
1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417
    """
    This funciton get the last step of sequence.

    .. code-block:: text

       x is a 1-level LoDTensor:
         x.lod = [[0, 2, 5, 7]]
         x.data = [1, 3, 2, 4, 6, 5, 1]
         x.dims = [7, 1]

       then output is a Tensor:
         out.dim = [3, 1]
         with condition len(x.lod[-1]) - 1 == out.dims[0]
         out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1)
F
fengjiayi 已提交
1418

L
Luo Tao 已提交
1419 1420 1421 1422 1423 1424 1425 1426 1427
    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
1428

Y
yangyaming 已提交
1429
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1430 1431 1432
                              dtype='float32', lod_level=1)
             x_last_step = fluid.layers.sequence_last_step(input=x)
    """
1433 1434 1435
    return sequence_pool(input=input, pool_type="last")


Y
Yu Yang 已提交
1436 1437 1438 1439 1440
def pool2d(input,
           pool_size,
           pool_type,
           pool_stride=None,
           pool_padding=None,
C
caoying03 已提交
1441
           global_pooling=False,
C
chengduoZH 已提交
1442
           use_cudnn=True,
C
caoying03 已提交
1443
           name=None):
Y
Yu Yang 已提交
1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461
    """
    This function adds the operator for pooling in 2 dimensions, using the
    pooling configurations mentioned in input parameters.
    """
    if pool_padding is None:
        pool_padding = [0, 0]
    if pool_stride is None:
        pool_stride = [1, 1]
    if pool_type not in ["max", "avg"]:
        raise ValueError(
            "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",
            str(pool_type))
    if isinstance(pool_size, int):
        pool_size = [pool_size, pool_size]
    if isinstance(pool_stride, int):
        pool_stride = [pool_stride, pool_stride]
    if isinstance(pool_padding, int):
        pool_padding = [pool_padding, pool_padding]
C
chengduoZH 已提交
1462 1463
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477

    helper = LayerHelper('pool2d', **locals())
    dtype = helper.input_dtype()
    pool_out = helper.create_tmp_variable(dtype)

    helper.append_op(
        type="pool2d",
        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
1478 1479
            "paddings": pool_padding,
            "use_cudnn": use_cudnn
Y
Yu Yang 已提交
1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491
        })

    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
1492
               data_layout='NCHW',
1493 1494 1495
               name=None,
               moving_mean_name=None,
               moving_variance_name=None):
Y
Yu Yang 已提交
1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521
    """
    This function helps create an operator to implement
    the BatchNorm layer using the configurations from the input parameters.
    """
    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(
1522
        attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
Y
Yu Yang 已提交
1523 1524

    mean = helper.create_global_variable(
1525
        name=moving_mean_name,
Q
QI JUN 已提交
1526 1527 1528 1529
        dtype=input.dtype,
        shape=param_shape,
        persistable=True,
        stop_gradient=True)
Y
Yu Yang 已提交
1530 1531 1532
    helper.set_variable_initializer(var=mean, initializer=Constant(0.0))

    variance = helper.create_global_variable(
1533
        name=moving_variance_name,
Q
QI JUN 已提交
1534 1535 1536 1537
        dtype=input.dtype,
        shape=param_shape,
        persistable=True,
        stop_gradient=True)
Y
Yu Yang 已提交
1538 1539 1540 1541 1542 1543 1544
    helper.set_variable_initializer(var=variance, initializer=Constant(1.0))

    # 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
1545 1546
    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
1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572

    batch_norm_out = helper.create_tmp_variable(dtype)

    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
        },
        attrs={"momentum": momentum,
               "epsilon": epsilon,
               "is_test": is_test})

    return helper.append_activation(batch_norm_out)


G
guosheng 已提交
1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 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 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643
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):
    """
    **Layer Normalization**

    Assume feature vectors exist on dimensions 
    :attr:`begin_norm_axis ... rank(input)` and calculate the moment statistics
    along these dimensions for each feature vector :math:`a` with size
    :math:`H`, then normalize each feature vector using the corresponding
    statistics. After that, apply learnable gain and bias on the normalized
    tensor to scale and shift if :attr:`scale` and :attr:`shift` are set.

    Refer to `Layer Normalization <https://arxiv.org/pdf/1607.06450v1.pdf>`_

    The formula is as follows:

    .. math::

        \\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)

    Args:
        input(Variable): The input tensor variable.
        scale(bool): Whether to learn the adaptive gain :math:`g` after 
            normalization.
        shift(bool): Whether to learn the adaptive bias :math:`b` after 
            normalization.
        begin_norm_axis(bool): The normalization will be performed along 
            dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`.
        epsilon(float): The small value added to the variance to prevent 
            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.

    Returns:
        Variable: A tensor variable with the same shape as the input.

    Examples:
        .. code-block:: python

            data = fluid.layers.data(
              name='data', shape=[3, 32, 32], dtype='float32')
            x = fluid.layers.layer_norm(input=data, begin_norm_axis=1)
    """
    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
1644
    if shift:
G
guosheng 已提交
1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668
        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)


C
caoying03 已提交
1669
def beam_search_decode(ids, scores, name=None):
Y
Yu Yang 已提交
1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691
    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
        })

    return sentence_ids, sentence_scores


def conv2d_transpose(input,
                     num_filters,
                     output_size=None,
                     filter_size=None,
                     padding=None,
                     stride=None,
C
chengduoZH 已提交
1692
                     dilation=None,
C
caoying03 已提交
1693
                     param_attr=None,
C
chengduoZH 已提交
1694
                     use_cudnn=True,
C
caoying03 已提交
1695
                     name=None):
Y
Yu Yang 已提交
1696
    """
1697 1698 1699 1700 1701 1702 1703 1704
    **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
1705 1706
    layer, please refer to the following explanation and references
    `therein <http://www.matthewzeiler.com/wp-content/uploads/2017/07/cvpr2010.pdf>`_.
1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718

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

    .. math::

        Out = W \\ast X

    In the above equation:

    * :math:`X`: Input value, a tensor with NCHW format.
    * :math:`W`: Filter value, a tensor with MCHW format.
    * :math:`\\ast` : Convolution transpose operation.
1719 1720
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be
                   different.
Y
Yu Yang 已提交
1721

1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734
    Example:

        - Input:

          Input shape: $(N, C_{in}, H_{in}, W_{in})$

          Filter shape: $(C_{in}, C_{out}, H_f, W_f)$

        - Output:

          Output shape: $(N, C_{out}, H_{out}, W_{out})$

        Where
Y
Yu Yang 已提交
1735

1736 1737 1738 1739
        .. 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
1740 1741

    Args:
1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760
       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.
1761 1762
       param_attr(ParamAttr): The parameters to the Conv2d_transpose Layer.
                              Default: None
1763 1764 1765 1766
       use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
           library is installed. Default: True
       name(str|None): A name for this layer(optional). If set None, the layer
           will be named automatically.
Y
Yu Yang 已提交
1767 1768

    Returns:
1769 1770 1771
       Variable: The tensor variable storing the convolution transpose result.

    Raises:
1772 1773
       ValueError: If the shapes of input, filter_size, stride, padding and
                   groups mismatch.
1774 1775 1776 1777

    Examples:
       .. code-block:: python

1778 1779 1780 1781
          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
1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795
    """
    helper = LayerHelper("conv2d_transpose", **locals())
    if not isinstance(input, Variable):
        raise TypeError("Input of conv2d_transpose must be Variable")
    input_channel = input.shape[1]

    op_attr = dict()

    if isinstance(padding, int):
        op_attr['paddings'] = [padding, padding]
    elif padding is not None:
        op_attr['paddings'] = padding

    if isinstance(stride, int):
C
chengduoZH 已提交
1796
        op_attr['strides'] = [stride, stride]
Y
Yu Yang 已提交
1797 1798 1799
    elif stride is not None:
        op_attr['strides'] = stride

C
chengduoZH 已提交
1800 1801 1802 1803 1804
    if isinstance(dilation, int):
        op_attr['dilations'] = [dilation, dilation]
    elif dilation is not None:
        op_attr['dilations'] = dilation

C
chengduoZH 已提交
1805 1806 1807 1808
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
    op_attr['use_cudnn'] = use_cudnn

Y
Yu Yang 已提交
1809 1810 1811 1812 1813 1814 1815 1816
    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]

        padding = op_attr.get('paddings', [0, 0])
        stride = op_attr.get('strides', [1, 1])
C
chengduoZH 已提交
1817
        dilation = op_attr.get('dilations', [1, 1])
Y
Yu Yang 已提交
1818 1819 1820

        h_in = input.shape[2]
        w_in = input.shape[3]
C
chengduoZH 已提交
1821 1822 1823 1824 1825

        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
1826
        filter_size = [filter_size_h, filter_size_w]
C
chengduoZH 已提交
1827

Y
Yu Yang 已提交
1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843
    elif isinstance(filter_size, int):
        filter_size = [filter_size, filter_size]

    filter_shape = [input_channel, num_filters] + filter_size
    img_filter = helper.create_parameter(
        dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)

    out = helper.create_tmp_variable(dtype=input.dtype)
    helper.append_op(
        type='conv2d_transpose',
        inputs={'Input': [input],
                'Filter': [img_filter]},
        outputs={'Output': out},
        attrs=op_attr)

    return out
Y
yangyaming 已提交
1844 1845


C
caoying03 已提交
1846
def sequence_expand(x, y, name=None):
1847 1848
    """Sequence Expand Layer. This layer will expand the input variable **x**
    according to LoD information of **y**. And the following examples will
Y
yangyaming 已提交
1849
    explain how sequence_expand works:
1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877

    .. code-block:: text

        * Case 1
            x is a LoDTensor:
                x.lod = [[0,       2, 3],
                         [0, 1,    3, 4]]
                x.data = [a, b, c, d]
                x.dims = [4, 1]

            y is a LoDTensor:
                y.lod = [[0,    2,    4],
                         [0, 3, 6, 7, 8]]

            with condition len(y.lod[-1]) - 1 == x.dims[0]

            then output is a 2-level LoDTensor:
                out.lod = [[0,                2,    4],
                           [0,       3,       6, 7, 8]]
                out.data = [a, a, a, b, b, b, c, d]
                out.dims = [8, 1]

        * Case 2
            x is a Tensor:
                x.data = [a, b, c]
                x.dims = [3, 1]

            y is a LoDTensor:
Y
yangyaming 已提交
1878
                y.lod = [[0, 2, 3, 6]]
1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889

            with condition len(y.lod[-1]) - 1 == x.dims[0]

            then output is a 1-level LoDTensor:
                out.lod = [[0,    2, 3,      6]]
                out.data = [a, a, b, c, c, c]
                out.dims = [6, 1]

    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
        y (Variable): The input variable which is a LoDTensor.
C
caoying03 已提交
1890 1891
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
1892 1893 1894 1895 1896 1897 1898 1899 1900 1901

    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
1902
            out = layers.sequence_expand(x=x, y=y)
1903
    """
Y
yangyaming 已提交
1904
    helper = LayerHelper('sequence_expand', input=x, **locals())
1905 1906 1907
    dtype = helper.input_dtype()
    tmp = helper.create_tmp_variable(dtype)
    helper.append_op(
Y
yangyaming 已提交
1908 1909
        type='sequence_expand', inputs={'X': x,
                                        'Y': y}, outputs={'Out': tmp})
1910
    return tmp
1911 1912


Q
Qiao Longfei 已提交
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 1940 1941 1942 1943 1944
def beam_search(pre_ids, ids, scores, beam_size, end_id, level=0):
    '''
    This function implements the beam search algorithm.
    '''
    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,
            '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


Y
yangyaming 已提交
1945 1946 1947 1948
def lstm_unit(x_t,
              hidden_t_prev,
              cell_t_prev,
              forget_bias=0.0,
Y
yangyaming 已提交
1949
              param_attr=None,
C
caoying03 已提交
1950 1951
              bias_attr=None,
              name=None):
Y
yangyaming 已提交
1952 1953 1954 1955
    """Lstm unit layer. The equation of a lstm step is:

        .. math::

1956
            i_t & = \sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i)
Y
yangyaming 已提交
1957

1958
            f_t & = \sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + b_f)
Y
yangyaming 已提交
1959

1960
            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
1961

1962
            o_t & = \sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + b_o)
Y
yangyaming 已提交
1963 1964 1965

            h_t & = o_t tanh(c_t)

1966 1967 1968 1969 1970 1971
    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
1972 1973 1974

        .. math::

1975
            L_{i_t} = W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i
Y
yangyaming 已提交
1976 1977 1978 1979 1980 1981 1982 1983

    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
1984
    This layer has two outputs including :math:`h_t` and :math:`o_t`.
Y
yangyaming 已提交
1985 1986

    Args:
Y
yangyaming 已提交
1987 1988 1989 1990 1991 1992
        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
1993
        forget_bias (float): The forget bias of lstm unit.
Y
yangyaming 已提交
1994 1995
        param_attr (ParamAttr): The attributes of parameter weights, used to set
            initializer, name etc.
Y
yangyaming 已提交
1996 1997
        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
1998 1999
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
Y
yangyaming 已提交
2000 2001

    Returns:
Y
yangyaming 已提交
2002
        tuple: The hidden value and cell value of lstm unit.
Y
yangyaming 已提交
2003 2004

    Raises:
2005 2006 2007 2008
        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
2009 2010 2011 2012 2013 2014

    Examples:

        .. code-block:: python

             x_t = fluid.layers.fc(input=x_t_data, size=10)
2015
             prev_hidden = fluid.layers.fc(input=prev_hidden_data, size=30)
Y
yangyaming 已提交
2016
             prev_cell = fluid.layers.fc(input=prev_cell_data, size=30)
Y
yangyaming 已提交
2017
             hidden_value, cell_value = fluid.layers.lstm_unit(x_t=x_t,
Y
yangyaming 已提交
2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033
                                                    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
2034
        raise ValueError("The 1st dimensions of x_t, hidden_t_prev and "
2035 2036 2037 2038
                         "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
2039 2040
                         "cell_t_prev must be the same.")

Y
yangyaming 已提交
2041 2042 2043
    if bias_attr is None:
        bias_attr = ParamAttr()

Y
yangyaming 已提交
2044
    size = cell_t_prev.shape[1]
2045
    concat_out = concat(input=[x_t, hidden_t_prev], axis=1)
Y
yangyaming 已提交
2046 2047
    fc_out = fc(input=concat_out,
                size=4 * size,
Y
yangyaming 已提交
2048
                param_attr=param_attr,
2049
                bias_attr=bias_attr)
Y
yangyaming 已提交
2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061
    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
2062
    return h, c
G
guosheng 已提交
2063 2064


C
caoying03 已提交
2065
def reduce_sum(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
2066
    """
Y
yangyaming 已提交
2067
    Computes the sum of tensor elements over the given dimension.
G
guosheng 已提交
2068 2069 2070

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
2071 2072 2073 2074
        dim (int|None): The dimension along which the sum is performed. If
            :attr:`None`, sum 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))`. If :math:`dim < 0`,
G
guosheng 已提交
2075
            the dimension to reduce is :math:`rank + dim`.
Y
yangyaming 已提交
2076 2077
        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
2078
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
2079 2080
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
2081 2082 2083

    Returns:
        Variable: The reduced Tensor variable.
F
fengjiayi 已提交
2084

G
guosheng 已提交
2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108
    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]]
    """
    helper = LayerHelper('reduce_sum', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
        type='reduce_sum',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
            'dim': dim if dim != None else 0,
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
2109 2110


C
caoying03 已提交
2111
def reduce_mean(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
2112
    """
Y
yangyaming 已提交
2113
    Computes the mean of tensor elements over the given dimension.
G
guosheng 已提交
2114 2115 2116

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
2117 2118 2119 2120
        dim (int|None): The dimension along which the mean is computed. If
            :attr:`None`, compute the mean 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))`. If
G
guosheng 已提交
2121
            :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`.
Y
yangyaming 已提交
2122 2123
        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
2124
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
2125 2126
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
2127 2128 2129

    Returns:
        Variable: The reduced Tensor variable.
F
fengjiayi 已提交
2130

G
guosheng 已提交
2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154
    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]
            fluid.layers.reduce_mean(x, dim=1, keep_dim=True)  # [[0.475], [0.4]]
    """
    helper = LayerHelper('reduce_mean', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
        type='reduce_mean',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
            'dim': dim if dim != None else 0,
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
2155 2156


C
caoying03 已提交
2157
def reduce_max(input, dim=None, keep_dim=False, name=None):
2158
    """
Y
yangyaming 已提交
2159
    Computes the maximum of tensor elements over the given dimension.
2160 2161 2162

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
2163 2164 2165 2166
        dim (int|None): The dimension along which the maximum is computed.
            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))`.
2167
            If :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`.
Y
yangyaming 已提交
2168 2169
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
2170
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
2171 2172
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
2173 2174 2175

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
2176

2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202
    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]]
    """
    helper = LayerHelper('reduce_max', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
        type='reduce_max',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
            'dim': dim if dim != None else 0,
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out


C
caoying03 已提交
2203
def reduce_min(input, dim=None, keep_dim=False, name=None):
2204
    """
Y
yangyaming 已提交
2205
    Computes the minimum of tensor elements over the given dimension.
2206 2207 2208

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
2209 2210 2211 2212
        dim (int|None): The dimension along which the minimum is computed.
            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))`.
2213
            If :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`.
Y
yangyaming 已提交
2214 2215
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
2216
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
2217 2218
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
2219 2220 2221

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
2222

2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246
    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]]
    """
    helper = LayerHelper('reduce_min', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
        type='reduce_min',
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
            'dim': dim if dim != None else 0,
            'keep_dim': keep_dim,
            'reduce_all': True if dim == None else False
        })
    return out
G
guosheng 已提交
2247 2248


C
caoying03 已提交
2249
def split(input, num_or_sections, dim=-1, name=None):
G
guosheng 已提交
2250
    """
C
caoying03 已提交
2251
    Split the input tensor into multiple sub-tensors.
G
guosheng 已提交
2252 2253 2254

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
C
caoying03 已提交
2255 2256 2257 2258 2259
        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
2260
            :attr:`dim` dimension orderly.
C
caoying03 已提交
2261
        dim (int): The dimension along which to split. If :math:`dim < 0`, the
G
guosheng 已提交
2262
            dimension to split along is :math:`rank(input) + dim`.
C
caoying03 已提交
2263 2264
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306

    Returns:
        List: The list of segmented tensor variables.

    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]
            x0, x1, x2 = fluid.layers.split(x, num_or_sections=[2, 3, 4], dim=1)
            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
2307 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


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

    output = x / sqrt(max(sum(x**2), epsilon))

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

    Args:
       x(Variable|list): The input tensor to l2_normalize layer.
       axis(int): Dimension along which to normalize the input.
       epsilon(float): A lower bound value for `x`'s l2 norm. sqrt(epsilon) will
                       be used as the divisor if the l2 norm of `x` is less than
                       sqrt(epsilon).
       name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.


    Returns:
        Variable: The output tensor variable.

    Examples:
        .. code-block:: python

          data = fluid.layers.data(name="data",
                                   shape=(3, 17, 13),
                                   dtype="float32")
Y
ying 已提交
2340
          normed = fluid.layers.l2_normalize(x=data, axis=1)
C
caoying03 已提交
2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369
    """

    if len(x.shape) == 1: axis = 0

    helper = LayerHelper("l2_normalize", **locals())

    square = helper.create_tmp_variable(dtype=x.dtype)
    helper.append_op(type="square", inputs={"X": x}, outputs={"Out": square})

    reduced_sum = helper.create_tmp_variable(dtype=x.dtype)
    helper.append_op(
        type="reduce_sum",
        inputs={"X": square},
        outputs={"Out": reduced_sum},
        attrs={
            "dim": 1 if axis is None else axis,
            "keep_dim": True,
            "reduce_all": False
        })

    # TODO(caoying) A lower bound value epsilon for the norm is needed to
    # imporve the numeric stability of reciprocal. This requires a maximum_op.
    rsquare = helper.create_tmp_variable(dtype=x.dtype)
    helper.append_op(
        type="reciprocal", inputs={"X": reduced_sum}, outputs={"Out": rsquare})

    # TODO(caoying) the current elementwise_mul operator does not support a
    # general broadcast rule which broadcasts input(Y) to have the same
    # dimension with Input(X) starting from a specified dimension. So this
2370
    # exanpsion is requred. Once a general broadcast rule is spported, this
C
caoying03 已提交
2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387
    # expanding canbe removed.
    rsquare_expanded = helper.create_tmp_variable(dtype=x.dtype)
    expand_times = [1] * len(x.shape)
    expand_times[axis] = int(x.shape[axis])
    helper.append_op(
        type="expand",
        inputs={"X": rsquare},
        outputs={"Out": rsquare_expanded},
        attrs={"expand_times": expand_times})

    out = helper.create_tmp_variable(dtype=x.dtype)
    helper.append_op(
        type="elementwise_mul",
        inputs={"X": x,
                "Y": rsquare_expanded},
        outputs={"Out": out})
    return out
2388 2389


2390
def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
G
guosheng 已提交
2391
    """
Y
ying 已提交
2392 2393 2394 2395
    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
2396

C
chengduoZH 已提交
2397
    The actual behavior depends on the shapes of :math:`x`, :math:`y` and the
2398
    flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically:
G
guosheng 已提交
2399

2400 2401 2402 2403 2404
    - 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
2405
      :math:`[1, D]` in transposed form.
G
guosheng 已提交
2406

C
chengduoZH 已提交
2407
    - After transpose, the two tensors are 2-D or n-D and matrix multiplication
2408
      performs in the following way.
G
guosheng 已提交
2409

2410
      - If both are 2-D, they are multiplied like conventional matrices.
C
chengduoZH 已提交
2411
      - If either is n-D, it is treated as a stack of matrices residing in the
Y
ying 已提交
2412
        last two dimensions and a batched matrix multiply supporting broadcast
2413
        applies on the two tensors.
G
guosheng 已提交
2414

Y
ying 已提交
2415 2416
    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
2417
    removed after matrix multiplication.
G
guosheng 已提交
2418 2419 2420

    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
2421 2422 2423
        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.
2424
        name(str|None): A name for this layer(optional). If set None, the layer
2425
            will be named automatically.
G
guosheng 已提交
2426 2427

    Returns:
2428
        Variable: The product Tensor variable.
G
guosheng 已提交
2429

G
guosheng 已提交
2430 2431 2432
    Examples:
        .. code-block:: python

2433
            # Examples to clarify shapes of the inputs and output
C
chengduoZH 已提交
2434 2435
            # x: [B, ..., M, K], y: [B, ..., K, N]
            fluid.layers.matmul(x, y)  # out: [B, ..., M, N]
Y
ying 已提交
2436

2437 2438
            # x: [B, M, K], y: [B, K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
2439

2440 2441
            # x: [B, M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
2442

2443 2444
            # x: [M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [M, N]
Y
ying 已提交
2445 2446 2447 2448

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

2449 2450
            # x: [K], y: [K]
            fluid.layers.matmul(x, y)  # out: [1]
2451

Y
ying 已提交
2452
            # x: [M], y: [N]
2453
            fluid.layers.matmul(x, y, True, True)  # out: [M, N]
G
guosheng 已提交
2454
    """
Y
ying 已提交
2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466

    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
2467
            y_shape = y_shape + [1]
Y
ying 已提交
2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483

        # 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)

2484
    helper = LayerHelper('matmul', **locals())
Y
ying 已提交
2485
    out = helper.create_tmp_variable(dtype=x.dtype)
G
guosheng 已提交
2486
    helper.append_op(
2487 2488 2489 2490 2491 2492 2493
        type='matmul',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
        attrs={'transpose_X': transpose_x,
               'transpose_Y': transpose_y})
    return out
2494 2495


W
wanghaoshuang 已提交
2496 2497 2498 2499 2500
def edit_distance(input,
                  label,
                  normalized=False,
                  ignored_tokens=None,
                  name=None):
2501
    """
Y
ying 已提交
2502 2503 2504 2505 2506 2507 2508 2509 2510
    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
2511

Y
ying 已提交
2512
    "kitten" -> "sitten" -> "sittin" -> "sitting"
W
wanghaoshuang 已提交
2513

Y
ying 已提交
2514 2515 2516 2517
    Input(Hyps) is a LoDTensor consisting of all the hypothesis strings with
    the total number denoted by `batch_size`, and the separation is specified
    by the LoD information. And the `batch_size` reference strings are arranged
    in order in the same way in the LoDTensor Input(Refs).
W
wanghaoshuang 已提交
2518

Y
ying 已提交
2519 2520 2521
    Output(Out) contains the `batch_size` results and each stands for the edit
    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
2522

2523 2524 2525 2526 2527
    Args:

        input(Variable): The indices for hypothesis strings.

        label(Variable): The indices for reference strings.
W
wanghaoshuang 已提交
2528

Y
ying 已提交
2529 2530
        normalized(bool): Indicated whether to normalize the edit distance by
                          the length of reference string.
2531

Y
ying 已提交
2532 2533
        ignored_tokens(list of int): Tokens that should be removed before
                                     calculating edit distance.
2534

W
wanghaoshuang 已提交
2535
    Returns:
W
wanghaoshuang 已提交
2536
        Variable: sequence-to-sequence edit distance in shape [batch_size, 1].
W
wanghaoshuang 已提交
2537 2538 2539 2540 2541

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
2542 2543
            y = fluid.layers.data(name='y', shape=[7], dtype='float32')

2544
            cost = fluid.layers.edit_distance(input=x,label=y)
2545
    """
2546
    helper = LayerHelper("edit_distance", **locals())
2547

2548
    # remove some tokens from input and labels
W
wanghaoshuang 已提交
2549
    if ignored_tokens is not None and len(ignored_tokens) > 0:
2550 2551 2552 2553 2554 2555 2556
        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
2557
            attrs={"tokens": ignored_tokens})
2558 2559 2560 2561 2562 2563
        input = erased_input

        helper.append_op(
            type="sequence_erase",
            inputs={"X": [label]},
            outputs={"Out": [erase_label]},
W
wanghaoshuang 已提交
2564
            attrs={"tokens": ignored_tokens})
2565 2566
        label = erased_label

2567 2568
    # edit distance op
    edit_distance_out = helper.create_tmp_variable(dtype="int64")
2569
    sequence_num = helper.create_tmp_variable(dtype="int64")
2570 2571 2572 2573
    helper.append_op(
        type="edit_distance",
        inputs={"Hyps": [input],
                "Refs": [label]},
2574 2575
        outputs={"Out": [edit_distance_out],
                 "SequenceNum": [sequence_num]},
2576 2577
        attrs={"normalized": normalized})

2578
    return edit_distance_out, sequence_num
2579 2580 2581 2582 2583


def ctc_greedy_decoder(input, blank, name=None):
    """
    This op is used to decode sequences by greedy policy by below steps:
Y
ying 已提交
2584 2585 2586 2587
    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.
2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616

    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]]

        input.lod = [[0, 4, 8]]

        Then:

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

        output.lod = [[0, 2, 3]]

    Args:

Y
ying 已提交
2617 2618 2619 2620 2621 2622
        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).
2623

Y
ying 已提交
2624 2625 2626
        blank(int): the blank label index of Connectionist Temporal
                    Classification (CTC) loss, which is in thehalf-opened
                    interval [0, num_classes + 1).
2627 2628

    Returns:
2629
        Variable: CTC greedy decode result. If all the sequences in result were
2630
        empty, the result LoDTensor will be [-1] with LoD [[0]] and dims [1, 1].
2631 2632 2633 2634 2635

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
W
wanghaoshuang 已提交
2636

2637
            cost = fluid.layers.ctc_greedy_decoder(input=x, blank=0)
W
wanghaoshuang 已提交
2638
    """
2639
    helper = LayerHelper("ctc_greedy_decoder", **locals())
2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654
    # top 1 op
    topk_out = helper.create_tmp_variable(dtype=input.dtype)
    topk_indices = helper.create_tmp_variable(dtype="int64")
    helper.append_op(
        type="top_k",
        inputs={"X": [input]},
        outputs={"Out": [topk_out],
                 "Indices": [topk_indices]},
        attrs={"k": 1})

    # 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
2655
        outputs={"Output": [ctc_out]},
2656 2657
        attrs={"merge_repeated": True,
               "blank": blank})
2658
    return ctc_out
2659 2660


W
wanghaoshuang 已提交
2661 2662
def warpctc(input, label, blank=0, norm_by_times=False, **kwargs):
    """
2663 2664
    An operator integrating the open source Warp-CTC library
    (https://github.com/baidu-research/warp-ctc)
W
wanghaoshuang 已提交
2665
    to compute Connectionist Temporal Classification (CTC) loss.
2666 2667
    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
2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680
    input tensor.

    Args:
       input(Variable): (LodTensor, default: LoDTensor<float>),
         the unscaled 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).
       label(Variable): (LodTensor, default: LoDTensor<int>), the ground truth
         of variable-length sequence, 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.
2681
       blank: (int, default: 0), the blank label index of Connectionist
W
wanghaoshuang 已提交
2682 2683
         Temporal Classification (CTC) loss, which is in the
         half-opened interval [0, num_classes + 1).
2684
       norm_by_times: (bool, default: false), whether to normalize
W
wanghaoshuang 已提交
2685
       the gradients by the number of time-step, which is also the
2686 2687
       sequence's length. There is no need to normalize the gradients
       if warpctc layer was follewed by a mean_op.
W
wanghaoshuang 已提交
2688 2689

    Returns:
2690 2691
        Variable: The Connectionist Temporal Classification (CTC) loss,
        which is a 2-D Tensor of the shape [batch_size, 1].
W
wanghaoshuang 已提交
2692 2693 2694

    Examples:
        .. code-block:: python
2695 2696 2697 2698
            y = layers.data(
                name='y', shape=[11, 8], dtype='float32', lod_level=1)
            y_predict = layers.data(
                name='y_predict', shape=[11, 1], dtype='float32')
W
wanghaoshuang 已提交
2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713
            cost = layers.warpctc(input=y_predict, label=y)

    """
    helper = LayerHelper('warpctc', **kwargs)
    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
2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767


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]]
            x.data = [[1, 2], [3, 4],
                      [5, 6], [7, 8], [9, 10], [11, 12]]
            x.dims = [6, 2]

        set new_dim = 4

        then out is a LoDTensor:
            out.lod  = [[0, 1, 3]]
            out.data = [[1, 2, 3, 4],
                        [5, 6, 7, 8], [9, 10, 11, 12]]
            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:
       input (Variable): (LodTensor, default: LoDTensor<float>), a 2-D LoDTensor
                with shape being [N, M] where M for dimension.
       new_dim (int): New dimension which the input LoDTensor is reshaped to.

    Returns:
        Variable: Reshaped LoDTensor according to new dimension.

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[5, 20],
                              dtype='float32', lod_level=1)
            x_reshaped = layers.sequence_reshape(input=x, new_dim=10)
    """
    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
2768 2769


2770
@autodoc()
Y
Yang Yu 已提交
2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796
def nce(input,
        label,
        num_total_classes,
        sample_weight=None,
        param_attr=None,
        bias_attr=None,
        num_neg_samples=None):
    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
2797 2798 2799 2800 2801 2802 2803 2804 2805
    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
2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821

    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
2822
    return cost / (num_neg_samples + 1)
2823 2824


Y
fix ci.  
ying 已提交
2825
def transpose(x, perm, name=None):
Y
ying 已提交
2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844
    """
    **transpose Layer**

    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:
       input (Variable): (Tensor), A Tensor.
       perm (list): A permutation of the dimensions of `input`.

    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
2845
            x_transposed = layers.transpose(x, perm=[1, 0, 2])
Y
ying 已提交
2846 2847
    """

Y
fix ci.  
ying 已提交
2848
    if len(perm) != len(x.shape):
Y
ying 已提交
2849 2850 2851
        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
2852 2853 2854 2855 2856 2857
    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
2858 2859

    helper = LayerHelper('transpose', **locals())
Y
fix ci.  
ying 已提交
2860
    out = helper.create_tmp_variable(x.dtype)
Y
ying 已提交
2861 2862
    helper.append_op(
        type='transpose',
Y
fix ci.  
ying 已提交
2863
        inputs={'X': [x]},
Y
ying 已提交
2864 2865 2866
        outputs={'Out': [out]},
        attrs={'axis': perm})
    return out
2867 2868


2869
def im2sequence(input, filter_size=1, stride=1, padding=0, name=None):
2870
    """
2871 2872 2873 2874 2875 2876 2877
    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:
2878 2879 2880 2881 2882 2883 2884 2885 2886 2887

    .. 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
2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905

        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.

2906 2907 2908
        name (int): The name of this layer. It is optional.

    Returns:
W
wanghaoshuang 已提交
2909 2910 2911 2912 2913
        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.
2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942

    Examples:

    As an example:

        .. 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
2943 2944 2945
            filter = [2, 2]
            stride = [1, 1]
            padding = [0, 0]
2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965

            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.]]

            output.dims = {8, 9}

            output.lod = [[0, 4, 8]]

        The simple usage is:

        .. code-block:: python

2966 2967
            output = fluid.layers.im2sequence(
                input=layer, stride=[1, 1], filter_size=[2, 2])
2968 2969

    """
W
wanghaoshuang 已提交
2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980

    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])

2981
    helper = LayerHelper('im2sequence', **locals())
2982 2983
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
2984
        type='im2sequence',
2985 2986 2987
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
wanghaoshuang 已提交
2988 2989 2990
            'kernels': filter_size,
            'strides': stride,
            'paddings': padding,
2991 2992
        })
    return out
2993 2994


2995 2996 2997 2998
def row_conv(input, future_context_size, param_attr=None, act=None):
    """Row Conv Operator. This layer will apply lookahead convolution to
    **input**. The input variable should be a 2D LoDTensor with shape [T, D].
    Parameters with shape [future_context_size + 1, D] will be created. The math
Y
yangyaming 已提交
2999
    equation of row convolution is as follows:
3000 3001 3002 3003 3004 3005 3006

    .. math::
        Out_{i} = \sum_{j = i} ^ {i + \\tau} X_{j} \odot W_{i - j}

    In the above equation:

    * :math:`Out_{i}`: The i-th row of output variable with shape [1, D].
Y
yangyaming 已提交
3007
    * :math:`\\tau`: Future context size.
3008 3009 3010 3011 3012 3013 3014 3015 3016 3017
    * :math:`X_{j}`: The j-th row of input variable with shape [1, D].
    * :math:`W_{i-j}`: The (i-j)-th row of parameters with shape [1, D].

    More details about row_conv please refer to the paper \
    (http://www.cs.cmu.edu/~dyogatam/papers/wang+etal.iclrworkshop2016.pdf) and
    the design document \
    (https://github.com/PaddlePaddle/Paddle/issues/2228#issuecomment-303903645).

    Args:
        input (Variable): Input variable, a 2D LoDTensor with shape [T, D].
Y
yangyaming 已提交
3018 3019
        future_context_size (int): Future context size. Please note, the shape
            of convolution kernel is [future_context_size + 1, D].
3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044
        param_attr (ParamAttr): Attributes of parameters, including
            name, initializer etc.
        act (str): Non-linear activation to be applied to output variable.

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

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[16],
                            dtype='float32', lod_level=1)
            out = fluid.layers.row_conv(input=x, future_context_size=2)
    """
    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
3045
    return helper.append_activation(out)
3046 3047


3048 3049 3050 3051
def multiplex(inputs, index):
    """
    **Multiplex Layer**

Y
yangyaming 已提交
3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066
    Referring to the given index variable, this layer selects rows from the
    input variables to construct a multiplex variable. Assuming that there are
    :math:`m` input variables and :math:`I_i` represents the i-th input
    variable and :math:`i` is in [0, :math:`m`). All input variables are
    tensors with same shape [:math:`d_0`, :math:`d_1`, ..., :math:`d_R`].
    Please note that rank of the input tensor should be at least 2. Each input
    variable will be treated as a 2-D matrix with shape [:math:`M`, :math:`N`]
    where :math:`M` for :math:`d_0` and :math:`N` for :math:`d_1` * :math:`d_2`
    * ... * :math:`d_R`. Let :math:`I_i[j]` be the j-th row of the i-th input
    variable. The given index variable should be a 2-D tensor with shape
    [:math:`M`, 1]. Let `ID[i]` be the i-th index value of the index variable.
    Then the output variable will be a tensor with shape [:math:`d_0`,
    :math:`d_1`, ..., :math:`d_R`]. If we treat the output tensor as a 2-D
    matrix with shape [:math:`M`, :math:`N`] and let :math:`O[i]` be the i-th
    row of the matrix, then `O[i]` is equal to :math:`I_{ID[i]}[i]`.
3067 3068

    Args:
Y
yangyaming 已提交
3069 3070
       inputs (list): A list of variables to gather from. All variables have the
                same shape and the rank is at least 2.
3071
       index (Variable): Tensor<int32>, index variable which is a 2-D tensor
Y
yangyaming 已提交
3072
                with shape [M, 1] where M is the batch size.
3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085

    Returns:
        Variable: Multiplex variable gathered from input variables.

    Examples:
        .. code-block:: python

            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)
    """
    helper = LayerHelper('multiplex', **locals())
Y
yangyaming 已提交
3086 3087 3088 3089 3090 3091

    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)
3092 3093 3094 3095 3096 3097
    helper.append_op(
        type='multiplex',
        inputs={'X': inputs,
                'Ids': index},
        outputs={'Out': [out]})
    return out
C
chengduoZH 已提交
3098 3099


C
chengduoZH 已提交
3100 3101 3102 3103
def reshape_with_axis(input, axis):
    """
    **ReshapeWithAxis Layer**

C
chengduoZH 已提交
3104
    ReshapeWithAxis is used to merge adjacent dimensions according to axis.
C
chengduoZH 已提交
3105 3106 3107

    Args:
       input(variable): The input tensor.
C
chengduoZH 已提交
3108
       axis(list): The axis which is used to merge the adjacent dimensions.
C
chengduoZH 已提交
3109 3110 3111 3112 3113 3114 3115 3116

    Returns:
        Variable: A tensor variable.

    Examples:
        .. code-block:: python

          x = fluid.layers.data(name="data", shape=[3, 32, 32], dtype="float32")
C
chengduoZH 已提交
3117
          reshaped = fluid.layers.reshape_with_axis(input=x, axis=[2])
C
chengduoZH 已提交
3118 3119 3120 3121 3122 3123 3124 3125 3126
          reshaped.shape
            >> [-1, 1024]
          reshaped = fluid.layers.reshape_with_axis(input=x, axis=[1,3])
          reshaped.shape
            >> [-1, 96, 32]
    """
    assert isinstance(axis, list), "axis should be list."
    assert len(input.shape) > len(
        axis), "the length of axis should be litter than input.shape's."
C
chengduoZH 已提交
3127
    input_shape = input.shape
C
chengduoZH 已提交
3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139
    temp = 0
    for ax in axis:
        assert ax < len(input.shape) and ax > 0, \
            'The data of Axis should be between 1 and len(input.shape)'
        assert ax > temp, 'Axis should be incremented sequence'
        temp = ax
    axis += [len(input.shape)]

    new_shape = []
    for i in range(len(axis) - 1):
        new_shape += [reduce(mul, input_shape[axis[i]:axis[i + 1]], 1)]
    new_shape = [-1] + new_shape
C
chengduoZH 已提交
3140 3141 3142 3143 3144 3145 3146

    helper = LayerHelper('reshape', **locals())
    out = helper.create_tmp_variable(helper.input_dtype())
    helper.append_op(
        type='reshape',
        inputs={'X': [input]},
        outputs={'Out': [out]},
C
chengduoZH 已提交
3147
        attrs={'shape': new_shape})
C
chengduoZH 已提交
3148 3149 3150
    return out


C
chengduoZH 已提交
3151
def flatten(input, axis=1):
C
chengduoZH 已提交
3152
    """
C
chengduoZH 已提交
3153 3154
    **Flatten Layer**
    ReshapeWithAxis is used to merge adjacent dimensions according to axis.
C
chengduoZH 已提交
3155
    Args:
C
chengduoZH 已提交
3156 3157
       input(variable): The input tensor.
       axis(int):
C
chengduoZH 已提交
3158
    Returns:
C
chengduoZH 已提交
3159
        Variable: A tensor variable.
C
chengduoZH 已提交
3160 3161
    Examples:
        .. code-block:: python
C
chengduoZH 已提交
3162 3163 3164 3165
          x = fluid.layers.data(name="data", shape=[3, 32, 32], dtype="float32")
          reshaped = fluid.layers.reshape_with_axis(input=x, axis=2)
          reshaped.shape
            >> [-1, 1024]
C
chengduoZH 已提交
3166
    """
C
chengduoZH 已提交
3167 3168 3169
    assert len(input.shape) > axis and axis > 0, \
        "the axis should be litter than input.shape's."
    input_shape = input.shape
C
chengduoZH 已提交
3170

C
chengduoZH 已提交
3171
    new_shape = [-1, reduce(mul, input_shape[axis:len(input_shape)], 1)]
C
chengduoZH 已提交
3172

C
chengduoZH 已提交
3173 3174 3175 3176 3177 3178 3179 3180
    helper = LayerHelper('reshape', **locals())
    out = helper.create_tmp_variable(helper.input_dtype())
    helper.append_op(
        type='reshape',
        inputs={'X': [input]},
        outputs={'Out': [out]},
        attrs={'shape': new_shape})
    return out