nn.py 119.6 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
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 17 18 19 20
"""
All layers just related to the neural network.
"""

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
import utils
Y
Yu Yang 已提交
25 26

__all__ = [
Y
ying 已提交
27 28 29
    'fc',
    'embedding',
    'dynamic_lstm',
Y
Yibing Liu 已提交
30
    'dynamic_lstmp',
G
guosheng 已提交
31
    'dynamic_gru',
Y
ying 已提交
32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51
    'gru_unit',
    'linear_chain_crf',
    'crf_decoding',
    'cos_sim',
    'cross_entropy',
    'square_error_cost',
    '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',
52
    'reduce_prod',
Y
ying 已提交
53 54 55 56
    'sequence_first_step',
    'sequence_last_step',
    'dropout',
    'split',
57 58
    'ctc_greedy_decoder',
    'edit_distance',
Y
ying 已提交
59 60 61 62
    'l2_normalize',
    'matmul',
    'warpctc',
    'sequence_reshape',
63
    'transpose',
64
    'im2sequence',
65
    'nce',
Q
Qiao Longfei 已提交
66
    'beam_search',
67
    'row_conv',
68
    'multiplex',
G
guosheng 已提交
69
    'layer_norm',
70 71
    'softmax_with_cross_entropy',
    'smooth_l1',
72
    'one_hot',
Y
Yu Yang 已提交
73
    'autoincreased_step_counter',
Y
Yu Yang 已提交
74 75 76 77 78 79 80 81 82
]


def fc(input,
       size,
       num_flatten_dims=1,
       param_attr=None,
       bias_attr=None,
       act=None,
83
       name=None):
Y
Yu Yang 已提交
84
    """
85
    **Fully Connected Layer**
Y
Yu Yang 已提交
86

C
caoying03 已提交
87
    The fully connected layer can take multiple tensors as its inputs. It
R
ranqiu 已提交
88 89 90 91 92 93
    creates a variable called weights for each input tensor, which represents
    a fully connected weight matrix from each input unit to each output unit.
    The fully connected layer multiplies each input tensor with its coresponding
    weight to produce an output Tensor. If multiple input tensors are given,
    the results of multiple multiplications will be sumed up. If bias_attr is
    not None, a bias variable will be created and added to the output. Finally,
Y
ying 已提交
94
    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).
108
    * :math:`Act`: The activation function.
C
caoying03 已提交
109
    * :math:`Out`: The output tensor.
Y
Yu Yang 已提交
110 111

    Args:
R
ranqiu 已提交
112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129
        input (Variable|list of Variable): The input tensor(s) of this layer, and the dimension of
            the input tensor(s) is at least 2.
        size(int): The number of output units in this layer.
        num_flatten_dims (int, default 1): The fc layer can accept an input tensor with more than
            two dimensions. If this happens, the multidimensional tensor will first be flattened
            into a 2-dimensional matrix. The parameter `num_flatten_dims` determines how the input
            tensor is flattened: the first `num_flatten_dims` (inclusive, index starts from 1)
            dimensions will be flatten to form the first dimension of the final matrix (height of
            the matrix), and the rest `rank(X) - num_flatten_dims` dimensions are flattened to
            form the second dimension of the final matrix (width of the matrix). For example, suppose
            `X` is a 6-dimensional tensor with a shape [2, 3, 4, 5, 6], and `num_flatten_dims` = 3.
            Then, the flattened matrix will have a shape [2 x 3 x 4, 5 x 6] = [24, 30].
        param_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for learnable
            parameters/weights of this layer.
        bias_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for the bias
            of this layer. If it is set to None, no bias will be added to the output units.
        act (str, default None): Activation to be applied to the output of this layer.
        name (str, default None): The name of this layer.
Y
Yu Yang 已提交
130

131
    Returns:
R
ranqiu 已提交
132
        A tensor variable storing the transformation result.
133 134

    Raises:
C
caoying03 已提交
135
        ValueError: If rank of the input tensor is less than 2.
136 137 138 139

    Examples:
        .. code-block:: python

C
caoying03 已提交
140
          data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
141
          fc = fluid.layers.fc(input=data, size=1000, act="tanh")
Y
Yu Yang 已提交
142
    """
C
caoying03 已提交
143

C
caoying03 已提交
144
    helper = LayerHelper("fc", **locals())
Y
Yu Yang 已提交
145 146 147 148 149 150 151 152 153

    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
154

Y
Yu Yang 已提交
155 156 157 158 159
        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
160 161
            inputs={"X": input_var,
                    "Y": w},
Y
Yu Yang 已提交
162
            outputs={"Out": tmp},
C
caoying03 已提交
163 164
            attrs={"x_num_col_dims": num_flatten_dims,
                   "y_num_col_dims": 1})
Y
Yu Yang 已提交
165 166 167 168 169 170 171 172 173 174
        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
175
    pre_activation = helper.append_bias_op(pre_bias, dim_start=num_flatten_dims)
Y
Yu Yang 已提交
176 177 178 179
    # add activation
    return helper.append_activation(pre_activation)


180 181 182 183 184 185
def embedding(input,
              size,
              is_sparse=False,
              padding_idx=None,
              param_attr=None,
              dtype='float32'):
Y
Yu Yang 已提交
186
    """
187 188
    **Embedding Layer**

189
    This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in
190 191
    a lookup table. The result of this lookup is the embedding of each ID in the
    :attr:`input`.
192 193 194

    All the input variables are passed in as local variables to the LayerHelper
    constructor.
Y
Yu Yang 已提交
195 196

    Args:
197 198 199 200 201 202 203
        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
204 205
            with zeros whenever lookup encounters it in :attr:`input`. If
            :math:`padding_idx < 0`, the padding_idx to use in lookup is
206 207
            :math:`size[0] + dim`.
        param_attr(ParamAttr): Parameters for this layer
208
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc
Y
Yu Yang 已提交
209

210 211 212
    Returns:
        Variable: The tensor variable storing the embeddings of the \
                  supplied inputs.
Y
Yu Yang 已提交
213

214 215
    Examples:
        .. code-block:: python
Y
Yu Yang 已提交
216

C
chengduoZH 已提交
217
          dict_size = len(dataset.ids)
218
          data = fluid.layers.data(name='ids', shape=[32, 32], dtype='float32')
C
chengduoZH 已提交
219
          fc = fluid.layers.embedding(input=data, size=[dict_size, 16])
Y
Yu Yang 已提交
220 221 222 223 224 225
    """

    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)
226 227
    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
228 229 230 231 232
    helper.append_op(
        type='lookup_table',
        inputs={'Ids': input,
                'W': w},
        outputs={'Out': tmp},
233 234
        attrs={'is_sparse': is_sparse,
               'padding_idx': padding_idx})
Y
Yu Yang 已提交
235 236 237 238 239 240 241 242 243 244 245 246 247
    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',
248 249
                 dtype='float32',
                 name=None):
Y
Yibing Liu 已提交
250 251 252 253 254 255
    """
    **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
256
    .. math::
Y
Yibing Liu 已提交
257

258
        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
259

260
        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
261

262
        \\tilde{c_t} & = act_g(W_{cx}x_t + W_{ch}h_{t-1} + b_c)
Y
Yibing Liu 已提交
263

264 265 266
        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
267

Y
Yibing Liu 已提交
268
        h_t & = o_t \odot act_h(c_t)
Y
Yibing Liu 已提交
269

270
    where the :math:`W` terms denote weight matrices (e.g. :math:`W_{xi}` is
271
    the matrix of weights from the input gate to the input), :math:`W_{ic}, \
272 273 274
    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
275
    gate bias vector), :math:`\sigma` is the non-linear activations, such as
276 277
    logistic sigmoid function, and :math:`i, f, o` and :math:`c` are the input
    gate, forget gate, output gate, and cell activation vectors, respectively,
278 279
    all of which have the same size as the cell output activation vector :math:`h`.

280 281 282 283
    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
284 285 286
    the previous hidden state.

    Set `use_peepholes` to `False` to disable peephole connection. The formula
Y
Yibing Liu 已提交
287 288 289
    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
290 291 292
    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
293 294

    Args:
295 296 297 298
        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
299 300
                         mini-batch, D is the hidden size.
        size(int): 4 * hidden size.
301
        param_attr(ParamAttr|None): The parameter attribute for the learnable
302
                               hidden-hidden weights.
Y
Yibing Liu 已提交
303 304 305

                               - Weights = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}
306 307 308
                               - The shape is (D x 4D), where D is the hidden
                                 size.
        bias_attr(ParamAttr|None): The bias attribute for the learnable bias
309 310 311
                              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
312

313
                              1. `use_peepholes = False`
Y
Yibing Liu 已提交
314
                                - Biases = {:math:`b_c, b_i, b_f, b_o`}.
315
                                - The shape is (1 x 4D).
316
                              2. `use_peepholes = True`
Y
Yibing Liu 已提交
317 318
                                - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
                                                 W_{fc}, W_{oc}`}.
319
                                - The shape is (1 x 7D).
320
        use_peepholes(bool): Whether to enable diagonal/peephole connections,
Y
Yibing Liu 已提交
321 322
                             default `True`.
        is_reverse(bool): Whether to compute reversed LSTM, default `False`.
323 324
        gate_activation(str): The activation for input gate, forget gate and
                              output gate. Choices = ["sigmoid", "tanh", "relu",
Y
Yibing Liu 已提交
325
                              "identity"], default "sigmoid".
326
        cell_activation(str): The activation for cell output. Choices = ["sigmoid",
Y
Yibing Liu 已提交
327 328 329 330 331
                              "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".
332 333
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
Y
Yibing Liu 已提交
334 335

    Returns:
Y
Yibing Liu 已提交
336 337
        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
338

Y
Yibing Liu 已提交
339
    Examples:
Y
Yibing Liu 已提交
340 341
        .. code-block:: python

Y
Yibing Liu 已提交
342 343
            hidden_dim = 512
            forward_proj = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
344
                                           act=None, bias_attr=None)
Y
Yibing Liu 已提交
345 346
            forward, _ = fluid.layers.dynamic_lstm(
                input=forward_proj, size=hidden_dim * 4, use_peepholes=False)
Y
Yibing Liu 已提交
347
    """
348

Y
Yu Yang 已提交
349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384
    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
385 386 387 388 389 390 391 392 393 394 395
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',
396 397
                  dtype='float32',
                  name=None):
Y
Yibing Liu 已提交
398 399 400
    """
    **Dynamic LSTMP Layer**

401 402 403 404 405 406
    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
407 408 409 410 411

    The formula is as follows:

    .. math::

412
        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
413

414
        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
415

416
        \\tilde{c_t} & = act_g(W_{cx}x_t + W_{cr}r_{t-1} + b_c)
Y
Yibing Liu 已提交
417

418
        o_t & = \sigma(W_{ox}x_{t} + W_{or}r_{t-1} + W_{oc}c_t + b_o)
Y
Yibing Liu 已提交
419

420
        c_t & = f_t \odot c_{t-1} + i_t \odot \\tilde{c_t}
Y
Yibing Liu 已提交
421

422
        h_t & = o_t \odot act_h(c_t)
Y
Yibing Liu 已提交
423

424
        r_t & = \overline{act_h}(W_{rh}h_t)
Y
Yibing Liu 已提交
425

Y
Yibing Liu 已提交
426 427 428 429 430 431
    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, \
432
          we use vectors to reprenset these diagonal weight matrices.
Y
Yibing Liu 已提交
433
    * :math:`b`: Denotes bias vectors (e.g. :math:`b_i` is the input gate \
434
          bias vector).
Y
Yibing Liu 已提交
435 436 437
    * :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 \
438
          the same size as the cell output activation vector :math:`h`.
Y
Yibing Liu 已提交
439
    * :math:`h`: The hidden state.
440
    * :math:`r`: The recurrent projection of the hidden state.
Y
Yibing Liu 已提交
441 442
    * :math:`\\tilde{c_t}`: The candidate hidden state, whose \
          computation is based on the current input and previous hidden state.
443
    * :math:`\odot`: The element-wise product of the vectors.
Y
Yibing Liu 已提交
444
    * :math:`act_g` and :math:`act_h`: The cell input and cell output \
445
          activation functions and `tanh` is usually used for them.
Y
Yibing Liu 已提交
446 447
    * :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
448 449 450 451

    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.
452

Y
Yibing Liu 已提交
453 454 455 456 457 458 459 460 461 462 463 464
    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.
465
        param_attr(ParamAttr|None): The parameter attribute for the learnable
Y
Yibing Liu 已提交
466 467
                               hidden-hidden weight and projection weight.

468 469
                               - Hidden-hidden weight = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}.
470 471
                               - 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
472 473
                                 size.
                               - Projection weight = {:math:`W_{rh}`}.
474 475
                               - 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
476 477 478 479 480 481
                              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`}.
482
                                - The shape is (1 x 4D).
Y
Yibing Liu 已提交
483 484 485
                              2. `use_peepholes = True`
                                - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
                                                 W_{fc}, W_{oc}`}.
486
                                - The shape is (1 x 7D).
Y
Yibing Liu 已提交
487 488 489 490 491 492 493 494 495 496 497 498 499 500 501
        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".
502 503
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
Y
Yibing Liu 已提交
504 505

    Returns:
506 507
        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
508 509 510 511 512
               (T x D), and both LoD is the same with the `input`.

    Examples:
        .. code-block:: python

Y
Yibing Liu 已提交
513
            hidden_dim, proj_dim = 512, 256
Y
Yibing Liu 已提交
514 515
            fc_out = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
                                     act=None, bias_attr=None)
516 517 518
            proj_out, _ = fluid.layers.dynamic_lstmp(input=fc_out,
                                                     size=hidden_dim * 4,
                                                     proj_size=proj_dim,
Y
Yibing Liu 已提交
519 520 521 522
                                                     use_peepholes=False,
                                                     is_reverse=True,
                                                     cell_activation="tanh",
                                                     proj_activation="tanh")
Y
Yibing Liu 已提交
523
    """
524

Y
Yibing Liu 已提交
525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570
    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
571 572 573 574 575 576 577 578 579 580 581
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**

582
    Refer to `Empirical Evaluation of Gated Recurrent Neural Networks on
G
guosheng 已提交
583
    Sequence Modeling <https://arxiv.org/abs/1412.3555>`_
584

G
guosheng 已提交
585 586 587 588 589 590 591 592 593
    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)
594

G
guosheng 已提交
595
        h_t & = (1-u_t) \odot h_{t-1} + u_t \odot \\tilde{h_t}
596

G
guosheng 已提交
597
    The :math:`\odot` is the element-wise product of the vectors. :math:`act_g`
598 599
    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
600 601 602 603
    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
604
    to use fully-connect layer before GRU layer.
G
guosheng 已提交
605 606

    Args:
607 608
        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
609
            Variable is a matrix with shape :math:`(T \\times 3D)`, where
610
            :math:`T` is the total time steps in this mini-batch, :math:`D`
G
guosheng 已提交
611 612
            is the hidden size.
        size(int): The dimension of the gru cell.
613
        param_attr(ParamAttr|None): The parameter attribute for the learnable
G
guosheng 已提交
614 615
            hidden-hidden weight matrix. Note:

616
            - The shape of the weight matrix is :math:`(T \\times 3D)`, where
G
guosheng 已提交
617
              :math:`D` is the hidden size.
618
            - All elements in the weight matrix can be divided into two parts.
G
guosheng 已提交
619
              The first part are weights of the update gate and reset gate with
620
              shape :math:`(D \\times 2D)`, and the second part are weights for
G
guosheng 已提交
621
              candidate hidden state with shape :math:`(D \\times D)`.
622
        bias_attr(ParamAttr): The parameter attribute for learnable the
G
guosheng 已提交
623
            hidden-hidden bias.
624
        is_reverse(bool): Whether to compute reversed GRU, default
G
guosheng 已提交
625 626 627
            :attr:`False`.
        gate_activation(str): The activation for update gate and reset gate.
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "sigmoid".
628
        activation(str): The activation for candidate hidden state.
G
guosheng 已提交
629 630 631
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh".

    Returns:
G
guosheng 已提交
632 633
        Variable: The hidden state of GRU. The shape is :math:`(T \\times D)`, \
            and lod is the same with the input.
634

G
guosheng 已提交
635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677
    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
678 679 680 681 682 683
def gru_unit(input,
             hidden,
             size,
             weight=None,
             bias=None,
             activation='tanh',
684
             gate_activation='sigmoid'):
Y
Yu Yang 已提交
685
    """
686
    GRU unit layer. The equation of a gru step is:
Y
Yu Yang 已提交
687

688 689
        .. math::
            u_t & = actGate(xu_{t} + W_u h_{t-1} + b_u)
Y
Yu Yang 已提交
690

691
            r_t & = actGate(xr_{t} + W_r h_{t-1} + b_r)
Y
Yu Yang 已提交
692

693
            m_t & = actNode(xm_t + W_c dot(r_t, h_{t-1}) + b_m)
694

695
            h_t & = dot((1-u_t), m_t) + dot(u_t, h_{t-1})
696 697

    The inputs of gru unit includes :math:`z_t`, :math:`h_{t-1}`. In terms
698 699 700
    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
701 702
    connected layer has to be applied, such that :math:`z_t = W_{fc}x_t`.

703 704
    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
705 706 707
    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`.
708 709 710 711 712 713 714

    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
715 716 717 718
        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
719

720 721 722 723 724 725
    Returns:
        tuple: The hidden value, reset-hidden value and gate values.

    Examples:

        .. code-block:: python
Y
Yu Yang 已提交
726

727
             # assuming we have x_t_data and prev_hidden of size=10
728
             x_t = fluid.layers.fc(input=x_t_data, size=30)
729 730
             hidden_val, r_h_val, gate_val = fluid.layers.gru_unit(input=x_t,
                                                    hidden = prev_hidden)
Y
Yu Yang 已提交
731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750

    """
    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
751

Y
Yu Yang 已提交
752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778
    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


779
def linear_chain_crf(input, label, param_attr=None):
Y
Yu Yang 已提交
780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804
    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


805
def crf_decoding(input, param_attr, label=None):
Y
Yu Yang 已提交
806 807 808 809 810 811 812 813 814 815 816 817 818
    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


F
fengjiayi 已提交
819
def cos_sim(X, Y):
Y
Yu Yang 已提交
820 821 822 823
    """
    This function performs the cosine similarity between two tensors
    X and Y and returns that as the output.
    """
F
fengjiayi 已提交
824
    helper = LayerHelper('cos_sim', **locals())
Y
Yu Yang 已提交
825 826 827 828 829 830 831 832 833 834 835 836 837
    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


F
fengjiayi 已提交
838
def dropout(x, dropout_prob, is_test=False, seed=None):
839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866
    """
    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)
    """

F
fengjiayi 已提交
867
    helper = LayerHelper('dropout', **locals())
868 869 870 871 872 873 874
    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]},
875 876 877 878 879 880
        attrs={
            'dropout_prob': dropout_prob,
            'is_test': is_test,
            'fix_seed': seed is not None,
            'seed': seed if seed is not None else 0
        })
881 882 883
    return out


F
fengjiayi 已提交
884
def cross_entropy(input, label, soft_label=False):
Y
Yu Yang 已提交
885
    """
Y
Yibing Liu 已提交
886 887
    **Cross Entropy Layer**

888 889 890
    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
891 892

    1) One-hot cross-entropy:
F
fengjiayi 已提交
893
        `soft_label = False`, `Label[i, 0]` indicates the class index for sample i:
Y
yangyaming 已提交
894

Y
Yibing Liu 已提交
895
        .. math::
Y
yangyaming 已提交
896

Y
Yibing Liu 已提交
897 898 899
            Y[i] = -\log(X[i, Label[i]])

    2) Soft-label cross-entropy:
F
fengjiayi 已提交
900 901
        `soft_label = True`, `Label[i, j]` indicates the soft label of class j
        for sample i:
Y
Yibing Liu 已提交
902 903 904 905 906

        .. math::

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

Y
Yibing Liu 已提交
907
       Please make sure that in this case the summation of each row of `label`
Y
Yibing Liu 已提交
908 909 910
       equals one.

    3) One-hot cross-entropy with vecterized `label`:
F
fengjiayi 已提交
911 912
         As a special case of 2), when each row of 'label' has only one
         non-zero element which is equal to 1, soft-label cross-entropy degenerates
Y
Yibing Liu 已提交
913
         to a one-hot cross-entropy with one-hot label representation.
Y
yangyaming 已提交
914

Y
Yibing Liu 已提交
915
    Args:
Y
yangyaming 已提交
916
        input (Variable|list):  a 2-D tensor with shape [N x D], where N is the
917 918 919 920
                                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
921
        label (Variable|list): the ground truth which is a 2-D tensor. When
922 923 924 925
                               `soft_label` is set to `False`, `label` is a
                               tensor<int64> with shape [N x 1]. When
                               `soft_label` is set to `True`, `label` is a
                               tensor<float/double> with shape [N x D].
F
fengjiayi 已提交
926
        soft_label (bool): a flag indicating whether to
927 928
                                           interpretate the given labels as soft
                                           labels, default `False`.
Y
Yibing Liu 已提交
929 930 931 932 933

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

    Raises:
934 935 936 937 938
        `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
939 940 941 942 943 944

    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
945
    """
F
fengjiayi 已提交
946
    helper = LayerHelper('cross_entropy', **locals())
Y
Yu Yang 已提交
947 948 949 950 951 952
    out = helper.create_tmp_variable(dtype=input.dtype)
    helper.append_op(
        type='cross_entropy',
        inputs={'X': [input],
                'Label': [label]},
        outputs={'Y': [out]},
F
fengjiayi 已提交
953
        attrs={"soft_label": soft_label})
Y
Yu Yang 已提交
954 955 956
    return out


F
fengjiayi 已提交
957
def square_error_cost(input, label):
Y
Yu Yang 已提交
958
    """
959 960
    **Square error cost layer**

961 962
    This layer accepts input predictions and target label and returns the
    squared error cost.
Y
ying 已提交
963

964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980
    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
981
        Variable: The tensor variable storing the element-wise squared error \
982
                  difference of input and label.
983 984 985 986 987 988 989 990

    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
991
    """
F
fengjiayi 已提交
992
    helper = LayerHelper('square_error_cost', **locals())
Y
Yu Yang 已提交
993 994 995 996 997 998 999 1000 1001
    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
1002 1003
        type='square', inputs={'X': [minus_out]},
        outputs={'Out': [square_out]})
Y
Yu Yang 已提交
1004 1005 1006 1007 1008 1009 1010
    return square_out


def chunk_eval(input,
               label,
               chunk_scheme,
               num_chunk_types,
F
fengjiayi 已提交
1011
               excluded_chunk_types=None):
Y
Yu Yang 已提交
1012
    """
Y
yangyaming 已提交
1013
    This function computes and outputs the precision, recall and
1014
    F1-score of chunk detection.
Y
Yu Yang 已提交
1015
    """
F
fengjiayi 已提交
1016
    helper = LayerHelper("chunk_eval", **locals())
Y
Yu Yang 已提交
1017 1018 1019 1020 1021

    # prepare output
    precision = helper.create_tmp_variable(dtype="float32")
    recall = helper.create_tmp_variable(dtype="float32")
    f1_score = helper.create_tmp_variable(dtype="float32")
1022 1023 1024
    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
1025 1026 1027 1028 1029 1030 1031 1032

    helper.append_op(
        type="chunk_eval",
        inputs={"Inference": [input],
                "Label": [label]},
        outputs={
            "Precision": [precision],
            "Recall": [recall],
1033 1034 1035 1036
            "F1-Score": [f1_score],
            "NumInferChunks": [num_infer_chunks],
            "NumLabelChunks": [num_label_chunks],
            "NumCorrectChunks": [num_correct_chunks]
Y
Yu Yang 已提交
1037 1038 1039
        },
        attrs={
            "num_chunk_types": num_chunk_types,
G
guosheng 已提交
1040 1041
            "chunk_scheme": chunk_scheme,
            "excluded_chunk_types": excluded_chunk_types or []
Y
Yu Yang 已提交
1042
        })
1043 1044
    return (precision, recall, f1_score, num_infer_chunks, num_label_chunks,
            num_correct_chunks)
Y
Yu Yang 已提交
1045 1046 1047 1048 1049 1050 1051 1052 1053


def sequence_conv(input,
                  num_filters,
                  filter_size=3,
                  filter_stride=1,
                  padding=None,
                  bias_attr=None,
                  param_attr=None,
1054
                  act=None):
Y
Yu Yang 已提交
1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090
    """
    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,
C
chengduoZH 已提交
1091 1092
           stride=1,
           padding=0,
Y
Yu Yang 已提交
1093 1094 1095
           groups=None,
           param_attr=None,
           bias_attr=None,
C
chengduoZH 已提交
1096
           use_cudnn=True,
1097
           use_mkldnn=False,
C
chengduoZH 已提交
1098
           act=None):
Y
Yu Yang 已提交
1099
    """
C
chengduoZH 已提交
1100 1101 1102
    **Convlution2D Layer**

    The convolution2D layer calculates the output based on the input, filter
1103 1104 1105
    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
1106 1107
    The details of convolution layer, please refer UFLDL's `convolution,
    <http://ufldl.stanford.edu/tutorial/supervised/FeatureExtractionUsingConvolution/>`_ .
1108 1109 1110
    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
1111

1112
    For each input :math:`X`, the equation is:
C
refine  
chengduoZH 已提交
1113

C
chengduoZH 已提交
1114 1115
    .. math::

C
refine  
chengduoZH 已提交
1116
        Out = \sigma (W \\ast X + b)
C
chengduoZH 已提交
1117

C
chengduoZH 已提交
1118
    In the above equation:
C
chengduoZH 已提交
1119

1120 1121 1122 1123 1124
    * :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.
1125 1126
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be
                   different.
C
chengduoZH 已提交
1127 1128 1129

    Example:

1130 1131 1132
        - Input:

          Input shape: $(N, C_{in}, H_{in}, W_{in})$
C
refine  
chengduoZH 已提交
1133

1134
          Filter shape: $(C_{out}, C_{in}, H_f, W_f)$
C
refine  
chengduoZH 已提交
1135

1136 1137
        - Output:
          Output shape: $(N, C_{out}, H_{out}, W_{out})$
C
refine  
chengduoZH 已提交
1138

C
chengduoZH 已提交
1139
        Where
1140 1141

        .. math::
C
chengduoZH 已提交
1142

C
chengduoZH 已提交
1143 1144
        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
1145 1146

    Args:
1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168
       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
1169 1170

    Returns:
G
guosheng 已提交
1171
        Variable: The tensor variable storing the convolution and \
C
chengduoZH 已提交
1172 1173
                  non-linearity activation result.

C
refine  
chengduoZH 已提交
1174
    Raises:
1175 1176
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
C
refine  
chengduoZH 已提交
1177

C
chengduoZH 已提交
1178 1179 1180
    Examples:
        .. code-block:: python

1181 1182 1183 1184
          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
1185 1186 1187 1188 1189
    """
    if stride is None:
        stride = [1, 1]

    num_channels = input.shape[1]
1190 1191

    l_type = 'conv2d'
X
xzl 已提交
1192 1193
    if (num_channels == groups and num_filters % num_channels == 0 and
            not use_cudnn):
1194
        l_type = 'depthwise_conv2d'
1195 1196 1197 1198

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

Y
Yu Yang 已提交
1199 1200 1201 1202 1203 1204 1205
    if groups is None:
        num_filter_channels = num_channels
    else:
        if num_channels % groups != 0:
            raise ValueError("num_channels must be divisible by groups.")
        num_filter_channels = num_channels / groups

C
chengduoZH 已提交
1206 1207 1208
    filter_size = utils.convert_to_list(filter_size, 2, 'filter_size')
    stride = utils.convert_to_list(stride, 2, 'stride')
    padding = utils.convert_to_list(padding, 2, 'padding')
C
chengduoZH 已提交
1209

C
chengduoZH 已提交
1210 1211
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228

    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(
1229
        type=l_type,
Y
Yu Yang 已提交
1230 1231 1232 1233 1234
        inputs={
            'Input': input,
            'Filter': filter_param,
        },
        outputs={"Output": pre_bias},
C
chengduoZH 已提交
1235 1236 1237 1238
        attrs={
            'strides': stride,
            'paddings': padding,
            'groups': groups,
1239 1240
            'use_cudnn': use_cudnn,
            'use_mkldnn': use_mkldnn
C
chengduoZH 已提交
1241
        })
Y
Yu Yang 已提交
1242 1243 1244 1245 1246 1247

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

    return helper.append_activation(pre_act)


F
fengjiayi 已提交
1248
def sequence_pool(input, pool_type):
Y
Yu Yang 已提交
1249
    """
Y
yangyaming 已提交
1250 1251 1252
    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
1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277

    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
1278

L
Luo Tao 已提交
1279 1280
    Args:
        input(variable): The input variable which is a LoDTensor.
Y
yangyaming 已提交
1281
        pool_type (string): The pooling type of sequence_pool.
L
Luo Tao 已提交
1282 1283 1284 1285 1286 1287 1288 1289
            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
1290

Y
yangyaming 已提交
1291
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1292 1293 1294 1295 1296
                              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
1297
    """
F
fengjiayi 已提交
1298
    helper = LayerHelper('sequence_pool', **locals())
Y
Yu Yang 已提交
1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309
    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
1310 1311 1312 1313 1314
    # 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
1315 1316 1317
    return pool_out


F
fengjiayi 已提交
1318
def sequence_first_step(input):
L
Luo Tao 已提交
1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332
    """
    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
1333

L
Luo Tao 已提交
1334 1335 1336 1337 1338 1339 1340 1341 1342
    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
1343

Y
yangyaming 已提交
1344
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1345 1346 1347
                              dtype='float32', lod_level=1)
             x_first_step = fluid.layers.sequence_first_step(input=x)
    """
1348 1349 1350
    return sequence_pool(input=input, pool_type="first")


F
fengjiayi 已提交
1351
def sequence_last_step(input):
L
Luo Tao 已提交
1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365
    """
    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
1366

L
Luo Tao 已提交
1367 1368 1369 1370 1371 1372 1373 1374 1375
    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
1376

Y
yangyaming 已提交
1377
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1378 1379 1380
                              dtype='float32', lod_level=1)
             x_last_step = fluid.layers.sequence_last_step(input=x)
    """
1381 1382 1383
    return sequence_pool(input=input, pool_type="last")


Y
Yu Yang 已提交
1384
def pool2d(input,
C
chengduoZH 已提交
1385 1386
           pool_size=-1,
           pool_type="max",
C
chengduoZH 已提交
1387 1388
           pool_stride=1,
           pool_padding=0,
C
caoying03 已提交
1389
           global_pooling=False,
C
chengduoZH 已提交
1390
           use_cudnn=True,
1391
           ceil_mode=False,
1392
           use_mkldnn=False,
C
caoying03 已提交
1393
           name=None):
Y
Yu Yang 已提交
1394 1395 1396 1397 1398 1399 1400 1401
    """
    This function adds the operator for pooling in 2 dimensions, using the
    pooling configurations mentioned in input parameters.
    """
    if pool_type not in ["max", "avg"]:
        raise ValueError(
            "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",
            str(pool_type))
C
chengduoZH 已提交
1402

C
chengduoZH 已提交
1403 1404 1405 1406 1407
    if global_pooling is False and pool_size == -1:
        raise ValueError(
            "When the global_pooling is False, pool_size must be passed "
            "and be a valid value. Received pool_size: " + str(pool_size))

C
chengduoZH 已提交
1408 1409 1410 1411
    pool_size = utils.convert_to_list(pool_size, 2, 'pool_size')
    pool_padding = utils.convert_to_list(pool_padding, 2, 'pool_padding')
    pool_stride = utils.convert_to_list(pool_stride, 2, 'pool_stride')

C
chengduoZH 已提交
1412 1413
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427

    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
1428
            "paddings": pool_padding,
1429
            "use_cudnn": use_cudnn,
1430 1431
            "ceil_mode": ceil_mode,
            "use_mkldnn": use_mkldnn
Y
Yu Yang 已提交
1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443
        })

    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
1444
               data_layout='NCHW',
1445 1446 1447
               name=None,
               moving_mean_name=None,
               moving_variance_name=None):
Y
Yu Yang 已提交
1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473
    """
    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(
1474
        attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
Y
Yu Yang 已提交
1475

1476 1477 1478
    mean = helper.create_parameter(
        attr=ParamAttr(
            name=moving_mean_name, initializer=Constant(0.0), trainable=False),
Q
QI JUN 已提交
1479
        shape=param_shape,
1480 1481 1482 1483 1484 1485 1486 1487
        dtype=input.dtype)
    mean.stop_gradient = True

    variance = helper.create_parameter(
        attr=ParamAttr(
            name=moving_variance_name,
            initializer=Constant(1.0),
            trainable=False),
Q
QI JUN 已提交
1488
        shape=param_shape,
1489 1490
        dtype=input.dtype)
    variance.stop_gradient = True
Y
Yu Yang 已提交
1491 1492 1493 1494 1495 1496

    # 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
1497 1498
    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
1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524

    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
1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536
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**

1537
    Assume feature vectors exist on dimensions
G
guosheng 已提交
1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557
    :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.
1558
        scale(bool): Whether to learn the adaptive gain :math:`g` after
G
guosheng 已提交
1559
            normalization.
1560
        shift(bool): Whether to learn the adaptive bias :math:`b` after
G
guosheng 已提交
1561
            normalization.
1562
        begin_norm_axis(bool): The normalization will be performed along
G
guosheng 已提交
1563
            dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`.
1564
        epsilon(float): The small value added to the variance to prevent
G
guosheng 已提交
1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595
            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
1596
    if shift:
G
guosheng 已提交
1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620
        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
1621
def beam_search_decode(ids, scores, name=None):
Y
Yu Yang 已提交
1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641
    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,
C
chengduoZH 已提交
1642 1643 1644
                     padding=0,
                     stride=1,
                     dilation=1,
C
caoying03 已提交
1645
                     param_attr=None,
C
chengduoZH 已提交
1646
                     use_cudnn=True,
C
caoying03 已提交
1647
                     name=None):
Y
Yu Yang 已提交
1648
    """
1649 1650 1651 1652 1653 1654 1655 1656
    **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
1657 1658
    layer, please refer to the following explanation and references
    `therein <http://www.matthewzeiler.com/wp-content/uploads/2017/07/cvpr2010.pdf>`_.
1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670

    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.
1671 1672
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be
                   different.
Y
Yu Yang 已提交
1673

1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686
    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
1687

1688 1689 1690 1691
        .. 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
1692 1693

    Args:
1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712
       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.
1713 1714
       param_attr(ParamAttr): The parameters to the Conv2d_transpose Layer.
                              Default: None
1715 1716 1717 1718
       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
1719 1720

    Returns:
1721 1722 1723
       Variable: The tensor variable storing the convolution transpose result.

    Raises:
1724 1725
       ValueError: If the shapes of input, filter_size, stride, padding and
                   groups mismatch.
1726 1727 1728 1729

    Examples:
       .. code-block:: python

1730 1731 1732 1733
          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
1734 1735 1736 1737 1738 1739
    """
    helper = LayerHelper("conv2d_transpose", **locals())
    if not isinstance(input, Variable):
        raise TypeError("Input of conv2d_transpose must be Variable")
    input_channel = input.shape[1]

C
chengduoZH 已提交
1740 1741 1742
    padding = utils.convert_to_list(padding, 2, 'padding')
    stride = utils.convert_to_list(stride, 2, 'stride')
    dilation = utils.convert_to_list(dilation, 2, 'dilation')
C
chengduoZH 已提交
1743

C
chengduoZH 已提交
1744 1745 1746
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")

Y
Yu Yang 已提交
1747 1748 1749 1750 1751 1752 1753 1754
    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]

        h_in = input.shape[2]
        w_in = input.shape[3]
C
chengduoZH 已提交
1755 1756 1757 1758 1759

        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
1760
        filter_size = [filter_size_h, filter_size_w]
C
chengduoZH 已提交
1761 1762 1763
    else:
        filter_size = utils.convert_to_list(filter_size, 2,
                                            'conv2d_transpose.filter_size')
Y
Yu Yang 已提交
1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774

    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},
C
chengduoZH 已提交
1775 1776 1777 1778 1779 1780
        attrs={
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
            'use_cudnn': use_cudnn
        })
Y
Yu Yang 已提交
1781 1782

    return out
Y
yangyaming 已提交
1783 1784


C
caoying03 已提交
1785
def sequence_expand(x, y, name=None):
1786 1787
    """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
1788
    explain how sequence_expand works:
1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816

    .. 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
1817
                y.lod = [[0, 2, 3, 6]]
1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828

            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
1829 1830
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
1831 1832 1833 1834 1835 1836 1837 1838 1839 1840

    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
1841
            out = layers.sequence_expand(x=x, y=y)
1842
    """
Y
yangyaming 已提交
1843
    helper = LayerHelper('sequence_expand', input=x, **locals())
1844 1845 1846
    dtype = helper.input_dtype()
    tmp = helper.create_tmp_variable(dtype)
    helper.append_op(
Y
yangyaming 已提交
1847 1848
        type='sequence_expand', inputs={'X': x,
                                        'Y': y}, outputs={'Out': tmp})
1849
    return tmp
1850 1851


Q
Qiao Longfei 已提交
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 1878 1879 1880 1881 1882 1883
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
1884 1885 1886 1887
def lstm_unit(x_t,
              hidden_t_prev,
              cell_t_prev,
              forget_bias=0.0,
Y
yangyaming 已提交
1888
              param_attr=None,
C
caoying03 已提交
1889 1890
              bias_attr=None,
              name=None):
Y
yangyaming 已提交
1891 1892 1893 1894
    """Lstm unit layer. The equation of a lstm step is:

        .. math::

1895
            i_t & = \sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i)
Y
yangyaming 已提交
1896

1897
            f_t & = \sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + b_f)
Y
yangyaming 已提交
1898

1899
            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
1900

1901
            o_t & = \sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + b_o)
Y
yangyaming 已提交
1902 1903 1904

            h_t & = o_t tanh(c_t)

1905 1906 1907 1908 1909 1910
    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
1911 1912 1913

        .. math::

1914
            L_{i_t} = W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i
Y
yangyaming 已提交
1915 1916 1917 1918 1919 1920 1921 1922

    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
1923
    This layer has two outputs including :math:`h_t` and :math:`o_t`.
Y
yangyaming 已提交
1924 1925

    Args:
Y
yangyaming 已提交
1926 1927 1928 1929 1930 1931
        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
1932
        forget_bias (float): The forget bias of lstm unit.
Y
yangyaming 已提交
1933 1934
        param_attr (ParamAttr): The attributes of parameter weights, used to set
            initializer, name etc.
Y
yangyaming 已提交
1935 1936
        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
1937 1938
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
Y
yangyaming 已提交
1939 1940

    Returns:
Y
yangyaming 已提交
1941
        tuple: The hidden value and cell value of lstm unit.
Y
yangyaming 已提交
1942 1943

    Raises:
1944 1945 1946 1947
        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
1948 1949 1950 1951 1952 1953

    Examples:

        .. code-block:: python

             x_t = fluid.layers.fc(input=x_t_data, size=10)
1954
             prev_hidden = fluid.layers.fc(input=prev_hidden_data, size=30)
Y
yangyaming 已提交
1955
             prev_cell = fluid.layers.fc(input=prev_cell_data, size=30)
Y
yangyaming 已提交
1956
             hidden_value, cell_value = fluid.layers.lstm_unit(x_t=x_t,
Y
yangyaming 已提交
1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972
                                                    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
1973
        raise ValueError("The 1st dimensions of x_t, hidden_t_prev and "
1974 1975 1976 1977
                         "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
1978 1979
                         "cell_t_prev must be the same.")

Y
yangyaming 已提交
1980 1981 1982
    if bias_attr is None:
        bias_attr = ParamAttr()

Y
yangyaming 已提交
1983
    size = cell_t_prev.shape[1]
1984
    concat_out = concat(input=[x_t, hidden_t_prev], axis=1)
Y
yangyaming 已提交
1985 1986
    fc_out = fc(input=concat_out,
                size=4 * size,
Y
yangyaming 已提交
1987
                param_attr=param_attr,
1988
                bias_attr=bias_attr)
Y
yangyaming 已提交
1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000
    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
2001
    return h, c
G
guosheng 已提交
2002 2003


C
caoying03 已提交
2004
def reduce_sum(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
2005
    """
Y
yangyaming 已提交
2006
    Computes the sum of tensor elements over the given dimension.
G
guosheng 已提交
2007 2008 2009

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
2010 2011 2012 2013
        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
2014
            the dimension to reduce is :math:`rank + dim`.
2015
        keep_dim (bool|False): Whether to reserve the reduced dimension in the
Y
yangyaming 已提交
2016
            output Tensor. The result tensor will have one fewer dimension
G
guosheng 已提交
2017
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
2018 2019
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
2020 2021 2022

    Returns:
        Variable: The reduced Tensor variable.
F
fengjiayi 已提交
2023

G
guosheng 已提交
2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047
    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
2048 2049


C
caoying03 已提交
2050
def reduce_mean(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
2051
    """
Y
yangyaming 已提交
2052
    Computes the mean of tensor elements over the given dimension.
G
guosheng 已提交
2053 2054 2055

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
2056 2057 2058 2059
        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
2060
            :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`.
Y
yangyaming 已提交
2061 2062
        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
2063
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
2064 2065
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
2066 2067 2068

    Returns:
        Variable: The reduced Tensor variable.
F
fengjiayi 已提交
2069

G
guosheng 已提交
2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093
    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
2094 2095


C
caoying03 已提交
2096
def reduce_max(input, dim=None, keep_dim=False, name=None):
2097
    """
Y
yangyaming 已提交
2098
    Computes the maximum of tensor elements over the given dimension.
2099 2100 2101

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
2102 2103 2104 2105
        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))`.
2106
            If :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`.
Y
yangyaming 已提交
2107 2108
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
2109
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
2110 2111
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
2112 2113 2114

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
2115

2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141
    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
2142
def reduce_min(input, dim=None, keep_dim=False, name=None):
2143
    """
Y
yangyaming 已提交
2144
    Computes the minimum of tensor elements over the given dimension.
2145 2146 2147

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
2148 2149 2150 2151
        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))`.
2152
            If :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`.
Y
yangyaming 已提交
2153 2154
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
2155
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
2156 2157
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
2158 2159 2160

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
2161

2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185
    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
2186 2187


2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201
def reduce_prod(input, dim=None, keep_dim=False, name=None):
    """
    Computes the product of tensor elements over the given dimension.

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
        dim (int|None): The dimension along which the product is performed. If
            :attr:`None`, multipy 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`,
            the dimension to reduce is :math:`rank + dim`.
        keep_dim (bool|False): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
            than the :attr:`input` unless :attr:`keep_dim` is true.
Z
zhouhanqing 已提交
2202 2203
        name(str|None): A name for this layer(optional). If set None, the 
            layer will be named automatically.
2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217

    Returns:
        Variable: The reduced Tensor variable.

    Examples:
        .. code-block:: python

            # x is a Tensor variable with following elements:
            #    [[0.2, 0.3, 0.5, 0.9]
            #     [0.1, 0.2, 0.6, 0.7]]
            # Each example is followed by the correspending output tensor.
            fluid.layers.reduce_prod(x)  # [0.0002268]
            fluid.layers.reduce_prod(x, dim=0)  # [0.02, 0.06, 0.3, 0.63]
            fluid.layers.reduce_prod(x, dim=-1)  # [0.027, 0.0084]
Z
zhouhanqing 已提交
2218 2219
            fluid.layers.reduce_prod(x, dim=1, 
                                     keep_dim=True)  # [[0.027], [0.0084]]
2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234
    """
    helper = LayerHelper('reduce_prod', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
        type='reduce_prod',
        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
2235
def split(input, num_or_sections, dim=-1, name=None):
G
guosheng 已提交
2236
    """
C
caoying03 已提交
2237
    Split the input tensor into multiple sub-tensors.
G
guosheng 已提交
2238 2239 2240

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
C
caoying03 已提交
2241 2242 2243 2244 2245
        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
2246
            :attr:`dim` dimension orderly.
C
caoying03 已提交
2247
        dim (int): The dimension along which to split. If :math:`dim < 0`, the
G
guosheng 已提交
2248
            dimension to split along is :math:`rank(input) + dim`.
C
caoying03 已提交
2249 2250
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 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

    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
2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325


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
2326
          normed = fluid.layers.l2_normalize(x=data, axis=1)
C
caoying03 已提交
2327 2328
    """

F
fengjiayi 已提交
2329 2330
    if len(x.shape) == 1:
        axis = 0
C
caoying03 已提交
2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356

    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
2357
    # exanpsion is requred. Once a general broadcast rule is spported, this
C
caoying03 已提交
2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374
    # 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
2375 2376


2377
def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
G
guosheng 已提交
2378
    """
Y
ying 已提交
2379 2380 2381 2382
    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
2383

C
chengduoZH 已提交
2384
    The actual behavior depends on the shapes of :math:`x`, :math:`y` and the
2385
    flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically:
G
guosheng 已提交
2386

2387 2388 2389 2390 2391
    - 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
2392
      :math:`[1, D]` in transposed form.
G
guosheng 已提交
2393

C
chengduoZH 已提交
2394
    - After transpose, the two tensors are 2-D or n-D and matrix multiplication
2395
      performs in the following way.
G
guosheng 已提交
2396

2397
      - If both are 2-D, they are multiplied like conventional matrices.
C
chengduoZH 已提交
2398
      - If either is n-D, it is treated as a stack of matrices residing in the
Y
ying 已提交
2399
        last two dimensions and a batched matrix multiply supporting broadcast
2400
        applies on the two tensors.
G
guosheng 已提交
2401

Y
ying 已提交
2402 2403
    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
2404
    removed after matrix multiplication.
G
guosheng 已提交
2405 2406 2407

    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
2408 2409 2410
        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.
2411
        name(str|None): A name for this layer(optional). If set None, the layer
2412
            will be named automatically.
G
guosheng 已提交
2413 2414

    Returns:
2415
        Variable: The product Tensor variable.
G
guosheng 已提交
2416

G
guosheng 已提交
2417 2418 2419
    Examples:
        .. code-block:: python

2420
            # Examples to clarify shapes of the inputs and output
C
chengduoZH 已提交
2421 2422
            # x: [B, ..., M, K], y: [B, ..., K, N]
            fluid.layers.matmul(x, y)  # out: [B, ..., M, N]
Y
ying 已提交
2423

2424 2425
            # x: [B, M, K], y: [B, K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
2426

2427 2428
            # x: [B, M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
2429

2430 2431
            # x: [M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [M, N]
Y
ying 已提交
2432 2433 2434 2435

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

2436 2437
            # x: [K], y: [K]
            fluid.layers.matmul(x, y)  # out: [1]
2438

Y
ying 已提交
2439
            # x: [M], y: [N]
2440
            fluid.layers.matmul(x, y, True, True)  # out: [M, N]
G
guosheng 已提交
2441
    """
Y
ying 已提交
2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453

    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
2454
            y_shape = y_shape + [1]
Y
ying 已提交
2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470

        # 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)

2471
    helper = LayerHelper('matmul', **locals())
Y
ying 已提交
2472
    out = helper.create_tmp_variable(dtype=x.dtype)
G
guosheng 已提交
2473
    helper.append_op(
2474 2475 2476 2477 2478 2479 2480
        type='matmul',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
        attrs={'transpose_X': transpose_x,
               'transpose_Y': transpose_y})
    return out
2481 2482


W
wanghaoshuang 已提交
2483
def edit_distance(input, label, normalized=True, ignored_tokens=None,
W
wanghaoshuang 已提交
2484
                  name=None):
2485
    """
Y
ying 已提交
2486 2487 2488 2489 2490 2491 2492 2493 2494
    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
2495

Y
ying 已提交
2496
    "kitten" -> "sitten" -> "sittin" -> "sitting"
W
wanghaoshuang 已提交
2497

Y
ying 已提交
2498 2499 2500 2501
    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
2502

Y
ying 已提交
2503 2504 2505
    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
2506

2507 2508 2509 2510 2511
    Args:

        input(Variable): The indices for hypothesis strings.

        label(Variable): The indices for reference strings.
W
wanghaoshuang 已提交
2512

Y
ying 已提交
2513 2514
        normalized(bool): Indicated whether to normalize the edit distance by
                          the length of reference string.
2515

Y
ying 已提交
2516 2517
        ignored_tokens(list of int): Tokens that should be removed before
                                     calculating edit distance.
2518

W
wanghaoshuang 已提交
2519
    Returns:
W
wanghaoshuang 已提交
2520
        Variable: sequence-to-sequence edit distance in shape [batch_size, 1].
W
wanghaoshuang 已提交
2521 2522 2523 2524 2525

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
2526 2527
            y = fluid.layers.data(name='y', shape=[7], dtype='float32')

2528
            cost = fluid.layers.edit_distance(input=x,label=y)
2529
    """
2530
    helper = LayerHelper("edit_distance", **locals())
2531

2532
    # remove some tokens from input and labels
W
wanghaoshuang 已提交
2533
    if ignored_tokens is not None and len(ignored_tokens) > 0:
2534 2535 2536 2537 2538 2539 2540
        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
2541
            attrs={"tokens": ignored_tokens})
2542 2543 2544 2545 2546 2547
        input = erased_input

        helper.append_op(
            type="sequence_erase",
            inputs={"X": [label]},
            outputs={"Out": [erase_label]},
W
wanghaoshuang 已提交
2548
            attrs={"tokens": ignored_tokens})
2549 2550
        label = erased_label

2551 2552
    # edit distance op
    edit_distance_out = helper.create_tmp_variable(dtype="int64")
2553
    sequence_num = helper.create_tmp_variable(dtype="int64")
2554 2555 2556 2557
    helper.append_op(
        type="edit_distance",
        inputs={"Hyps": [input],
                "Refs": [label]},
2558 2559
        outputs={"Out": [edit_distance_out],
                 "SequenceNum": [sequence_num]},
2560 2561
        attrs={"normalized": normalized})

2562
    return edit_distance_out, sequence_num
2563 2564 2565 2566 2567


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
2568 2569 2570 2571
    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.
2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600

    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
2601 2602 2603 2604 2605 2606
        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).
2607

Y
ying 已提交
2608 2609 2610
        blank(int): the blank label index of Connectionist Temporal
                    Classification (CTC) loss, which is in thehalf-opened
                    interval [0, num_classes + 1).
2611 2612

    Returns:
2613
        Variable: CTC greedy decode result. If all the sequences in result were
2614
        empty, the result LoDTensor will be [-1] with LoD [[0]] and dims [1, 1].
2615 2616 2617 2618 2619

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
W
wanghaoshuang 已提交
2620

2621
            cost = fluid.layers.ctc_greedy_decoder(input=x, blank=0)
W
wanghaoshuang 已提交
2622
    """
2623
    helper = LayerHelper("ctc_greedy_decoder", **locals())
2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638
    # 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
2639
        outputs={"Output": [ctc_out]},
2640 2641
        attrs={"merge_repeated": True,
               "blank": blank})
2642
    return ctc_out
2643 2644


F
fengjiayi 已提交
2645
def warpctc(input, label, blank=0, norm_by_times=False):
W
wanghaoshuang 已提交
2646
    """
2647 2648
    An operator integrating the open source Warp-CTC library
    (https://github.com/baidu-research/warp-ctc)
W
wanghaoshuang 已提交
2649
    to compute Connectionist Temporal Classification (CTC) loss.
2650 2651
    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
2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664
    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.
2665
       blank: (int, default: 0), the blank label index of Connectionist
W
wanghaoshuang 已提交
2666 2667
         Temporal Classification (CTC) loss, which is in the
         half-opened interval [0, num_classes + 1).
2668
       norm_by_times: (bool, default: false), whether to normalize
W
wanghaoshuang 已提交
2669
       the gradients by the number of time-step, which is also the
2670 2671
       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
2672 2673

    Returns:
2674 2675
        Variable: The Connectionist Temporal Classification (CTC) loss,
        which is a 2-D Tensor of the shape [batch_size, 1].
W
wanghaoshuang 已提交
2676 2677 2678

    Examples:
        .. code-block:: python
2679 2680 2681 2682
            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
2683 2684 2685
            cost = layers.warpctc(input=y_predict, label=y)

    """
F
fengjiayi 已提交
2686
    helper = LayerHelper('warpctc', **locals())
W
wanghaoshuang 已提交
2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697
    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
2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 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


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
2752 2753


2754
@autodoc()
Y
Yang Yu 已提交
2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780
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
2781 2782 2783 2784 2785 2786 2787 2788 2789
    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
2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805

    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
2806
    return cost / (num_neg_samples + 1)
2807 2808


Y
fix ci.  
ying 已提交
2809
def transpose(x, perm, name=None):
Y
ying 已提交
2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828
    """
    **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
2829
            x_transposed = layers.transpose(x, perm=[1, 0, 2])
Y
ying 已提交
2830 2831
    """

Y
fix ci.  
ying 已提交
2832
    if len(perm) != len(x.shape):
Y
ying 已提交
2833 2834 2835
        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
2836 2837 2838 2839 2840 2841
    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
2842 2843

    helper = LayerHelper('transpose', **locals())
Y
fix ci.  
ying 已提交
2844
    out = helper.create_tmp_variable(x.dtype)
Y
ying 已提交
2845 2846
    helper.append_op(
        type='transpose',
Y
fix ci.  
ying 已提交
2847
        inputs={'X': [x]},
Y
ying 已提交
2848 2849 2850
        outputs={'Out': [out]},
        attrs={'axis': perm})
    return out
2851 2852


2853
def im2sequence(input, filter_size=1, stride=1, padding=0, name=None):
2854
    """
2855 2856 2857 2858 2859 2860 2861
    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:
2862 2863 2864 2865 2866 2867 2868 2869 2870 2871

    .. 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
2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889

        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.

2890 2891 2892
        name (int): The name of this layer. It is optional.

    Returns:
W
wanghaoshuang 已提交
2893 2894 2895 2896 2897
        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.
2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926

    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
2927 2928 2929
            filter = [2, 2]
            stride = [1, 1]
            padding = [0, 0]
2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949

            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

2950 2951
            output = fluid.layers.im2sequence(
                input=layer, stride=[1, 1], filter_size=[2, 2])
2952 2953

    """
W
wanghaoshuang 已提交
2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964

    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])

2965
    helper = LayerHelper('im2sequence', **locals())
2966 2967
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
2968
        type='im2sequence',
2969 2970 2971
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
wanghaoshuang 已提交
2972 2973 2974
            'kernels': filter_size,
            'strides': stride,
            'paddings': padding,
2975 2976
        })
    return out
2977 2978


2979 2980 2981 2982
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
2983
    equation of row convolution is as follows:
2984 2985 2986 2987 2988 2989 2990

    .. 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
2991
    * :math:`\\tau`: Future context size.
2992 2993 2994 2995 2996 2997 2998 2999 3000 3001
    * :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
3002 3003
        future_context_size (int): Future context size. Please note, the shape
            of convolution kernel is [future_context_size + 1, D].
3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028
        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
3029
    return helper.append_activation(out)
3030 3031


3032 3033 3034 3035
def multiplex(inputs, index):
    """
    **Multiplex Layer**

Y
yangyaming 已提交
3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050
    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]`.
3051 3052

    Args:
Y
yangyaming 已提交
3053 3054
       inputs (list): A list of variables to gather from. All variables have the
                same shape and the rank is at least 2.
3055
       index (Variable): Tensor<int32>, index variable which is a 2-D tensor
Y
yangyaming 已提交
3056
                with shape [M, 1] where M is the batch size.
3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069

    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
3070 3071 3072 3073 3074 3075

    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)
3076 3077 3078 3079 3080 3081
    helper.append_op(
        type='multiplex',
        inputs={'X': inputs,
                'Ids': index},
        outputs={'Out': [out]})
    return out
3082 3083 3084 3085 3086


def softmax_with_cross_entropy(logits, label, soft_label=False):
    """
    **Softmax With Cross Entropy Operator.**
3087

3088 3089 3090 3091
    Cross entropy loss with softmax is used as the output layer extensively. This
    operator computes the softmax normalized values for each row of the input
    tensor, after which cross-entropy loss is computed. This provides a more
    numerically stable gradient.
3092

3093 3094 3095
    Because this operator performs a softmax on logits internally, it expects
    unscaled logits. This operator should not be used with the output of
    softmax operator since that would produce incorrect results.
3096

3097 3098 3099
    When the attribute soft_label is set false, this operators expects mutually
    exclusive hard labels, each sample in a batch is in exactly one class with a
    probability of 1.0. Each sample in the batch will have a single label.
3100

3101
    The equation is as follows:
3102

3103
    1) Hard label (one-hot label, so every sample has exactly one class)
3104

3105 3106 3107 3108
    .. math::

        loss_j =  -\\text{logit}_{label_j} +
        \\log\\left(\\sum_{i=0}^{K}\\exp(\\text{logit}_i)\\right), j = 1,..., K
3109

3110 3111 3112
    2) Soft label (each sample can have a distribution over all classes)

    .. math::
3113

3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157
        loss_j =  -\\sum_{i=0}^{K}\\text{label}_i
        \\left(\\text{logit}_i - \\log\\left(\\sum_{i=0}^{K}
        \\exp(\\text{logit}_i)\\right)\\right), j = 1,...,K

    Args:
        logits (Variable): The unscaled log probabilities, which is a 2-D tensor
            with shape [N x K]. N is the batch_size, and K is the class number.
        label (Variable): The ground truth which is a 2-D tensor. If soft_label
            is set to false, Label is a Tensor<int64> with shape [N x 1]. If
            soft_label is set to true, Label is a Tensor<float/double> with
        soft_label (bool): A flag to indicate whether to interpretate the given
            labels as soft labels. By default, `soft_label` is set to False.
    Returns:
        Variable: The cross entropy loss is a 2-D tensor with shape [N x 1].

    Examples:
        .. code-block:: python

            data = fluid.layers.data(name='data', shape=[128], dtype='float32')
            label = fluid.layers.data(name='label', shape=[1], dtype='int64')
            fc = fluid.layers.fc(input=data, size=100)
            out = fluid.layers.softmax_with_cross_entropy(logits=fc, label=label)
    """
    helper = LayerHelper('softmax_with_cross_entropy', **locals())
    softmax = helper.create_tmp_variable(dtype=logits.dtype)
    loss = helper.create_tmp_variable(dtype=logits.dtype)
    helper.append_op(
        type='softmax_with_cross_entropy',
        inputs={'Logits': logits,
                'Label': label},
        outputs={'Softmax': softmax,
                 'Loss': loss},
        attrs={'soft_label': soft_label})
    return loss


def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None):
    """
    **Smooth L1 Loss Operator. **

    This operator computes the smooth l1 loss for X and Y.
    The operator takes the first dimension of X and Y as batch size.
    For each instance, it computes the smooth l1 loss element by element first
    and then sums all the losses. So the shape of Out is [batch_size, 1].
3158

3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182 3183
    Args:
        x (Variable): A tensor with rank at least 2. The input value of smooth
            l1 loss op with shape [batch_size, dim1, ..., dimN].
        y (Variable): A tensor with rank at least 2. The target value of smooth
            l1 loss op with same shape as x.
        inside_weight (Variable|None):  A tensor with rank at least 2. This
            input is optional and should have same shape with x. If provided,
            the result of (x - y) will be multiplied by this tensor element by
            element.
        outside_weight (Variable|None): A tensor with rank at least 2. This
            input is optional and should have same shape with x. If provided,
            the out smooth l1 loss will be multiplied by this tensor element
            by element.
        sigma (float|None): Hyper parameter of smooth l1 loss op. A float scalar
            with default value 1.0.
    Returns:
        Variable: A tensor with rank be 2. The output smooth l1 loss with
            shape [batch_size, 1].

    Examples:
        .. code-block:: python

            data = fluid.layers.data(name='data', shape=[128], dtype='float32')
            label = fluid.layers.data(name='label', shape=[100], dtype='int64')
            fc = fluid.layers.fc(input=data, size=100)
F
fengjiayi 已提交
3184
            out = fluid.layers.smooth_l1(x=fc, y=label)
3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200
    """
    helper = LayerHelper('smooth_l1_loss', **locals())
    diff = helper.create_tmp_variable(dtype=x.dtype)
    loss = helper.create_tmp_variable(dtype=x.dtype)
    helper.append_op(
        type='smooth_l1_loss',
        inputs={
            'X': x,
            'Y': y,
            'InsideWeight': inside_weight,
            'OutsideWeight': outside_weight
        },
        outputs={'Diff': diff,
                 'Out': loss},
        attrs={'sigma': sigma})
    return loss
3201 3202 3203 3204 3205 3206 3207 3208 3209


def one_hot(input, depth):
    """
    One Hot Operator. This operator creates the one-hot representations for input
    index values. The following example will help to explain the function of this
    operator.

    Args:
F
fengjiayi 已提交
3210
        input(variable):  A Tensor/LodTensor of indices, last dimension must be 1.
3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237
        depth(scalar): an interger defining the depth of the one hot dimension.

    Returns:
         The one-hot tensor or LodTensor, same as input.

    Examples:
        X is a LoDTensor:
          X.lod = [[0, 1, 4]]
          X.shape = [4, 1]
          X.data = [[1], [1], [3], [0]]
        set depth = 4
        Out is a LoDTensor:
          Out.lod = [[0, 1, 4]]
          Out.shape = [4, 4]
          Out.data = [[0., 1., 0., 0.],
                      [0., 1., 0., 0.],
                      [0., 0., 0., 1.],
                      [1., 0., 0., 0.]]
    """
    helper = LayerHelper("one_hot", **locals())
    one_hot_out = helper.create_tmp_variable(dtype='float32')
    helper.append_op(
        type="one_hot",
        inputs={'X': input},
        attrs={'depth': depth},
        outputs={'Out': one_hot_out})
    return one_hot_out
Y
Yu Yang 已提交
3238 3239


Y
Yu Yang 已提交
3240
def autoincreased_step_counter(counter_name=None, begin=1, step=1):
Y
Yu Yang 已提交
3241
    """
Y
Yu Yang 已提交
3242
    NOTE: The counter will be automatically increased by 1 every mini-batch
Y
Yu Yang 已提交
3243
    Return the run counter of the main program, which is started with 1.
Y
Yu Yang 已提交
3244 3245 3246 3247 3248 3249

    Args:
        counter_name(str): The counter name, default is '@STEP_COUNTER@'.
        begin(int): The first value of this counter.
        step(int): The increment step between each execution.

Y
Yu Yang 已提交
3250 3251 3252
    Returns(Variable): The global run counter.
    """
    helper = LayerHelper('global_step_counter')
Y
Yu Yang 已提交
3253 3254
    if counter_name is None:
        counter_name = '@STEP_COUNTER@'
Y
Yu Yang 已提交
3255 3256 3257 3258 3259
    counter, is_new_var = helper.create_or_get_global_variable(
        name=counter_name, dtype='int64', shape=[1], persistable=True)
    if is_new_var:
        helper.set_variable_initializer(
            counter, initializer=Constant(
Y
Yu Yang 已提交
3260
                value=begin - 1, force_cpu=True))
Y
Yu Yang 已提交
3261 3262 3263
        helper.main_program.global_block().prepend_op(
            type='increment',
            inputs={'X': [counter]},
Y
Yu Yang 已提交
3264 3265
            outputs={'Out': [counter]},
            attrs={'step': float(step)})
Y
Yu Yang 已提交
3266 3267 3268
        counter.stop_gradient = True

    return counter