nn.py 137.0 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
    'gru_unit',
    'linear_chain_crf',
    'crf_decoding',
    'cos_sim',
    'cross_entropy',
    'square_error_cost',
    'chunk_eval',
    'sequence_conv',
    'conv2d',
    'sequence_pool',
42 43
    'sequence_softmax',
    'softmax',
Y
ying 已提交
44 45 46 47 48 49 50 51 52 53
    'pool2d',
    'batch_norm',
    'beam_search_decode',
    'conv2d_transpose',
    'sequence_expand',
    'lstm_unit',
    'reduce_sum',
    'reduce_mean',
    'reduce_max',
    'reduce_min',
54
    'reduce_prod',
Y
ying 已提交
55 56 57 58
    'sequence_first_step',
    'sequence_last_step',
    'dropout',
    'split',
59 60
    'ctc_greedy_decoder',
    'edit_distance',
Y
ying 已提交
61 62 63 64
    'l2_normalize',
    'matmul',
    'warpctc',
    'sequence_reshape',
65
    'transpose',
66
    'im2sequence',
67
    'nce',
Q
Qiao Longfei 已提交
68
    'beam_search',
69
    'row_conv',
70
    'multiplex',
G
guosheng 已提交
71
    'layer_norm',
72 73
    'softmax_with_cross_entropy',
    'smooth_l1',
74
    'one_hot',
Y
Yu Yang 已提交
75
    'autoincreased_step_counter',
C
caoying03 已提交
76
    'reshape',
Y
yangyaming 已提交
77
    'lod_reset',
D
dragonwarrior 已提交
78
    'lrn',
G
guosheng 已提交
79
    'pad',
80
    'label_smooth',
Y
Yu Yang 已提交
81 82 83 84 85 86 87 88
]


def fc(input,
       size,
       num_flatten_dims=1,
       param_attr=None,
       bias_attr=None,
89
       use_mkldnn=False,
Y
Yu Yang 已提交
90
       act=None,
91
       name=None):
Y
Yu Yang 已提交
92
    """
93
    **Fully Connected Layer**
Y
Yu Yang 已提交
94

C
caoying03 已提交
95
    The fully connected layer can take multiple tensors as its inputs. It
R
ranqiu 已提交
96 97 98 99 100 101
    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
102
    if activation is not None, it will be applied to the output as well.
C
caoying03 已提交
103

C
caoying03 已提交
104
    This process can be formulated as follows:
105 106 107

    .. math::

108
        Out = Act({\sum_{i=0}^{N-1}X_iW_i + b})
109 110 111

    In the above equation:

C
caoying03 已提交
112 113 114 115
    * :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).
116
    * :math:`Act`: The activation function.
C
caoying03 已提交
117
    * :math:`Out`: The output tensor.
Y
Yu Yang 已提交
118 119

    Args:
R
ranqiu 已提交
120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136
        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.
M
mozga-intel 已提交
137 138
        use_mkldnn(bool): Use mkldnn kernel or not, it is valid only when the mkldnn
            library is installed. Default: False
R
ranqiu 已提交
139
        name (str, default None): The name of this layer.
Y
Yu Yang 已提交
140

141
    Returns:
R
ranqiu 已提交
142
        A tensor variable storing the transformation result.
143 144

    Raises:
C
caoying03 已提交
145
        ValueError: If rank of the input tensor is less than 2.
146 147 148 149

    Examples:
        .. code-block:: python

C
caoying03 已提交
150
          data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
151
          fc = fluid.layers.fc(input=data, size=1000, act="tanh")
Y
Yu Yang 已提交
152
    """
C
caoying03 已提交
153

C
caoying03 已提交
154
    helper = LayerHelper("fc", **locals())
Y
Yu Yang 已提交
155 156 157 158

    dtype = helper.input_dtype()

    mul_results = []
159 160 161
    if use_mkldnn:
        tmp = helper.create_tmp_variable(dtype)
        input_shape = input.shape
Y
Yu Yang 已提交
162 163 164
        param_shape = [
            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
        ] + [size]
Y
ying 已提交
165

Y
Yu Yang 已提交
166
        w = helper.create_parameter(
167 168 169 170
            attr=helper.param_attr,
            shape=param_shape,
            dtype=dtype,
            is_bias=False)
M
mozga-intel 已提交
171 172 173
        if bias_attr is None or bias_attr is False:
            bias_attr = False
        else:
174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192
            bias_attr = True
        helper.append_op(
            type="fc",
            inputs={"Input": input,
                    "W": w},
            outputs={"Out": tmp},
            attrs={"use_mkldnn": use_mkldnn,
                   "bias_attr": bias_attr})
        return helper.append_activation(tmp)
    else:
        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]

            w = helper.create_parameter(
                attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False)
            tmp = helper.create_tmp_variable(dtype)
M
mozga-intel 已提交
193 194 195 196 197 198 199 200 201
            helper.append_op(
                type="mul",
                inputs={"X": input_var,
                        "Y": w},
                outputs={"Out": tmp},
                attrs={
                    "x_num_col_dims": num_flatten_dims,
                    "y_num_col_dims": 1,
                })
202 203 204 205
            mul_results.append(tmp)

        if len(mul_results) == 1:
            pre_bias = mul_results[0]
M
mozga-intel 已提交
206
        else:
207
            pre_bias = helper.create_tmp_variable(dtype)
M
mozga-intel 已提交
208
            helper.append_op(
209 210 211 212 213 214 215 216
                type="sum",
                inputs={"X": mul_results},
                outputs={"Out": pre_bias})
        # add bias
        pre_activation = helper.append_bias_op(
            pre_bias, dim_start=num_flatten_dims)
        # add activation
        return helper.append_activation(pre_activation)
Y
Yu Yang 已提交
217 218


219 220 221
def embedding(input,
              size,
              is_sparse=False,
222
              is_distributed=False,
223 224 225
              padding_idx=None,
              param_attr=None,
              dtype='float32'):
Y
Yu Yang 已提交
226
    """
227 228
    **Embedding Layer**

229
    This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in
230 231
    a lookup table. The result of this lookup is the embedding of each ID in the
    :attr:`input`.
232 233 234

    All the input variables are passed in as local variables to the LayerHelper
    constructor.
Y
Yu Yang 已提交
235 236

    Args:
237 238 239 240 241 242 243
        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
244 245
            with zeros whenever lookup encounters it in :attr:`input`. If
            :math:`padding_idx < 0`, the padding_idx to use in lookup is
246 247
            :math:`size[0] + dim`.
        param_attr(ParamAttr): Parameters for this layer
248
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc
Y
Yu Yang 已提交
249

250 251 252
    Returns:
        Variable: The tensor variable storing the embeddings of the \
                  supplied inputs.
Y
Yu Yang 已提交
253

254 255
    Examples:
        .. code-block:: python
Y
Yu Yang 已提交
256

C
chengduoZH 已提交
257
          dict_size = len(dataset.ids)
258
          data = fluid.layers.data(name='ids', shape=[32, 32], dtype='float32')
C
chengduoZH 已提交
259
          fc = fluid.layers.embedding(input=data, size=[dict_size, 16])
Y
Yu Yang 已提交
260 261 262 263 264 265
    """

    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)
266 267
    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
268 269 270 271 272
    helper.append_op(
        type='lookup_table',
        inputs={'Ids': input,
                'W': w},
        outputs={'Out': tmp},
273 274 275 276 277
        attrs={
            'is_sparse': is_sparse,
            'is_distributed': is_distributed,
            'padding_idx': padding_idx
        })
Y
Yu Yang 已提交
278 279 280 281 282 283 284 285 286 287 288 289 290
    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',
291 292
                 dtype='float32',
                 name=None):
Y
Yibing Liu 已提交
293 294 295 296 297 298
    """
    **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
299
    .. math::
Y
Yibing Liu 已提交
300

301
        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
302

303
        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
304

305
        \\tilde{c_t} & = act_g(W_{cx}x_t + W_{ch}h_{t-1} + b_c)
Y
Yibing Liu 已提交
306

307 308 309
        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
310

Y
Yibing Liu 已提交
311
        h_t & = o_t \odot act_h(c_t)
Y
Yibing Liu 已提交
312

313
    where the :math:`W` terms denote weight matrices (e.g. :math:`W_{xi}` is
314
    the matrix of weights from the input gate to the input), :math:`W_{ic}, \
315 316 317
    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
318
    gate bias vector), :math:`\sigma` is the non-linear activations, such as
319 320
    logistic sigmoid function, and :math:`i, f, o` and :math:`c` are the input
    gate, forget gate, output gate, and cell activation vectors, respectively,
321 322
    all of which have the same size as the cell output activation vector :math:`h`.

323 324 325 326
    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
327 328 329
    the previous hidden state.

    Set `use_peepholes` to `False` to disable peephole connection. The formula
Y
Yibing Liu 已提交
330 331 332
    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
333 334 335
    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
336 337

    Args:
338 339 340 341
        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
342 343
                         mini-batch, D is the hidden size.
        size(int): 4 * hidden size.
344
        param_attr(ParamAttr|None): The parameter attribute for the learnable
345
                               hidden-hidden weights.
Y
Yibing Liu 已提交
346 347 348

                               - Weights = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}
349 350 351
                               - The shape is (D x 4D), where D is the hidden
                                 size.
        bias_attr(ParamAttr|None): The bias attribute for the learnable bias
352 353 354
                              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
355

356
                              1. `use_peepholes = False`
Y
Yibing Liu 已提交
357
                                - Biases = {:math:`b_c, b_i, b_f, b_o`}.
358
                                - The shape is (1 x 4D).
359
                              2. `use_peepholes = True`
Y
Yibing Liu 已提交
360 361
                                - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
                                                 W_{fc}, W_{oc}`}.
362
                                - The shape is (1 x 7D).
363
        use_peepholes(bool): Whether to enable diagonal/peephole connections,
Y
Yibing Liu 已提交
364 365
                             default `True`.
        is_reverse(bool): Whether to compute reversed LSTM, default `False`.
366 367
        gate_activation(str): The activation for input gate, forget gate and
                              output gate. Choices = ["sigmoid", "tanh", "relu",
Y
Yibing Liu 已提交
368
                              "identity"], default "sigmoid".
369
        cell_activation(str): The activation for cell output. Choices = ["sigmoid",
Y
Yibing Liu 已提交
370 371 372 373 374
                              "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".
375 376
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
Y
Yibing Liu 已提交
377 378

    Returns:
Y
Yibing Liu 已提交
379 380
        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
381

Y
Yibing Liu 已提交
382
    Examples:
Y
Yibing Liu 已提交
383 384
        .. code-block:: python

Y
Yibing Liu 已提交
385 386
            hidden_dim = 512
            forward_proj = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
387
                                           act=None, bias_attr=None)
Y
Yibing Liu 已提交
388 389
            forward, _ = fluid.layers.dynamic_lstm(
                input=forward_proj, size=hidden_dim * 4, use_peepholes=False)
Y
Yibing Liu 已提交
390
    """
391

Y
Yu Yang 已提交
392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427
    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
428 429 430 431 432 433 434 435 436 437 438
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',
439 440
                  dtype='float32',
                  name=None):
Y
Yibing Liu 已提交
441 442 443
    """
    **Dynamic LSTMP Layer**

444 445 446 447 448 449
    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
450 451 452 453 454

    The formula is as follows:

    .. math::

455
        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
456

457
        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
458

459
        \\tilde{c_t} & = act_g(W_{cx}x_t + W_{cr}r_{t-1} + b_c)
Y
Yibing Liu 已提交
460

461
        o_t & = \sigma(W_{ox}x_{t} + W_{or}r_{t-1} + W_{oc}c_t + b_o)
Y
Yibing Liu 已提交
462

463
        c_t & = f_t \odot c_{t-1} + i_t \odot \\tilde{c_t}
Y
Yibing Liu 已提交
464

465
        h_t & = o_t \odot act_h(c_t)
Y
Yibing Liu 已提交
466

467
        r_t & = \overline{act_h}(W_{rh}h_t)
Y
Yibing Liu 已提交
468

Y
Yibing Liu 已提交
469 470 471 472 473 474
    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, \
475
          we use vectors to reprenset these diagonal weight matrices.
Y
Yibing Liu 已提交
476
    * :math:`b`: Denotes bias vectors (e.g. :math:`b_i` is the input gate \
477
          bias vector).
Y
Yibing Liu 已提交
478 479 480
    * :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 \
481
          the same size as the cell output activation vector :math:`h`.
Y
Yibing Liu 已提交
482
    * :math:`h`: The hidden state.
483
    * :math:`r`: The recurrent projection of the hidden state.
Y
Yibing Liu 已提交
484 485
    * :math:`\\tilde{c_t}`: The candidate hidden state, whose \
          computation is based on the current input and previous hidden state.
486
    * :math:`\odot`: The element-wise product of the vectors.
Y
Yibing Liu 已提交
487
    * :math:`act_g` and :math:`act_h`: The cell input and cell output \
488
          activation functions and `tanh` is usually used for them.
Y
Yibing Liu 已提交
489 490
    * :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
491 492 493 494

    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.
495

Y
Yibing Liu 已提交
496 497 498 499 500 501 502 503 504 505 506 507
    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.
508
        param_attr(ParamAttr|None): The parameter attribute for the learnable
Y
Yibing Liu 已提交
509 510
                               hidden-hidden weight and projection weight.

511 512
                               - Hidden-hidden weight = {:math:`W_{ch}, W_{ih}, \
                                                W_{fh}, W_{oh}`}.
513 514
                               - 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
515 516
                                 size.
                               - Projection weight = {:math:`W_{rh}`}.
517 518
                               - 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
519 520 521 522 523 524
                              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`}.
525
                                - The shape is (1 x 4D).
Y
Yibing Liu 已提交
526 527 528
                              2. `use_peepholes = True`
                                - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
                                                 W_{fc}, W_{oc}`}.
529
                                - The shape is (1 x 7D).
Y
Yibing Liu 已提交
530 531 532 533 534 535 536 537 538 539 540 541 542 543 544
        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".
545 546
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.
Y
Yibing Liu 已提交
547 548

    Returns:
549 550
        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
551 552 553 554 555
               (T x D), and both LoD is the same with the `input`.

    Examples:
        .. code-block:: python

Y
Yibing Liu 已提交
556
            hidden_dim, proj_dim = 512, 256
Y
Yibing Liu 已提交
557 558
            fc_out = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
                                     act=None, bias_attr=None)
559 560 561
            proj_out, _ = fluid.layers.dynamic_lstmp(input=fc_out,
                                                     size=hidden_dim * 4,
                                                     proj_size=proj_dim,
Y
Yibing Liu 已提交
562 563 564 565
                                                     use_peepholes=False,
                                                     is_reverse=True,
                                                     cell_activation="tanh",
                                                     proj_activation="tanh")
Y
Yibing Liu 已提交
566
    """
567

Y
Yibing Liu 已提交
568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613
    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
614 615 616 617 618 619 620 621 622 623 624
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**

625
    Refer to `Empirical Evaluation of Gated Recurrent Neural Networks on
G
guosheng 已提交
626
    Sequence Modeling <https://arxiv.org/abs/1412.3555>`_
627

G
guosheng 已提交
628 629 630 631 632 633 634 635 636
    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)
637

G
guosheng 已提交
638
        h_t & = (1-u_t) \odot h_{t-1} + u_t \odot \\tilde{h_t}
639

G
guosheng 已提交
640
    The :math:`\odot` is the element-wise product of the vectors. :math:`act_g`
641 642
    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
643 644 645 646
    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
647
    to use fully-connect layer before GRU layer.
G
guosheng 已提交
648 649

    Args:
650 651
        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
652
            Variable is a matrix with shape :math:`(T \\times 3D)`, where
653
            :math:`T` is the total time steps in this mini-batch, :math:`D`
G
guosheng 已提交
654 655
            is the hidden size.
        size(int): The dimension of the gru cell.
656
        param_attr(ParamAttr|None): The parameter attribute for the learnable
G
guosheng 已提交
657 658
            hidden-hidden weight matrix. Note:

659
            - The shape of the weight matrix is :math:`(T \\times 3D)`, where
G
guosheng 已提交
660
              :math:`D` is the hidden size.
661
            - All elements in the weight matrix can be divided into two parts.
G
guosheng 已提交
662
              The first part are weights of the update gate and reset gate with
663
              shape :math:`(D \\times 2D)`, and the second part are weights for
G
guosheng 已提交
664
              candidate hidden state with shape :math:`(D \\times D)`.
665
        bias_attr(ParamAttr): The parameter attribute for learnable the
G
guosheng 已提交
666
            hidden-hidden bias.
667
        is_reverse(bool): Whether to compute reversed GRU, default
G
guosheng 已提交
668 669 670
            :attr:`False`.
        gate_activation(str): The activation for update gate and reset gate.
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "sigmoid".
671
        activation(str): The activation for candidate hidden state.
G
guosheng 已提交
672 673 674
            Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh".

    Returns:
G
guosheng 已提交
675 676
        Variable: The hidden state of GRU. The shape is :math:`(T \\times D)`, \
            and lod is the same with the input.
677

G
guosheng 已提交
678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720
    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
721 722 723 724 725 726
def gru_unit(input,
             hidden,
             size,
             weight=None,
             bias=None,
             activation='tanh',
727
             gate_activation='sigmoid'):
Y
Yu Yang 已提交
728
    """
729
    GRU unit layer. The equation of a gru step is:
Y
Yu Yang 已提交
730

731 732
        .. math::
            u_t & = actGate(xu_{t} + W_u h_{t-1} + b_u)
Y
Yu Yang 已提交
733

734
            r_t & = actGate(xr_{t} + W_r h_{t-1} + b_r)
Y
Yu Yang 已提交
735

736
            m_t & = actNode(xm_t + W_c dot(r_t, h_{t-1}) + b_m)
737

738
            h_t & = dot((1-u_t), m_t) + dot(u_t, h_{t-1})
739 740

    The inputs of gru unit includes :math:`z_t`, :math:`h_{t-1}`. In terms
741 742 743
    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
744 745
    connected layer has to be applied, such that :math:`z_t = W_{fc}x_t`.

746 747
    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
748 749 750
    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`.
751 752 753 754 755 756 757

    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
758 759 760 761
        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
762

763 764 765 766 767 768
    Returns:
        tuple: The hidden value, reset-hidden value and gate values.

    Examples:

        .. code-block:: python
Y
Yu Yang 已提交
769

770
             # assuming we have x_t_data and prev_hidden of size=10
771
             x_t = fluid.layers.fc(input=x_t_data, size=30)
772 773
             hidden_val, r_h_val, gate_val = fluid.layers.gru_unit(input=x_t,
                                                    hidden = prev_hidden)
Y
Yu Yang 已提交
774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793

    """
    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
794

Y
Yu Yang 已提交
795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821
    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


822
def linear_chain_crf(input, label, param_attr=None):
Y
Yu Yang 已提交
823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847
    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


848
def crf_decoding(input, param_attr, label=None):
Y
Yu Yang 已提交
849 850 851 852 853 854 855 856 857 858 859 860 861
    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
862
def cos_sim(X, Y):
Y
Yu Yang 已提交
863 864 865 866
    """
    This function performs the cosine similarity between two tensors
    X and Y and returns that as the output.
    """
F
fengjiayi 已提交
867
    helper = LayerHelper('cos_sim', **locals())
Y
Yu Yang 已提交
868 869 870 871 872 873 874 875 876 877 878 879 880
    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
881
def dropout(x, dropout_prob, is_test=False, seed=None):
882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909
    """
    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
910
    helper = LayerHelper('dropout', **locals())
911 912 913 914 915 916 917
    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]},
918 919 920 921 922 923
        attrs={
            'dropout_prob': dropout_prob,
            'is_test': is_test,
            'fix_seed': seed is not None,
            'seed': seed if seed is not None else 0
        })
924 925 926
    return out


F
fengjiayi 已提交
927
def cross_entropy(input, label, soft_label=False):
Y
Yu Yang 已提交
928
    """
Y
Yibing Liu 已提交
929 930
    **Cross Entropy Layer**

931 932 933
    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
934 935

    1) One-hot cross-entropy:
F
fengjiayi 已提交
936
        `soft_label = False`, `Label[i, 0]` indicates the class index for sample i:
Y
yangyaming 已提交
937

Y
Yibing Liu 已提交
938
        .. math::
Y
yangyaming 已提交
939

Y
Yibing Liu 已提交
940 941 942
            Y[i] = -\log(X[i, Label[i]])

    2) Soft-label cross-entropy:
F
fengjiayi 已提交
943 944
        `soft_label = True`, `Label[i, j]` indicates the soft label of class j
        for sample i:
Y
Yibing Liu 已提交
945 946 947 948 949

        .. math::

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

Y
Yibing Liu 已提交
950
       Please make sure that in this case the summation of each row of `label`
Y
Yibing Liu 已提交
951 952 953
       equals one.

    3) One-hot cross-entropy with vecterized `label`:
F
fengjiayi 已提交
954 955
         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
956
         to a one-hot cross-entropy with one-hot label representation.
Y
yangyaming 已提交
957

Y
Yibing Liu 已提交
958
    Args:
Y
yangyaming 已提交
959
        input (Variable|list):  a 2-D tensor with shape [N x D], where N is the
960 961 962 963
                                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
964
        label (Variable|list): the ground truth which is a 2-D tensor. When
965 966 967 968
                               `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
969
        soft_label (bool): a flag indicating whether to
970 971
                                           interpretate the given labels as soft
                                           labels, default `False`.
Y
Yibing Liu 已提交
972 973 974 975 976

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

    Raises:
977 978 979 980 981
        `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
982 983 984 985 986 987

    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
988
    """
F
fengjiayi 已提交
989
    helper = LayerHelper('cross_entropy', **locals())
Y
Yu Yang 已提交
990 991 992 993 994 995
    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
996
        attrs={"soft_label": soft_label})
Y
Yu Yang 已提交
997 998 999
    return out


F
fengjiayi 已提交
1000
def square_error_cost(input, label):
Y
Yu Yang 已提交
1001
    """
1002 1003
    **Square error cost layer**

1004 1005
    This layer accepts input predictions and target label and returns the
    squared error cost.
Y
ying 已提交
1006

1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023
    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
1024
        Variable: The tensor variable storing the element-wise squared error \
1025
                  difference of input and label.
1026 1027 1028 1029 1030 1031 1032 1033

    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
1034
    """
F
fengjiayi 已提交
1035
    helper = LayerHelper('square_error_cost', **locals())
Y
Yu Yang 已提交
1036 1037 1038 1039 1040 1041 1042 1043 1044
    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
1045 1046
        type='square', inputs={'X': [minus_out]},
        outputs={'Out': [square_out]})
Y
Yu Yang 已提交
1047 1048 1049 1050 1051 1052 1053
    return square_out


def chunk_eval(input,
               label,
               chunk_scheme,
               num_chunk_types,
F
fengjiayi 已提交
1054
               excluded_chunk_types=None):
Y
Yu Yang 已提交
1055
    """
Y
yangyaming 已提交
1056
    This function computes and outputs the precision, recall and
1057
    F1-score of chunk detection.
Y
Yu Yang 已提交
1058
    """
F
fengjiayi 已提交
1059
    helper = LayerHelper("chunk_eval", **locals())
Y
Yu Yang 已提交
1060 1061 1062 1063 1064

    # prepare output
    precision = helper.create_tmp_variable(dtype="float32")
    recall = helper.create_tmp_variable(dtype="float32")
    f1_score = helper.create_tmp_variable(dtype="float32")
1065 1066 1067
    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
1068 1069 1070 1071 1072 1073 1074 1075

    helper.append_op(
        type="chunk_eval",
        inputs={"Inference": [input],
                "Label": [label]},
        outputs={
            "Precision": [precision],
            "Recall": [recall],
1076 1077 1078 1079
            "F1-Score": [f1_score],
            "NumInferChunks": [num_infer_chunks],
            "NumLabelChunks": [num_label_chunks],
            "NumCorrectChunks": [num_correct_chunks]
Y
Yu Yang 已提交
1080 1081 1082
        },
        attrs={
            "num_chunk_types": num_chunk_types,
G
guosheng 已提交
1083 1084
            "chunk_scheme": chunk_scheme,
            "excluded_chunk_types": excluded_chunk_types or []
Y
Yu Yang 已提交
1085
        })
1086 1087
    return (precision, recall, f1_score, num_infer_chunks, num_label_chunks,
            num_correct_chunks)
Y
Yu Yang 已提交
1088 1089 1090 1091 1092 1093 1094 1095 1096


def sequence_conv(input,
                  num_filters,
                  filter_size=3,
                  filter_stride=1,
                  padding=None,
                  bias_attr=None,
                  param_attr=None,
1097
                  act=None):
Y
Yu Yang 已提交
1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130
    """
    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)


1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154
def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=True):
    helper = LayerHelper('sequence_softmax', **locals())
    dtype = helper.input_dtype()
    softmax_out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type="sequence_softmax",
        inputs={"X": input},
        outputs={"Out": softmax_out},
        attrs={"use_cudnn": use_cudnn})
    return softmax_out


def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True):
    helper = LayerHelper('softmax', **locals())
    dtype = helper.input_dtype()
    softmax_out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type="softmax",
        inputs={"X": input},
        outputs={"Out": softmax_out},
        attrs={"use_cudnn": use_cudnn})
    return softmax_out


Y
Yu Yang 已提交
1155 1156 1157
def conv2d(input,
           num_filters,
           filter_size,
C
chengduoZH 已提交
1158 1159
           stride=1,
           padding=0,
1160
           dilation=1,
Y
Yu Yang 已提交
1161 1162 1163
           groups=None,
           param_attr=None,
           bias_attr=None,
C
chengduoZH 已提交
1164
           use_cudnn=True,
1165
           use_mkldnn=False,
1166 1167
           act=None,
           name=None):
Y
Yu Yang 已提交
1168
    """
C
chengduoZH 已提交
1169 1170 1171
    **Convlution2D Layer**

    The convolution2D layer calculates the output based on the input, filter
1172 1173 1174
    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
1175 1176
    The details of convolution layer, please refer UFLDL's `convolution,
    <http://ufldl.stanford.edu/tutorial/supervised/FeatureExtractionUsingConvolution/>`_ .
1177 1178 1179
    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
1180

1181
    For each input :math:`X`, the equation is:
C
refine  
chengduoZH 已提交
1182

C
chengduoZH 已提交
1183 1184
    .. math::

C
refine  
chengduoZH 已提交
1185
        Out = \sigma (W \\ast X + b)
C
chengduoZH 已提交
1186

C
chengduoZH 已提交
1187
    In the above equation:
C
chengduoZH 已提交
1188

1189 1190 1191 1192 1193
    * :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.
1194 1195
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be
                   different.
C
chengduoZH 已提交
1196 1197 1198

    Example:

1199 1200 1201
        - Input:

          Input shape: $(N, C_{in}, H_{in}, W_{in})$
C
refine  
chengduoZH 已提交
1202

1203
          Filter shape: $(C_{out}, C_{in}, H_f, W_f)$
C
refine  
chengduoZH 已提交
1204

1205 1206
        - Output:
          Output shape: $(N, C_{out}, H_{out}, W_{out})$
C
refine  
chengduoZH 已提交
1207

C
chengduoZH 已提交
1208
        Where
1209 1210

        .. math::
C
chengduoZH 已提交
1211

C
chengduoZH 已提交
1212 1213
        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
1214 1215

    Args:
1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227
       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.
1228 1229 1230
       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.
1231 1232 1233 1234 1235 1236 1237 1238 1239 1240
       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
1241 1242
       name(str|None): A name for this layer(optional). If set None, the layer
           will be named automatically.
C
chengduoZH 已提交
1243 1244

    Returns:
G
guosheng 已提交
1245
        Variable: The tensor variable storing the convolution and \
C
chengduoZH 已提交
1246 1247
                  non-linearity activation result.

C
refine  
chengduoZH 已提交
1248
    Raises:
1249 1250
        ValueError: If the shapes of input, filter_size, stride, padding and
                    groups mismatch.
C
refine  
chengduoZH 已提交
1251

C
chengduoZH 已提交
1252 1253 1254
    Examples:
        .. code-block:: python

1255 1256 1257 1258
          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
1259 1260 1261 1262 1263
    """
    if stride is None:
        stride = [1, 1]

    num_channels = input.shape[1]
1264 1265

    l_type = 'conv2d'
X
xzl 已提交
1266 1267
    if (num_channels == groups and num_filters % num_channels == 0 and
            not use_cudnn):
1268
        l_type = 'depthwise_conv2d'
1269 1270 1271 1272

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

Y
Yu Yang 已提交
1273 1274 1275 1276 1277 1278 1279
    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
1280 1281 1282
    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')
1283
    dilation = utils.convert_to_list(dilation, 2, 'dilation')
C
chengduoZH 已提交
1284

C
chengduoZH 已提交
1285 1286
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303

    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(
1304
        type=l_type,
Y
Yu Yang 已提交
1305 1306 1307 1308 1309
        inputs={
            'Input': input,
            'Filter': filter_param,
        },
        outputs={"Output": pre_bias},
C
chengduoZH 已提交
1310 1311 1312
        attrs={
            'strides': stride,
            'paddings': padding,
1313
            'dilations': dilation,
C
chengduoZH 已提交
1314
            'groups': groups,
1315 1316
            'use_cudnn': use_cudnn,
            'use_mkldnn': use_mkldnn
C
chengduoZH 已提交
1317
        })
Y
Yu Yang 已提交
1318 1319 1320 1321 1322 1323

    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
1324
def sequence_pool(input, pool_type):
Y
Yu Yang 已提交
1325
    """
Y
yangyaming 已提交
1326 1327 1328
    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
1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353

    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
1354

L
Luo Tao 已提交
1355 1356
    Args:
        input(variable): The input variable which is a LoDTensor.
Y
yangyaming 已提交
1357
        pool_type (string): The pooling type of sequence_pool.
L
Luo Tao 已提交
1358 1359 1360 1361 1362 1363 1364 1365
            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
1366

Y
yangyaming 已提交
1367
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1368 1369 1370 1371 1372
                              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
1373
    """
F
fengjiayi 已提交
1374
    helper = LayerHelper('sequence_pool', **locals())
Y
Yu Yang 已提交
1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385
    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
1386 1387 1388 1389 1390
    # 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
1391 1392 1393
    return pool_out


F
fengjiayi 已提交
1394
def sequence_first_step(input):
L
Luo Tao 已提交
1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408
    """
    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
1409

L
Luo Tao 已提交
1410 1411 1412 1413 1414 1415 1416 1417 1418
    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
1419

Y
yangyaming 已提交
1420
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1421 1422 1423
                              dtype='float32', lod_level=1)
             x_first_step = fluid.layers.sequence_first_step(input=x)
    """
1424 1425 1426
    return sequence_pool(input=input, pool_type="first")


F
fengjiayi 已提交
1427
def sequence_last_step(input):
L
Luo Tao 已提交
1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441
    """
    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
1442

L
Luo Tao 已提交
1443 1444 1445 1446 1447 1448 1449 1450 1451
    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
1452

Y
yangyaming 已提交
1453
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
1454 1455 1456
                              dtype='float32', lod_level=1)
             x_last_step = fluid.layers.sequence_last_step(input=x)
    """
1457 1458 1459
    return sequence_pool(input=input, pool_type="last")


Y
Yu Yang 已提交
1460
def pool2d(input,
C
chengduoZH 已提交
1461 1462
           pool_size=-1,
           pool_type="max",
C
chengduoZH 已提交
1463 1464
           pool_stride=1,
           pool_padding=0,
C
caoying03 已提交
1465
           global_pooling=False,
C
chengduoZH 已提交
1466
           use_cudnn=True,
1467
           ceil_mode=False,
1468
           use_mkldnn=False,
C
caoying03 已提交
1469
           name=None):
Y
Yu Yang 已提交
1470 1471 1472 1473 1474 1475 1476 1477
    """
    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
1478

C
chengduoZH 已提交
1479 1480 1481 1482 1483
    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
1484 1485 1486 1487
    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
1488 1489
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")
Y
Yu Yang 已提交
1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503

    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
1504
            "paddings": pool_padding,
1505
            "use_cudnn": use_cudnn,
1506 1507
            "ceil_mode": ceil_mode,
            "use_mkldnn": use_mkldnn
Y
Yu Yang 已提交
1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519
        })

    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
1520
               data_layout='NCHW',
Y
Yang Yang 已提交
1521
               in_place=False,
1522 1523
               name=None,
               moving_mean_name=None,
W
wanghaoshuang 已提交
1524
               moving_variance_name=None,
W
wanghaoshuang 已提交
1525
               do_model_average_for_mean_and_var=False):
Y
Yu Yang 已提交
1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551
    """
    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(
1552
        attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
Y
Yu Yang 已提交
1553

1554 1555
    mean = helper.create_parameter(
        attr=ParamAttr(
W
wanghaoshuang 已提交
1556 1557 1558
            name=moving_mean_name,
            initializer=Constant(0.0),
            trainable=False,
W
wanghaoshuang 已提交
1559
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
1560
        shape=param_shape,
1561 1562 1563 1564 1565 1566 1567
        dtype=input.dtype)
    mean.stop_gradient = True

    variance = helper.create_parameter(
        attr=ParamAttr(
            name=moving_variance_name,
            initializer=Constant(1.0),
W
wanghaoshuang 已提交
1568
            trainable=False,
W
wanghaoshuang 已提交
1569
            do_model_average=do_model_average_for_mean_and_var),
Q
QI JUN 已提交
1570
        shape=param_shape,
1571 1572
        dtype=input.dtype)
    variance.stop_gradient = True
Y
Yu Yang 已提交
1573 1574 1575 1576 1577 1578

    # 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
1579 1580
    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
1581

Y
Yang Yang 已提交
1582
    batch_norm_out = input if in_place else helper.create_tmp_variable(dtype)
Y
Yu Yang 已提交
1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606

    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
1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618
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**

1619
    Assume feature vectors exist on dimensions
G
guosheng 已提交
1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639
    :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.
1640
        scale(bool): Whether to learn the adaptive gain :math:`g` after
G
guosheng 已提交
1641
            normalization.
1642
        shift(bool): Whether to learn the adaptive bias :math:`b` after
G
guosheng 已提交
1643
            normalization.
1644
        begin_norm_axis(bool): The normalization will be performed along
G
guosheng 已提交
1645
            dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`.
1646
        epsilon(float): The small value added to the variance to prevent
G
guosheng 已提交
1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677
            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
1678
    if shift:
G
guosheng 已提交
1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702
        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
1703
def beam_search_decode(ids, scores, name=None):
Y
Yu Yang 已提交
1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723
    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
1724 1725 1726
                     padding=0,
                     stride=1,
                     dilation=1,
C
caoying03 已提交
1727
                     param_attr=None,
1728
                     bias_attr=None,
C
chengduoZH 已提交
1729
                     use_cudnn=True,
1730
                     act=None,
C
caoying03 已提交
1731
                     name=None):
Y
Yu Yang 已提交
1732
    """
1733 1734 1735 1736 1737 1738 1739 1740
    **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
1741 1742
    layer, please refer to the following explanation and references
    `therein <http://www.matthewzeiler.com/wp-content/uploads/2017/07/cvpr2010.pdf>`_.
1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754

    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.
1755 1756
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be
                   different.
Y
Yu Yang 已提交
1757

1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770
    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
1771

1772 1773 1774 1775
        .. 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
1776 1777

    Args:
1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796
       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.
1797 1798
       param_attr(ParamAttr): The parameters to the Conv2d_transpose Layer.
                              Default: None
1799
       bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None
1800 1801
       use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
           library is installed. Default: True
1802
       act(str): Activation type. Default: None
1803 1804
       name(str|None): A name for this layer(optional). If set None, the layer
           will be named automatically.
Y
Yu Yang 已提交
1805 1806

    Returns:
1807 1808 1809
       Variable: The tensor variable storing the convolution transpose result.

    Raises:
1810 1811
       ValueError: If the shapes of input, filter_size, stride, padding and
                   groups mismatch.
1812 1813 1814 1815

    Examples:
       .. code-block:: python

1816 1817 1818 1819
          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
1820 1821 1822 1823 1824 1825
    """
    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
1826 1827 1828
    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
1829

C
chengduoZH 已提交
1830 1831 1832
    if not isinstance(use_cudnn, bool):
        raise ValueError("use_cudnn should be True or False")

Y
Yu Yang 已提交
1833 1834 1835 1836 1837 1838 1839 1840
    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
1841 1842 1843 1844 1845

        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
1846
        filter_size = [filter_size_h, filter_size_w]
C
chengduoZH 已提交
1847 1848 1849
    else:
        filter_size = utils.convert_to_list(filter_size, 2,
                                            'conv2d_transpose.filter_size')
Y
Yu Yang 已提交
1850 1851 1852 1853 1854

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

1855
    pre_bias = helper.create_tmp_variable(dtype=input.dtype)
Y
Yu Yang 已提交
1856 1857 1858 1859
    helper.append_op(
        type='conv2d_transpose',
        inputs={'Input': [input],
                'Filter': [img_filter]},
1860
        outputs={'Output': pre_bias},
C
chengduoZH 已提交
1861 1862 1863 1864 1865 1866
        attrs={
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
            'use_cudnn': use_cudnn
        })
Y
Yu Yang 已提交
1867

1868 1869
    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
    out = helper.append_activation(pre_act)
Y
Yu Yang 已提交
1870
    return out
Y
yangyaming 已提交
1871 1872


Y
yangyaming 已提交
1873
def sequence_expand(x, y, ref_level=-1, name=None):
1874
    """Sequence Expand Layer. This layer will expand the input variable **x**
Y
yangyaming 已提交
1875 1876 1877 1878
    according to specified level lod of **y**. Please note that lod level of
    **x** is at most 1 and rank of **x** is at least 2. When rank of **x**
    is greater than 2, then it would be viewed as a 2-D tensor.
    Following examples will explain how sequence_expand works:
1879 1880 1881 1882 1883

    .. code-block:: text

        * Case 1
            x is a LoDTensor:
Y
yangyaming 已提交
1884 1885
                x.lod  = [[0,   2,        4]]
                x.data = [[a], [b], [c], [d]]
1886 1887 1888 1889 1890 1891
                x.dims = [4, 1]

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

Y
yangyaming 已提交
1892
            ref_level: 0
1893

Y
yangyaming 已提交
1894 1895 1896
            then output is a 1-level LoDTensor:
                out.lod =  [[0,   2,        4,        6,        8]]
                out.data = [[a], [b], [a], [b], [c], [d], [c], [d]]
1897 1898 1899 1900
                out.dims = [8, 1]

        * Case 2
            x is a Tensor:
Y
yangyaming 已提交
1901
                x.data = [[a], [b], [c]]
1902 1903 1904
                x.dims = [3, 1]

            y is a LoDTensor:
Y
yangyaming 已提交
1905
                y.lod = [[0, 2, 2, 5]]
1906

Y
yangyaming 已提交
1907
            ref_level: -1
1908

Y
yangyaming 已提交
1909 1910 1911
            then output is a Tensor:
                out.data = [[a], [a], [c], [c], [c]]
                out.dims = [5, 1]
1912 1913 1914
    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
        y (Variable): The input variable which is a LoDTensor.
Y
yangyaming 已提交
1915 1916
        ref_level (int): Lod level of `y` to be referred by `x`. If set to -1,
                         refer the last level of lod.
C
caoying03 已提交
1917
        name(str|None): A name for this layer(optional). If set None, the layer
Y
yangyaming 已提交
1918
                        will be named automatically.
1919 1920 1921 1922 1923 1924 1925 1926 1927 1928

    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
1929
            out = layers.sequence_expand(x=x, y=y, ref_level=0)
1930
    """
Y
yangyaming 已提交
1931
    helper = LayerHelper('sequence_expand', input=x, **locals())
1932 1933 1934
    dtype = helper.input_dtype()
    tmp = helper.create_tmp_variable(dtype)
    helper.append_op(
Y
yangyaming 已提交
1935 1936 1937 1938 1939
        type='sequence_expand',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': tmp},
        attrs={'ref_level': ref_level})
1940
    return tmp
1941 1942


Q
Qiao Longfei 已提交
1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974
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
1975 1976 1977 1978
def lstm_unit(x_t,
              hidden_t_prev,
              cell_t_prev,
              forget_bias=0.0,
Y
yangyaming 已提交
1979
              param_attr=None,
C
caoying03 已提交
1980 1981
              bias_attr=None,
              name=None):
Y
yangyaming 已提交
1982 1983 1984 1985
    """Lstm unit layer. The equation of a lstm step is:

        .. math::

1986
            i_t & = \sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i)
Y
yangyaming 已提交
1987

1988
            f_t & = \sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + b_f)
Y
yangyaming 已提交
1989

1990
            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
1991

1992
            o_t & = \sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + b_o)
Y
yangyaming 已提交
1993 1994 1995

            h_t & = o_t tanh(c_t)

1996 1997 1998 1999 2000 2001
    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
2002 2003 2004

        .. math::

2005
            L_{i_t} = W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i
Y
yangyaming 已提交
2006 2007 2008 2009 2010 2011 2012 2013

    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
2014
    This layer has two outputs including :math:`h_t` and :math:`o_t`.
Y
yangyaming 已提交
2015 2016

    Args:
Y
yangyaming 已提交
2017 2018 2019 2020 2021 2022
        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
2023
        forget_bias (float): The forget bias of lstm unit.
Y
yangyaming 已提交
2024 2025
        param_attr (ParamAttr): The attributes of parameter weights, used to set
            initializer, name etc.
Y
yangyaming 已提交
2026 2027
        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
2028 2029
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
Y
yangyaming 已提交
2030 2031

    Returns:
Y
yangyaming 已提交
2032
        tuple: The hidden value and cell value of lstm unit.
Y
yangyaming 已提交
2033 2034

    Raises:
2035 2036 2037 2038
        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
2039 2040 2041 2042 2043 2044

    Examples:

        .. code-block:: python

             x_t = fluid.layers.fc(input=x_t_data, size=10)
2045
             prev_hidden = fluid.layers.fc(input=prev_hidden_data, size=30)
Y
yangyaming 已提交
2046
             prev_cell = fluid.layers.fc(input=prev_cell_data, size=30)
Y
yangyaming 已提交
2047
             hidden_value, cell_value = fluid.layers.lstm_unit(x_t=x_t,
Y
yangyaming 已提交
2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063
                                                    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
2064
        raise ValueError("The 1st dimensions of x_t, hidden_t_prev and "
2065 2066 2067 2068
                         "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
2069 2070
                         "cell_t_prev must be the same.")

Y
yangyaming 已提交
2071 2072 2073
    if bias_attr is None:
        bias_attr = ParamAttr()

Y
yangyaming 已提交
2074
    size = cell_t_prev.shape[1]
2075
    concat_out = concat(input=[x_t, hidden_t_prev], axis=1)
Y
yangyaming 已提交
2076 2077
    fc_out = fc(input=concat_out,
                size=4 * size,
Y
yangyaming 已提交
2078
                param_attr=param_attr,
2079
                bias_attr=bias_attr)
Y
yangyaming 已提交
2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091
    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
2092
    return h, c
G
guosheng 已提交
2093 2094


C
caoying03 已提交
2095
def reduce_sum(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
2096
    """
Y
yangyaming 已提交
2097
    Computes the sum of tensor elements over the given dimension.
G
guosheng 已提交
2098 2099 2100

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
2101 2102 2103 2104
        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
2105
            the dimension to reduce is :math:`rank + dim`.
2106
        keep_dim (bool|False): Whether to reserve the reduced dimension in the
Y
yangyaming 已提交
2107
            output Tensor. The result tensor will have one fewer dimension
G
guosheng 已提交
2108
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
2109 2110
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
2111 2112 2113

    Returns:
        Variable: The reduced Tensor variable.
F
fengjiayi 已提交
2114

G
guosheng 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
    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
2139 2140


C
caoying03 已提交
2141
def reduce_mean(input, dim=None, keep_dim=False, name=None):
G
guosheng 已提交
2142
    """
Y
yangyaming 已提交
2143
    Computes the mean of tensor elements over the given dimension.
G
guosheng 已提交
2144 2145 2146

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
2147 2148 2149 2150
        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
2151
            :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`.
Y
yangyaming 已提交
2152 2153
        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
2154
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
2155 2156
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
2157 2158 2159

    Returns:
        Variable: The reduced Tensor variable.
F
fengjiayi 已提交
2160

G
guosheng 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
    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
2185 2186


C
caoying03 已提交
2187
def reduce_max(input, dim=None, keep_dim=False, name=None):
2188
    """
Y
yangyaming 已提交
2189
    Computes the maximum of tensor elements over the given dimension.
2190 2191 2192

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
2193 2194 2195 2196
        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))`.
2197
            If :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`.
Y
yangyaming 已提交
2198 2199
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
2200
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
2201 2202
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
2203 2204 2205

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
2206

2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232
    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
2233
def reduce_min(input, dim=None, keep_dim=False, name=None):
2234
    """
Y
yangyaming 已提交
2235
    Computes the minimum of tensor elements over the given dimension.
2236 2237 2238

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
2239 2240 2241 2242
        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))`.
2243
            If :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`.
Y
yangyaming 已提交
2244 2245
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
2246
            than the :attr:`input` unless :attr:`keep_dim` is true.
C
caoying03 已提交
2247 2248
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
2249 2250 2251

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
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
    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
2277 2278


2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292
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.
Y
yangyaming 已提交
2293
        name(str|None): A name for this layer(optional). If set None, the
Z
zhouhanqing 已提交
2294
            layer will be named automatically.
2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308

    Returns:
        Variable: The reduced Tensor variable.

    Examples:
        .. code-block:: python

            # x is a Tensor variable with following elements:
            #    [[0.2, 0.3, 0.5, 0.9]
            #     [0.1, 0.2, 0.6, 0.7]]
            # Each example is followed by the correspending output tensor.
            fluid.layers.reduce_prod(x)  # [0.0002268]
            fluid.layers.reduce_prod(x, dim=0)  # [0.02, 0.06, 0.3, 0.63]
            fluid.layers.reduce_prod(x, dim=-1)  # [0.027, 0.0084]
Y
yangyaming 已提交
2309
            fluid.layers.reduce_prod(x, dim=1,
Z
zhouhanqing 已提交
2310
                                     keep_dim=True)  # [[0.027], [0.0084]]
2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325
    """
    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
2326
def split(input, num_or_sections, dim=-1, name=None):
G
guosheng 已提交
2327
    """
C
caoying03 已提交
2328
    Split the input tensor into multiple sub-tensors.
G
guosheng 已提交
2329 2330 2331

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
C
caoying03 已提交
2332 2333 2334 2335 2336
        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
2337
            :attr:`dim` dimension orderly.
C
caoying03 已提交
2338
        dim (int): The dimension along which to split. If :math:`dim < 0`, the
G
guosheng 已提交
2339
            dimension to split along is :math:`rank(input) + dim`.
C
caoying03 已提交
2340 2341
        name(str|None): A name for this layer(optional). If set None, the layer
                       will be named automatically.
G
guosheng 已提交
2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383

    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
2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416


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
2417
          normed = fluid.layers.l2_normalize(x=data, axis=1)
C
caoying03 已提交
2418 2419
    """

F
fengjiayi 已提交
2420 2421
    if len(x.shape) == 1:
        axis = 0
C
caoying03 已提交
2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447

    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
2448
    # exanpsion is requred. Once a general broadcast rule is spported, this
C
caoying03 已提交
2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465
    # 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
2466 2467


2468
def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
G
guosheng 已提交
2469
    """
Y
ying 已提交
2470 2471 2472 2473
    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
2474

C
chengduoZH 已提交
2475
    The actual behavior depends on the shapes of :math:`x`, :math:`y` and the
2476
    flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically:
G
guosheng 已提交
2477

2478 2479 2480 2481 2482
    - 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
2483
      :math:`[1, D]` in transposed form.
G
guosheng 已提交
2484

C
chengduoZH 已提交
2485
    - After transpose, the two tensors are 2-D or n-D and matrix multiplication
2486
      performs in the following way.
G
guosheng 已提交
2487

2488
      - If both are 2-D, they are multiplied like conventional matrices.
C
chengduoZH 已提交
2489
      - If either is n-D, it is treated as a stack of matrices residing in the
Y
ying 已提交
2490
        last two dimensions and a batched matrix multiply supporting broadcast
2491
        applies on the two tensors.
G
guosheng 已提交
2492

Y
ying 已提交
2493 2494
    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
2495
    removed after matrix multiplication.
G
guosheng 已提交
2496 2497 2498

    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
2499 2500 2501
        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.
2502
        name(str|None): A name for this layer(optional). If set None, the layer
2503
            will be named automatically.
G
guosheng 已提交
2504 2505

    Returns:
2506
        Variable: The product Tensor variable.
G
guosheng 已提交
2507

G
guosheng 已提交
2508 2509 2510
    Examples:
        .. code-block:: python

2511
            # Examples to clarify shapes of the inputs and output
C
chengduoZH 已提交
2512 2513
            # x: [B, ..., M, K], y: [B, ..., K, N]
            fluid.layers.matmul(x, y)  # out: [B, ..., M, N]
Y
ying 已提交
2514

2515 2516
            # x: [B, M, K], y: [B, K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
2517

2518 2519
            # x: [B, M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [B, M, N]
Y
ying 已提交
2520

2521 2522
            # x: [M, K], y: [K, N]
            fluid.layers.matmul(x, y)  # out: [M, N]
Y
ying 已提交
2523 2524 2525 2526

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

2527 2528
            # x: [K], y: [K]
            fluid.layers.matmul(x, y)  # out: [1]
2529

Y
ying 已提交
2530
            # x: [M], y: [N]
2531
            fluid.layers.matmul(x, y, True, True)  # out: [M, N]
G
guosheng 已提交
2532
    """
Y
ying 已提交
2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544

    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
2545
            y_shape = y_shape + [1]
Y
ying 已提交
2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561

        # 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)

2562
    helper = LayerHelper('matmul', **locals())
Y
ying 已提交
2563
    out = helper.create_tmp_variable(dtype=x.dtype)
G
guosheng 已提交
2564
    helper.append_op(
2565 2566 2567 2568 2569 2570 2571
        type='matmul',
        inputs={'X': x,
                'Y': y},
        outputs={'Out': out},
        attrs={'transpose_X': transpose_x,
               'transpose_Y': transpose_y})
    return out
2572 2573


W
wanghaoshuang 已提交
2574
def edit_distance(input, label, normalized=True, ignored_tokens=None,
W
wanghaoshuang 已提交
2575
                  name=None):
2576
    """
Y
ying 已提交
2577 2578 2579 2580 2581 2582 2583 2584 2585
    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
2586

Y
ying 已提交
2587
    "kitten" -> "sitten" -> "sittin" -> "sitting"
W
wanghaoshuang 已提交
2588

Y
ying 已提交
2589 2590 2591 2592
    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
2593

Y
ying 已提交
2594 2595 2596
    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
2597

2598 2599 2600 2601 2602
    Args:

        input(Variable): The indices for hypothesis strings.

        label(Variable): The indices for reference strings.
W
wanghaoshuang 已提交
2603

Y
ying 已提交
2604 2605
        normalized(bool): Indicated whether to normalize the edit distance by
                          the length of reference string.
2606

Y
ying 已提交
2607 2608
        ignored_tokens(list of int): Tokens that should be removed before
                                     calculating edit distance.
2609

W
wanghaoshuang 已提交
2610
    Returns:
W
wanghaoshuang 已提交
2611
        Variable: sequence-to-sequence edit distance in shape [batch_size, 1].
W
wanghaoshuang 已提交
2612 2613 2614 2615 2616

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
2617 2618
            y = fluid.layers.data(name='y', shape=[7], dtype='float32')

2619
            cost = fluid.layers.edit_distance(input=x,label=y)
2620
    """
2621
    helper = LayerHelper("edit_distance", **locals())
2622

2623
    # remove some tokens from input and labels
W
wanghaoshuang 已提交
2624
    if ignored_tokens is not None and len(ignored_tokens) > 0:
2625 2626 2627 2628 2629 2630 2631
        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
2632
            attrs={"tokens": ignored_tokens})
2633 2634 2635 2636 2637 2638
        input = erased_input

        helper.append_op(
            type="sequence_erase",
            inputs={"X": [label]},
            outputs={"Out": [erase_label]},
W
wanghaoshuang 已提交
2639
            attrs={"tokens": ignored_tokens})
2640 2641
        label = erased_label

2642 2643
    # edit distance op
    edit_distance_out = helper.create_tmp_variable(dtype="int64")
2644
    sequence_num = helper.create_tmp_variable(dtype="int64")
2645 2646 2647 2648
    helper.append_op(
        type="edit_distance",
        inputs={"Hyps": [input],
                "Refs": [label]},
2649 2650
        outputs={"Out": [edit_distance_out],
                 "SequenceNum": [sequence_num]},
2651 2652
        attrs={"normalized": normalized})

2653
    return edit_distance_out, sequence_num
2654 2655 2656 2657 2658


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
2659 2660 2661 2662
    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.
2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691

    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
2692 2693 2694 2695 2696 2697
        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).
2698

Y
ying 已提交
2699 2700 2701
        blank(int): the blank label index of Connectionist Temporal
                    Classification (CTC) loss, which is in thehalf-opened
                    interval [0, num_classes + 1).
2702 2703

    Returns:
2704
        Variable: CTC greedy decode result. If all the sequences in result were
2705
        empty, the result LoDTensor will be [-1] with LoD [[0]] and dims [1, 1].
2706 2707 2708 2709 2710

    Examples:
        .. code-block:: python

            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
W
wanghaoshuang 已提交
2711

2712
            cost = fluid.layers.ctc_greedy_decoder(input=x, blank=0)
W
wanghaoshuang 已提交
2713
    """
2714
    helper = LayerHelper("ctc_greedy_decoder", **locals())
2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729
    # 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
2730
        outputs={"Output": [ctc_out]},
2731 2732
        attrs={"merge_repeated": True,
               "blank": blank})
2733
    return ctc_out
2734 2735


F
fengjiayi 已提交
2736
def warpctc(input, label, blank=0, norm_by_times=False):
W
wanghaoshuang 已提交
2737
    """
2738 2739
    An operator integrating the open source Warp-CTC library
    (https://github.com/baidu-research/warp-ctc)
W
wanghaoshuang 已提交
2740
    to compute Connectionist Temporal Classification (CTC) loss.
2741 2742
    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
2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755
    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.
2756
       blank: (int, default: 0), the blank label index of Connectionist
W
wanghaoshuang 已提交
2757 2758
         Temporal Classification (CTC) loss, which is in the
         half-opened interval [0, num_classes + 1).
2759
       norm_by_times: (bool, default: false), whether to normalize
W
wanghaoshuang 已提交
2760
       the gradients by the number of time-step, which is also the
2761 2762
       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
2763 2764

    Returns:
2765 2766
        Variable: The Connectionist Temporal Classification (CTC) loss,
        which is a 2-D Tensor of the shape [batch_size, 1].
W
wanghaoshuang 已提交
2767 2768 2769

    Examples:
        .. code-block:: python
2770 2771 2772 2773
            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
2774 2775 2776
            cost = layers.warpctc(input=y_predict, label=y)

    """
F
fengjiayi 已提交
2777
    helper = LayerHelper('warpctc', **locals())
W
wanghaoshuang 已提交
2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788
    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
2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842


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
2843 2844


2845
@autodoc()
Y
Yang Yu 已提交
2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871
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
2872 2873 2874 2875 2876 2877 2878 2879 2880
    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
2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896

    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
2897
    return cost / (num_neg_samples + 1)
2898 2899


Y
fix ci.  
ying 已提交
2900
def transpose(x, perm, name=None):
Y
ying 已提交
2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919
    """
    **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
2920
            x_transposed = layers.transpose(x, perm=[1, 0, 2])
Y
ying 已提交
2921 2922
    """

Y
fix ci.  
ying 已提交
2923
    if len(perm) != len(x.shape):
Y
ying 已提交
2924 2925 2926
        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
2927 2928 2929 2930 2931 2932
    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
2933 2934

    helper = LayerHelper('transpose', **locals())
Y
fix ci.  
ying 已提交
2935
    out = helper.create_tmp_variable(x.dtype)
Y
ying 已提交
2936 2937
    helper.append_op(
        type='transpose',
Y
fix ci.  
ying 已提交
2938
        inputs={'X': [x]},
Y
ying 已提交
2939 2940 2941
        outputs={'Out': [out]},
        attrs={'axis': perm})
    return out
2942 2943


2944
def im2sequence(input, filter_size=1, stride=1, padding=0, name=None):
2945
    """
2946 2947 2948 2949 2950 2951 2952
    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:
2953 2954 2955 2956 2957 2958 2959 2960 2961 2962

    .. 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
2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980

        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.

2981 2982 2983
        name (int): The name of this layer. It is optional.

    Returns:
W
wanghaoshuang 已提交
2984 2985 2986 2987 2988
        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.
2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017

    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
3018 3019 3020
            filter = [2, 2]
            stride = [1, 1]
            padding = [0, 0]
3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040

            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

3041 3042
            output = fluid.layers.im2sequence(
                input=layer, stride=[1, 1], filter_size=[2, 2])
3043 3044

    """
W
wanghaoshuang 已提交
3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055

    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])

3056
    helper = LayerHelper('im2sequence', **locals())
3057 3058
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
3059
        type='im2sequence',
3060 3061 3062
        inputs={'X': input},
        outputs={'Out': out},
        attrs={
W
wanghaoshuang 已提交
3063 3064 3065
            'kernels': filter_size,
            'strides': stride,
            'paddings': padding,
3066 3067
        })
    return out
3068 3069


3070 3071 3072 3073
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
3074
    equation of row convolution is as follows:
3075 3076 3077 3078 3079 3080 3081

    .. 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
3082
    * :math:`\\tau`: Future context size.
3083 3084 3085 3086 3087 3088 3089 3090 3091 3092
    * :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
3093 3094
        future_context_size (int): Future context size. Please note, the shape
            of convolution kernel is [future_context_size + 1, D].
3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119
        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
3120
    return helper.append_activation(out)
3121 3122


3123 3124 3125 3126
def multiplex(inputs, index):
    """
    **Multiplex Layer**

Y
yangyaming 已提交
3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141
    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]`.
3142 3143

    Args:
Y
yangyaming 已提交
3144 3145
       inputs (list): A list of variables to gather from. All variables have the
                same shape and the rank is at least 2.
3146
       index (Variable): Tensor<int32>, index variable which is a 2-D tensor
Y
yangyaming 已提交
3147
                with shape [M, 1] where M is the batch size.
3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160

    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
3161 3162 3163 3164 3165 3166

    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)
3167 3168 3169 3170 3171 3172
    helper.append_op(
        type='multiplex',
        inputs={'X': inputs,
                'Ids': index},
        outputs={'Out': [out]})
    return out
3173 3174 3175 3176 3177


def softmax_with_cross_entropy(logits, label, soft_label=False):
    """
    **Softmax With Cross Entropy Operator.**
3178

3179 3180 3181 3182
    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.
3183

3184 3185 3186
    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.
3187

3188 3189 3190
    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.
3191

3192
    The equation is as follows:
3193

3194
    1) Hard label (one-hot label, so every sample has exactly one class)
3195

3196 3197 3198 3199
    .. math::

        loss_j =  -\\text{logit}_{label_j} +
        \\log\\left(\\sum_{i=0}^{K}\\exp(\\text{logit}_i)\\right), j = 1,..., K
3200

3201 3202 3203
    2) Soft label (each sample can have a distribution over all classes)

    .. math::
3204

3205 3206 3207 3208 3209 3210 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 3238 3239 3240 3241 3242 3243 3244 3245 3246 3247 3248
        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].
3249

3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274
    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
3275
            out = fluid.layers.smooth_l1(x=fc, y=label)
3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291
    """
    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
3292 3293 3294 3295 3296 3297 3298 3299 3300


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
3301
        input(variable):  A Tensor/LodTensor of indices, last dimension must be 1.
3302 3303 3304 3305 3306 3307
        depth(scalar): an interger defining the depth of the one hot dimension.

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

    Examples:
C
caoying03 已提交
3308 3309
        .. code-block:: python

3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330
        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
3331 3332


Y
Yu Yang 已提交
3333
def autoincreased_step_counter(counter_name=None, begin=1, step=1):
Y
Yu Yang 已提交
3334
    """
Y
Yu Yang 已提交
3335
    NOTE: The counter will be automatically increased by 1 every mini-batch
Y
Yu Yang 已提交
3336
    Return the run counter of the main program, which is started with 1.
Y
Yu Yang 已提交
3337 3338 3339 3340 3341 3342

    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
3343 3344 3345
    Returns(Variable): The global run counter.
    """
    helper = LayerHelper('global_step_counter')
Y
Yu Yang 已提交
3346 3347
    if counter_name is None:
        counter_name = '@STEP_COUNTER@'
Y
Yu Yang 已提交
3348 3349 3350 3351 3352
    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
3353
                value=begin - 1, force_cpu=True))
Y
Yu Yang 已提交
3354 3355 3356
        helper.main_program.global_block().prepend_op(
            type='increment',
            inputs={'X': [counter]},
Y
Yu Yang 已提交
3357 3358
            outputs={'Out': [counter]},
            attrs={'step': float(step)})
Y
Yu Yang 已提交
3359 3360 3361
        counter.stop_gradient = True

    return counter
Y
yangyaming 已提交
3362 3363


3364
def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None):
C
caoying03 已提交
3365
    """
C
caoying03 已提交
3366 3367
    Gives a new shape to the input Tensor without changing its data.

3368 3369 3370 3371 3372
    The target shape can be given by :attr:`shape` or :attr:`actual_shape`.
    :attr:`shape` is a list of integer while :attr:`actual_shape` is a tensor
    variable. :attr:`actual_shape` has a higher priority than :attr:`shape`
    if it is provided, while :attr:`shape` still should be set correctly to
    gurantee shape inference in compile-time.
C
caoying03 已提交
3373

3374
    Some tricks exist when specifying the target shape.
C
caoying03 已提交
3375

3376 3377 3378 3379
    1. -1 means the value of this dimension is inferred from the total element
    number of x and remaining dimensions. Thus one and only one dimension can
    be set -1.

3380
    2. 0 means the actual dimension value is going to be copied from the
3381 3382 3383 3384
    corresponding dimension of x. The indice of 0s in shape can not exceed
    Rank(X).

    Here are some examples to explain it.
C
caoying03 已提交
3385 3386

    1. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
W
wanghaoshuang 已提交
3387
    is [6, 8], the reshape operator will transform x into a 2-D tensor with
3388
    shape [6, 8] and leaving x's data unchanged.
C
caoying03 已提交
3389

3390
    2. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
3391 3392
    specified is [2, 3, -1, 2], the reshape operator will transform x into a
    4-D tensor with shape [2, 3, 4, 2] and leaving x's data unchanged. In this
W
wanghaoshuang 已提交
3393 3394
    case, one dimension of the target shape is set to -1, the value of this
    dimension is inferred from the total element number of x and remaining
3395
    dimensions.
C
caoying03 已提交
3396

3397
    3. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape
3398 3399 3400 3401
    is [-1, 0, 3, 2], the reshape operator will transform x into a 4-D tensor
    with shape [2, 4, 3, 2] and leaving x's data unchanged. In this case,
    besides -1, 0 means the actual dimension value is going to be copied from
    the corresponding dimension of x.
C
caoying03 已提交
3402 3403 3404 3405 3406

    Args:
        input(variable): The input tensor.
        shape(list): The new shape. At most one dimension of the new shape can
                     be -1.
3407 3408 3409 3410 3411
        actual_shape(variable): An optional input. If provided, reshape
                                according to this given shape rather than
                                :attr:`shape` specifying shape. That is to
                                say :attr:`actual_shape` has a higher priority
                                than :attr:`shape`.
C
caoying03 已提交
3412 3413 3414 3415 3416 3417 3418 3419 3420
        act (str): The non-linear activation to be applied to output variable.
        inplace(bool): If this flag is set true, a new output tensor is created
                       whose data is copied from input x, otherwise the output
                       shares data with input without copying.

    Returns(variable): The output tensor.

    Examples:
        .. code-block:: python
G
guosheng 已提交
3421

3422
            data = fluid.layers.data(
3423
                name='data', shape=[2, 4, 6], dtype='float32')
C
caoying03 已提交
3424
            reshaped = fluid.layers.reshape(
3425
                x=data, shape=[-1, 0, 3, 2], act='tanh', inplace=True)
C
caoying03 已提交
3426 3427 3428 3429 3430
    """

    if not (isinstance(shape, list) or isinstance(shape, tuple)):
        raise ValueError("Input shape must be a python lsit or tuple.")

3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445
    # Validate the shape
    unk_dim_idx = -1
    for dim_idx, dim_size in enumerate(shape):
        if dim_size == -1:
            assert unk_dim_idx == -1, (
                "Only one dimension in shape can be unknown.")
            unk_dim_idx = dim_idx
        elif dim_size == 0:
            assert dim_idx < len(x.shape), (
                "The indice of 0s in shape can not exceed Rank(X).")
        else:
            assert dim_size > 0, (
                "Each dimension size given in shape must not be negtive "
                "except one unknown dimension.")

C
caoying03 已提交
3446 3447 3448 3449
    helper = LayerHelper("reshape", **locals())
    reshaped = helper.create_tmp_variable(dtype=x.dtype)
    helper.append_op(
        type="reshape",
3450 3451 3452
        inputs={"X": x,
                "Shape": actual_shape}
        if isinstance(actual_shape, Variable) else {"X": x},
C
caoying03 已提交
3453 3454 3455 3456 3457
        attrs={"shape": shape,
               "inplace": inplace},
        outputs={"Out": reshaped})

    return helper.append_activation(reshaped)
3458 3459


Y
yangyaming 已提交
3460
def lod_reset(x, y=None, target_lod=None):
Y
yangyaming 已提交
3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551 3552
    """
    LoD Reset Operator. Set LoD of **x** to a new one specified by **y** or
    **target_lod**. When **y** provided, **y.lod** would be considered as target
    LoD first, otherwise **y.data** would be considered as target LoD. If **y**
    is not provided, target LoD should be specified by **target_lod**.
    If target LoD is specified by **Y.data** or **target_lod**, only one level
    LoD is supported.

    .. code-block:: text

        * Example 1:

            Given a 1-level LoDTensor x:
                x.lod =  [[ 0,     2,                   5      6 ]]
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            target_lod: [0, 4, 6]

            then we get a 1-level LoDTensor:
                out.lod =  [[ 0,                   4,            6 ]]
                out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                out.dims = [6, 1]

        * Example 2:

            Given a 1-level LoDTensor x:
                x.lod =  [[ 0,     2,                   5      6 ]]
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a Tensor:
                y.data = [[0, 2, 6]]
                y.dims = [1, 3]

            then we get a 1-level LoDTensor:
                out.lod =  [[ 0,     2,                          6 ]]
                out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                out.dims = [6, 1]

        * Example 3:

            Given a 1-level LoDTensor x:
                x.lod =  [[ 0,      2,                   5     6 ]]
                x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                x.dims = [6, 1]

            y is a 2-level LoDTensor:
                y.lod =  [[0, 2, 4], [0, 2, 5, 6]]
                y.data = [[1.1], [2.1], [3.1], [4.1], [5.1], [6.1]]
                y.dims = [6, 1]

            then we get a 2-level LoDTensor:
                out.lod =  [[0, 2, 4], [0, 2, 5, 6]]
                out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]]
                out.dims = [6, 1]

    Args:
        x (Variable): Input variable which could be a Tensor or LodTensor.
        y (Variable|None): If provided, output's LoD would be derived from y.
        target_lod (list|tuple|None): One level LoD which should be considered
                                      as target LoD when y not provided.

    Returns:
        Variable: Output variable with LoD specified by this operator.

    Raises:
        ValueError: If y and target_lod are both None.

    Examples:
        .. code-block:: python

            x = layers.data(name='x', shape=[10])
            y = layers.data(name='y', shape=[10, 20], lod_level=2)
            out = layers.lod_reset(x=x, y=y)
    """
    helper = LayerHelper("lod_reset", **locals())
    out = helper.create_tmp_variable(dtype=x.dtype)
    if y is not None:
        helper.append_op(
            type="lod_reset", inputs={'X': x,
                                      'Y': y}, outputs={'Out': out})
    elif target_lod is not None:
        helper.append_op(
            type="lod_reset",
            inputs={'X': x},
            attrs={'target_lod': target_lod},
            outputs={'Out': out})
    else:
        raise ValueError("y and target_lod should not be both None.")

    return out
D
dragonwarrior 已提交
3553 3554 3555 3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622


def lrn(input, n=5, k=1.0, alpha=1e-4, beta=0.75, name=None):
    """
    Local Response Normalization Layer. This layer performs a type of
    "lateral inhibition" by normalizing over local input regions.

    The formula is as follows:

    .. math::

        Output(i, x, y) = Input(i, x, y) / \left(
        k + \alpha \sum\limits^{\min(C, c + n/2)}_{j = \max(0, c - n/2)}
        (Input(j, x, y))^2 \right)^{\beta}

    In the above equation:

    * :math:`n`: The number of channels to sum over.
    * :math:`k`: The offset (avoid being divided by 0).
    * :math:`alpha`: The scaling parameter.
    * :math:`beta`: The exponent parameter.

    Refer to `ImageNet Classification with Deep Convolutional Neural Networks
    <https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf>`_

    Args:
        input (Variable): The input tensor of this layer, and the dimension of input tensor must be 4.
        n (int, default 5): The number of channels to sum over.
        k (float, default 1.0): An offset (usually positive to avoid dividing by 0).
        alpha (float, default 1e-4): The scaling parameter.
        beta (float, default 0.75): The exponent.
        name (str, default None): A name for this operation.

    Raises:
        ValueError: If rank of the input tensor is not 4.

    Returns:
        A tensor variable storing the transformation result.

    Examples:
        .. code-block:: python

          data = fluid.layers.data(name="data", shape=[3, 112, 112], dtype="float32")
          lrn = fluid.layers.lrn(input=data)
    """
    helper = LayerHelper('lrn', **locals())
    dtype = helper.input_dtype()
    input_shape = input.shape
    dims = len(input_shape)

    if dims != 4:
        raise ValueError(
            "dims of input must be 4(not %d), and it's order must be NCHW" %
            (dims))

    mid_out = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
    lrn_out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type="lrn",
        inputs={"X": input},
        outputs={
            "Out": lrn_out,
            "MidOut": mid_out,
        },
        attrs={"n": n,
               "k": k,
               "alpha": alpha,
               "beta": beta})

    return lrn_out
G
guosheng 已提交
3623 3624 3625 3626


def pad(x, paddings, pad_value=0., name=None):
    """
G
guosheng 已提交
3627
    Pads a tensor with a constant value given by :attr:`pad_value`, and the
W
wanghaoshuang 已提交
3628
    padded width is specified by :attr:`paddings`.
G
guosheng 已提交
3629

G
guosheng 已提交
3630 3631 3632 3633
    Specifically, the number of values padded before the contents of :attr:`x`
    in dimension :attr:`i` is indicated by :attr:`paddings[i]`, and the number
    of values padded after the contents of :attr:`x` in dimension :attr:`i` is
    indicated by :attr:`paddings[i+1]`.
G
guosheng 已提交
3634 3635 3636 3637 3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648 3649 3650 3651 3652 3653 3654 3655

    See below for an example.

    .. code-block:: text

        Given:
            x = [[1, 2], [3, 4]]

            paddings = [0, 1, 1, 2]

            pad_value = 0

        Return:

            out = [[0, 1, 2, 0, 0]
                   [0, 3, 4, 0, 0]
                   [0, 0, 0, 0, 0]]

    Args:
        x (Variable): The input tensor variable.
        paddings (list): A list of integers. Its elements specify the padded
                         width before and after for each dimension in turn.
W
wanghaoshuang 已提交
3656
                         The length of :attr:paddings must be
G
guosheng 已提交
3657 3658 3659 3660 3661 3662 3663 3664 3665 3666
                         :math:`rank(x) \\times 2`.
        pad_value (float): The constant value used to pad.
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        Variable: The padded tensor variable.

    Examples:
        .. code-block:: python
G
guosheng 已提交
3667

G
guosheng 已提交
3668 3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681
            # x is a rank 2 tensor variable.
            out = fluid.layers.pad(
                x=x, paddings=[0, 1, 1, 2], pad_value=0.)
    """
    helper = LayerHelper('pad', input=x, **locals())
    dtype = helper.input_dtype()
    out = helper.create_tmp_variable(dtype)
    helper.append_op(
        type='pad',
        inputs={'X': x},
        outputs={'Out': out},
        attrs={'paddings': paddings,
               'pad_value': float(pad_value)})
    return out
3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713 3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746


def label_smooth(label,
                 prior_dist=None,
                 epsilon=0.1,
                 dtype="float32",
                 name=None):
    """
    Label smoothing is a mechanism to regularize the classifier layer and is
    called label-smoothing regularization (LSR). 
    
    Label smoothing is proposed to encourage the model to be less confident,
    since optimizing the log-likelihood of the correct label directly may
    cause overfitting and reduce the ability of the model to adapt. Label
    smoothing replaces the ground-truth label :math:`y` with the weighted sum
    of itself and some fixed distribution :math:`\mu`. For class :math:`k`,
    i.e.

    .. math::

        \\tilde{y_k} = (1 - \epsilon) * y_k + \epsilon * \mu_k,

    where :math:`1 - \epsilon` and :math:`\epsilon` are the weights
    respectively, and :math:`\\tilde{y}_k` is the smoothed label. Usually
    uniform distribution is used for :math:`\mu`.

    See more details about label smoothing in https://arxiv.org/abs/1512.00567.

    Args:
        label(Variable): The input variable containing the label data. The
                          label data should use one-hot representation.
        prior_dist(Variable): The prior distribution to be used to smooth
                              labels. If not provided, an uniform distribution
                              is used. The shape of :attr:`prior_dist` should
                              be :math:`(1, class\_num)`. 
        epsilon(float): The weight used to mix up the original ground-truth
                        distribution and the fixed distribution.
        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, 
                                                  float_64, int etc.
        name(str|None): A name for this layer(optional). If set None, the layer
                        will be named automatically.

    Returns:
        Variable: The tensor variable containing the smoothed labels.

    Examples:
        .. code-block:: python

            label = layers.data(name="label", shape=[1], dtype="float32")
            one_hot_label = layers.one_hot(input=label, depth=10)
            smooth_label = layers.label_smooth(
                label=one_hot_label, epsilon=0.1, dtype="float32")
    """
    if epsilon > 1. or epsilon < 0.:
        raise ValueError("The value of epsilon must be between 0 and 1.")
    helper = LayerHelper("label_smooth", **locals())
    label.stop_gradient = True
    smooth_label = helper.create_tmp_variable(dtype)
    helper.append_op(
        type="label_smooth",
        inputs={"X": label,
                "PriorDist": prior_dist} if prior_dist else {"X": label},
        outputs={"Out": smooth_label},
        attrs={"epsilon": float(epsilon)})
    return smooth_label