nn.py 55.1 KB
Newer Older
Y
Yu Yang 已提交
1 2 3 4 5 6 7
"""
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
8
from ..param_attr import ParamAttr
Y
yangyaming 已提交
9
from tensor import concat
Y
Yu Yang 已提交
10 11 12 13 14

__all__ = [
    'fc', 'embedding', 'dynamic_lstm', 'gru_unit', 'linear_chain_crf',
    'crf_decoding', 'cos_sim', 'cross_entropy', 'square_error_cost', 'accuracy',
    'chunk_eval', 'sequence_conv', 'conv2d', 'sequence_pool', 'pool2d',
15
    'batch_norm', 'beam_search_decode', 'conv2d_transpose', 'sequence_expand',
16
    'lstm_unit', 'reduce_sum', 'reduce_mean', 'reduce_max', 'reduce_min',
W
wanghaoshuang 已提交
17
    'sequence_first_step', 'sequence_last_step', 'dropout', 'warpctc'
Y
Yu Yang 已提交
18 19 20 21 22 23 24 25 26
]


def fc(input,
       size,
       num_flatten_dims=1,
       param_attr=None,
       bias_attr=None,
       act=None,
27
       name=None):
Y
Yu Yang 已提交
28
    """
29
    **Fully Connected Layer**
Y
Yu Yang 已提交
30

C
caoying03 已提交
31 32 33 34 35 36 37 38 39
    The fully connected layer can take multiple tensors as its inputs. It
    creates a variable (one for each input tensor) called weights for each input
    tensor, which represents a fully connected weight matrix from each input
    unit to each output unit. The fully connected layer multiplies each input
    tensor with its coresponding weight to produce an output Tensor. If
    multiple input tensors are given, the results of multiple multiplications
    will be sumed up. If bias_attr is not None, a biases variable will be
    created and added to the output. Finally, if activation is not None,
    it will be applied to the output as well.
C
caoying03 已提交
40

C
caoying03 已提交
41
    This process can be formulated as follows:
42 43 44

    .. math::

C
caoying03 已提交
45
        Out = Act({\sum_{i=0}^{N-1}W_iX_i + b})
46 47 48

    In the above equation:

C
caoying03 已提交
49 50 51 52
    * :math:`N`: Number of the input.
    * :math:`X_i`: The input tensor.
    * :math:`W`: The weights created by this layer.
    * :math:`b`: The bias parameter created by this layer (if needed).
C
caoying03 已提交
53 54
    * :math:`Act`: The activation funtion.
    * :math:`Out`: The output tensor.
Y
Yu Yang 已提交
55 56

    Args:
C
caoying03 已提交
57 58 59 60 61 62 63 64 65 66
       input(Variable|list): The input tensor(s) to the fully connected layer.
       size(int): The number of output units in the fully connected layer.
       num_flatten_dims(int): The fc layer can accept an input tensor with more
                              than two dimensions. If this happens, the
                              multidimensional tensor will first be flattened
                              into a 2-dimensional matrix. The parameter
                              `num_flatten_dims` determines how the input tensor
                              is flattened: the first `num_flatten_dims`
                              dimensions will be flatten to form the first
                              dimension of the final matrix (height of the
E
emailweixu 已提交
67
                              matrix), and the rest `rank(X) - num_flatten_dims`
C
caoying03 已提交
68 69 70 71
                              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
E
emailweixu 已提交
72
                              `num_flatten_dims` = 3. Then, the flattened matrix
C
caoying03 已提交
73
                              will have a shape [2 x 3 x 4, 5 x 6] = [24, 30].
E
emailweixu 已提交
74
                              By default, `num_flatten_dims` is set to 1.
C
caoying03 已提交
75 76 77 78 79 80 81 82 83 84 85 86 87 88 89
       param_attr(ParamAttr|list): The parameter attribute for learnable
                                   parameters/weights of the fully connected
                                   layer.
       param_initializer(ParamAttr|list): The initializer used for the
                                          weight/parameter. If set None,
                                          XavierInitializer() will be used.
       bias_attr(ParamAttr|list): The parameter attribute for the bias parameter
                                  for this layer. If set None, no bias will be
                                  added to the output units.
       bias_initializer(ParamAttr|list): The initializer used for the bias.
                                        If set None, then ConstantInitializer()
                                        will be used.
       act(str): Activation to be applied to the output of the fully connected
                 layer.
       name(str): Name/alias of the fully connected layer.
Y
Yu Yang 已提交
90 91


92
    Returns:
C
caoying03 已提交
93
        Variable: The output tensor variable.
94 95

    Raises:
C
caoying03 已提交
96
        ValueError: If rank of the input tensor is less than 2.
97 98 99 100

    Examples:
        .. code-block:: python

C
caoying03 已提交
101
          data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
102
          fc = fluid.layers.fc(input=data, size=1000, act="tanh")
Y
Yu Yang 已提交
103
    """
C
caoying03 已提交
104

C
caoying03 已提交
105
    helper = LayerHelper("fc", **locals())
Y
Yu Yang 已提交
106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124

    dtype = helper.input_dtype()

    mul_results = []
    for input_var, param_attr in helper.iter_inputs_and_params():
        input_shape = input_var.shape
        param_shape = [
            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
        ] + [size]
        w = helper.create_parameter(
            attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False)
        tmp = helper.create_tmp_variable(dtype)
        helper.append_op(
            type="mul",
            inputs={
                "X": input_var,
                "Y": w,
            },
            outputs={"Out": tmp},
C
caoying03 已提交
125 126
            attrs={"x_num_col_dims": num_flatten_dims,
                   "y_num_col_dims": 1})
Y
Yu Yang 已提交
127 128 129 130 131 132 133 134 135 136 137 138 139 140 141
        mul_results.append(tmp)

    # sum
    if len(mul_results) == 1:
        pre_bias = mul_results[0]
    else:
        pre_bias = helper.create_tmp_variable(dtype)
        helper.append_op(
            type="sum", inputs={"X": mul_results}, outputs={"Out": pre_bias})
    # add bias
    pre_activation = helper.append_bias_op(pre_bias)
    # add activation
    return helper.append_activation(pre_activation)


142
def embedding(input, size, is_sparse=False, param_attr=None, dtype='float32'):
Y
Yu Yang 已提交
143
    """
144 145 146 147 148 149 150
    **Embedding Layer**

    This layer is used to lookup a vector of IDs, provided by *input*, in a lookup table.
    The result of this lookup is the embedding of each ID in the *input*.

    All the input variables are passed in as local variables to the LayerHelper
    constructor.
Y
Yu Yang 已提交
151 152

    Args:
153
       input(Variable): Input to the function
Y
yangyaming 已提交
154
       size(tuple|list|None): Shape of the look up table parameter
155 156 157
       is_sparse(bool): Boolean flag that specifying whether the input is sparse
       param_attr(ParamAttr): Parameters for this layer
       dtype(np.dtype|core.DataType|str): The type of data : float32, float_16, int etc
Y
Yu Yang 已提交
158

159 160 161
    Returns:
        Variable: The tensor variable storing the embeddings of the \
                  supplied inputs.
Y
Yu Yang 已提交
162

163 164
    Examples:
        .. code-block:: python
Y
Yu Yang 已提交
165

C
chengduoZH 已提交
166
          dict_size = len(dataset.ids)
167
          data = fluid.layers.data(name='ids', shape=[32, 32], dtype='float32')
C
chengduoZH 已提交
168
          fc = fluid.layers.embedding(input=data, size=[dict_size, 16])
Y
Yu Yang 已提交
169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193
    """

    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)
    helper.append_op(
        type='lookup_table',
        inputs={'Ids': input,
                'W': w},
        outputs={'Out': tmp},
        attrs={'is_sparse': is_sparse})
    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',
194
                 dtype='float32'):
Y
Yu Yang 已提交
195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236
    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


def gru_unit(input,
             hidden,
             size,
             weight=None,
             bias=None,
             activation='tanh',
237
             gate_activation='sigmoid'):
Y
Yu Yang 已提交
238
    """
239
    GRU unit layer. The equation of a gru step is:
Y
Yu Yang 已提交
240

241 242
        .. math::
            u_t & = actGate(xu_{t} + W_u h_{t-1} + b_u)
Y
Yu Yang 已提交
243

244
            r_t & = actGate(xr_{t} + W_r h_{t-1} + b_r)
Y
Yu Yang 已提交
245

246
            m_t & = actNode(xm_t + W_c dot(r_t, h_{t-1}) + b_m)
247

248
            h_t & = dot((1-u_t), m_t) + dot(u_t, h_{t-1})
249 250

    The inputs of gru unit includes :math:`z_t`, :math:`h_{t-1}`. In terms
W
wanghaoshuang 已提交
251 252 253
    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
254 255
    connected layer has to be applied, such that :math:`z_t = W_{fc}x_t`.

W
wanghaoshuang 已提交
256 257
    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
258 259 260
    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`.
261 262 263 264 265 266 267 268 269

    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
        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
270

271 272 273 274 275 276
    Returns:
        tuple: The hidden value, reset-hidden value and gate values.

    Examples:

        .. code-block:: python
Y
Yu Yang 已提交
277

278
             # assuming we have x_t_data and prev_hidden of size=10
W
wanghaoshuang 已提交
279
             x_t = fluid.layers.fc(input=x_t_data, size=30)
280 281
             hidden_val, r_h_val, gate_val = fluid.layers.gru_unit(input=x_t,
                                                    hidden = prev_hidden)
Y
Yu Yang 已提交
282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301

    """
    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
302

Y
Yu Yang 已提交
303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329
    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


330
def linear_chain_crf(input, label, param_attr=None):
Y
Yu Yang 已提交
331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355
    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


356
def crf_decoding(input, param_attr, label=None):
Y
Yu Yang 已提交
357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388
    helper = LayerHelper('crf_decoding', **locals())
    transition = helper.get_parameter(param_attr.name)
    viterbi_path = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
        type='crf_decoding',
        inputs={"Emission": [input],
                "Transition": transition,
                "Label": label},
        outputs={"ViterbiPath": [viterbi_path]})

    return viterbi_path


def cos_sim(X, Y, **kwargs):
    """
    This function performs the cosine similarity between two tensors
    X and Y and returns that as the output.
    """
    helper = LayerHelper('cos_sim', **kwargs)
    out = helper.create_tmp_variable(dtype=X.dtype)
    xnorm = helper.create_tmp_variable(dtype=X.dtype)
    ynorm = helper.create_tmp_variable(dtype=X.dtype)
    helper.append_op(
        type='cos_sim',
        inputs={'X': [X],
                'Y': [Y]},
        outputs={'Out': [out],
                 'XNorm': [xnorm],
                 'YNorm': [ynorm]})
    return out


389 390 391 392 393 394 395 396 397 398 399 400 401 402 403
def dropout(x, dropout_prob, is_test=False, seed=0, **kwargs):
    helper = LayerHelper('dropout', **kwargs)
    out = helper.create_tmp_variable(dtype=x.dtype)
    mask = helper.create_tmp_variable(dtype=x.dtype, stop_gradient=True)
    helper.append_op(
        type='dropout',
        inputs={'X': [x]},
        outputs={'Out': [out],
                 'Mask': [mask]},
        attrs={'dropout_prob': dropout_prob,
               'is_test': is_test,
               'seed': seed})
    return out


Y
Yu Yang 已提交
404 405
def cross_entropy(input, label, **kwargs):
    """
Y
Yibing Liu 已提交
406 407 408 409 410 411
    **Cross Entropy Layer**

    This layer computes the cross entropy between `input` and `label`. It supports
    both standard cross-entropy and soft-label cross-entropy loss computation.

    1) One-hot cross-entropy:
Y
Yibing Liu 已提交
412
	`soft_label = False`, `Label[i, 0]` indicates the class index for sample i:
Y
yangyaming 已提交
413

Y
Yibing Liu 已提交
414
        .. math::
Y
yangyaming 已提交
415

Y
Yibing Liu 已提交
416 417 418
            Y[i] = -\log(X[i, Label[i]])

    2) Soft-label cross-entropy:
Y
Yibing Liu 已提交
419
	`soft_label = True`, `Label[i, j]` indicates the soft label of class j
Y
Yibing Liu 已提交
420 421 422 423 424 425
	for sample i:

        .. math::

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

Y
Yibing Liu 已提交
426
       Please make sure that in this case the summation of each row of `label`
Y
Yibing Liu 已提交
427 428 429 430
       equals one.

    3) One-hot cross-entropy with vecterized `label`:
	 As a special case of 2), when each row of 'label' has only one
Y
Yibing Liu 已提交
431 432
	 non-zero element which is equal to 1, soft-label cross-entropy degenerates
         to a one-hot cross-entropy with one-hot label representation.
Y
yangyaming 已提交
433

Y
Yibing Liu 已提交
434
    Args:
Y
yangyaming 已提交
435 436
        input (Variable|list):  a 2-D tensor with shape [N x D], where N is the
            batch size and D is the number of classes. This input is a probability
Y
Yibing Liu 已提交
437 438
            computed by the previous operator, which is almost always the result
            of a softmax operator.
Y
yangyaming 已提交
439 440 441
        label (Variable|list): the ground truth which is a 2-D tensor. When
              `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
Y
Yibing Liu 已提交
442
              tensor<float/double> with shape [N x D].
Y
Yibing Liu 已提交
443
        soft_label (bool, via `**kwargs`): a flag indicating whether to interpretate
Y
Yibing Liu 已提交
444
              the given labels as soft labels, default `False`.
Y
Yibing Liu 已提交
445 446 447 448 449

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

    Raises:
Y
yangyaming 已提交
450
        `ValueError`: 1) the 1st dimension of `input` and `label` are not equal; 2) when \
Y
Yibing Liu 已提交
451 452
              `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
453 454 455 456 457 458

    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
459 460 461 462 463 464 465 466 467 468 469 470 471 472
    """
    helper = LayerHelper('cross_entropy', **kwargs)
    out = helper.create_tmp_variable(dtype=input.dtype)
    helper.append_op(
        type='cross_entropy',
        inputs={'X': [input],
                'Label': [label]},
        outputs={'Y': [out]},
        attrs=kwargs)
    return out


def square_error_cost(input, label, **kwargs):
    """
473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502
    **Square error cost layer**

    This layer accepts input predictions and target label and returns the squared error cost.
    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:
        Variable: The tensor variable storing the element-wise squared error difference \
                  of input and label.

    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
503 504 505 506 507 508 509 510 511 512 513
    """
    helper = LayerHelper('square_error_cost', **kwargs)
    minus_out = helper.create_tmp_variable(dtype=input.dtype)
    helper.append_op(
        type='elementwise_sub',
        inputs={'X': [input],
                'Y': [label]},
        outputs={'Out': [minus_out]})

    square_out = helper.create_tmp_variable(dtype=input.dtype)
    helper.append_op(
F
fengjiayi 已提交
514 515
        type='square', inputs={'X': [minus_out]},
        outputs={'Out': [square_out]})
Y
Yu Yang 已提交
516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559
    return square_out


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


def chunk_eval(input,
               label,
               chunk_scheme,
               num_chunk_types,
               excluded_chunk_types=None,
               **kwargs):
    """
Y
yangyaming 已提交
560
    This function computes and outputs the precision, recall and
561
    F1-score of chunk detection.
Y
Yu Yang 已提交
562 563 564 565 566 567 568
    """
    helper = LayerHelper("chunk_eval", **kwargs)

    # prepare output
    precision = helper.create_tmp_variable(dtype="float32")
    recall = helper.create_tmp_variable(dtype="float32")
    f1_score = helper.create_tmp_variable(dtype="float32")
569 570 571
    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
572 573 574 575 576 577 578 579

    helper.append_op(
        type="chunk_eval",
        inputs={"Inference": [input],
                "Label": [label]},
        outputs={
            "Precision": [precision],
            "Recall": [recall],
580 581 582 583
            "F1-Score": [f1_score],
            "NumInferChunks": [num_infer_chunks],
            "NumLabelChunks": [num_label_chunks],
            "NumCorrectChunks": [num_correct_chunks]
Y
Yu Yang 已提交
584 585 586
        },
        attrs={
            "num_chunk_types": num_chunk_types,
G
guosheng 已提交
587 588
            "chunk_scheme": chunk_scheme,
            "excluded_chunk_types": excluded_chunk_types or []
Y
Yu Yang 已提交
589
        })
590
    return precision, recall, f1_score, num_infer_chunks, num_label_chunks, num_correct_chunks
Y
Yu Yang 已提交
591 592 593 594 595 596 597 598 599


def sequence_conv(input,
                  num_filters,
                  filter_size=3,
                  filter_stride=1,
                  padding=None,
                  bias_attr=None,
                  param_attr=None,
600
                  act=None):
Y
Yu Yang 已提交
601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641
    """
    This function creates the op for sequence_conv, using the inputs and
    other convolutional configurations for the filters and stride as given
    in the input parameters to the function.
    """

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

    helper = LayerHelper('sequence_conv', **locals())
    dtype = helper.input_dtype()
    filter_shape = [filter_size * input.shape[1], num_filters]
    filter_param = helper.create_parameter(
        attr=helper.param_attr, shape=filter_shape, dtype=dtype)
    pre_bias = helper.create_tmp_variable(dtype)

    helper.append_op(
        type='sequence_conv',
        inputs={
            'X': [input],
            'Filter': [filter_param],
        },
        outputs={"Out": pre_bias},
        attrs={
            'contextStride': filter_stride,
            'contextStart': -int(filter_size / 2),
            'contextLength': filter_size
        })
    pre_act = helper.append_bias_op(pre_bias)
    return helper.append_activation(pre_act)


def conv2d(input,
           num_filters,
           filter_size,
           stride=None,
           padding=None,
           groups=None,
           param_attr=None,
           bias_attr=None,
C
chengduoZH 已提交
642
           act=None):
Y
Yu Yang 已提交
643
    """
C
chengduoZH 已提交
644 645 646 647 648 649 650 651
    **Convlution2D Layer**

    The convolution2D layer calculates the output based on the input, filter
    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.
    The details of convolution layer, please refer UFLDL's `convolution,
    <http://ufldl.stanford.edu/tutorial/supervised/FeatureExtractionUsingConvolution/>`_ .
C
refine  
chengduoZH 已提交
652
    If bias attribution and activation type are provided, bias is added to the output of the convolution,
C
chengduoZH 已提交
653 654 655
    and the corresponding activation function is applied to the final result.
    For each input :math:`X`, the equation is:

C
refine  
chengduoZH 已提交
656

C
chengduoZH 已提交
657 658
    .. math::

C
refine  
chengduoZH 已提交
659
        Out = \sigma (W \\ast X + b)
C
chengduoZH 已提交
660

C
chengduoZH 已提交
661
    In the above equation:
C
chengduoZH 已提交
662 663 664

        * :math:`X`: Input value, a tensor with NCHW format.
        * :math:`W`: Filter value, a tensor with MCHW format.
C
chengduoZH 已提交
665
        * :math:`\\ast`: Convolution operation.
C
refine  
chengduoZH 已提交
666
        * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
C
chengduoZH 已提交
667
        * :math:`\\sigma`: Activation function.
C
chengduoZH 已提交
668 669 670 671
        * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.

    Example:

C
chengduoZH 已提交
672 673
        Input:
            Input shape: $(N, C_{in}, H_{in}, W_{in})$
C
refine  
chengduoZH 已提交
674

C
chengduoZH 已提交
675
            Filter shape: $(C_{out}, C_{in}, H_f, W_f)$
C
refine  
chengduoZH 已提交
676

C
chengduoZH 已提交
677 678
        Output:
            Output shape: $(N, C_{out}, H_{out}, W_{out})$
C
chengduoZH 已提交
679
        Where
C
chengduoZH 已提交
680
    .. math::
C
chengduoZH 已提交
681

C
chengduoZH 已提交
682 683
        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
684 685

    Args:
C
chengduoZH 已提交
686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705
        input(Variable): The input image with [N, C, H, W] format.
        num_filters(int): The number of filter. It is as same as the output
            image channel.
        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
            it must contain two integers, (filter_size_H, filter_size_W).
            Otherwise, the filter will be a square.
        stride(int|tuple): The stride size. If stride is a tuple, it must
            contain two integers, (stride_H, stride_W). Otherwise, the
            stride_H = stride_W = stride. Default: stride = 1.
        padding(int|tuple): The padding size. If padding is a tuple, it must
            contain two integers, (padding_H, padding_W). Otherwise, the
            padding_H = padding_W = padding. Default: padding = 0.
        groups(int): The groups number of the Conv2d Layer. According to grouped
            convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
            the first half of the filters is only connected to the first half
            of the input channels, while the second half of the filters is only
            connected to the second half of the input channels. Default: groups=1
        param_attr(ParamAttr): The parameters to the Conv2d Layer. Default: None
        bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None
        act(str): Activation type. Default: None
C
chengduoZH 已提交
706 707 708 709 710

    Returns:
        Variable: The tensor variable storing the convolution and \
                  non-linearity activation result.

C
refine  
chengduoZH 已提交
711 712 713
    Raises:
        ValueError: If the shapes of input, filter_size, stride, padding and groups mismatch.

C
chengduoZH 已提交
714 715 716
    Examples:
        .. code-block:: python

C
refine  
chengduoZH 已提交
717
          data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32')
C
chengduoZH 已提交
718
          conv2d = fluid.layers.conv2d(input=data, num_filters=2, filter_size=3, act="relu")
Y
Yu Yang 已提交
719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773
    """

    if stride is None:
        stride = [1, 1]
    helper = LayerHelper('conv2d', **locals())
    dtype = helper.input_dtype()

    num_channels = input.shape[1]
    if groups is None:
        num_filter_channels = num_channels
    else:
        if num_channels % groups != 0:
            raise ValueError("num_channels must be divisible by groups.")
        num_filter_channels = num_channels / groups

    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]

    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(
        type='conv2d_cudnn',
        inputs={
            'Input': input,
            'Filter': filter_param,
        },
        outputs={"Output": pre_bias},
        attrs={'strides': stride,
               'paddings': padding,
               'groups': groups})

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

    return helper.append_activation(pre_act)


def sequence_pool(input, pool_type, **kwargs):
    """
Y
yangyaming 已提交
774 775 776
    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
777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801

    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
802

L
Luo Tao 已提交
803 804
    Args:
        input(variable): The input variable which is a LoDTensor.
Y
yangyaming 已提交
805
        pool_type (string): The pooling type of sequence_pool.
L
Luo Tao 已提交
806 807 808 809 810 811 812 813
            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
814

Y
yangyaming 已提交
815
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
816 817 818 819 820
                              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
821 822 823 824 825 826 827 828 829 830 831 832 833
    """
    helper = LayerHelper('sequence_pool', input=input, **kwargs)
    dtype = helper.input_dtype()
    pool_out = helper.create_tmp_variable(dtype)
    max_index = helper.create_tmp_variable(dtype)

    helper.append_op(
        type="sequence_pool",
        inputs={"X": input},
        outputs={"Out": pool_out,
                 "MaxIndex": max_index},
        attrs={"pooltype": pool_type.upper()})

Y
yangyaming 已提交
834 835 836 837 838
    # 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
839 840 841
    return pool_out


842
def sequence_first_step(input, **kwargs):
L
Luo Tao 已提交
843 844 845 846 847 848 849 850 851 852 853 854 855 856
    """
    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
857

L
Luo Tao 已提交
858 859 860 861 862 863 864 865 866
    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
867

Y
yangyaming 已提交
868
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
869 870 871
                              dtype='float32', lod_level=1)
             x_first_step = fluid.layers.sequence_first_step(input=x)
    """
872 873 874 875
    return sequence_pool(input=input, pool_type="first")


def sequence_last_step(input, **kwargs):
L
Luo Tao 已提交
876 877 878 879 880 881 882 883 884 885 886 887 888 889
    """
    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
890

L
Luo Tao 已提交
891 892 893 894 895 896 897 898 899
    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
900

Y
yangyaming 已提交
901
             x = fluid.layers.data(name='x', shape=[7, 1],
L
Luo Tao 已提交
902 903 904
                              dtype='float32', lod_level=1)
             x_last_step = fluid.layers.sequence_last_step(input=x)
    """
905 906 907
    return sequence_pool(input=input, pool_type="last")


Y
Yu Yang 已提交
908 909 910 911 912
def pool2d(input,
           pool_size,
           pool_type,
           pool_stride=None,
           pool_padding=None,
913
           global_pooling=False):
Y
Yu Yang 已提交
914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958
    """
    This function adds the operator for pooling in 2 dimensions, using the
    pooling configurations mentioned in input parameters.
    """
    if pool_padding is None:
        pool_padding = [0, 0]
    if pool_stride is None:
        pool_stride = [1, 1]
    if pool_type not in ["max", "avg"]:
        raise ValueError(
            "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",
            str(pool_type))
    if isinstance(pool_size, int):
        pool_size = [pool_size, pool_size]
    if isinstance(pool_stride, int):
        pool_stride = [pool_stride, pool_stride]
    if isinstance(pool_padding, int):
        pool_padding = [pool_padding, pool_padding]

    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,
            "paddings": pool_padding
        })

    return pool_out


def batch_norm(input,
               act=None,
               is_test=False,
               momentum=0.9,
               epsilon=1e-05,
               param_attr=None,
               bias_attr=None,
959
               data_layout='NCHW'):
Y
Yu Yang 已提交
960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988
    """
    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(
        attr=helper.param_attr, shape=param_shape, dtype=dtype, is_bias=True)

    mean = helper.create_global_variable(
Q
QI JUN 已提交
989 990 991 992
        dtype=input.dtype,
        shape=param_shape,
        persistable=True,
        stop_gradient=True)
Y
Yu Yang 已提交
993 994 995
    helper.set_variable_initializer(var=mean, initializer=Constant(0.0))

    variance = helper.create_global_variable(
Q
QI JUN 已提交
996 997 998 999
        dtype=input.dtype,
        shape=param_shape,
        persistable=True,
        stop_gradient=True)
Y
Yu Yang 已提交
1000 1001 1002 1003 1004 1005 1006
    helper.set_variable_initializer(var=variance, initializer=Constant(1.0))

    # create output
    # mean and mean_out share the same memory
    mean_out = mean
    # variance and variance out share the same memory
    variance_out = variance
Q
QI JUN 已提交
1007 1008
    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
1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034

    batch_norm_out = helper.create_tmp_variable(dtype)

    helper.append_op(
        type="batch_norm",
        inputs={
            "X": input,
            "Scale": scale,
            "Bias": bias,
            "Mean": mean,
            "Variance": variance
        },
        outputs={
            "Y": batch_norm_out,
            "MeanOut": mean_out,
            "VarianceOut": variance_out,
            "SavedMean": saved_mean,
            "SavedVariance": saved_variance
        },
        attrs={"momentum": momentum,
               "epsilon": epsilon,
               "is_test": is_test})

    return helper.append_activation(batch_norm_out)


1035
def beam_search_decode(ids, scores):
Y
Yu Yang 已提交
1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057
    helper = LayerHelper('beam_search_decode', **locals())
    sentence_ids = helper.create_tmp_variable(dtype=ids.dtype)
    sentence_scores = helper.create_tmp_variable(dtype=ids.dtype)

    helper.append_op(
        type="beam_search_decode",
        inputs={"Ids": ids,
                "Scores": scores},
        outputs={
            "SentenceIds": sentence_ids,
            "SentenceScores": sentence_scores
        })

    return sentence_ids, sentence_scores


def conv2d_transpose(input,
                     num_filters,
                     output_size=None,
                     filter_size=None,
                     padding=None,
                     stride=None,
C
chengduoZH 已提交
1058
                     dilation=None,
1059
                     param_attr=None):
Y
Yu Yang 已提交
1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081
    """
    The transpose of conv2d layer.

    This layer is also known as deconvolution layer.

    Args:
        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.
        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.
        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.
C
chengduoZH 已提交
1082 1083 1084
        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.
Y
Yu Yang 已提交
1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104
        param_attr: Parameter Attribute.
        main_program(Program): the main program
        startup_program(Program): the startup program

    Returns:
        Variable: Output image.
    """
    helper = LayerHelper("conv2d_transpose", **locals())
    if not isinstance(input, Variable):
        raise TypeError("Input of conv2d_transpose must be Variable")
    input_channel = input.shape[1]

    op_attr = dict()

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

    if isinstance(stride, int):
C
chengduoZH 已提交
1105
        op_attr['strides'] = [stride, stride]
Y
Yu Yang 已提交
1106 1107 1108
    elif stride is not None:
        op_attr['strides'] = stride

C
chengduoZH 已提交
1109 1110 1111 1112 1113
    if isinstance(dilation, int):
        op_attr['dilations'] = [dilation, dilation]
    elif dilation is not None:
        op_attr['dilations'] = dilation

Y
Yu Yang 已提交
1114 1115 1116 1117 1118 1119 1120 1121
    if filter_size is None:
        if output_size is None:
            raise ValueError("output_size must be set when filter_size is None")
        if isinstance(output_size, int):
            output_size = [output_size, output_size]

        padding = op_attr.get('paddings', [0, 0])
        stride = op_attr.get('strides', [1, 1])
C
chengduoZH 已提交
1122
        dilation = op_attr.get('dilations', [1, 1])
Y
Yu Yang 已提交
1123 1124 1125

        h_in = input.shape[2]
        w_in = input.shape[3]
C
chengduoZH 已提交
1126 1127 1128 1129 1130

        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
1131
        filter_size = [filter_size_h, filter_size_w]
C
chengduoZH 已提交
1132

Y
Yu Yang 已提交
1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148
    elif isinstance(filter_size, int):
        filter_size = [filter_size, filter_size]

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

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

    return out
Y
yangyaming 已提交
1149 1150


1151
def sequence_expand(x, y):
1152 1153
    """Sequence Expand Layer. This layer will expand the input variable **x**
    according to LoD information of **y**. And the following examples will
Y
yangyaming 已提交
1154
    explain how sequence_expand works:
1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182

    .. code-block:: text

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

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

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

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

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

            y is a LoDTensor:
Y
yangyaming 已提交
1183
                y.lod = [[0, 2, 3, 6]]
1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204

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

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

    Args:
        x (Variable): The input variable which is a Tensor or LoDTensor.
        y (Variable): The input variable which is a LoDTensor.

    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
1205
            out = layers.sequence_expand(x=x, y=y)
1206
    """
Y
yangyaming 已提交
1207
    helper = LayerHelper('sequence_expand', input=x, **locals())
1208 1209 1210
    dtype = helper.input_dtype()
    tmp = helper.create_tmp_variable(dtype)
    helper.append_op(
Y
yangyaming 已提交
1211 1212
        type='sequence_expand', inputs={'X': x,
                                        'Y': y}, outputs={'Out': tmp})
1213
    return tmp
1214 1215


Y
yangyaming 已提交
1216 1217 1218 1219
def lstm_unit(x_t,
              hidden_t_prev,
              cell_t_prev,
              forget_bias=0.0,
Y
yangyaming 已提交
1220
              param_attr=None,
1221
              bias_attr=None):
Y
yangyaming 已提交
1222 1223 1224 1225
    """Lstm unit layer. The equation of a lstm step is:

        .. math::

1226
            i_t & = \sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i)
Y
yangyaming 已提交
1227

1228
            f_t & = \sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + b_f)
Y
yangyaming 已提交
1229

1230
            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
1231

1232
            o_t & = \sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + b_o)
Y
yangyaming 已提交
1233 1234 1235

            h_t & = o_t tanh(c_t)

1236 1237 1238 1239 1240 1241
    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
1242 1243 1244

        .. math::

1245
            L_{i_t} = W_{x_i}x_{t} + W_{h_i}h_{t-1} + b_i
Y
yangyaming 已提交
1246 1247 1248 1249 1250 1251 1252 1253

    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
1254
    This layer has two outputs including :math:`h_t` and :math:`o_t`.
Y
yangyaming 已提交
1255 1256

    Args:
Y
yangyaming 已提交
1257 1258 1259 1260 1261 1262
        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
1263
        forget_bias (float): The forget bias of lstm unit.
Y
yangyaming 已提交
1264 1265
        param_attr (ParamAttr): The attributes of parameter weights, used to set
            initializer, name etc.
Y
yangyaming 已提交
1266 1267
        bias_attr (ParamAttr): The attributes of bias weights, if not False,
            bias weights will be created and be set to default value.
Y
yangyaming 已提交
1268 1269

    Returns:
Y
yangyaming 已提交
1270
        tuple: The hidden value and cell value of lstm unit.
Y
yangyaming 已提交
1271 1272 1273 1274

    Raises:
        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** \
1275 1276
                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
1277 1278 1279 1280 1281 1282

    Examples:

        .. code-block:: python

             x_t = fluid.layers.fc(input=x_t_data, size=10)
1283
             prev_hidden = fluid.layers.fc(input=prev_hidden_data, size=30)
Y
yangyaming 已提交
1284
             prev_cell = fluid.layers.fc(input=prev_cell_data, size=30)
Y
yangyaming 已提交
1285
             hidden_value, cell_value = fluid.layers.lstm_unit(x_t=x_t,
Y
yangyaming 已提交
1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301
                                                    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
1302
        raise ValueError("The 1st dimensions of x_t, hidden_t_prev and "
1303 1304 1305 1306
                         "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
1307 1308
                         "cell_t_prev must be the same.")

Y
yangyaming 已提交
1309 1310 1311
    if bias_attr is None:
        bias_attr = ParamAttr()

Y
yangyaming 已提交
1312
    size = cell_t_prev.shape[1]
1313
    concat_out = concat(input=[x_t, hidden_t_prev], axis=1)
Y
yangyaming 已提交
1314 1315
    fc_out = fc(input=concat_out,
                size=4 * size,
Y
yangyaming 已提交
1316
                param_attr=param_attr,
1317
                bias_attr=bias_attr)
Y
yangyaming 已提交
1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329
    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
1330
    return h, c
G
guosheng 已提交
1331 1332 1333 1334


def reduce_sum(input, dim=None, keep_dim=False):
    """
Y
yangyaming 已提交
1335
    Computes the sum of tensor elements over the given dimension.
G
guosheng 已提交
1336 1337 1338

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
1339 1340 1341 1342
        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
1343
            the dimension to reduce is :math:`rank + dim`.
Y
yangyaming 已提交
1344 1345
        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
1346 1347 1348 1349
            than the :attr:`input` unless :attr:`keep_dim` is true.

    Returns:
        Variable: The reduced Tensor variable.
F
fengjiayi 已提交
1350

G
guosheng 已提交
1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374
    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
1375 1376 1377 1378


def reduce_mean(input, dim=None, keep_dim=False):
    """
Y
yangyaming 已提交
1379
    Computes the mean of tensor elements over the given dimension.
G
guosheng 已提交
1380 1381 1382

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
1383 1384 1385 1386
        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
1387
            :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`.
Y
yangyaming 已提交
1388 1389
        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
1390 1391 1392 1393
            than the :attr:`input` unless :attr:`keep_dim` is true.

    Returns:
        Variable: The reduced Tensor variable.
F
fengjiayi 已提交
1394

G
guosheng 已提交
1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418
    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
1419 1420 1421 1422


def reduce_max(input, dim=None, keep_dim=False):
    """
Y
yangyaming 已提交
1423
    Computes the maximum of tensor elements over the given dimension.
1424 1425 1426

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
1427 1428 1429 1430
        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))`.
1431
            If :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`.
Y
yangyaming 已提交
1432 1433
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
1434 1435 1436 1437
            than the :attr:`input` unless :attr:`keep_dim` is true.

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
1438

1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466
    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


def reduce_min(input, dim=None, keep_dim=False):
    """
Y
yangyaming 已提交
1467
    Computes the minimum of tensor elements over the given dimension.
1468 1469 1470

    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
yangyaming 已提交
1471 1472 1473 1474
        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))`.
1475
            If :math:`dim < 0`, the dimension to reduce is :math:`rank + dim`.
Y
yangyaming 已提交
1476 1477
        keep_dim (bool): Whether to reserve the reduced dimension in the
            output Tensor. The result tensor will have one fewer dimension
1478 1479 1480 1481
            than the :attr:`input` unless :attr:`keep_dim` is true.

    Returns:
        Variable: The reduced Tensor variable.
Y
yangyaming 已提交
1482

1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506
    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
W
wanghaoshuang 已提交
1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 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 1552 1553 1554 1555 1556 1557


def warpctc(input, label, blank=0, norm_by_times=False, **kwargs):
    """
    An operator integrating the open source warp-ctc  library
    to compute Connectionist Temporal Classification (CTC) loss.
    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
    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.
       blank: (int, default: 0), the blank label of Connectionist
         Temporal Classification (CTC) loss, which is in the
         half-opened interval [0, num_classes + 1).
       norm_by_times: (bool, default: false), whether to
         normalize the gradients by the number of time-step,
         which is also the sequence's length.

    Returns:
        Variable: The Connectionist Temporal Classification (CTC) loss, which is a 2-D Tensor of the shape [batch_size, 1].

    Examples:
        .. code-block:: python

            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')
            cost = layers.warpctc(input=y_predict, label=y)

    """
    helper = LayerHelper('warpctc', **kwargs)
    loss_out = helper.create_tmp_variable(dtype=input.dtype)
    grad_out = helper.create_tmp_variable(dtype=input.dtype)
    helper.append_op(
        type='warpctc',
        inputs={'Logits': [input],
                'Label': [label]},
        outputs={'WarpCTCGrad': [grad_out],
                 'Loss': [loss_out]},
        attrs={'blank': blank,
               'norm_by_times': norm_by_times})
    return loss_out