layers.py 150.6 KB
Newer Older
Z
zhangjinchao01 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
# Copyright (c) 2016 Baidu, Inc. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import functools
16
import collections
Z
zhangjinchao01 已提交
17 18 19 20 21 22 23 24

from paddle.trainer.config_parser import *
from .activations import LinearActivation, SigmoidActivation, TanhActivation, \
    ReluActivation, IdentityActivation, SoftmaxActivation
from .evaluators import *
from .poolings import MaxPooling, AvgPooling, BasePoolingType
from .attrs import *
from .default_decorators import *
25

Z
zhangjinchao01 已提交
26 27 28 29 30 31
try:
    import cPickle as pickle
except ImportError:
    import pickle
import copy

Q
qijun 已提交
32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67
__all__ = [
    "full_matrix_projection",
    "AggregateLevel",
    "ExpandLevel",
    "identity_projection",
    "dotmul_projection",
    "dotmul_operator",
    "repeat_layer",
    "table_projection",
    "mixed_layer",
    "data_layer",
    "embedding_layer",
    "fc_layer",
    "grumemory",
    "pooling_layer",
    "lstmemory",
    "last_seq",
    "first_seq",
    "cos_sim",
    "hsigmoid",
    "conv_projection",
    "regression_cost",
    'classification_cost',
    "LayerOutput",
    'img_conv_layer',
    'img_pool_layer',
    'batch_norm_layer',
    'img_cmrnorm_layer',
    'addto_layer',
    'concat_layer',
    'lstm_step_layer',
    'recurrent_group',
    'memory',
    'StaticInput',
    'expand_layer',
    'scaling_layer',
X
xuwei06 已提交
68
    'scaling_projection',
Q
qijun 已提交
69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93
    'power_layer',
    'interpolation_layer',
    'bilinear_interp_layer',
    'trans_layer',
    'sum_to_one_norm_layer',
    'get_output_layer',
    'LayerType',
    'context_projection',
    'beam_search',
    'maxid_layer',
    'GeneratedInput',
    'SubsequenceInput',
    'gru_step_layer',
    'recurrent_layer',
    'BaseGeneratedInput',
    'conv_operator',
    'conv_shift_layer',
    'tensor_layer',
    'selective_fc_layer',
    'sampling_id_layer',
    'slope_intercept_layer',
    'trans_full_matrix_projection',
    'linear_comb_layer',
    'convex_comb_layer',
    'ctc_layer',
94
    'warp_ctc_layer',
Q
qijun 已提交
95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110
    'crf_layer',
    'crf_decoding_layer',
    'nce_layer',
    'cross_entropy_with_selfnorm',
    'cross_entropy',
    'multi_binary_label_cross_entropy',
    'sum_cost',
    'rank_cost',
    'lambda_cost',
    'huber_cost',
    'block_expand_layer',
    'maxout_layer',
    'out_prod_layer',
    'print_layer',
    'spp_layer',
]
Z
zhangjinchao01 已提交
111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127


class LayerType(object):
    """
    Layer type enumerations.
    """

    DATA = "data"
    MIXED_LAYER = "mixed"
    LSTMEMORY = "lstmemory"
    GRUMEMORY = "gated_recurrent"
    SEQUENCE_LAST_INSTANCE = "seqlastins"
    SEQUENCE_FIRST_INSTANCE = "seqfirstins"
    POOLING_MAX = "max"
    POOLING_AVG = 'average'
    FC_LAYER = "fc"
    COST = 'cost'
128 129
    COSINE_SIM_VEC = 'cos_vm'
    COSINE_SIM = 'cos'
Z
zhangjinchao01 已提交
130 131
    HSIGMOID = 'hsigmoid'
    CONV_LAYER = "conv"
132
    CONVTRANS_LAYER = "convt"
133 134 135
    EXCONV_LAYER = "exconv"
    EXCONVTRANS_LAYER = "exconvt"
    CUDNNCONV_LAYER = "cudnn_conv"
Z
zhangjinchao01 已提交
136 137 138 139 140 141 142 143 144 145 146 147 148 149 150
    POOL_LAYER = "pool"
    BATCH_NORM_LAYER = 'batch_norm'
    NORM_LAYER = 'norm'
    SUM_TO_ONE_NORM_LAYER = 'sum_to_one_norm'
    ADDTO_LAYER = 'addto'

    CONCAT_LAYER = 'concat'
    CONCAT_PROJ_LAYER = 'concat2'

    LSTM_STEP_LAYER = 'lstm_step'
    GRU_STEP_LAYER = 'gru_step'
    GET_OUTPUT_LAYER = 'get_output'

    EXPAND_LAYER = 'expand'
    INTERPOLATION_LAYER = 'interpolation'
L
liaogang 已提交
151
    BILINEAR_INTERP_LAYER = 'bilinear_interp'
Z
zhangjinchao01 已提交
152 153 154
    POWER_LAYER = 'power'
    SCALING_LAYER = 'scaling'
    TRANS_LAYER = 'trans'
H
Haonan 已提交
155
    OUT_PROD_LAYER = 'out_prod'
X
xuwei06 已提交
156
    FEATURE_MAP_EXPAND_LAYER = 'featmap_expand'
Z
zhangjinchao01 已提交
157 158 159 160 161 162 163 164 165 166 167

    MEMORY = 'memory'
    MAXID_LAYER = 'maxid'
    EOSID_LAYER = 'eos_id'
    RECURRENT_LAYER = 'recurrent'

    CONV_SHIFT_LAYER = "conv_shift"
    TENSOR_LAYER = "tensor"
    SEL_FC_LAYER = "selective_fc"
    SAMPLING_ID_LAYER = "sampling_id"
    SLOPE_INTERCEPT_LAYER = "slope_intercept"
168
    LINEAR_COMBINATION_LAYER = "convex_comb"
Z
zhangjinchao01 已提交
169
    BLOCK_EXPAND = "blockexpand"
170
    MAXOUT = "maxout"
Q
qijun 已提交
171
    SPP_LAYER = "spp"
Z
zhangjinchao01 已提交
172

173 174
    PRINT_LAYER = "print"

Z
zhangjinchao01 已提交
175
    CTC_LAYER = "ctc"
176
    WARP_CTC_LAYER = "warp_ctc"
Z
zhangjinchao01 已提交
177 178
    CRF_LAYER = "crf"
    CRF_DECODING_LAYER = "crf_decoding"
179
    NCE_LAYER = 'nce'
Z
zhangjinchao01 已提交
180 181 182 183 184 185 186 187

    RANK_COST = "rank-cost"
    LAMBDA_COST = "lambda_cost"
    HUBER = "huber"
    CROSS_ENTROPY = "multi-class-cross-entropy"
    CROSS_ENTROPY_WITH_SELFNORM = "multi_class_cross_entropy_with_selfnorm"
    SOFT_BIN_CLASS_CROSS_ENTROPY = "soft_binary_class_cross_entropy"
    MULTI_BIN_LABEL_CROSS_ENTROPY = "multi_binary_label_cross_entropy"
X
xuwei06 已提交
188
    SUM_COST = "sum_cost"
Z
zhangjinchao01 已提交
189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233

    @staticmethod
    def is_layer_type(type_name):
        """
        If type_name is a layer type.

        :param type_name: layer type name. Because layer type enumerations are
                          strings.
        :type type_name: basestring
        :return: True if is a layer_type
        :rtype: bool
        """
        for key in dir(LayerType):
            if key.isupper():
                att = getattr(LayerType, key)
                if isinstance(att, basestring) and type_name == att:
                    return True
        return False


class AggregateLevel(object):
    EACH_TIMESTEP = 'non-seq'
    EACH_SEQUENCE = 'seq'


class LayerOutput(object):
    """
    LayerOutput is output for layer function. It is used internally by several
    reasons.

    - Check layer connection make sense.

        - FC(Softmax) => Cost(MSE Error) is not good for example.

    - Tracking layer connection.

    - Pass to layer methods as input.

    :param name: Layer output name.
    :type name: basestring
    :param layer_type: Current Layer Type. One of LayerType enumeration.
    :type layer_type: basestring
    :param activation: Layer Activation.
    :type activation: BaseActivation.
    :param parents: Layer's parents.
234
    :type parents: list|tuple|collections.Sequence
Z
zhangjinchao01 已提交
235 236
    """

Q
qijun 已提交
237 238 239 240 241 242 243 244 245
    def __init__(self,
                 name,
                 layer_type,
                 parents=None,
                 activation=None,
                 num_filters=None,
                 img_norm_type=None,
                 size=None,
                 outputs=None,
246
                 reverse=None):
Z
zhangjinchao01 已提交
247 248
        assert isinstance(name, basestring)
        assert isinstance(layer_type, basestring)
X
xuwei06 已提交
249
        assert size is not None
Z
zhangjinchao01 已提交
250 251 252
        assert LayerType.is_layer_type(layer_type)
        self.name = name
        self.layer_type = layer_type
253 254
        if parents is not None and type(parents) != list:
            parents = [parents]
Z
zhangjinchao01 已提交
255 256 257 258 259 260 261 262
        self.parents = [] if parents is None else parents
        self.activation = activation
        self.num_filters = num_filters
        self.img_norm_type = img_norm_type
        self.size = size
        if outputs is None:
            outputs = ['default']
        self.outputs = outputs
263
        self.reverse = reverse
Z
zhangjinchao01 已提交
264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279

    def __repr__(self):
        """
        Disable __repr__ for debug reason. Will be implemented when release
        """
        assert False, "this method should not be invoked"

    def __str__(self):
        """
        Disable __str__ for debug reason. Will be implemented when release
        """
        assert False, "this method should not be invoked"


ERROR_CLIPPING = 'error_clipping_threshold'
DROPOUT = 'drop_rate'
280
DEVICE = 'device'
Z
zhangjinchao01 已提交
281 282 283


def layer_support(*attrs):
284
    attrs_list = list(attrs)
285
    attrs_list.append(DEVICE)
Q
qijun 已提交
286

Z
zhangjinchao01 已提交
287 288 289
    def decorator(method):
        @functools.wraps(method)
        def wrapper(*args, **kwargs):
290
            for attr in attrs_list:
Z
zhangjinchao01 已提交
291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345
                for each in args:
                    if isinstance(each, ExtraLayerAttribute):
                        setattr(each, '_'.join(['can', attr]), True)
                for key in kwargs:
                    val = kwargs[key]
                    if isinstance(val, ExtraLayerAttribute):
                        setattr(val, '_'.join(['can', attr]), True)
            for each in args:
                if isinstance(each, ExtraLayerAttribute):
                    each.check(method.__name__)
            for key in kwargs:
                val = kwargs[key]
                if isinstance(val, ExtraLayerAttribute):
                    val.check(method.__name__)
            return method(*args, **kwargs)

        return wrapper

    return decorator


@wrap_param_attr_default()
def full_matrix_projection(input, size=0, param_attr=None):
    """
    Full Matrix Projection. It performs full matrix multiplication.

    ..  math::
        out.row[i] += in.row[i] * weight

    There are two styles of usage.

    1. When used in mixed_layer like this, you can only set the input:

    .. code-block:: python

       with mixed_layer(size=100) as m:
           m += full_matrix_projection(input=layer)

    2. When used as an independant object like this, you must set the size:

    .. code-block:: python

       proj = full_matrix_projection(input=layer,
                                     size=100,
                                     param_attr=ParamAttr(name='_proj'))

    :param input: input layer
    :type input: LayerOutput
    :param size: The parameter size. Means the width of parameter.
    :type size: int
    :param param_attr: Parameter config, None if use default.
    :type param_attr: ParameterAttribute
    :return: A FullMatrixProjection Object.
    :rtype: FullMatrixProjection
    """
Q
qijun 已提交
346 347
    proj = FullMatrixProjection(
        input_layer_name=input.name, size=size, **param_attr.attr)
Z
zhangjinchao01 已提交
348 349 350 351
    proj.origin = input
    return proj


352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381
@wrap_param_attr_default()
def trans_full_matrix_projection(input, size=0, param_attr=None):
    """
    Different from full_matrix_projection, this projection performs matrix
    multiplication, using transpose of weight.

    ..  math::
        out.row[i] += in.row[i] * w^\mathrm{T}

    :math:`w^\mathrm{T}` means transpose of weight.
    The simply usage is:

    .. code-block:: python

       proj = trans_full_matrix_projection(input=layer,
                                           size=100,
                                           param_attr=ParamAttr(
                                                name='_proj',
                                                initial_mean=0.0,
                                                initial_std=0.01))

    :param input: input layer
    :type input: LayerOutput
    :param size: The parameter size. Means the width of parameter.
    :type size: int
    :param param_attr: Parameter config, None if use default.
    :type param_attr: ParameterAttribute
    :return: A TransposedFullMatrixProjection Object.
    :rtype: TransposedFullMatrixProjection
    """
Q
qijun 已提交
382 383
    proj = TransposedFullMatrixProjection(
        input_layer_name=input.name, size=size, **param_attr.attr)
384 385 386 387
    proj.origin = input
    return proj


Z
zhangjinchao01 已提交
388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426
@wrap_param_attr_default()
def table_projection(input, size=0, param_attr=None):
    """
    Table Projection. It selects rows from parameter where row\_id
    is in input\_ids.

    .. math::
       out.row[i] += table.row[ids[i]]

    where :math:`out` is output, :math:`table` is parameter, :math:`ids` is input\_ids,
    and :math:`i` is row\_id.

    There are two styles of usage.

    1. When used in mixed_layer like this, you can only set the input:

    .. code-block:: python

       with mixed_layer(size=100) as m:
           m += table_projection(input=layer)

    2. When used as an independant object like this, you must set the size:

    .. code-block:: python

       proj = table_projection(input=layer,
                               size=100,
                               param_attr=ParamAttr(name='_proj'))


    :param input: Input layer, which must contains id fields.
    :type input: LayerOutput
    :param size: The parameter size. Means the width of parameter.
    :type size: int
    :param param_attr: Parameter config, None if use default.
    :type param_attr: ParameterAttribute
    :return: A TableProjection Object.
    :rtype: TableProjection
    """
Q
qijun 已提交
427 428
    proj = TableProjection(
        input_layer_name=input.name, size=size, **param_attr.attr)
Z
zhangjinchao01 已提交
429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463
    proj.origin = input
    return proj


def identity_projection(input, offset=None):
    """
    1. IdentityProjection if offset=None. It performs:

    .. math::
       out.row[i] += in.row[i]

    The example usage is:

    .. code-block:: python

       proj = identity_projection(input=layer)


    2. IdentityOffsetProjection if offset!=None. It likes IdentityProjection,
    but layer size may be smaller than input size.
    It select dimesions [offset, offset+layer_size) from input:

    .. math::
       out.row[i] += in.row[i + \\textrm{offset}]

    The example usage is:

    .. code-block:: python

       proj = identity_projection(input=layer,
                                  offset=10)

    Note that both of two projections should not have any parameter.

    :param input: Input Layer.
464
    :type input: LayerOutput
Z
zhangjinchao01 已提交
465 466
    :param offset: Offset, None if use default.
    :type offset: int
X
xuwei06 已提交
467
    :return: A IdentityProjection or IdentityOffsetProjection object
Z
zhangjinchao01 已提交
468 469 470 471 472 473
    :rtype: IdentityProjection or IdentityOffsetProjection
    """
    if offset is None:
        proj = IdentityProjection(input_layer_name=input.name)
        proj.origin = input
    else:
Q
qijun 已提交
474 475
        proj = IdentityOffsetProjection(
            input_layer_name=input.name, offset=offset)
Z
zhangjinchao01 已提交
476 477 478 479
        proj.origin = input
    return proj


X
xuwei06 已提交
480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501
@wrap_param_attr_default()
def scaling_projection(input, param_attr=None):
    """
    scaling_projection multiplies the input with a scalar parameter and add to
    the output.

    .. math::
       out += w * in

    The example usage is:

    .. code-block:: python

       proj = scaling_projection(input=layer)

    :param input: Input Layer.
    :type input: LayerOutput
    :param param_attr: Parameter config, None if use default.
    :type param_attr: ParameterAttribute
    :return: A ScalingProjection object
    :rtype: ScalingProjection
    """
L
Luo Tao 已提交
502
    proj = ScalingProjection(input_layer_name=input.name, **param_attr.attr)
X
xuwei06 已提交
503 504 505 506
    proj.origin = input
    return proj


Z
zhangjinchao01 已提交
507
@wrap_param_attr_default()
508
def dotmul_projection(input, param_attr=None):
Z
zhangjinchao01 已提交
509
    """
510
    DotMulProjection with a layer as input.
Z
zhangjinchao01 已提交
511 512 513 514 515 516 517 518 519 520 521 522 523
    It performs element-wise multiplication with weight.

    ..  math::
        out.row[i] += in.row[i] .* weight

    where :math:`.*` means element-wise multiplication.

    The example usage is:

    .. code-block:: python

       proj = dotmul_projection(input=layer)

524 525 526 527 528 529 530
    :param input: Input layer.
    :type input: LayerOutput
    :param param_attr: Parameter config, None if use default.
    :type param_attr: ParameterAttribute
    :return: A DotMulProjection Object.
    :rtype: DotMulProjection
    """
Q
qijun 已提交
531 532
    proj = DotMulProjection(
        input_layer_name=input.name, size=input.size, **param_attr.attr)
533
    proj.origin = input
534
    return proj
Z
zhangjinchao01 已提交
535

536 537

def dotmul_operator(a=None, b=None, scale=1, **kwargs):
538 539
    """
    DotMulOperator takes two inputs and performs element-wise multiplication:
540

Z
zhangjinchao01 已提交
541
    .. math::
542 543
       out.row[i] += scale * (x.row[i] .* y.row[i])

Z
zhangjinchao01 已提交
544 545
    where :math:`.*` means element-wise multiplication, and
    scale is a config scalar, its default value is one.
546

Z
zhangjinchao01 已提交
547
    The example usage is:
548

Z
zhangjinchao01 已提交
549
    .. code-block:: python
550 551 552

       op = dotmul_operator(x=layer1, y=layer2, scale=0.5)

553 554 555 556
    :param a: Input layer1
    :type a: LayerOutput
    :param b: Input layer2
    :type b: LayerOutput
Z
zhangjinchao01 已提交
557 558
    :param scale: config scalar, default value is one.
    :type scale: float
559 560
    :return: A DotMulOperator Object.
    :rtype: DotMulOperator
Z
zhangjinchao01 已提交
561
    """
562 563 564
    if 'x' in kwargs or 'y' in kwargs:
        logger.warning('x and y arguments for dotmul_operator is deprecated. '
                       'Please use a and b as parameter.')
Q
qijun 已提交
565
    a = kwargs.get('x', a)  # For Backward capacity.
566 567 568 569 570 571
    b = kwargs.get('y', b)
    assert isinstance(a, LayerOutput)
    assert isinstance(b, LayerOutput)
    if a.size is not None and b.size is not None:
        assert a.size == b.size

Q
qijun 已提交
572
    op = DotMulOperator(input_layer_names=[a.name, b.name], scale=scale)
573
    op.origin = [a, b]
574
    return op
Z
zhangjinchao01 已提交
575

576

Z
zhangjinchao01 已提交
577
@wrap_bias_attr_default(['padding_attr'])
Q
qijun 已提交
578 579 580
def context_projection(input,
                       context_len,
                       context_start=None,
Z
zhangjinchao01 已提交
581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616
                       padding_attr=False):
    """
    Context Projection.

    It just simply reorganizes input sequence, combines "context_len" sequence
    to one context from context_start. "context_start" will be set to
    -(context_len - 1) / 2 by default. If context position out of sequence
    length, padding will be filled as zero if padding_attr = False, otherwise
    it is trainable.

    For example, origin sequence is [A B C D E F G], context len is 3, then
    after context projection and not set padding_attr, sequence will
    be [ 0AB ABC BCD CDE DEF EFG FG0 ].

    :param input: Input Sequence.
    :type input: LayerOutput
    :param context_len: context length.
    :type context_len: int
    :param context_start: context start position. Default is
                          -(context_len - 1)/2
    :type context_start: int
    :param padding_attr: Padding Parameter Attribute. If false, it means padding
                         always be zero. Otherwise Padding is learnable, and
                         parameter attribute is set by this parameter.
    :type padding_attr: bool|ParameterAttribute
    :return: Projection
    :rtype: Projection
    """
    context_start = -(
        context_len - 1) / 2 if context_start is None else context_start

    extra_dict = dict()
    trainable = isinstance(padding_attr, ParameterAttribute)
    if trainable:
        extra_dict = padding_attr.attr

Q
qijun 已提交
617 618 619 620 621 622
    proj = ContextProjection(
        input_layer_name=input.name,
        context_length=context_len,
        context_start=context_start,
        trainable_padding=trainable,
        **extra_dict)
Z
zhangjinchao01 已提交
623 624 625 626 627 628 629 630 631 632 633 634 635
    proj.origin = input
    return proj


class MixedLayerType(LayerOutput):
    """
    The internal object for trainer_helpers.
    """

    class AddToSealedMixedLayerException(Exception):
        def __init__(self):
            Exception.__init__(self)

Q
qijun 已提交
636
    def __init__(self, name, size, act, bias_attr, layer_attr, parents=None):
Z
zhangjinchao01 已提交
637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652
        """
        Ctor.
        :param name: layer name.
        :type name: basestring
        :param size: layer size.
        :type size: int
        :param act: activation type.
        :type act: BaseActivation
        :param bias_attr: The Bias Attribute. If no bias, then pass False or
                          something not type of ParameterAttribute. None will
                          get a default Bias.
        :type bias_attr: ParameterAttribute or None means has bias. Any other
                         type means no bias.
        :param layer_attr: Extra Layer Attribute.
        :type layer_attr: ExtraLayerAttribute or None
        """
Q
qijun 已提交
653 654 655 656 657 658 659
        LayerOutput.__init__(
            self,
            name,
            LayerType.MIXED_LAYER,
            parents,
            size=size,
            activation=act)
Z
zhangjinchao01 已提交
660 661 662 663 664
        self.bias_attr = bias_attr
        self.layer_attr = layer_attr
        self.inputs = []
        self.finalized = False

665
    def __iadd__(self, other):
Z
zhangjinchao01 已提交
666 667 668 669 670 671 672 673
        """
        + += operator
        :param other: Other projection.
        :type other: Projection
        :return: self.
        :rtype: MixedLayerType
        """
        if not self.finalized:
674
            assert isinstance(other, Projection) or isinstance(other, Operator)
Z
zhangjinchao01 已提交
675
            self.inputs.append(other)
676 677 678 679
            if isinstance(other, Projection):
                self.parents.append(other.origin)
            else:
                self.parents.extend(other.origin)
Z
zhangjinchao01 已提交
680 681 682 683 684 685 686 687 688 689 690
            return self
        else:
            raise MixedLayerType.AddToSealedMixedLayerException()

    def __enter__(self):
        assert len(self.inputs) == 0
        return self

    def __exit__(self, *args, **kwargs):
        del args, kwargs  # unused parameter to suppress warning
        assert len(self.inputs) != 0
691
        ml = MixedLayer(
Z
zhangjinchao01 已提交
692 693 694 695 696
            name=self.name,
            size=self.size,
            active_type=self.activation.name,
            bias=ParamAttr.to_bias(self.bias_attr),
            inputs=self.inputs,
Q
qijun 已提交
697
            **ExtraLayerAttribute.to_kwargs(self.layer_attr))
698 699 700
        # update the size which might be computed inside MixedLayer
        # according to the operator's output size
        self.size = ml.config.size
Z
zhangjinchao01 已提交
701 702 703 704 705 706


@wrap_name_default("mixed")
@wrap_act_default(act=LinearActivation())
@wrap_bias_attr_default(has_bias=False)
@layer_support(ERROR_CLIPPING, DROPOUT)
Q
qijun 已提交
707 708 709 710 711
def mixed_layer(size=0,
                input=None,
                name=None,
                act=None,
                bias_attr=False,
Z
zhangjinchao01 已提交
712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755
                layer_attr=None):
    """
    Mixed Layer. A mixed layer will add all inputs together, then activate.
    Each inputs is a projection or operator.

    There are two styles of usages.

    1. When not set inputs parameter, use mixed_layer like this:

    .. code-block:: python

       with mixed_layer(size=256) as m:
           m += full_matrix_projection(input=layer1)
           m += identity_projection(input=layer2)

    2. You can also set all inputs when invoke mixed_layer as follows:

    .. code-block:: python

       m = mixed_layer(size=256,
                       input=[full_matrix_projection(input=layer1),
                              full_matrix_projection(input=layer2)])

    :param name: mixed layer name. Can be referenced by other layer.
    :type name: basestring
    :param size: layer size.
    :type size: int
    :param input: inputs layer. It is an optional parameter. If set,
                  then this function will just return layer's name.
    :param act: Activation Type.
    :type act: BaseActivation
    :param bias_attr: The Bias Attribute. If no bias, then pass False or
                      something not type of ParameterAttribute. None will get a
                      default Bias.
    :type bias_attr: ParameterAttribute or None or bool
    :param layer_attr: The extra layer config. Default is None.
    :type layer_attr: ExtraLayerAttribute
    :return: MixedLayerType object can add inputs or layer name.
    :rtype: MixedLayerType
    """

    if input is None:
        return MixedLayerType(name, size, act, bias_attr, layer_attr)
    else:
Q
qijun 已提交
756 757 758 759 760 761
        with mixed_layer(
                name=name,
                size=size,
                act=act,
                bias_attr=bias_attr,
                layer_attr=layer_attr) as m:
762
            if isinstance(input, collections.Sequence):
Z
zhangjinchao01 已提交
763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787
                for each in input:
                    m += each
            else:
                m += input
        return m


@layer_support()
def data_layer(name, size, layer_attr=None):
    """
    Define DataLayer For NeuralNetwork.

    The example usage is:

    ..  code-block:: python

        data = data_layer(name="input",
                          size=1000)

    :param name: Name of this data layer.
    :type name: basestring
    :param size: Size of this data layer.
    :type size: int
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
788
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
789 790
    :rtype: LayerOutput
    """
Q
qijun 已提交
791 792 793 794 795
    Layer(
        type=LayerType.DATA,
        name=name,
        size=size,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817

    return LayerOutput(name, LayerType.DATA, size=size)


@wrap_name_default("embedding")
@wrap_param_attr_default()
@layer_support(ERROR_CLIPPING)
def embedding_layer(input, size, name=None, param_attr=None, layer_attr=None):
    """
    Define a embedding Layer.

    :param name: Name of this embedding layer.
    :type name: basestring
    :param input: The input layer for this embedding. NOTE: must be Index Data.
    :type input: LayerOutput
    :param size: The embedding dimension.
    :type size: int
    :param param_attr: The embedding parameter attribute. See ParameterAttribute
                      for details.
    :type param_attr: ParameterAttribute|None
    :param layer_attr: Extra layer Config. Default is None.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
818
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
819 820
    :rtype: LayerOutput
    """
Q
qijun 已提交
821 822 823 824 825 826
    with mixed_layer(
            name=name,
            size=size,
            act=LinearActivation(),
            bias_attr=False,
            layer_attr=layer_attr) as mix:
Z
zhangjinchao01 已提交
827 828 829 830 831 832 833 834 835
        mix += table_projection(input=input, size=size, param_attr=param_attr)
    return mix


@wrap_name_default()
@wrap_param_attr_default()
@wrap_bias_attr_default()
@wrap_act_default()
@layer_support(ERROR_CLIPPING, DROPOUT)
Q
qijun 已提交
836 837 838 839 840 841 842
def fc_layer(input,
             size,
             act=None,
             name=None,
             param_attr=None,
             bias_attr=None,
             layer_attr=None):
Z
zhangjinchao01 已提交
843 844 845 846 847 848 849 850 851 852 853 854
    """
    Helper for declare fully connected layer.

    The example usage is:

    .. code-block:: python

       fc = fc_layer(input=layer,
                     size=1024,
                     act=LinearActivation(),
                     bias_attr=False)

L
luotao02 已提交
855
    which is equal to:
Z
zhangjinchao01 已提交
856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877

    .. code-block:: python

       with mixed_layer(size=1024) as fc:
           fc += full_matrix_projection(input=layer)

    :param name: The Layer Name.
    :type name: basestring
    :param input: The input layer. Could be a list/tuple of input layer.
    :type input: LayerOutput|list|tuple
    :param size: The layer dimension.
    :type size: int
    :param act: Activation Type. Default is tanh.
    :type act: BaseActivation
    :param param_attr: The Parameter Attribute|list.
    :type param_attr: ParameterAttribute
    :param bias_attr: The Bias Attribute. If no bias, then pass False or
                      something not type of ParameterAttribute. None will get a
                      default Bias.
    :type bias_attr: ParameterAttribute|None|Any
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
878
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
879 880 881 882
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
883
        assert not isinstance(param_attr, collections.Sequence)
Z
zhangjinchao01 已提交
884 885
        param_attr = [param_attr]
    else:
886
        if isinstance(param_attr, collections.Sequence):
Z
zhangjinchao01 已提交
887 888 889 890
            assert len(input) == len(param_attr)
        else:
            param_attr = [copy.deepcopy(param_attr) for _ in range(len(input))]

891
    assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
892 893

    Layer(
Q
qijun 已提交
894 895 896
        inputs=[
            Input(ipt.name, **attr.attr) for ipt, attr in zip(input, param_attr)
        ],
Z
zhangjinchao01 已提交
897 898 899 900 901
        name=name,
        type=LayerType.FC_LAYER,
        size=size,
        bias=ParamAttr.to_bias(bias_attr),
        active_type=act.name,
Q
qijun 已提交
902 903 904
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.FC_LAYER, input, activation=act, size=size)
Z
zhangjinchao01 已提交
905

906

907 908 909 910
@wrap_name_default("print")
def print_layer(input, name=None):
    """
    Print the output value of input layers. This layer is useful for debugging.
911 912 913 914 915

    :param name: The Layer Name.
    :type name: basestring
    :param input: The input layer. Could be a list/tuple of input layer.
    :type input: LayerOutput|list|tuple
916
    :return: LayerOutput
917
    """
918 919 920 921 922
    if isinstance(input, LayerOutput):
        input = [input]
    assert isinstance(input, collections.Sequence)  # list or tuple
    for each in input:
        assert isinstance(each, LayerOutput)
923 924 925 926

    Layer(
        name=name,
        type=LayerType.PRINT_LAYER,
Q
qijun 已提交
927
        inputs=[l.name for l in input], )
928
    # this layer don't return anything, can not be input of other layer.
929

Z
zhangjinchao01 已提交
930 931 932 933 934

@wrap_name_default("seq_pooling")
@wrap_bias_attr_default(has_bias=False)
@wrap_param_default(['pooling_type'], default_factory=lambda _: MaxPooling())
@layer_support()
Q
qijun 已提交
935 936 937 938
def pooling_layer(input,
                  pooling_type=None,
                  name=None,
                  bias_attr=None,
Z
zhangjinchao01 已提交
939 940 941 942 943 944 945 946 947 948 949 950 951
                  agg_level=AggregateLevel.EACH_TIMESTEP,
                  layer_attr=None):
    """
    Pooling layer for sequence inputs, not used for Image.

    The example usage is:

    .. code-block:: python

       seq_pool = pooling_layer(input=layer,
                                pooling_type=AvgPooling(),
                                agg_level=AggregateLevel.EACH_SEQUENCE)

C
caoying03 已提交
952 953
    :param agg_level: AggregateLevel.EACH_TIMESTEP or
                      AggregateLevel.EACH_SEQUENCE
Z
zhangjinchao01 已提交
954 955 956 957 958 959 960 961 962 963 964 965
    :type agg_level: AggregateLevel
    :param name: layer name.
    :type name: basestring
    :param input: input layer name.
    :type input: LayerOutput
    :param pooling_type: Type of pooling, MaxPooling(default), AvgPooling,
                         SumPooling, SquareRootNPooling.
    :type pooling_type: BasePoolingType|None
    :param bias_attr: Bias parameter attribute. False if no bias.
    :type bias_attr: ParameterAttribute|None|False
    :param layer_attr: The Extra Attributes for layer, such as dropout.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
966
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
967 968 969
    :rtype: LayerType
    """
    extra_dict = dict()
970
    # noinspection PyUnresolvedReferences
Z
zhangjinchao01 已提交
971 972
    if isinstance(pooling_type, AvgPooling):
        extra_dict['average_strategy'] = pooling_type.strategy
973 974 975 976
    elif isinstance(pooling_type, MaxPooling) and \
                    pooling_type.output_max_index is not None:
        assert isinstance(pooling_type.output_max_index, bool)
        extra_dict['output_max_index'] = pooling_type.output_max_index
Z
zhangjinchao01 已提交
977 978 979 980 981 982 983 984
    extra_dict.update(ExtraLayerAttribute.to_kwargs(layer_attr))

    Layer(
        name=name,
        type=pooling_type.name,
        inputs=[Input(input.name)],
        bias=ParamAttr.to_bias(bias_attr),
        trans_type=agg_level,
Q
qijun 已提交
985
        **extra_dict)
Z
zhangjinchao01 已提交
986

Q
qijun 已提交
987 988
    return LayerOutput(
        name, pooling_type.name, parents=[input], size=input.size)
Z
zhangjinchao01 已提交
989

Q
qijun 已提交
990

Z
zhangjinchao01 已提交
991 992
@wrap_bias_attr_default()
@wrap_param_attr_default()
Q
qijun 已提交
993
@wrap_act_default(param_names=['gate_act'], act=SigmoidActivation())
Z
zhangjinchao01 已提交
994 995 996
@wrap_act_default(param_names=["act", 'state_act'], act=TanhActivation())
@wrap_name_default("lstmemory")
@layer_support(DROPOUT)
Q
qijun 已提交
997 998 999 1000 1001 1002 1003 1004 1005
def lstmemory(input,
              name=None,
              reverse=False,
              act=None,
              gate_act=None,
              size=None,
              state_act=None,
              bias_attr=None,
              param_attr=None,
Z
zhangjinchao01 已提交
1006 1007 1008 1009 1010 1011 1012 1013
              layer_attr=None):
    """
    Long Short-term Memory Cell.

    The memory cell was implemented as follow equations.

    ..  math::

L
luotao02 已提交
1014
        i_t & = \\sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i)
Z
zhangjinchao01 已提交
1015

L
luotao02 已提交
1016
        f_t & = \\sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f)
Z
zhangjinchao01 已提交
1017

L
luotao02 已提交
1018
        c_t & = f_tc_{t-1} + i_t tanh (W_{xc}x_t+W_{hc}h_{t-1} + b_c)
Z
zhangjinchao01 已提交
1019

L
luotao02 已提交
1020
        o_t & = \\sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o)
Z
zhangjinchao01 已提交
1021

L
luotao02 已提交
1022
        h_t & = o_t tanh(c_t)
Z
zhangjinchao01 已提交
1023 1024


C
caoying03 已提交
1025
    NOTE: In PaddlePaddle's implementation, the multiplications
Z
zhangjinchao01 已提交
1026
    :math:`W_{xi}x_{t}` , :math:`W_{xf}x_{t}`,
C
caoying03 已提交
1027 1028 1029 1030
    :math:`W_{xc}x_t`, :math:`W_{xo}x_{t}` are not done in the lstmemory layer,
    so an additional mixed_layer with full_matrix_projection or a fc_layer must
    be included in the configuration file to complete the input-to-hidden
    mappings before lstmemory is called.
Z
zhangjinchao01 已提交
1031

C
caoying03 已提交
1032
    NOTE: This is a low level user interface. You can use network.simple_lstm
Z
zhangjinchao01 已提交
1033 1034
    to config a simple plain lstm layer.

C
caoying03 已提交
1035 1036 1037 1038
    Please refer to **Generating Sequences With Recurrent Neural Networks** for
    more details about LSTM.

    Link_ goes as below.
Z
zhangjinchao01 已提交
1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061

    .. _Link: http://arxiv.org/abs/1308.0850

    :param name: The lstmemory layer name.
    :type name: basestring
    :param input: input layer name.
    :type input: LayerOutput
    :param reverse: is sequence process reversed or not.
    :type reverse: bool
    :param act: activation type, TanhActivation by default. :math:`h_t`
    :type act: BaseActivation
    :param gate_act: gate activation type, SigmoidActivation by default.
    :type gate_act: BaseActivation
    :param state_act: state activation type, TanhActivation by default.
    :type state_act: BaseActivation

    :param bias_attr: Bias attribute. None means default bias. False means no
                      bias.
    :type bias_attr: ParameterAttribute|None|False
    :param param_attr: Parameter Attribute.
    :type param_attr: ParameterAttribute|None|False
    :param layer_attr: Extra Layer attribute
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
1062
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1063 1064 1065 1066 1067 1068
    :rtype: LayerOutput
    """

    assert gate_act.support_hppl
    assert state_act.support_hppl
    assert act.support_hppl
1069 1070 1071 1072 1073 1074 1075 1076 1077 1078
    assert input.size is not None and input.size % 4 == 0
    if size is not None:
        if input.size / 4 == size:
            plog = logger.warning
        else:
            plog = logger.fatal

        plog("NOTE: The lstmemory layer[%s]'s size is set by previous input "
             "layer. The lstm size should be equal with input layer size/4. The"
             " size which is set explicitly will be ignored." % name)
Z
zhangjinchao01 已提交
1079

Q
qijun 已提交
1080 1081 1082 1083 1084 1085 1086 1087 1088 1089
    Layer(
        name=name,
        type=LayerType.LSTMEMORY,
        active_type=act.name,
        active_state_type=state_act.name,
        active_gate_type=gate_act.name,
        reversed=reverse,
        bias=ParamAttr.to_bias(bias_attr),
        inputs=[Input(input.name, **param_attr.attr)],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
1090

Q
qijun 已提交
1091 1092 1093 1094 1095
    return LayerOutput(
        name,
        LayerType.LSTMEMORY, [input],
        size=input.size / 4,
        reverse=reverse)
1096

Z
zhangjinchao01 已提交
1097 1098 1099

@wrap_bias_attr_default()
@wrap_param_attr_default()
Q
qijun 已提交
1100
@wrap_act_default(param_names=['gate_act'], act=SigmoidActivation())
Z
zhangjinchao01 已提交
1101 1102 1103
@wrap_act_default(param_names=["act"], act=TanhActivation())
@wrap_name_default("gru")
@layer_support(DROPOUT)
Q
qijun 已提交
1104 1105 1106 1107 1108 1109 1110 1111
def grumemory(input,
              name=None,
              reverse=False,
              act=None,
              gate_act=None,
              size=None,
              bias_attr=None,
              param_attr=None,
Z
zhangjinchao01 已提交
1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132
              layer_attr=None):
    """
    Gate Recurrent Unit Layer.

    The memory cell was implemented as follow equations.

    1. update gate :math:`z`: defines how much of the previous memory to
    keep around or the unit updates its activations. The update gate
    is computed by:

    ..  math::

        z_t = \\sigma(W_{z}x_{t} + U_{z}h_{t-1} + b_z)

    2. reset gate :math:`r`: determines how to combine the new input with the
    previous memory. The reset gate is computed similarly to the update gate:

    ..  math::

        r_t = \\sigma(W_{r}x_{t} + U_{r}h_{t-1} + b_r)

C
caoying03 已提交
1133 1134
    3. The candidate activation :math:`\\tilde{h_t}` is computed similarly to
    that of the traditional recurrent unit:
Z
zhangjinchao01 已提交
1135 1136 1137 1138 1139

    ..  math::

        {\\tilde{h_t}} = tanh(W x_{t} + U (r_{t} \odot h_{t-1}) + b)

C
caoying03 已提交
1140 1141 1142
    4. The hidden activation :math:`h_t` of the GRU at time t is a linear
    interpolation between the previous activation :math:`h_{t-1}` and the
    candidate activation :math:`\\tilde{h_t}`:
Z
zhangjinchao01 已提交
1143 1144 1145 1146 1147

    ..  math::

        h_t = (1 - z_t) h_{t-1} + z_t {\\tilde{h_t}}

C
caoying03 已提交
1148
    NOTE: In PaddlePaddle's implementation, the multiplication operations
Z
zhangjinchao01 已提交
1149
    :math:`W_{r}x_{t}`, :math:`W_{z}x_{t}` and :math:`W x_t` are not computed in
C
caoying03 已提交
1150 1151 1152
    gate_recurrent layer. Consequently, an additional mixed_layer with
    full_matrix_projection or a fc_layer must be included before grumemory
    is called.
Z
zhangjinchao01 已提交
1153

C
caoying03 已提交
1154 1155 1156
    More details can be found by referring to `Empirical Evaluation of Gated
    Recurrent Neural Networks on Sequence Modeling.
    <https://arxiv.org/abs/1412.3555>`_
Z
zhangjinchao01 已提交
1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167

    The simple usage is:

    .. code-block:: python

       gru = grumemory(input)

    :param name: The gru layer name.
    :type name: None|basestring
    :param input: input layer.
    :type input: LayerOutput.
1168
    :param reverse: Whether sequence process is reversed or not.
Z
zhangjinchao01 已提交
1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183
    :type reverse: bool
    :param act: activation type, TanhActivation by default. This activation
                affects the :math:`{\\tilde{h_t}}`.
    :type act: BaseActivation
    :param gate_act: gate activation type, SigmoidActivation by default.
                     This activation affects the :math:`z_t` and :math:`r_t`. It is the
                     :math:`\\sigma` in the above formula.
    :type gate_act: BaseActivation
    :param bias_attr: Bias attribute. None means default bias. False means no
                      bias.
    :type bias_attr: ParameterAttribute|None|False
    :param param_attr: Parameter Attribute.
    :type param_attr: ParameterAttribute|None|False
    :param layer_attr: Extra Layer attribute
    :type layer_attr: ExtraLayerAttribute|None
1184 1185 1186
    :param size: Stub parameter of size, but actually not used. If set this size
                 will get a warning.
    :type size: None
D
dangqingqing 已提交
1187
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1188 1189 1190 1191
    :rtype: LayerOutput
    """
    assert act.support_hppl
    assert gate_act.support_hppl
1192 1193 1194 1195 1196 1197 1198 1199 1200
    assert input.size is not None and input.size % 3 == 0
    if size is not None:
        if input.size / 3 == size:
            plog = logger.warning
        else:
            plog = logger.fatal
        plog("NOTE: the gru memory layer's size is set by previous input layer,"
             " and should be input size / 3. Set size explicitly will be "
             "ignored.")
Z
zhangjinchao01 已提交
1201

Q
qijun 已提交
1202 1203 1204 1205 1206 1207 1208 1209 1210
    Layer(
        name=name,
        type=LayerType.GRUMEMORY,
        active_type=act.name,
        active_gate_type=gate_act.name,
        reversed=reverse,
        bias=ParamAttr.to_bias(bias_attr),
        inputs=[Input(input.name, **param_attr.attr)],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
1211

Q
qijun 已提交
1212 1213 1214 1215 1216
    return LayerOutput(
        name,
        LayerType.GRUMEMORY, [input],
        size=input.size / 3,
        reverse=reverse)
1217

Z
zhangjinchao01 已提交
1218 1219 1220

@wrap_name_default()
@layer_support()
Q
qijun 已提交
1221 1222 1223
def last_seq(input,
             name=None,
             agg_level=AggregateLevel.EACH_TIMESTEP,
Z
zhangjinchao01 已提交
1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234
             layer_attr=None):
    """
    Get Last Timestamp Activation of a sequence.

    :param agg_level: Aggregated level
    :param name: Layer name.
    :type name: basestring
    :param input: Input layer name.
    :type input: LayerOutput
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1235
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1236 1237
    :rtype: LayerOutput
    """
1238 1239 1240 1241 1242 1243
    if input.reverse is not None and input.reverse:
        logger.warning("You are getting the last instance of a sequence that"
                       " is a output of a REVERSED layer. There is no time"
                       " series information at all. Maybe you want to use"
                       " first_seq instead.")

Z
zhangjinchao01 已提交
1244 1245 1246 1247 1248
    Layer(
        name=name,
        type=LayerType.SEQUENCE_LAST_INSTANCE,
        inputs=[input.name],
        trans_type=agg_level,
Q
qijun 已提交
1249 1250 1251 1252 1253 1254
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.SEQUENCE_LAST_INSTANCE,
        parents=[input],
        size=input.size)
Z
zhangjinchao01 已提交
1255 1256 1257 1258


@wrap_name_default()
@layer_support()
Q
qijun 已提交
1259 1260 1261
def first_seq(input,
              name=None,
              agg_level=AggregateLevel.EACH_TIMESTEP,
Z
zhangjinchao01 已提交
1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272
              layer_attr=None):
    """
    Get First Timestamp Activation of a sequence.

    :param agg_level: aggregation level
    :param name: Layer name.
    :type name: basestring
    :param input: Input layer name.
    :type input: LayerOutput
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1273
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1274 1275
    :rtype: LayerOutput
    """
1276 1277 1278 1279 1280 1281 1282

    if input.reverse is not None and not input.reverse:
        logger.warning('You are getting the first instance for a time series,'
                       ' and it is a normal recurrent layer output. There is no'
                       ' time series information at all. Maybe you want to use'
                       ' last_seq instead.')

Z
zhangjinchao01 已提交
1283 1284 1285 1286 1287
    Layer(
        name=name,
        type=LayerType.SEQUENCE_FIRST_INSTANCE,
        inputs=[input.name],
        trans_type=agg_level,
Q
qijun 已提交
1288 1289 1290 1291 1292 1293
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.SEQUENCE_FIRST_INSTANCE,
        parents=[input],
        size=input.size)
Z
zhangjinchao01 已提交
1294 1295 1296 1297 1298 1299


class ExpandLevel(object):
    FROM_TIMESTEP = AggregateLevel.EACH_TIMESTEP
    FROM_SEQUENCE = AggregateLevel.EACH_SEQUENCE

1300

Z
zhangjinchao01 已提交
1301 1302
@wrap_name_default()
@layer_support()
Q
qijun 已提交
1303 1304
def expand_layer(input,
                 expand_as,
Z
zhangjinchao01 已提交
1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333
                 name=None,
                 bias_attr=False,
                 expand_level=ExpandLevel.FROM_TIMESTEP,
                 layer_attr=None):
    """
    A layer for "Expand Dense data or (sequence data where the length of each
    sequence is one) to sequence data."

    The example usage is:

    .. code-block:: python

       expand = expand_layer(input=layer1,
                             expand_as=layer2,
                             expand_level=ExpandLevel.FROM_TIMESTEP)

    :param input: Input layer
    :type input: LayerOutput
    :param expand_as: Expand as this layer's sequence info.
    :type expand_as: LayerOutput
    :param name: Layer name.
    :type name: basestring
    :param bias_attr: Bias attribute. None means default bias. False means no
                      bias.
    :type bias_attr: ParameterAttribute|None|False
    :param expand_level: whether input layer is timestep(default) or sequence.
    :type expand_level: ExpandLevel
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1334
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1335 1336 1337 1338 1339 1340 1341 1342 1343
    :rtype: LayerOutput
    """

    Layer(
        inputs=[input.name, expand_as.name],
        name=name,
        bias=ParamAttr.to_bias(bias_attr=bias_attr),
        type=LayerType.EXPAND_LAYER,
        trans_type=expand_level,
Q
qijun 已提交
1344 1345 1346 1347 1348 1349
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name=name,
        size=input.size,
        layer_type=LayerType.EXPAND_LAYER,
        parents=[input, expand_as])
Z
zhangjinchao01 已提交
1350 1351


X
xuwei06 已提交
1352 1353
@wrap_name_default()
@layer_support()
Q
qijun 已提交
1354
def repeat_layer(input, num_repeats, name=None, layer_attr=None):
X
xuwei06 已提交
1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384
    """
    A layer for repeating the input for num_repeats times. This is equivalent
    to apply concat_layer() with num_repeats same input.

    .. math::
       y  = [x, x, \cdots, x]

    The example usage is:

    .. code-block:: python

       expand = repeat_layer(layer, 4)

    :param input: Input layer
    :type input: LayerOutput
    :param num_repeats: Repeat the input so many times
    :type num_repeats: int
    :param name: Layer name.
    :type name: basestring
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
    :return: LayerOutput object.
    :rtype: LayerOutput
    """

    l = Layer(
        inputs=[input.name],
        name=name,
        num_filters=num_repeats,
        type=LayerType.FEATURE_MAP_EXPAND_LAYER,
Q
qijun 已提交
1385 1386 1387 1388 1389 1390 1391
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name=name,
        size=l.config.size,
        layer_type=LayerType.FEATURE_MAP_EXPAND_LAYER,
        parents=[input])

X
xuwei06 已提交
1392

Z
zhangjinchao01 已提交
1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420
@wrap_name_default()
@layer_support()
def interpolation_layer(input, weight, name=None, layer_attr=None):
    """
    This layer is for linear interpolation with two inputs,
    which is used in NEURAL TURING MACHINE.

    .. math::
       y.row[i] = w[i] * x_1.row[i] + (1 - w[i]) * x_2.row[i]

    where :math:`x_1` and :math:`x_2` are two (batchSize x dataDim) inputs,
    :math:`w` is (batchSize x 1) weight vector, and :math:`y` is
    (batchSize x dataDim) output.

    The example usage is:

    .. code-block:: python

       interpolation = interpolation_layer(input=[layer1, layer2], weight=layer3)

    :param input: Input layer.
    :type input: list|tuple
    :param weight: Weight layer.
    :type weight: LayerOutput
    :param name: Layer name.
    :type name: basestring
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1421
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1422 1423
    :rtype: LayerOutput
    """
1424
    assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
1425
    assert len(input) == 2
1426 1427 1428 1429 1430 1431 1432
    assert isinstance(input[0], LayerOutput) and isinstance(input[1],
                                                            LayerOutput)
    if input[0].size is not None and input[1].size is not None:
        assert input[0].size == input[1].size
    assert isinstance(weight, LayerOutput)
    if weight.size is not None:
        assert weight.size == 1
Z
zhangjinchao01 已提交
1433 1434 1435 1436
    Layer(
        name=name,
        type=LayerType.INTERPOLATION_LAYER,
        inputs=[weight.name, input[0].name, input[1].name],
Q
qijun 已提交
1437 1438 1439 1440 1441 1442
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.INTERPOLATION_LAYER,
        parents=[weight, input[0], input[1]],
        size=input[0].size)
Z
zhangjinchao01 已提交
1443 1444


L
liaogang 已提交
1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460
@wrap_name_default()
@layer_support()
def bilinear_interp_layer(input,
                          out_size_x=None,
                          out_size_y=None,
                          name=None,
                          layer_attr=None):
    """
    This layer is to implement bilinear interpolation on conv layer output.

    Please refer to Wikipedia: https://en.wikipedia.org/wiki/Bilinear_interpolation

    The simple usage is:

    .. code-block:: python

L
liaogang 已提交
1461
       bilinear = bilinear_interp_layer(input=layer1, out_size_x=64, out_size_y=64)
X
xuwei06 已提交
1462

L
liaogang 已提交
1463
    :param   input:        A input layer.
L
liaogang 已提交
1464
    :type    input:        LayerOutput.
L
liaogang 已提交
1465
    :param   out_size_x:   bilinear interpolation output width.
X
xuwei06 已提交
1466
    :type    out_size_x:   int|None
L
liaogang 已提交
1467
    :param   out_size_y:   bilinear interpolation output height.
L
liaogang 已提交
1468
    :type    out_size_y:   int|None
L
liaogang 已提交
1469
    :param   name:         The layer's name, which cna not be specified.
L
liaogang 已提交
1470
    :type    name:         None|basestring
L
liaogang 已提交
1471
    :param   layer_attr:   Extra Layer attribute.
L
liaogang 已提交
1472 1473 1474 1475 1476 1477 1478
    :type    layer_attr:   ExtraLayerAttribute
    :return: LayerOutput object.
    :rtype:  LayerOutput
    """
    assert input.layer_type == LayerType.CONV_LAYER
    assert isinstance(input.activation, LinearActivation)
    assert out_size_x > 0 and out_size_y > 0
L
liaogang 已提交
1479
    assert input.num_filters is not None
L
liaogang 已提交
1480
    num_channels = input.num_filters
Q
qijun 已提交
1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497
    l = Layer(
        name=name,
        inputs=Input(
            input.name,
            bilinear_interp=BilinearInterp(
                out_size_x=out_size_x,
                out_size_y=out_size_y,
                num_channels=num_channels)),
        type=LayerType.BILINEAR_INTERP_LAYER,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.BILINEAR_INTERP_LAYER,
        parents=[input],
        num_filters=num_channels,
        size=l.config.size)

L
liaogang 已提交
1498

Z
zhangjinchao01 已提交
1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525
@wrap_name_default()
@layer_support()
def power_layer(input, weight, name=None, layer_attr=None):
    """
    This layer applies a power function to a vector element-wise,
    which is used in NEURAL TURING MACHINE.

    .. math::
       y = x^w

    where :math:`x` is a input vector, :math:`w` is scalar weight,
    and :math:`y` is a output vector.

    The example usage is:

    .. code-block:: python

       power = power_layer(input=layer1, weight=layer2)

    :param input: Input layer.
    :type input: LayerOutput
    :param weight: Weight layer.
    :type weight: LayerOutput
    :param name: Layer name.
    :type name: basestring
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1526
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1527 1528
    :rtype: LayerOutput
    """
1529 1530 1531
    assert isinstance(input, LayerOutput) and isinstance(weight, LayerOutput)
    if weight.size is not None:
        assert weight.size == 1
Z
zhangjinchao01 已提交
1532 1533 1534
    Layer(
        name=name,
        type=LayerType.POWER_LAYER,
1535
        inputs=[weight.name, input.name],
Q
qijun 已提交
1536 1537 1538
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.POWER_LAYER, parents=[input, weight], size=input.size)
Z
zhangjinchao01 已提交
1539 1540 1541 1542 1543 1544


@wrap_name_default()
@layer_support()
def scaling_layer(input, weight, name=None, layer_attr=None):
    """
1545
    A layer for multiplying input vector by weight scalar.
Z
zhangjinchao01 已提交
1546 1547

    .. math::
1548
       y  = w x
Z
zhangjinchao01 已提交
1549

1550 1551 1552 1553 1554
    where :math:`x` is size=dataDim input, :math:`w` is size=1 weight,
    and :math:`y` is size=dataDim output.

    Note that the above computation is for one sample. Multiple samples are
    processed in one batch.
Z
zhangjinchao01 已提交
1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569

    The example usage is:

    .. code-block:: python

       scale = scaling_layer(input=layer1, weight=layer2)

    :param input: Input layer.
    :type input: LayerOutput
    :param weight: Weight layer.
    :type weight: LayerOutput
    :param name: Layer name.
    :type name: basestring
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1570
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1571 1572
    :rtype: LayerOutput
    """
1573 1574 1575
    assert isinstance(weight, LayerOutput) and isinstance(input, LayerOutput)
    if weight.size is not None:
        assert weight.size == 1
Z
zhangjinchao01 已提交
1576 1577 1578 1579
    Layer(
        name=name,
        type=LayerType.SCALING_LAYER,
        inputs=[weight.name, input.name],
Q
qijun 已提交
1580 1581 1582
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.SCALING_LAYER, parents=[weight, input], size=input.size)
Z
zhangjinchao01 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 1607


@wrap_name_default()
@layer_support()
def trans_layer(input, name=None, layer_attr=None):
    """
    A layer for transposition.

    .. math::
       y = x^\mathrm{T}

    where :math:`x` is (M x N) input, and :math:`y` is (N x M) output.

    The example usage is:

    .. code-block:: python

       trans = trans_layer(input=layer)

    :param input: Input layer.
    :type input: LayerOutput
    :param name: Layer name.
    :type name: basestring
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1608
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1609 1610 1611 1612 1613 1614
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.TRANS_LAYER,
        inputs=[input.name],
Q
qijun 已提交
1615 1616 1617
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.TRANS_LAYER, parents=[input], size=input.size)
Z
zhangjinchao01 已提交
1618 1619 1620 1621 1622 1623 1624 1625 1626


@wrap_name_default()
@layer_support()
def cos_sim(a, b, scale=5, size=1, name=None, layer_attr=None):
    """
    Cosine Similarity Layer. The cosine similarity equation is here.

    ..  math::
D
dangqingqing 已提交
1627
        similarity = cos(\\theta) = {\\mathbf{a} \\cdot \\mathbf{b}
1628 1629 1630 1631 1632
        \\over \\|\\mathbf{a}\\| \\|\\mathbf{b}\\|}

    The size of a is M, size of b is M*N,
    Similarity will be calculated N times by step M. The output size is
    N. The scale will be multiplied to similarity.
Z
zhangjinchao01 已提交
1633

1634 1635
    Note that the above computation is for one sample. Multiple samples are
    processed in one batch.
Z
zhangjinchao01 已提交
1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648

    :param name: layer name
    :type name: basestring
    :param a: input layer a
    :type a: LayerOutput
    :param b: input layer b
    :type b: LayerOutput
    :param scale: scale for cosine value. default is 5.
    :type scale: float
    :param size: layer size. NOTE size_a * size should equal size_b.
    :type size: int
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
1649
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1650 1651
    :rtype: LayerOutput
    """
1652
    assert isinstance(a, LayerOutput) and isinstance(b, LayerOutput)
1653 1654 1655 1656 1657 1658
    if size == 1:
        Layer(
            name=name,
            type=LayerType.COSINE_SIM,
            cos_scale=scale,
            inputs=[a.name, b.name],
Q
qijun 已提交
1659
            **ExtraLayerAttribute.to_kwargs(layer_attr))
1660
    else:
1661 1662
        if a.size is not None and b.size is not None:
            assert size == b.size / a.size
1663 1664 1665 1666 1667 1668
        Layer(
            name=name,
            type=LayerType.COSINE_SIM_VEC,
            size=size,
            cos_scale=scale,
            inputs=[a.name, b.name],
Q
qijun 已提交
1669
            **ExtraLayerAttribute.to_kwargs(layer_attr))
X
xuwei06 已提交
1670
    return LayerOutput(name, LayerType.COSINE_SIM, parents=[a, b], size=size)
Z
zhangjinchao01 已提交
1671

1672

Z
zhangjinchao01 已提交
1673 1674
@wrap_name_default()
@wrap_bias_attr_default(has_bias=True)
1675
@wrap_param_attr_default()
Z
zhangjinchao01 已提交
1676
@layer_support()
Q
qijun 已提交
1677 1678 1679 1680 1681 1682 1683
def hsigmoid(input,
             label,
             num_classes,
             name=None,
             bias_attr=None,
             param_attr=None,
             layer_attr=None):
Z
zhangjinchao01 已提交
1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704
    """
    Organize the classes into a binary tree. At each node, a sigmoid function
    is used to calculate the probability of belonging to the right branch.
    This idea is from "F. Morin, Y. Bengio (AISTATS 05):
    Hierarchical Probabilistic Neural Network Language Model."

    The example usage is:

    ..  code-block:: python

        cost = hsigmoid(input=[layer1, layer2],
                        label=data_layer,
                        num_classes=3)

    :param input: Input layers. It could be a LayerOutput or list/tuple of
                 LayerOutput.
    :type input: LayerOutput|list|tuple
    :param label: Label layer.
    :type label: LayerOutput
    :param num_classes: number of classes.
    :type num_classes: int
L
luotao02 已提交
1705 1706
    :param name: layer name
    :type name: basestring
Z
zhangjinchao01 已提交
1707 1708 1709 1710 1711
    :param bias_attr: Bias attribute. None means default bias.
                      False means no bias.
    :type bias_attr: ParameterAttribute|False
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
1712
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1713 1714 1715 1716
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
1717 1718 1719 1720 1721 1722 1723 1724 1725
        if not isinstance(param_attr, collections.Sequence):
            param_attr = [param_attr]
    else:
        if not isinstance(param_attr, collections.Sequence):
            param_attr = [param_attr] * len(input)
        else:
            assert len(param_attr) == len(input)

    assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
1726 1727 1728 1729 1730
    assert isinstance(label, LayerOutput)
    assert label.layer_type == LayerType.DATA

    ipts_for_layer = []
    parents = []
1731
    for each_input, each_param_attr in zip(input, param_attr):
Z
zhangjinchao01 已提交
1732
        assert isinstance(each_input, LayerOutput)
1733
        ipts_for_layer.append(Input(each_input.name, **each_param_attr.attr))
Z
zhangjinchao01 已提交
1734 1735 1736 1737
        parents.append(each_input)
    ipts_for_layer.append(label.name)
    parents.append(label)

X
xuwei06 已提交
1738
    l = Layer(
Z
zhangjinchao01 已提交
1739 1740 1741 1742 1743
        name=name,
        type=LayerType.HSIGMOID,
        num_classes=num_classes,
        bias=ParamAttr.to_bias(bias_attr),
        inputs=ipts_for_layer,
Q
qijun 已提交
1744 1745 1746
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.HSIGMOID, parents=parents, size=l.config.size)
Z
zhangjinchao01 已提交
1747

1748

Z
zhangjinchao01 已提交
1749 1750 1751 1752 1753
@wrap_name_default("conv")
@wrap_param_attr_default()
@wrap_bias_attr_default()
@wrap_act_default(act=ReluActivation())
@layer_support(DROPOUT)
Q
qijun 已提交
1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769
def img_conv_layer(input,
                   filter_size,
                   num_filters,
                   name=None,
                   num_channels=None,
                   act=None,
                   groups=1,
                   stride=1,
                   padding=0,
                   bias_attr=None,
                   param_attr=None,
                   shared_biases=True,
                   layer_attr=None,
                   filter_size_y=None,
                   stride_y=None,
                   padding_y=None,
1770 1771
                   trans=False,
                   layer_type=None):
Z
zhangjinchao01 已提交
1772 1773 1774 1775 1776 1777 1778
    """
    Convolution layer for image. Paddle only support square input currently and
    thus input image's width equals height.

    The details of convolution layer, please refer UFLDL's `convolution
    <http://ufldl.stanford.edu/tutorial/supervised/
    FeatureExtractionUsingConvolution/>`_ .
X
xuwei06 已提交
1779 1780

    Convolution Transpose (deconv) layer for image. Paddle only support square
1781 1782
    input currently and thus input image's width equals height.

X
xuwei06 已提交
1783
    The details of convolution transpose layer,
1784 1785 1786
    please refer to the following explanation and references therein
    <http://datascience.stackexchange.com/questions/6107/
    what-are-deconvolutional-layers/>`_ .
Z
zhangjinchao01 已提交
1787 1788 1789 1790
    The num_channel means input image's channel number. It may be 1 or 3 when
    input is raw pixels of image(mono or RGB), or it may be the previous layer's
    num_filters * num_group.

C
caoying03 已提交
1791 1792 1793
    There are several group of filter in PaddlePaddle implementation.
    Each group will process some channel of the inputs. For example, if an input
    num_channel = 256, group = 4, num_filter=32, the PaddlePaddle will create
Z
zhangjinchao01 已提交
1794
    32*4 = 128 filters to process inputs. The channels will be split into 4
C
caoying03 已提交
1795 1796
    pieces. First 256/4 = 64 channels will process by first 32 filters. The
    rest channels will be processed by rest group of filters.
Z
zhangjinchao01 已提交
1797 1798 1799 1800 1801

    :param name: Layer name.
    :type name: basestring
    :param input: Layer Input.
    :type input: LayerOutput
1802 1803 1804
    :param filter_size: The x dimension of a filter kernel. Or input a tuple for
                        two image dimension.
    :type filter_size: int|tuple|list
C
caoying03 已提交
1805 1806 1807
    :param filter_size_y: The y dimension of a filter kernel. Since PaddlePaddle
                        currently supports rectangular filters, the filter's
                        shape will be (filter_size, filter_size_y).
1808
    :type filter_size_y: int|None
Z
zhangjinchao01 已提交
1809 1810 1811 1812 1813
    :param num_filters: Each filter group's number of filter
    :param act: Activation type. Default is tanh
    :type act: BaseActivation
    :param groups: Group size of filters.
    :type groups: int
1814 1815 1816
    :param stride: The x dimension of the stride. Or input a tuple for two image
                   dimension.
    :type stride: int|tuple|list
Z
zhangjinchao01 已提交
1817 1818
    :param stride_y: The y dimension of the stride.
    :type stride_y: int
1819 1820 1821
    :param padding: The x dimension of the padding. Or input a tuple for two
                    image dimension
    :type padding: int|tuple|list
Z
zhangjinchao01 已提交
1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835
    :param padding_y: The y dimension of the padding.
    :type padding_y: int
    :param bias_attr: Convolution bias attribute. None means default bias.
                      False means no bias.
    :type bias_attr: ParameterAttribute|False
    :param num_channels: number of input channels. If None will be set
                        automatically from previous output.
    :type num_channels: int
    :param param_attr: Convolution param attribute. None means default attribute
    :type param_attr: ParameterAttribute
    :param shared_biases: Is biases will be shared between filters or not.
    :type shared_biases: bool
    :param layer_attr: Layer Extra Attribute.
    :type layer_attr: ExtraLayerAttribute
1836 1837
    :param trans: true if it is a convTransLayer, false if it is a convLayer
    :type trans: bool
1838 1839 1840 1841
    :param layer_type: specify the layer_type, default is None. If trans=True,
                       layer_type has to be "exconvt", otherwise layer_type 
                       has to be either "exconv" or "cudnn_conv"
    :type layer_type: String
D
dangqingqing 已提交
1842
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1843 1844 1845 1846 1847
    :rtype: LayerOutput
    """
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters
1848

Z
zhangjinchao01 已提交
1849
    if filter_size_y is None:
1850 1851 1852 1853 1854 1855
        if isinstance(filter_size, collections.Sequence):
            assert len(filter_size) == 2
            filter_size, filter_size_y = filter_size
        else:
            filter_size_y = filter_size

Z
zhangjinchao01 已提交
1856
    if stride_y is None:
1857 1858 1859 1860 1861 1862
        if isinstance(stride, collections.Sequence):
            assert len(stride) == 2
            stride, stride_y = stride
        else:
            stride_y = stride

Z
zhangjinchao01 已提交
1863
    if padding_y is None:
1864 1865 1866 1867 1868 1869 1870 1871
        if isinstance(padding, collections.Sequence):
            assert len(padding) == 2
            padding, padding_y = padding
        else:
            padding_y = padding

    if param_attr.attr.get('initial_smart'):
        # special initial for conv layers.
Q
qijun 已提交
1872
        init_w = (2.0 / (filter_size**2 * num_channels))**0.5
1873 1874 1875 1876
        param_attr.attr["initial_mean"] = 0.0
        param_attr.attr["initial_std"] = init_w
        param_attr.attr["initial_strategy"] = 0
        param_attr.attr["initial_smart"] = False
1877 1878 1879 1880 1881 1882 1883 1884 1885
    
    if layer_type:
        if trans:
            assert layer_type in ["exconvt"]
        else:
            assert layer_type in ["exconv", "cudnn_conv"]
        lt = layer_type
    else:
        lt = LayerType.CONVTRANS_LAYER if trans else LayerType.CONV_LAYER
Q
qijun 已提交
1886

X
xuwei06 已提交
1887
    l = Layer(
Z
zhangjinchao01 已提交
1888
        name=name,
Q
qijun 已提交
1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900
        inputs=Input(
            input.name,
            conv=Conv(
                filter_size=filter_size,
                padding=padding,
                stride=stride,
                channels=num_channels,
                groups=groups,
                filter_size_y=filter_size_y,
                padding_y=padding_y,
                stride_y=stride_y),
            **param_attr.attr),
Z
zhangjinchao01 已提交
1901 1902 1903 1904
        active_type=act.name,
        num_filters=num_filters,
        bias=ParamAttr.to_bias(bias_attr),
        shared_biases=shared_biases,
1905
        type=lt,
Q
qijun 已提交
1906 1907 1908 1909 1910 1911 1912 1913
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        lt,
        parents=[input],
        activation=act,
        num_filters=num_filters,
        size=l.config.size)
Z
zhangjinchao01 已提交
1914 1915 1916 1917


@wrap_name_default("pool")
@layer_support()
Q
qijun 已提交
1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928
def img_pool_layer(input,
                   pool_size,
                   name=None,
                   num_channels=None,
                   pool_type=None,
                   stride=1,
                   padding=0,
                   layer_attr=None,
                   pool_size_y=None,
                   stride_y=None,
                   padding_y=None,
1929
                   img_width=None):
Z
zhangjinchao01 已提交
1930 1931 1932 1933 1934 1935 1936
    """
    Image pooling Layer.

    The details of pooling layer, please refer ufldl's pooling_ .

    .. _pooling: http://ufldl.stanford.edu/tutorial/supervised/Pooling/

1937
    :param padding: pooling padding width.
Z
zhangjinchao01 已提交
1938
    :type padding: int
1939 1940
    :param padding_y: pooling padding height. It's equal to padding by default.
    :type padding_y: int|None
Z
zhangjinchao01 已提交
1941 1942 1943 1944
    :param name: name of pooling layer
    :type name: basestring.
    :param input: layer's input
    :type input: LayerOutput
1945
    :param pool_size: pooling window width
Z
zhangjinchao01 已提交
1946
    :type pool_size: int
1947 1948
    :param pool_size_y: pooling window height. It's eaqual to pool_size by default.
    :type pool_size_y: int|None
Z
zhangjinchao01 已提交
1949 1950
    :param num_channels: number of input channel.
    :type num_channels: int
1951
    :param pool_type: pooling type. MaxPooling or AvgPooling. Default is
Z
zhangjinchao01 已提交
1952 1953
                      MaxPooling.
    :type pool_type: BasePoolingType
1954
    :param stride: stride width of pooling.
Z
zhangjinchao01 已提交
1955
    :type stride: int
1956 1957
    :param stride_y: stride height of pooling. It is equal to stride by default.
    :type stride_y: int|None
Z
zhangjinchao01 已提交
1958 1959
    :param layer_attr: Extra Layer attribute.
    :type layer_attr: ExtraLayerAttribute
1960 1961 1962
    :param img_width: the width of input feature map. If it is None, the input feature
                      map should be square.
    :type img_width: int|None
D
dangqingqing 已提交
1963 1964
    :return: LayerOutput object.
    :rtype: LayerOutput
Z
zhangjinchao01 已提交
1965 1966 1967 1968 1969 1970 1971 1972 1973 1974
    """
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters

    if pool_type is None:
        pool_type = MaxPooling()
    elif isinstance(pool_type, AvgPooling):
        pool_type.name = 'avg'

1975 1976 1977 1978 1979 1980 1981 1982
    type_name = pool_type.name + '-projection' \
      if (isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)) \
      else pool_type.name

    pool_size_y = pool_size if pool_size_y is None else pool_size_y
    stride_y = stride if stride_y is None else stride_y
    padding_y = padding if padding_y is None else padding_y

X
xuwei06 已提交
1983
    l = Layer(
Z
zhangjinchao01 已提交
1984 1985
        name=name,
        type=LayerType.POOL_LAYER,
Q
qijun 已提交
1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007
        inputs=[
            Input(
                input.name,
                pool=Pool(
                    pool_type=type_name,
                    channels=num_channels,
                    size_x=pool_size,
                    start=None,
                    stride=stride,
                    padding=padding,
                    size_y=pool_size_y,
                    stride_y=stride_y,
                    padding_y=padding_y,
                    img_width=img_width))
        ],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.POOL_LAYER,
        parents=[input],
        num_filters=num_channels,
        size=l.config.size)
Z
zhangjinchao01 已提交
2008 2009


Q
qijun 已提交
2010 2011
@wrap_name_default("spp")
@layer_support()
Q
qijun 已提交
2012 2013 2014 2015 2016 2017 2018
def spp_layer(input,
              name=None,
              num_channels=None,
              pool_type=None,
              pyramid_height=None,
              img_width=None,
              layer_attr=None):
Q
qijun 已提交
2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054
    """
    Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition.
    The details please refer to
    `Kaiming He's paper <https://arxiv.org/abs/1406.4729>`_.

    :param name: layer name.
    :type name: basestring
    :param input: layer's input.
    :type input: LayerOutput
    :param num_channels: number of input channel.
    :type num_channels: int
    :param pool_type: Pooling type. MaxPooling or AveragePooling. Default is MaxPooling.
    :type scale: BasePoolingType
    :param pyramid_height: pyramid height.
    :type pyramid_height: int
    :param img_width: the width of input feature map. If it is None, the input feature
                      map should be square.
    :type img_width: int|None
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
    :return: LayerOutput object.
    :rtype: LayerOutput
    """
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters

    if pool_type is None:
        pool_type = MaxPooling()
    elif isinstance(pool_type, AvgPooling):
        pool_type.name = 'avg'

    type_name = pool_type.name
    if (isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)):
        type_name += '-projection'

Q
qijun 已提交
2055
    l = Layer(
Q
qijun 已提交
2056 2057
        name=name,
        type=LayerType.SPP_LAYER,
Q
qijun 已提交
2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075
        inputs=Input(
            input.name,
            spp=SpatialPyramidPool(
                pool_type=type_name,
                channels=num_channels,
                pyramid_height=pyramid_height,
                img_width=img_width)),
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        layer_type=LayerType.SPP_LAYER,
        parents=[input],
        num_filters=num_channels,
        size=l.config.size)


def __img_norm_layer__(name, input, size, norm_type, scale, power, num_channels,
                       blocked, layer_attr):
Z
zhangjinchao01 已提交
2076 2077 2078 2079
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters

X
xuwei06 已提交
2080
    l = Layer(
Q
qijun 已提交
2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099
        name=name,
        type=LayerType.NORM_LAYER,
        inputs=Input(
            input.name,
            norm=Norm(
                norm_type=norm_type,
                channels=num_channels,
                size=size,
                scale=scale,
                pow=power,
                blocked=blocked)),
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        layer_type=LayerType.NORM_LAYER,
        parents=[input],
        num_filters=num_channels,
        img_norm_type=norm_type,
        size=l.config.size)
Z
zhangjinchao01 已提交
2100 2101 2102 2103


@wrap_name_default("crmnorm")
@layer_support()
Q
qijun 已提交
2104 2105 2106 2107 2108 2109
def img_cmrnorm_layer(input,
                      size,
                      scale=0.0128,
                      power=0.75,
                      name=None,
                      num_channels=None,
2110
                      layer_attr=None):
Z
zhangjinchao01 已提交
2111
    """
2112
    Response normalization across feature maps.
D
dangqingqing 已提交
2113 2114
    The details please refer to
    `Alex's paper <http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf>`_.
Z
zhangjinchao01 已提交
2115 2116

    :param name: layer name.
D
dangqingqing 已提交
2117
    :type name: None|basestring
Z
zhangjinchao01 已提交
2118 2119
    :param input: layer's input.
    :type input: LayerOutput
2120
    :param size: Normalize in number of :math:`size` feature maps.
Z
zhangjinchao01 已提交
2121
    :type size: int
D
dangqingqing 已提交
2122
    :param scale: The hyper-parameter.
Z
zhangjinchao01 已提交
2123
    :type scale: float
D
dangqingqing 已提交
2124
    :param power: The hyper-parameter.
Z
zhangjinchao01 已提交
2125 2126 2127 2128 2129
    :type power: float
    :param num_channels: input layer's filers number or channels. If
                         num_channels is None, it will be set automatically.
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
2130
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2131 2132 2133
    :rtype: LayerOutput
    """
    return __img_norm_layer__(name, input, size, "cmrnorm-projection", scale,
2134
                              power, num_channels, 0, layer_attr)
Z
zhangjinchao01 已提交
2135 2136 2137 2138 2139 2140 2141 2142


@wrap_bias_attr_default()
@wrap_param_attr_default(default_factory=lambda _: ParamAttr(initial_mean=1.0,
                                                             initial_std=0.))
@wrap_act_default(act=ReluActivation())
@wrap_name_default("batch_norm")
@layer_support(DROPOUT)
Q
qijun 已提交
2143 2144 2145 2146 2147 2148 2149
def batch_norm_layer(input,
                     act=None,
                     name=None,
                     num_channels=None,
                     bias_attr=None,
                     param_attr=None,
                     layer_attr=None,
Z
zhangjinchao01 已提交
2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184
                     batch_norm_type=None,
                     moving_average_fraction=0.9,
                     use_global_stats=None):
    """
    Batch Normalization Layer. The notation of this layer as follow.

    :math:`x` is the input features over a mini-batch.

    ..  math::

        \\mu_{\\beta} &\\gets \\frac{1}{m} \\sum_{i=1}^{m} x_i \\qquad &//\\
        \ mini-batch\ mean \\\\
        \\sigma_{\\beta}^{2} &\\gets \\frac{1}{m} \\sum_{i=1}^{m}(x_i - \\
        \\mu_{\\beta})^2 \\qquad &//\ mini-batch\ variance \\\\
        \\hat{x_i} &\\gets \\frac{x_i - \\mu_\\beta} {\\sqrt{\\
        \\sigma_{\\beta}^{2} + \\epsilon}} \\qquad &//\ normalize \\\\
        y_i &\\gets \\gamma \\hat{x_i} + \\beta \\qquad &//\ scale\ and\ shift

    The details of batch normalization please refer to this
    `paper <http://arxiv.org/abs/1502.03167>`_.

    :param name: layer name.
    :type name: basestring
    :param input: batch normalization input. Better be linear activation.
                Because there is an activation inside batch_normalization.
    :type input: LayerOutput
    :param batch_norm_type: We have batch_norm and cudnn_batch_norm. batch_norm
                            supports both CPU and GPU. cudnn_batch_norm requires
                            cuDNN version greater or equal to v4 (>=v4). But
                            cudnn_batch_norm is faster and needs less memory
                            than batch_norm. By default (None), we will
                            automaticly select cudnn_batch_norm for GPU and
                            batch_norm for CPU. Otherwise, select batch norm
                            type based on the specified type. If you use cudnn_batch_norm,
                            we suggested you use latest version, such as v5.1.
2185
    :type batch_norm_type: None|string, None or "batch_norm" or "cudnn_batch_norm"
Z
zhangjinchao01 已提交
2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212
    :param act: Activation Type. Better be relu. Because batch
                     normalization will normalize input near zero.
    :type act: BaseActivation
    :param num_channels: num of image channels or previous layer's number of
                         filters. None will automatically get from layer's
                         input.
    :type num_channels: int
    :param bias_attr: :math:`\\beta`, better be zero when initialize. So the
                      initial_std=0, initial_mean=1 is best practice.
    :type bias_attr: ParameterAttribute
    :param param_attr: :math:`\\gamma`, better be one when initialize. So the
                       initial_std=0, initial_mean=1 is best practice.
    :type param_attr: ParameterAttribute
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
    :param use_global_stats: whether use moving mean/variance statistics
                             during testing peroid. If None or True,
                             it will use moving mean/variance statistics during
                             testing. If False, it will use the mean
                             and variance of current batch of test data for
                             testing.
    :type use_global_stats: bool|None.
    :param moving_average_fraction: Factor used in the moving average
                                   computation, referred to as facotr,
                                   :math:`runningMean = newMean*(1-factor)
                                   + runningMean*factor`
    :type moving_average_fraction: float.
D
dangqingqing 已提交
2213
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232
    :rtype: LayerOutput
    """
    if not isinstance(act, ReluActivation):
        logger.log(logging.WARN,
                   "%s is not recommend for batch normalization's activation, "
                   "maybe the relu is better" % act.name)

    if not isinstance(input.activation, LinearActivation):
        logger.log(logging.WARN,
                   "The activation should be inside batch normalization, the "
                   "previous layer's activation may be Linear")

    if num_channels is None:
        if input.num_filters is not None:
            num_channels = input.num_filters
        else:
            num_channels = input.size
    assert (batch_norm_type is None) or (batch_norm_type == "batch_norm") or \
           (batch_norm_type == "cudnn_batch_norm")
X
xuwei06 已提交
2233
    l = Layer(
Z
zhangjinchao01 已提交
2234
        name=name,
Q
qijun 已提交
2235 2236
        inputs=Input(
            input.name, image=Image(channels=num_channels), **param_attr.attr),
Z
zhangjinchao01 已提交
2237 2238 2239 2240 2241 2242
        active_type=act.name,
        type=LayerType.BATCH_NORM_LAYER,
        batch_norm_type=batch_norm_type,
        bias=ParamAttr.to_bias(bias_attr),
        moving_average_fraction=moving_average_fraction,
        use_global_stats=use_global_stats,
Q
qijun 已提交
2243
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
2244

Q
qijun 已提交
2245 2246 2247 2248 2249 2250 2251
    return LayerOutput(
        name=name,
        layer_type=LayerType.BATCH_NORM_LAYER,
        parents=[input],
        activation=act,
        num_filters=num_channels,
        size=l.config.size)
Z
zhangjinchao01 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 2277 2278


@wrap_name_default()
@layer_support()
def sum_to_one_norm_layer(input, name=None, layer_attr=None):
    """
    A layer for sum-to-one normalization,
    which is used in NEURAL TURING MACHINE.

    .. math::
       out[i] = \\frac {in[i]} {\sum_{k=1}^N in[k]}

    where :math:`in` is a (batchSize x dataDim) input vector,
    and :math:`out` is a (batchSize x dataDim) output vector.

    The example usage is:

    .. code-block:: python

       sum_to_one_norm = sum_to_one_norm_layer(input=layer)

    :param input: Input layer.
    :type input: LayerOutput
    :param name: Layer name.
    :type name: basestring
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
2279
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2280 2281 2282 2283 2284 2285
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.SUM_TO_ONE_NORM_LAYER,
        inputs=[input.name],
Q
qijun 已提交
2286 2287 2288
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.SUM_TO_ONE_NORM_LAYER, parents=[input], size=input.size)
Z
zhangjinchao01 已提交
2289 2290 2291 2292 2293 2294


@wrap_name_default("addto")
@wrap_act_default(act=LinearActivation())
@wrap_bias_attr_default(has_bias=False)
@layer_support(DROPOUT)
Q
qijun 已提交
2295
def addto_layer(input, act=None, name=None, bias_attr=None, layer_attr=None):
Z
zhangjinchao01 已提交
2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317
    """
    AddtoLayer.

    ..  math::

        y = f(\\sum_{i} x_i + b)

    where :math:`y` is output, :math:`x` is input, :math:`b` is bias,
    and :math:`f` is activation function.

    The example usage is:

    ..  code-block:: python

        addto = addto_layer(input=[layer1, layer2],
                            act=ReluActivation(),
                            bias_attr=False)

    This layer just simply add all input layers together, then activate the sum
    inputs. Each input of this layer should be the same size, which is also the
    output size of this layer.

C
caoying03 已提交
2318 2319 2320
    There is no weight matrix for each input, because it just a simple add
    operation. If you want a complicated operation before add, please use
    mixed_layer.
Z
zhangjinchao01 已提交
2321 2322

    It is a very good way to set dropout outside the layers. Since not all
C
caoying03 已提交
2323 2324
    PaddlePaddle layer support dropout, you can add an add_to layer, set
    dropout here.
Z
zhangjinchao01 已提交
2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338
    Please refer to dropout_layer for details.

    :param name: Layer name.
    :type name: basestring
    :param input: Input layers. It could be a LayerOutput or list/tuple of
                 LayerOutput.
    :type input: LayerOutput|list|tuple
    :param act: Activation Type, default is tanh.
    :type act: BaseActivation
    :param bias_attr: Bias attribute. If False, means no bias. None is default
                      bias.
    :type bias_attr: ParameterAttribute|bool
    :param layer_attr: Extra Layer attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
2339
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2340 2341 2342 2343 2344 2345
    :rtype: LayerOutput
    """
    num_filters = None
    if isinstance(input, LayerOutput):
        input = [input]

2346
    assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
2347 2348 2349 2350 2351 2352 2353
    ipts_for_layer = []
    for each_input in input:
        assert isinstance(each_input, LayerOutput)
        ipts_for_layer.append(Input(each_input.name))
        if each_input.num_filters is not None:
            num_filters = each_input.num_filters

X
xuwei06 已提交
2354
    l = Layer(
Q
qijun 已提交
2355 2356 2357
        name=name,
        type=LayerType.ADDTO_LAYER,
        inputs=ipts_for_layer,
Z
zhangjinchao01 已提交
2358 2359
        bias=ParamAttr.to_bias(bias_attr),
        active_type=act.name,
Q
qijun 已提交
2360
        **ExtraLayerAttribute.to_kwargs(layer_attr))
2361

Q
qijun 已提交
2362 2363 2364 2365 2366 2367 2368
    return LayerOutput(
        name,
        LayerType.ADDTO_LAYER,
        parents=input,
        activation=act,
        num_filters=num_filters,
        size=l.config.size)
Z
zhangjinchao01 已提交
2369 2370 2371 2372 2373


@wrap_act_default(act=IdentityActivation())
@wrap_name_default("concat")
@layer_support()
2374
def concat_layer(input, act=None, name=None, layer_attr=None, bias_attr=None):
Z
zhangjinchao01 已提交
2375 2376 2377 2378
    """
    Concat all input vector into one huge vector.
    Inputs can be list of LayerOutput or list of projection.

2379 2380 2381 2382 2383 2384
    The example usage is:

    ..  code-block:: python

        concat = concat_layer(input=[layer1, layer2])

Z
zhangjinchao01 已提交
2385 2386 2387
    :param name: Layer name.
    :type name: basestring
    :param input: input layers or projections
2388
    :type input: list|tuple|collections.Sequence
Z
zhangjinchao01 已提交
2389 2390 2391 2392
    :param act: Activation type.
    :type act: BaseActivation
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
2393
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2394 2395 2396 2397 2398 2399 2400 2401
    :rtype: LayerOutput
    """

    if isinstance(input, LayerOutput):
        input = [input]
    elif isinstance(input, Projection):
        input = [input]
    else:
2402
        assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
2403 2404

    def __is_type__(o, tp):
2405
        if not isinstance(o, collections.Sequence):
Z
zhangjinchao01 已提交
2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426
            if o == tp:
                return True
            elif len(o.__bases__) == 0:
                return False
            else:
                for bs in o.__bases__:
                    if __is_type__(bs, tp):
                        return True
                return False
        else:
            tmp = map(lambda _x: __is_type__(_x, tp), o)
            a = tmp[0]
            for b in tmp[1:]:
                assert a == b
            return a

    def __reduce_concat_type__(a, b):
        assert __is_type__([a, b], Projection) or __is_type__([a, b],
                                                              LayerOutput)
        return a

Q
qijun 已提交
2427 2428
    is_concat_layer = __is_type__(
        reduce(__reduce_concat_type__, map(type, input)), LayerOutput)
Z
zhangjinchao01 已提交
2429

Q
qijun 已提交
2430 2431
    layer_type = (LayerType.CONCAT_LAYER
                  if is_concat_layer else LayerType.CONCAT_PROJ_LAYER)
Z
zhangjinchao01 已提交
2432

2433 2434
    if layer_type == LayerType.CONCAT_LAYER:
        assert not bias_attr
2435

Z
zhangjinchao01 已提交
2436
    Layer(
Q
qijun 已提交
2437 2438
        name=name,
        type=layer_type,
Z
zhangjinchao01 已提交
2439 2440
        inputs=[x.name for x in input] if is_concat_layer else input,
        active_type=act.name,
2441
        bias=ParamAttr.to_bias(bias_attr),
Q
qijun 已提交
2442
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
2443 2444 2445 2446 2447 2448 2449 2450 2451

    sz = 0
    for each_input in input:
        if each_input.size is not None:
            sz += each_input.size
        else:
            sz = None
            break

Q
qijun 已提交
2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465
    return LayerOutput(
        name,
        layer_type=layer_type,
        parents=input if is_concat_layer else [x.origin for x in input],
        activation=act,
        size=sz)


def memory(name,
           size,
           is_seq=False,
           boot_layer=None,
           boot_bias=None,
           boot_bias_active_type=None,
Z
zhangjinchao01 已提交
2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500
           boot_with_const_id=None):
    """
    The memory layers is a layer cross each time step. Reference this output
    as previous time step layer :code:`name` 's output.

    The default memory is zero in first time step, previous time step's
    output in the rest time steps.

    If boot_bias, the first time step value is this bias and
    with activation.

    If boot_with_const_id, then the first time stop is a IndexSlot, the
    Arguments.ids()[0] is this :code:`cost_id`.

    If boot_layer is not null, the memory is just the boot_layer's output.
    Set :code:`is_seq` is true boot layer is sequence.


    The same name layer in recurrent group will set memory on each time
    step.

    :param name: memory's name.
    :type name: basestring
    :param size: size of memory.
    :type size: int
    :param is_seq: is sequence for boot_layer
    :type is_seq: bool
    :param boot_layer: boot layer of memory.
    :type boot_layer: LayerOutput|None
    :param boot_bias: boot layer's bias
    :type boot_bias: ParameterAttribute|None
    :param boot_bias_active_type: boot layer's active type.
    :type boot_bias_active_type: BaseActivation
    :param boot_with_const_id: boot layer's id.
    :type boot_with_const_id: int
D
dangqingqing 已提交
2501
    :return: LayerOutput object which is a memory.
Z
zhangjinchao01 已提交
2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512
    :rtype: LayerOutput
    """
    if boot_bias_active_type is None:
        boot_bias_active_type = LinearActivation()

    assert boot_bias is None or isinstance(boot_bias, ParameterAttribute)
    if isinstance(boot_bias, ParameterAttribute):
        boot_bias = ParamAttr.to_bias(boot_bias)

    assert boot_layer is None or isinstance(boot_layer, LayerOutput)

Q
qijun 已提交
2513 2514 2515 2516 2517 2518 2519 2520 2521
    agent_name = Memory(name, size, is_seq, boot_layer.name
                        if boot_layer is not None else None, boot_bias,
                        boot_bias_active_type.name, boot_with_const_id)

    lout = LayerOutput(
        name=agent_name,
        size=size,
        layer_type=LayerType.MEMORY,
        parents=[boot_layer] if boot_layer is not None else None)
Z
zhangjinchao01 已提交
2522 2523 2524 2525
    return lout


@wrap_bias_attr_default()
Q
qijun 已提交
2526 2527
@wrap_act_default(
    param_names=['gate_act', 'state_act'], act=SigmoidActivation())
Z
zhangjinchao01 已提交
2528 2529 2530
@wrap_act_default(act=TanhActivation())
@wrap_name_default('lstm_step')
@layer_support()
Q
qijun 已提交
2531 2532 2533 2534 2535 2536 2537 2538 2539
def lstm_step_layer(input,
                    state,
                    size,
                    act=None,
                    name=None,
                    gate_act=None,
                    state_act=None,
                    bias_attr=None,
                    layer_attr=None):
Z
zhangjinchao01 已提交
2540 2541 2542 2543 2544 2545
    """
    LSTM Step Layer. It used in recurrent_group. The lstm equations are shown
    as follow.

    ..  math::

L
luotao02 已提交
2546
        i_t & = \\sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i)
Z
zhangjinchao01 已提交
2547

L
luotao02 已提交
2548
        f_t & = \\sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f)
Z
zhangjinchao01 已提交
2549

L
luotao02 已提交
2550
        c_t & = f_tc_{t-1} + i_t tanh (W_{xc}x_t+W_{hc}h_{t-1} + b_c)
Z
zhangjinchao01 已提交
2551

L
luotao02 已提交
2552
        o_t & = \\sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o)
Z
zhangjinchao01 已提交
2553

L
luotao02 已提交
2554
        h_t & = o_t tanh(c_t)
Z
zhangjinchao01 已提交
2555 2556


L
luotao02 已提交
2557
    The input of lstm step is :math:`Wx_t + Wh_{t-1}`, and user should use
Z
zhangjinchao01 已提交
2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595
    :code:`mixed_layer` and :code:`full_matrix_projection` to calculate these
    input vector.

    The state of lstm step is :math:`c_{t-1}`. And lstm step layer will do

    ..  math::

        i_t = \\sigma(input + W_{ci}c_{t-1} + b_i)

        ...


    This layer contains two outputs. Default output is :math:`h_t`. The other
    output is :math:`o_t`, which name is 'state' and can use
    :code:`get_output_layer` to extract this output.

    :param name: Layer's name.
    :type name: basestring
    :param size: Layer's size. NOTE: lstm layer's size, should be equal as
                 :code:`input.size/4`, and should be equal as
                 :code:`state.size`.
    :type size: int
    :param input: input layer. :math:`Wx_t + Wh_{t-1}`
    :type input: LayerOutput
    :param state: State Layer. :math:`c_{t-1}`
    :type state: LayerOutput
    :param act: Activation type. Default is tanh
    :type act: BaseActivation
    :param gate_act: Gate Activation Type. Default is sigmoid, and should
                          be sigmoid only.
    :type gate_act: BaseActivation
    :param state_act: State Activation Type. Default is sigmoid, and should
                           be sigmoid only.
    :type state_act: BaseActivation
    :param bias_attr: Bias Attribute.
    :type bias_attr: ParameterAttribute
    :param layer_attr: layer's extra attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
2596
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2597 2598 2599 2600 2601 2602 2603 2604 2605
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.LSTM_STEP_LAYER,
        active_type=act.name,
        active_gate_type=gate_act.name,
        active_state_type=state_act.name,
        bias=ParamAttr.to_bias(bias_attr),
Q
qijun 已提交
2606 2607 2608
        size=size,
        inputs=[input.name, state.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
2609

Q
qijun 已提交
2610 2611 2612 2613 2614 2615 2616
    return LayerOutput(
        name=name,
        layer_type=LayerType.LSTM_STEP_LAYER,
        parents=[input, state],
        activation=act,
        size=size,
        outputs=['default', 'state'])
Z
zhangjinchao01 已提交
2617 2618 2619


@wrap_bias_attr_default()
Q
qijun 已提交
2620
@wrap_act_default(param_names=['gate_act'], act=SigmoidActivation())
Z
zhangjinchao01 已提交
2621 2622 2623
@wrap_act_default(act=TanhActivation())
@wrap_name_default('gru_step')
@layer_support()
Q
qijun 已提交
2624 2625 2626 2627 2628 2629 2630 2631
def gru_step_layer(input,
                   output_mem,
                   size=None,
                   act=None,
                   name=None,
                   gate_act=None,
                   bias_attr=None,
                   layer_attr=None):
Z
zhangjinchao01 已提交
2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642
    """

    :param input:
    :type input: LayerOutput
    :param output_mem:
    :param size:
    :param act:
    :param name:
    :param gate_act:
    :param bias_attr:
    :param layer_attr:
D
dangqingqing 已提交
2643
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2644 2645 2646 2647 2648 2649 2650 2651
    :rtype: LayerOutput
    """
    assert input.size % 3 == 0
    if size is None:
        size = input.size / 3
    Layer(
        name=name,
        type=LayerType.GRU_STEP_LAYER,
Q
qijun 已提交
2652
        inputs=[input.name, output_mem.name],
Z
zhangjinchao01 已提交
2653 2654 2655 2656
        bias=ParamAttr.to_bias(bias_attr),
        size=size,
        active_type=act.name,
        active_gate_type=gate_act.name,
Q
qijun 已提交
2657
        **ExtraAttr.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
2658
    return LayerOutput(
Q
qijun 已提交
2659 2660
        name=name,
        layer_type=LayerType.GRU_STEP_LAYER,
Z
zhangjinchao01 已提交
2661
        parents=[input, output_mem],
Q
qijun 已提交
2662 2663
        size=size,
        activation=act)
Z
zhangjinchao01 已提交
2664 2665 2666 2667 2668 2669


@wrap_name_default()
@layer_support()
def get_output_layer(input, arg_name, name=None, layer_attr=None):
    """
C
caoying03 已提交
2670 2671 2672 2673
    Get layer's output by name. In PaddlePaddle, a layer might return multiple
    values, but returns one layer's output. If the user wants to use another
    output besides the default one, please use get_output_layer first to get
    the output from input.
Z
zhangjinchao01 已提交
2674 2675 2676 2677 2678 2679 2680 2681 2682

    :param name: Layer's name.
    :type name: basestring
    :param input: get output layer's input. And this layer should contains
                   multiple outputs.
    :type input: LayerOutput
    :param arg_name: Output name from input.
    :type arg_name: basestring
    :param layer_attr: Layer's extra attribute.
D
dangqingqing 已提交
2683
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2684 2685 2686 2687 2688 2689 2690
    :rtype: LayerOutput
    """
    # GetOutputLayer
    assert arg_name in input.outputs, 'Get Output From an not existed input.' \
                                      ' The get output name is %s, which not' \
                                      ' in %s' % (
                                          arg_name, ",".join(input.outputs))
Q
qijun 已提交
2691 2692 2693 2694 2695 2696 2697
    Layer(
        name=name,
        type=LayerType.GET_OUTPUT_LAYER,
        inputs=[Input(
            input.name, input_layer_argument=arg_name)],
        size=input.size,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
2698

Q
qijun 已提交
2699 2700 2701 2702 2703
    return LayerOutput(
        name=name,
        layer_type=LayerType.GET_OUTPUT_LAYER,
        parents=[input],
        size=input.size)
Z
zhangjinchao01 已提交
2704 2705 2706 2707 2708 2709 2710


@wrap_name_default()
@wrap_act_default()
@wrap_bias_attr_default()
@wrap_param_attr_default()
@layer_support()
Q
qijun 已提交
2711 2712 2713 2714 2715 2716 2717
def recurrent_layer(input,
                    act=None,
                    bias_attr=None,
                    param_attr=None,
                    name=None,
                    reverse=False,
                    layer_attr=None):
Z
zhangjinchao01 已提交
2718
    """
2719 2720
    Simple recurrent unit layer. It is just a fully connect layer through both
    time and neural network.
Z
zhangjinchao01 已提交
2721

2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748
    For each sequence [start, end] it performs the following computation\:

    ..  math::

        out_{i} = act(in_{i})     \\      \\      \\text{for} \\ i = start \\\\
        out_{i} = act(in_{i} + out_{i-1} * W) \\ \\ \\text{for} \\ start < i <= end

    If reversed is true, the order is reversed\:

    ..  math::

        out_{i} = act(in_{i})           \\    \\   \\text{for} \\ i = end  \\\\
        out_{i} = act(in_{i} + out_{i+1} * W) \\ \\ \\text{for} \\ start <= i < end


    :param input: Input Layer
    :type input: LayerOutput
    :param act: activation.
    :type act: BaseActivation
    :param bias_attr: bias attribute.
    :type bias_attr: ParameterAttribute
    :param param_attr: parameter attribute.
    :type param_attr: ParameterAttribute
    :param name: name of the layer
    :type name: basestring
    :param layer_attr: Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
2749
    :return: LayerOutput object.
2750
    :rtype: LayerOutput
Z
zhangjinchao01 已提交
2751
    """
Q
qijun 已提交
2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766
    Layer(
        name=name,
        type=LayerType.RECURRENT_LAYER,
        inputs=Input(input.name, **param_attr.attr),
        active_type=act.name,
        bias=ParamAttr.to_bias(bias_attr),
        reversed=reverse,
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name=name,
        layer_type=LayerType.RECURRENT_LAYER,
        parents=[input],
        size=input.size,
        activation=act,
        reverse=reverse)
Z
zhangjinchao01 已提交
2767 2768 2769 2770 2771 2772 2773


class StaticInput(object):
    """
    StaticInput is only used in recurrent_group which defines a read-only memory
    that can be a sequence or non-sequence.
    """
2774

Z
zhangjinchao01 已提交
2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793
    def __init__(self, input, is_seq=False, size=None):
        assert isinstance(input, LayerOutput)
        self.input = input
        self.is_seq = is_seq
        assert input.size is not None or size is not None
        if size is not None:
            input.size = size


class SubsequenceInput(object):
    """
    Input sequence has sub-sequence, used in recurrent_group.

    The example usage is:

    .. code-block:: python

       input = SubsequenceInput(layer)
    """
2794

Z
zhangjinchao01 已提交
2795 2796 2797 2798 2799 2800 2801
    def __init__(self, input):
        assert isinstance(input, LayerOutput)
        assert input.size is not None
        self.input = input


@wrap_name_default("recurrent_group")
L
Luo Tao 已提交
2802 2803 2804 2805 2806
def recurrent_group(step,
                    input,
                    reverse=False,
                    name=None,
                    targetInlink=None,
L
Luo Tao 已提交
2807
                    is_generating=False):
Z
zhangjinchao01 已提交
2808
    """
C
caoying03 已提交
2809 2810 2811 2812 2813
    Recurrent layer group is an extremely flexible recurrent unit in
    PaddlePaddle. As long as the user defines the calculation done within a
    time step, PaddlePaddle will iterate such a recurrent calculation over
    sequence input. This is extremely usefull for attention based model, or
    Neural Turning Machine like models.
Z
zhangjinchao01 已提交
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 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857

    The basic usage (time steps) is:

    .. code-block:: python

       def step(input):
           output = fc_layer(input=layer,
                             size=1024,
                             act=LinearActivation(),
                             bias_attr=False)
           return output

       group = recurrent_group(input=layer,
                               step=step)

    You can see following configs for further usages:

    - time steps: lstmemory_group, paddle/gserver/tests/sequence_layer_group.conf, \
                  demo/seqToseq/seqToseq_net.py
    - sequence steps: paddle/gserver/tests/sequence_nest_layer_group.conf

    :param step: recurrent one time step function.The input of this function is
                 input of the group. The return of this function will be
                 recurrent group's return value.

                 The recurrent group scatter a sequence into time steps. And
                 for each time step, will invoke step function, and return
                 a time step result. Then gather each time step of output into
                 layer group's output.

    :type step: callable

    :param name: recurrent_group's name.
    :type name: basestring

    :param input: Input links array.

                  LayerOutput will be scattered into time steps.
                  SubsequenceInput will be scattered into sequence steps.
                  StaticInput will be imported to each time step, and doesn't change
                  through time. It's a mechanism to access layer outside step function.

    :type input: LayerOutput|StaticInput|SubsequenceInput|list|tuple

2858 2859
    :param reverse: If reverse is set true, the recurrent unit will process the
                    input sequence in a reverse order.
Z
zhangjinchao01 已提交
2860
    :type reverse: bool
2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871

    :param targetInlink: the input layer which share info with layer group's output

                         Param input specifies multiple input layers. For
                         SubsequenceInput inputs, config should assign one input
                         layer that share info(the number of sentences and the number
                         of words in each sentence) with all layer group's outputs.
                         targetInlink should be one of the layer group's input.

    :type targetInlink: LayerOutput|SubsequenceInput

L
Luo Tao 已提交
2872
    :param is_generating: If is generating, none of input type should be LayerOutput;
2873
                          else, for training or testing, one of the input type must
L
Luo Tao 已提交
2874
                          be LayerOutput.
L
Luo Tao 已提交
2875

L
Luo Tao 已提交
2876
    : type is_generating: bool
2877

D
dangqingqing 已提交
2878
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2879 2880 2881 2882 2883 2884 2885 2886 2887 2888
    :rtype: LayerOutput
    """
    model_type('recurrent_nn')

    def is_single_input(x):
        return isinstance(x, LayerOutput) or isinstance(x, StaticInput) \
               or isinstance(x, SubsequenceInput)

    if is_single_input(input):
        input = [input]
2889
    assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
2890 2891 2892 2893 2894 2895

    def is_in_links(x):
        return isinstance(x, LayerOutput) or isinstance(x, SubsequenceInput)

    in_links = filter(is_in_links, input)

2896 2897 2898 2899 2900 2901 2902 2903 2904
    def targetInlink_in_inlinks():
        for inlink in in_links:
            if isinstance(inlink, SubsequenceInput):
                if targetInlink == inlink.input:
                    return True
            elif targetInlink == inlink:
                return True
        return False

Q
qijun 已提交
2905
    assert (targetInlink == None or targetInlink_in_inlinks())
2906 2907 2908 2909
    targetInlinkName = None if targetInlink == None \
                            else targetInlink.name if isinstance(targetInlink, LayerOutput) \
                                                   else targetInlink.input.name

Z
zhangjinchao01 已提交
2910 2911 2912 2913 2914 2915 2916 2917 2918 2919
    contains_sub_seq = [False]

    def map_in_links(x):
        if isinstance(x, SubsequenceInput):
            contains_sub_seq[0] = True
            return Link(name=x.input.name, has_subseq=True)
        else:
            return x.name

    RecurrentLayerGroupWithoutOutLinksBegin(
Q
qijun 已提交
2920 2921
        name=name,
        in_links=map(map_in_links, in_links),
2922 2923
        seq_reversed=reverse,
        target_inlinkname=targetInlinkName)
Z
zhangjinchao01 已提交
2924
    in_args = []
2925
    has_LayerOutput = False
Z
zhangjinchao01 已提交
2926 2927 2928 2929
    for each_input in input:
        assert is_single_input(each_input)
        if isinstance(each_input, LayerOutput):
            in_args.append(each_input)
2930
            has_LayerOutput = True
Z
zhangjinchao01 已提交
2931 2932
        elif isinstance(each_input, SubsequenceInput):
            in_args.append(each_input.input)
2933
            has_LayerOutput = True
Z
zhangjinchao01 已提交
2934 2935
        else:
            mem_name = "__%s_memory__" % each_input.input.name
Q
qijun 已提交
2936 2937 2938 2939 2940 2941 2942 2943 2944
            mem = memory(
                name=mem_name,
                is_seq=each_input.is_seq,
                size=each_input.input.size,
                boot_layer=each_input.input)
            with mixed_layer(
                    name=mem_name,
                    size=each_input.input.size,
                    act=IdentityActivation()) as mix:
Z
zhangjinchao01 已提交
2945 2946 2947
                mix += identity_projection(mem)
            in_args.append(mem)

L
Luo Tao 已提交
2948
    assert (is_generating != has_LayerOutput)
L
Luo Tao 已提交
2949

Z
zhangjinchao01 已提交
2950 2951 2952 2953 2954 2955 2956
    layer_outs = step(*in_args)

    if isinstance(layer_outs, LayerOutput):
        layer_outs = [layer_outs]

    for ot in layer_outs:
        assert isinstance(ot, LayerOutput)
2957
        ot.reverse = reverse
Z
zhangjinchao01 已提交
2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969
        if contains_sub_seq[0]:
            RecurrentLayerGroupSetOutLink(Link(ot.name, has_subseq=True))
        else:
            RecurrentLayerGroupSetOutLink(ot.name)

    RecurrentLayerGroupEnd(name=name)

    if len(layer_outs) == 1:
        return layer_outs[0]
    else:
        return layer_outs

2970

Z
zhangjinchao01 已提交
2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987
class BaseGeneratedInput(object):
    def __init__(self):
        self.bos_id = None
        self.eos_id = None

    def before_real_step(self):
        raise NotImplementedError()

    def after_real_step(self, *args):
        raise NotImplementedError()


class GeneratedInput(BaseGeneratedInput):
    def after_real_step(self, input):
        return maxid_layer(input=input, name='__beam_search_predict__')

    def before_real_step(self):
Q
qijun 已提交
2988 2989 2990 2991 2992 2993 2994 2995 2996
        predict_id = memory(
            name='__beam_search_predict__',
            size=self.size,
            boot_with_const_id=self.bos_id)

        trg_emb = embedding_layer(
            input=predict_id,
            size=self.embedding_size,
            param_attr=ParamAttr(name=self.embedding_name))
Z
zhangjinchao01 已提交
2997 2998 2999
        return trg_emb

    def __init__(self, size, embedding_name, embedding_size):
3000
        super(GeneratedInput, self).__init__()
Z
zhangjinchao01 已提交
3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023
        self.size = size
        self.embedding_name = embedding_name
        self.embedding_size = embedding_size


@wrap_name_default()
def maxid_layer(input, name=None, layer_attr=None):
    """
    A layer for finding the id which has the maximal value for each sample.
    The result is stored in output.ids.

    The example usage is:

    .. code-block:: python

       maxid = maxid_layer(input=layer)

    :param input: Input layer name.
    :type input: LayerOutput
    :param name: Layer name.
    :type name: basestring
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
3024
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3025 3026 3027 3028
    :rtype: LayerOutput
    """

    assert isinstance(input, LayerOutput)
Q
qijun 已提交
3029 3030 3031 3032 3033 3034 3035 3036 3037 3038
    l = Layer(
        name=name,
        type='maxid',
        inputs=[input.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name=name,
        layer_type=LayerType.MAXID_LAYER,
        parents=[input],
        size=l.config.size)
Z
zhangjinchao01 已提交
3039

3040

H
Haonan 已提交
3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066
@wrap_name_default()
def out_prod_layer(input1, input2, name=None, layer_attr=None):
    """
    A layer for computing the outer product of two vectors
    The result is a matrix of size(input1) x size(input2)

    The example usage is:

    .. code-block:: python

       out_prod = out_prod_layer(input1=vec1, input2=vec2)

    :param name: Layer name.
    :type name: basestring
    :param input1: The first input layer name.
    :type input: LayerOutput
    :param input2: The second input layer name.
    :type input2: LayerOutput
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
    :return: LayerOutput object.
    :rtype: LayerOutput
    """

    assert isinstance(input1, LayerOutput)
    assert isinstance(input2, LayerOutput)
Q
qijun 已提交
3067 3068 3069 3070 3071 3072 3073 3074 3075 3076
    l = Layer(
        name=name,
        type=LayerType.OUT_PROD_LAYER,
        inputs=[input1.name, input2.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name=name,
        layer_type=LayerType.OUT_PROD_LAYER,
        parents=[input1, input2],
        size=l.config.size)
3077

Z
zhangjinchao01 已提交
3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093

@wrap_name_default()
def eos_layer(input, eos_id, name=None, layer_attr=None):
    """
    A layer for checking EOS for each sample:
    - output_id = (input_id == conf.eos_id)

    The result is stored in output\_.ids.
    It is used by recurrent layer group.

    The example usage is:

    .. code-block:: python

       eos = eos_layer(input=layer, eos_id=id)

L
luotao02 已提交
3094 3095
    :param name: Layer name.
    :type name: basestring
Z
zhangjinchao01 已提交
3096 3097 3098 3099 3100 3101
    :param input: Input layer name.
    :type input: LayerOutput
    :param eos_id: end id of sequence
    :type eos_id: int
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
3102
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3103 3104
    :rtype: LayerOutput
    """
Q
qijun 已提交
3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115
    l = Layer(
        name=name,
        type=LayerType.EOSID_LAYER,
        eos_id=eos_id,
        inputs=[input.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name=name,
        layer_type=LayerType.EOSID_LAYER,
        parents=[input],
        size=l.config.size)
Z
zhangjinchao01 已提交
3116 3117 3118


@wrap_name_default()
Q
qijun 已提交
3119 3120 3121 3122 3123 3124 3125
def beam_search(step,
                input,
                bos_id,
                eos_id,
                beam_size,
                max_length=500,
                name=None,
Z
zhangjinchao01 已提交
3126
                num_results_per_sample=None):
3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137
    """
    Beam search is a heuristic search algorithm used in sequence generation.
    It explores a graph by expanding the most promising nodes in a limited set
    to maintain tractability.

    The example usage is:

    .. code-block:: python

        def rnn_step(input):
            last_time_step_output = memory(name='rnn', size=512)
3138
            with mixed_layer(size=512, name='rnn') as simple_rnn:
3139 3140 3141 3142 3143 3144
                simple_rnn += full_matrix_projection(input)
                simple_rnn += last_time_step_output
            return simple_rnn

        beam_gen = beam_search(name="decoder",
                               step=rnn_step,
3145
                               input=[StaticInput(encoder_last)],
3146 3147
                               bos_id=0,
                               eos_id=1,
3148
                               beam_size=5)
3149 3150 3151 3152 3153 3154 3155 3156 3157

    Please see the following demo for more details:

    - machine translation : demo/seqToseq/translation/gen.conf \
                            demo/seqToseq/seqToseq_net.py

    :param name: Name of the recurrent unit that generates sequences.
    :type name: base string
    :param step: A callable function that defines the calculation in a time
3158
                 step, and it is applied to sequences with arbitrary length by
3159 3160 3161 3162 3163 3164
                 sharing a same set of weights.

                 You can refer to the first parameter of recurrent_group, or
                 demo/seqToseq/seqToseq_net.py for more details.
    :type step: callable
    :param input: Input data for the recurrent unit
3165
    :type input: list
3166 3167 3168
    :param bos_id: Index of the start symbol in the dictionary. The start symbol
                   is a special token for NLP task, which indicates the
                   beginning of a sequence. In the generation task, the start
3169
                   symbol is essential, since it is used to initialize the RNN
3170 3171 3172 3173 3174 3175 3176 3177
                   internal state.
    :type bos_id: int
    :param eos_id: Index of the end symbol in the dictionary. The end symbol is
                   a special token for NLP task, which indicates the end of a
                   sequence. The generation process will stop once the end
                   symbol is generated, or a pre-defined max iteration number
                   is exceeded.
    :type eos_id: int
3178 3179
    :param max_length: Max generated sequence length.
    :type max_length: int
3180 3181 3182 3183 3184 3185 3186 3187 3188 3189
    :param beam_size: Beam search for sequence generation is an iterative search
                      algorithm. To maintain tractability, every iteration only
                      only stores a predetermined number, called the beam_size,
                      of the most promising next words. The greater the beam
                      size, the fewer candidate words are pruned.
    :type beam_size: int
    :param num_results_per_sample: Number of the generated results per input
                                  sequence. This number must always be less than
                                  beam size.
    :type num_results_per_sample: int
3190 3191
    :return: The generated word index.
    :rtype: LayerOutput
3192 3193
    """

Z
zhangjinchao01 已提交
3194 3195 3196 3197 3198
    if num_results_per_sample is None:
        num_results_per_sample = beam_size
    if num_results_per_sample > beam_size:
        logger.warning("num_results_per_sample should be less than beam_size")

Q
qijun 已提交
3199
    if isinstance(input, StaticInput) or isinstance(input, BaseGeneratedInput):
Z
zhangjinchao01 已提交
3200 3201 3202 3203 3204 3205
        input = [input]

    generated_input_index = -1

    real_input = []
    for i, each_input in enumerate(input):
3206 3207
        assert isinstance(each_input, StaticInput) or isinstance(
            each_input, BaseGeneratedInput)
Z
zhangjinchao01 已提交
3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223
        if isinstance(each_input, BaseGeneratedInput):
            assert generated_input_index == -1
            generated_input_index = i
        else:
            real_input.append(each_input)

    assert generated_input_index != -1

    gipt = input[generated_input_index]
    assert isinstance(gipt, BaseGeneratedInput)

    gipt.bos_id = bos_id
    gipt.eos_id = eos_id

    def __real_step__(*args):
        eos_name = "__%s_eos_layer__" % name
Q
qijun 已提交
3224 3225 3226 3227 3228 3229
        RecurrentLayerGroupSetGenerator(
            Generator(
                eos_layer_name=eos_name,
                max_num_frames=max_length,
                beam_size=beam_size,
                num_results_per_sample=num_results_per_sample))
Z
zhangjinchao01 已提交
3230 3231 3232 3233 3234 3235 3236 3237 3238 3239

        args = list(args)
        args.insert(generated_input_index, gipt.before_real_step())

        predict = gipt.after_real_step(step(*args))

        eos_layer(input=predict, eos_id=eos_id, name=eos_name)

        return predict

Q
qijun 已提交
3240
    tmp = recurrent_group(
L
Luo Tao 已提交
3241 3242 3243 3244
        step=__real_step__,
        input=real_input,
        reverse=False,
        name=name,
L
Luo Tao 已提交
3245
        is_generating=True)
3246

Z
zhangjinchao01 已提交
3247 3248
    return tmp

Q
qijun 已提交
3249

3250 3251
def __cost_input__(input, label, weight=None):
    """
3252
    inputs and parents for cost layers.
3253 3254 3255 3256 3257 3258 3259 3260
    """
    ipts = [Input(input.name), Input(label.name)]
    parents = [input, label]
    if weight is not None:
        assert weight.layer_type == LayerType.DATA
        ipts.append(Input(weight.name))
        parents.append(weight)
    return ipts, parents
3261

Z
zhangjinchao01 已提交
3262 3263

@wrap_name_default()
L
luotao1 已提交
3264
@layer_support()
Q
qijun 已提交
3265
def regression_cost(input, label, weight=None, name=None, layer_attr=None):
Z
zhangjinchao01 已提交
3266 3267 3268 3269 3270 3271
    """
    Regression Layer.

    TODO(yuyang18): Complete this method.

    :param name: layer name.
3272
    :type name: basestring
Z
zhangjinchao01 已提交
3273
    :param input: Network prediction.
3274
    :type input: LayerOutput
Z
zhangjinchao01 已提交
3275
    :param label: Data label.
3276 3277 3278 3279
    :type label: LayerOutput
    :param weight: The weight affects the cost, namely the scale of cost.
                   It is an optional argument.
    :type weight: LayerOutput
L
luotao1 已提交
3280 3281
    :param layer_attr: layer's extra attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
3282
    :return: LayerOutput object.
3283
    :rtype: LayerOutput
Z
zhangjinchao01 已提交
3284
    """
3285 3286
    ipts, parents = __cost_input__(input, label, weight)

Q
qijun 已提交
3287 3288 3289 3290 3291
    Layer(
        inputs=ipts,
        type="square_error",
        name=name,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
X
xuwei06 已提交
3292
    return LayerOutput(name, LayerType.COST, parents=parents, size=1)
Z
zhangjinchao01 已提交
3293 3294 3295


@wrap_name_default("cost")
3296
@layer_support()
Q
qijun 已提交
3297 3298 3299 3300
def classification_cost(input,
                        label,
                        weight=None,
                        name=None,
3301 3302
                        evaluator=classification_error_evaluator,
                        layer_attr=None):
Z
zhangjinchao01 已提交
3303 3304 3305 3306 3307 3308 3309 3310 3311
    """
    classification cost Layer.

    :param name: layer name.
    :type name: basestring
    :param input: input layer name. network output.
    :type input: LayerOutput
    :param label: label layer name. data_layer often.
    :type label: LayerOutput
3312 3313 3314
    :param weight: The weight affects the cost, namely the scale of cost.
                   It is an optional argument.
    :type weight: LayerOutput
Z
zhangjinchao01 已提交
3315
    :param evaluator: Evaluator method.
3316 3317
    :param layer_attr: layer's extra attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
3318
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3319 3320 3321 3322 3323
    :rtype: LayerOutput
    """
    assert input.layer_type != LayerType.DATA
    assert isinstance(input.activation, SoftmaxActivation)
    assert label.layer_type == LayerType.DATA
3324 3325 3326

    ipts, parents = __cost_input__(input, label, weight)

Q
qijun 已提交
3327 3328 3329 3330 3331
    Layer(
        name=name,
        type="multi-class-cross-entropy",
        inputs=ipts,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
3332 3333 3334 3335 3336 3337 3338 3339 3340 3341

    def __add_evaluator__(e):
        assert callable(e)
        assert hasattr(e, 'is_evaluator')
        assert isinstance(e.is_evaluator, bool)
        assert e.is_evaluator
        assert hasattr(e, "for_classification")
        assert isinstance(e.for_classification, bool)
        assert e.for_classification

3342
        e(name=e.__name__, input=input, label=label, weight=weight)
Z
zhangjinchao01 已提交
3343

3344
    if not isinstance(evaluator, collections.Sequence):
Z
zhangjinchao01 已提交
3345 3346 3347 3348 3349
        evaluator = [evaluator]

    for each_evaluator in evaluator:
        __add_evaluator__(each_evaluator)

X
xuwei06 已提交
3350
    return LayerOutput(name, LayerType.COST, parents=parents, size=1)
Z
zhangjinchao01 已提交
3351

3352

Q
qijun 已提交
3353 3354 3355 3356 3357 3358 3359 3360 3361 3362
def conv_operator(img,
                  filter,
                  filter_size,
                  num_filters,
                  num_channels=None,
                  stride=1,
                  padding=0,
                  filter_size_y=None,
                  stride_y=None,
                  padding_y=None):
Z
zhangjinchao01 已提交
3363 3364 3365 3366 3367 3368 3369 3370 3371 3372
    """
    Different from img_conv_layer, conv_op is an Operator, which can be used
    in mixed_layer. And conv_op takes two inputs to perform convolution.
    The first input is the image and the second is filter kernel. It only
    support GPU mode.

    The example usage is:

    .. code-block:: python

3373 3374
       op = conv_operator(img=input1,
                          filter=input2,
3375
                          filter_size=3,
Z
zhangjinchao01 已提交
3376 3377 3378
                          num_filters=64,
                          num_channels=64)

3379 3380 3381 3382
    :param img: input image
    :type img: LayerOutput
    :param filter: input filter
    :type filter: LayerOutput
Z
zhangjinchao01 已提交
3383 3384
    :param filter_size: The x dimension of a filter kernel.
    :type filter_size: int
C
caoying03 已提交
3385 3386 3387
    :param filter_size_y: The y dimension of a filter kernel. Since
                        PaddlePaddle now supports rectangular filters,
                        the filter's shape can be (filter_size, filter_size_y).
Z
zhangjinchao01 已提交
3388
    :type filter_size_y: int
3389 3390
    :param num_filters: channel of output data.
    :type num_filters: int
3391 3392
    :param num_channels: channel of input data.
    :type num_channels: int
Z
zhangjinchao01 已提交
3393
    :param stride: The x dimension of the stride.
L
luotao02 已提交
3394
    :type stride: int
Z
zhangjinchao01 已提交
3395
    :param stride_y: The y dimension of the stride.
L
luotao02 已提交
3396
    :type stride_y: int
Z
zhangjinchao01 已提交
3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409
    :param padding: The x dimension of padding.
    :type padding: int
    :param padding_y: The y dimension of padding.
    :type padding_y: int
    :return: A ConvOperator Object.
    :rtype: ConvOperator
    """
    if filter_size_y is None:
        filter_size_y = filter_size
    if stride_y is None:
        stride_y = stride
    if padding_y is None:
        padding_y = padding
3410

3411 3412
    if num_channels is None:
        num_channels = img.num_filters
3413 3414 3415

    assert isinstance(filter, LayerOutput)
    if filter.size is not None:
3416
        filter.size = filter_size * filter_size_y * num_filters * num_channels
3417

Q
qijun 已提交
3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429
    op = ConvOperator(
        input_layer_names=[img.name, filter.name],
        num_filters=num_filters,
        conv_conf=Conv(
            filter_size=filter_size,
            padding=padding,
            stride=stride,
            channels=num_channels,
            filter_size_y=filter_size_y,
            padding_y=padding_y,
            stride_y=stride_y,
            groups=1))
3430
    op.origin = [img, filter]
Z
zhangjinchao01 已提交
3431 3432
    return op

Q
qijun 已提交
3433

3434
@wrap_param_attr_default()
Q
qijun 已提交
3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445
def conv_projection(input,
                    filter_size,
                    num_filters,
                    num_channels=None,
                    stride=1,
                    padding=0,
                    filter_size_y=None,
                    stride_y=None,
                    padding_y=None,
                    groups=1,
                    param_attr=None):
3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472
    """
    ConvProjection with a layer as input.
    It performs element-wise multiplication with weight.

    Different from img_conv_layer and conv_op, conv_projection is an Projection,
    which can be used in mixed_layer and conat_layer. It use cudnn to implement
    conv and only support GPU mode.

    The example usage is:

    .. code-block:: python

       proj = conv_projection(img=input1,
                              filter_size=3,
                              num_filters=64,
                              num_channels=64)

    :param input: input layer
    :type input: LayerOutput
    :param filter_size: The x dimension of a filter kernel.
    :type filter_size: int
    :param filter_size_y: The y dimension of a filter kernel. Since
                          PaddlePaddle now supports rectangular filters,
                          the filter's shape can be (filter_size, filter_size_y).
    :type filter_size_y: int
    :param num_filters: channel of output data.
    :type num_filters: int
3473 3474
    :param num_channels: channel of input data.
    :type num_channels: int
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
    :param stride: The x dimension of the stride.
    :type stride: int
    :param stride_y: The y dimension of the stride.
    :type stride_y: int
    :param padding: The x dimension of padding.
    :type padding: int
    :param padding_y: The y dimension of padding.
    :type padding_y: int
    :param groups: The group number.
    :type groups: int
    :param param_attr: Convolution param attribute. None means default attribute
    :type param_attr: ParameterAttribute
    :return: A DotMulProjection Object.
    :rtype: DotMulProjection
    """
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters

    if filter_size_y is None:
        if isinstance(filter_size, collections.Sequence):
            assert len(filter_size) == 2
            filter_size, filter_size_y = filter_size
        else:
            filter_size_y = filter_size

    if stride_y is None:
        if isinstance(stride, collections.Sequence):
            assert len(stride) == 2
            stride, stride_y = stride
        else:
            stride_y = stride

    if padding_y is None:
        if isinstance(padding, collections.Sequence):
            assert len(padding) == 2
            padding, padding_y = padding
        else:
            padding_y = padding

    if param_attr.attr.get('initial_smart'):
        # special initial for conv layers.
Q
qijun 已提交
3517
        init_w = (2.0 / (filter_size**2 * num_channels))**0.5
3518 3519 3520 3521 3522
        param_attr.attr["initial_mean"] = 0.0
        param_attr.attr["initial_std"] = init_w
        param_attr.attr["initial_strategy"] = 0
        param_attr.attr["initial_smart"] = False

Q
qijun 已提交
3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535
    proj = ConvProjection(
        input_layer_name=input.name,
        num_filters=num_filters,
        conv_conf=Conv(
            filter_size=filter_size,
            padding=padding,
            stride=stride,
            channels=num_channels,
            filter_size_y=filter_size_y,
            padding_y=padding_y,
            stride_y=stride_y,
            groups=groups),
        **param_attr.attr)
3536 3537 3538 3539

    proj.origin = input
    return proj

Z
zhangjinchao01 已提交
3540 3541

@wrap_name_default()
L
luotao1 已提交
3542 3543
@layer_support()
def conv_shift_layer(a, b, name=None, layer_attr=None):
Z
zhangjinchao01 已提交
3544 3545 3546 3547 3548 3549 3550 3551 3552 3553 3554
    """
    This layer performs cyclic convolution for two input. For example:
      - a[in]: contains M elements.
      - b[in]: contains N elements (N should be odd).
      - c[out]: contains M elements.

    .. math::

        c[i] = \sum_{j=-(N-1)/2}^{(N-1)/2}a_{i+j} * b_{j}

    In this formular:
3555 3556 3557 3558
     - a's index is computed modulo M. When it is negative, then get item from
       the right side (which is the end of array) to the left.
     - b's index is computed modulo N. When it is negative, then get item from
       the right size (which is the end of array) to the left.
Z
zhangjinchao01 已提交
3559 3560 3561 3562 3563

    The example usage is:

    .. code-block:: python

3564
       conv_shift = conv_shift_layer(input=[layer1, layer2])
Z
zhangjinchao01 已提交
3565 3566 3567

    :param name: layer name
    :type name: basestring
3568 3569 3570 3571
    :param a: Input layer a.
    :type a: LayerOutput
    :param b: input layer b
    :type b: LayerOutput
L
luotao1 已提交
3572 3573
    :param layer_attr: layer's extra attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
3574
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3575 3576
    :rtype: LayerOutput
    """
3577 3578
    assert isinstance(a, LayerOutput) and isinstance(b, LayerOutput)
    assert b.size is None or b.size % 2 == 1  # size of b must be odd.
Z
zhangjinchao01 已提交
3579 3580 3581
    Layer(
        name=name,
        type=LayerType.CONV_SHIFT_LAYER,
3582
        inputs=[a.name, b.name],
Q
qijun 已提交
3583
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
3584

Q
qijun 已提交
3585 3586
    return LayerOutput(
        name, LayerType.CONV_SHIFT_LAYER, parents=[a, b], size=a.size)
Z
zhangjinchao01 已提交
3587 3588 3589 3590 3591


@wrap_name_default()
@wrap_param_attr_default()
@wrap_bias_attr_default()
3592
@wrap_act_default(act=LinearActivation())
Z
zhangjinchao01 已提交
3593
@layer_support(ERROR_CLIPPING, DROPOUT)
Q
qijun 已提交
3594 3595 3596 3597 3598 3599 3600 3601
def tensor_layer(a,
                 b,
                 size,
                 act=None,
                 name=None,
                 param_attr=None,
                 bias_attr=None,
                 layer_attr=None):
Z
zhangjinchao01 已提交
3602 3603 3604 3605 3606
    """
    This layer performs tensor operation for two input.
    For example, each sample:

    .. math::
3607
       y_{i} = a * W_{i} * {b^\mathrm{T}}, i=0,1,...,K-1
Z
zhangjinchao01 已提交
3608 3609

    In this formular:
3610 3611
      - :math:`a`: the first input contains M elements.
      - :math:`b`: the second input contains N elements.
Z
zhangjinchao01 已提交
3612 3613
      - :math:`y_{i}`: the i-th element of y.
      - :math:`W_{i}`: the i-th learned weight, shape if [M, N]
3614
      - :math:`b^\mathrm{T}`: the transpose of :math:`b_{2}`.
Z
zhangjinchao01 已提交
3615 3616 3617 3618 3619

    The simple usage is:

    .. code-block:: python

3620
       tensor = tensor_layer(a=layer1, b=layer2, size=1000)
Z
zhangjinchao01 已提交
3621 3622 3623

    :param name: layer name
    :type name: basestring
3624 3625 3626 3627
    :param a: Input layer a.
    :type a: LayerOutput
    :param b: input layer b.
    :type b: LayerOutput
Z
zhangjinchao01 已提交
3628
    :param size: the layer dimension.
L
luotao02 已提交
3629
    :type size: int.
Z
zhangjinchao01 已提交
3630 3631 3632
    :param act: Activation Type. Default is tanh.
    :type act: BaseActivation
    :param param_attr: The Parameter Attribute.
3633
    :type param_attr: ParameterAttribute
Z
zhangjinchao01 已提交
3634 3635 3636 3637 3638 3639
    :param bias_attr: The Bias Attribute. If no bias, then pass False or
                      something not type of ParameterAttribute. None will get a
                      default Bias.
    :type bias_attr: ParameterAttribute|None|Any
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
3640
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3641 3642
    :rtype: LayerOutput
    """
3643
    assert isinstance(a, LayerOutput) and isinstance(b, LayerOutput)
Z
zhangjinchao01 已提交
3644 3645 3646 3647 3648 3649
    Layer(
        name=name,
        size=size,
        type=LayerType.TENSOR_LAYER,
        active_type=act.name,
        bias=ParamAttr.to_bias(bias_attr),
Q
qijun 已提交
3650 3651 3652 3653
        inputs=[Input(a.name, **param_attr.attr), Input(b.name)],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.TENSOR_LAYER, parents=[a, b], activation=act, size=size)
Z
zhangjinchao01 已提交
3654 3655 3656 3657 3658 3659


@wrap_name_default()
@wrap_param_attr_default()
@wrap_bias_attr_default()
@wrap_act_default()
L
luotao1 已提交
3660
@layer_support()
Q
qijun 已提交
3661 3662 3663 3664 3665
def selective_fc_layer(input,
                       select,
                       size,
                       act=None,
                       name=None,
Z
zhangjinchao01 已提交
3666 3667 3668
                       pass_generation=False,
                       has_selected_colums=True,
                       mul_ratio=0.02,
Q
qijun 已提交
3669 3670 3671
                       param_attr=None,
                       bias_attr=None,
                       layer_attr=None):
Z
zhangjinchao01 已提交
3672 3673 3674 3675 3676 3677 3678 3679 3680 3681
    """
    Selectived fully connected layer. Different from fc_layer, the output
    of this layer maybe sparse. It requires an additional input to indicate
    several selected columns for output. If the selected columns is not
    specified, selective_fc_layer acts exactly like fc_layer.

    The simple usage is:

    .. code-block:: python

3682
       sel_fc = selective_fc_layer(input=input, size=128, act=TanhActivation())
Z
zhangjinchao01 已提交
3683 3684 3685 3686 3687

    :param name: The Layer Name.
    :type name: basestring
    :param input: The input layer.
    :type input: LayerOutput|list|tuple
3688 3689 3690
    :param select: The select layer. The output of select layer should be a
                   sparse binary matrix, and treat as the mask of selective fc.
    :type select: LayerOutput
Z
zhangjinchao01 已提交
3691 3692 3693 3694 3695 3696 3697 3698 3699 3700 3701 3702
    :param size: The layer dimension.
    :type size: int
    :param act: Activation Type. Default is tanh.
    :type act: BaseActivation
    :param param_attr: The Parameter Attribute.
    :type param_attr: ParameterAttribute
    :param bias_attr: The Bias Attribute. If no bias, then pass False or
                      something not type of ParameterAttribute. None will get a
                      default Bias.
    :type bias_attr: ParameterAttribute|None|Any
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
3703
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3704 3705 3706 3707
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
3708
        assert not isinstance(param_attr, collections.Sequence)
Z
zhangjinchao01 已提交
3709 3710
        param_attr = [param_attr]
    else:
3711
        if isinstance(param_attr, collections.Sequence):
Z
zhangjinchao01 已提交
3712 3713 3714 3715
            assert len(input) == len(param_attr)
        else:
            param_attr = [copy.deepcopy(param_attr) for _ in range(len(input))]

3716 3717 3718 3719
    assert isinstance(input, collections.Sequence)
    assert isinstance(select, LayerOutput)
    if select.size is not None:
        assert select.size == size
Z
zhangjinchao01 已提交
3720
    Layer(
Q
qijun 已提交
3721 3722 3723
        inputs=[
            Input(ipt.name, **attr.attr) for ipt, attr in zip(input, param_attr)
        ] + [select.name],
Z
zhangjinchao01 已提交
3724 3725 3726
        name=name,
        type=LayerType.SEL_FC_LAYER,
        size=size,
3727
        bias=ParameterAttribute.to_bias(bias_attr),
Z
zhangjinchao01 已提交
3728 3729 3730 3731
        active_type=act.name,
        selective_fc_pass_generation=pass_generation,
        has_selected_colums=has_selected_colums,
        selective_fc_full_mul_ratio=mul_ratio,
Q
qijun 已提交
3732 3733 3734 3735 3736 3737 3738
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.SEL_FC_LAYER,
        list(input) + [select],
        activation=act,
        size=size)
Z
zhangjinchao01 已提交
3739 3740 3741


@wrap_name_default()
L
luotao1 已提交
3742 3743
@layer_support()
def sampling_id_layer(input, name=None, layer_attr=None):
Z
zhangjinchao01 已提交
3744 3745 3746 3747 3748 3749 3750 3751 3752 3753 3754 3755 3756 3757
    """
    A layer for sampling id from multinomial distribution from the input layer.
    Sampling one id for one sample.

    The simple usage is:

    .. code-block:: python

       samping_id = sampling_id_layer(input=input)

    :param input: The input layer.
    :type input: LayerOutput
    :param name: The Layer Name.
    :type name: basestring
L
luotao1 已提交
3758 3759
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
3760
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3761 3762
    :rtype: LayerOutput
    """
X
xuwei06 已提交
3763
    l = Layer(
Z
zhangjinchao01 已提交
3764 3765 3766
        name=name,
        type=LayerType.SAMPLING_ID_LAYER,
        inputs=[Input(input.name)],
Q
qijun 已提交
3767 3768 3769
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.SAMPLING_ID_LAYER, input, size=l.config.size)
Z
zhangjinchao01 已提交
3770 3771 3772


@wrap_name_default()
L
luotao1 已提交
3773
@layer_support()
Q
qijun 已提交
3774 3775 3776 3777
def slope_intercept_layer(input,
                          name=None,
                          slope=1.0,
                          intercept=0.0,
L
luotao1 已提交
3778
                          layer_attr=None):
Z
zhangjinchao01 已提交
3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798 3799
    """
    This layer for applying a slope and an intercept to the input
    element-wise. There is no activation and weight.

    ..  math::
        y = slope * x + intercept

    The simple usage is:

    .. code-block:: python

       scale = slope_intercept_layer(input=input, slope=-1.0, intercept=1.0)

    :param input: The input layer.
    :type input: LayerOutput
    :param name: The Layer Name.
    :type name: basestring
    :param slope: the scale factor.
    :type slope: float.
    :param intercept: the offset.
    :type intercept: float.
L
luotao1 已提交
3800 3801
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
3802
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3803 3804 3805 3806 3807 3808 3809 3810
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.SLOPE_INTERCEPT_LAYER,
        slope=slope,
        intercept=intercept,
        inputs=[Input(input.name)],
Q
qijun 已提交
3811 3812 3813
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.SLOPE_INTERCEPT_LAYER, input, size=input.size)
Z
zhangjinchao01 已提交
3814 3815 3816


@wrap_name_default()
L
luotao1 已提交
3817
@layer_support()
Q
qijun 已提交
3818
def linear_comb_layer(weights, vectors, size=None, name=None, layer_attr=None):
Z
zhangjinchao01 已提交
3819
    """
3820 3821 3822 3823
    A layer for weighted sum of vectors takes two inputs.
      - Input: size of weights is M
               size of vectors is M*N
      - Output: a vector of size=N
Z
zhangjinchao01 已提交
3824 3825 3826

    .. math::

3827
       z(i) = \sum_{j=0}^{M-1} x(j) y(i+Nj)
3828

3829 3830 3831 3832 3833
    where :math:`0 \le i \le N-1`

    Or in the matrix notation:

    .. math::
Z
zhangjinchao01 已提交
3834

3835
       z = x^\mathrm{T} Y
Z
zhangjinchao01 已提交
3836 3837

    In this formular:
3838 3839 3840 3841 3842 3843
      - :math:`x`: weights
      - :math:`y`: vectors.
      - :math:`z`: the output.

    Note that the above computation is for one sample. Multiple samples are
    processed in one batch.
Z
zhangjinchao01 已提交
3844 3845 3846 3847 3848

    The simple usage is:

    .. code-block:: python

3849
       linear_comb = linear_comb_layer(weights=weight, vectors=vectors,
Z
zhangjinchao01 已提交
3850 3851
                                       size=elem_dim)

3852 3853 3854 3855
    :param weights: The weight layer.
    :type weights: LayerOutput
    :param vectors: The vector layer.
    :type vectors: LayerOutput
Z
zhangjinchao01 已提交
3856 3857 3858 3859
    :param size: the dimension of this layer.
    :type size: int
    :param name: The Layer Name.
    :type name: basestring
L
luotao1 已提交
3860 3861
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
3862
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3863 3864
    :rtype: LayerOutput
    """
3865 3866 3867 3868
    assert isinstance(weights, LayerOutput) and isinstance(vectors, LayerOutput)
    if vectors.size is not None and weights.size is not None:
        assert vectors.size % weights.size == 0
        if size is None:
Q
qijun 已提交
3869
            size = vectors.size / weights.size
3870 3871
        else:
            assert size == vectors.size / weights.size
Z
zhangjinchao01 已提交
3872 3873
    Layer(
        name=name,
3874
        type=LayerType.LINEAR_COMBINATION_LAYER,
Z
zhangjinchao01 已提交
3875
        size=size,
3876
        inputs=[Input(weights.name), Input(vectors.name)],
Q
qijun 已提交
3877 3878 3879
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.LINEAR_COMBINATION_LAYER, [weights, vectors], size=size)
3880

3881

3882
convex_comb_layer = linear_comb_layer
Z
zhangjinchao01 已提交
3883

3884

Z
zhangjinchao01 已提交
3885
@wrap_name_default()
L
luotao1 已提交
3886
@layer_support()
Z
zhangjinchao01 已提交
3887 3888 3889 3890 3891 3892 3893
def block_expand_layer(input,
                       block_x=0,
                       block_y=0,
                       stride_x=0,
                       stride_y=0,
                       padding_x=0,
                       padding_y=0,
3894
                       num_channels=None,
L
luotao1 已提交
3895 3896
                       name=None,
                       layer_attr=None):
Z
zhangjinchao01 已提交
3897 3898
    """
    Expand feature map to minibatch matrix.
3899
       - matrix width is: block_y * block_x * num_channels
L
luotao02 已提交
3900
       - matirx height is: outputH * outputW
Z
zhangjinchao01 已提交
3901 3902 3903 3904 3905 3906 3907 3908 3909 3910

    .. math::

       outputH = 1 + (2 * padding_y + imgSizeH - block_y + stride_y - 1) / stride_y

       outputW = 1 + (2 * padding_x + imgSizeW - block_x + stride_x - 1) / stride_x

    The expand method is the same with ExpandConvLayer, but saved the transposed
    value. After expanding, output.sequenceStartPositions will store timeline.
    The number of time steps are outputH * outputW and the dimension of each
3911
    time step is block_y * block_x * num_channels. This layer can be used after
Z
zhangjinchao01 已提交
3912 3913
    convolution neural network, and before recurrent neural network.

3914 3915 3916 3917 3918
    The simple usage is:

    .. code-block:: python

       block_expand = block_expand_layer(input,
3919
                                         num_channels=128,
3920 3921 3922 3923 3924
                                         stride_x=1,
                                         stride_y=1,
                                         block_x=1,
                                         block_x=3)

Z
zhangjinchao01 已提交
3925 3926
    :param input: The input layer.
    :type input: LayerOutput
3927 3928
    :param num_channels: The channel number of input layer.
    :type num_channels: int|None
Z
zhangjinchao01 已提交
3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939 3940 3941 3942
    :param block_x: The width of sub block.
    :type block_x: int
    :param block_y: The width of sub block.
    :type block_y: int
    :param stride_x: The stride size in horizontal direction.
    :type stride_x: int
    :param stride_y: The stride size in vertical direction.
    :type stride_y: int
    :param padding_x: The padding size in horizontal direction.
    :type padding_x: int
    :param padding_y: The padding size in vertical direction.
    :type padding_y: int
    :param name: The name of this layer, which can not specify.
    :type name: None|basestring.
L
luotao1 已提交
3943 3944
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
3945
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3946 3947
    :rtype: LayerOutput
    """
3948 3949 3950
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters
Q
qijun 已提交
3951 3952 3953 3954 3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967
    l = Layer(
        name=name,
        inputs=Input(
            input.name,
            block_expand=BlockExpand(
                channels=num_channels,
                block_x=block_x,
                block_y=block_y,
                stride_x=stride_x,
                stride_y=stride_y,
                padding_x=padding_x,
                padding_y=padding_y)),
        type=LayerType.BLOCK_EXPAND,
        **ExtraLayerAttribute.to_kwargs(layer_attr))

    return LayerOutput(
        name, LayerType.BLOCK_EXPAND, parents=[input], size=l.config.size)
Z
zhangjinchao01 已提交
3968 3969


3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982 3983
@wrap_name_default()
@layer_support()
def maxout_layer(input,
                 groups,
                 num_channels=None,
                 size_x=None,
                 size_y=None,
                 name=None,
                 layer_attr=None):
    """
    A layer to do max out on conv layer output.
      - Input: output of a conv layer.
      - Output: feature map size same as input. Channel is (input channel) / groups.

3984
    So groups should be larger than 1, and the num of channels should be able
3985 3986
    to devided by groups.

3987
    Please refer to Paper:
3988 3989 3990 3991
      - Maxout Networks: http://www.jmlr.org/proceedings/papers/v28/goodfellow13.pdf
      - Multi-digit Number Recognition from Street View \
        Imagery using Deep Convolutional Neural Networks: \
        https://arxiv.org/pdf/1312.6082v4.pdf
3992

3993 3994 3995 3996 3997 3998 3999 4000 4001 4002 4003 4004 4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027
    The simple usage is:

    .. code-block:: python

       maxout = maxout_layer(input,
                             num_channels=128,
                             groups=4)

    :param input: The input layer.
    :type input: LayerOutput
    :param num_channels: The channel number of input layer. If None will be set
                     automatically from previous output.
    :type num_channels: int|None
    :param groups: The group number of input layer.
    :type groups: int
    :param size_x: conv output width. If None will be set
                   automatically from previous output.
    :type size_x: int|None
    :param size_y: conv output height. If None will be set
                   automatically from previous output.
    :type size_y: int|None
    :param name: The name of this layer, which can not specify.
    :type name: None|basestring.
    :param layer_attr: Extra Layer attribute.
    :type layer_attr: ExtraLayerAttribute
    :return: LayerOutput object.
    :rtype: LayerOutput
    """
    assert input.layer_type == LayerType.CONV_LAYER
    assert isinstance(input.activation, LinearActivation)
    assert groups > 1
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters
    assert num_channels % groups == 0
Q
qijun 已提交
4028 4029 4030 4031 4032 4033 4034 4035 4036
    l = Layer(
        name=name,
        inputs=Input(
            input.name, maxout=MaxOut(
                channels=num_channels, groups=groups)),
        type=LayerType.MAXOUT,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.MAXOUT, parents=[input], size=l.config.size)
4037 4038


Z
zhangjinchao01 已提交
4039
@wrap_name_default()
L
luotao1 已提交
4040
@layer_support()
Q
qijun 已提交
4041 4042 4043 4044 4045
def ctc_layer(input,
              label,
              size=None,
              name=None,
              norm_by_times=False,
L
luotao1 已提交
4046
              layer_attr=None):
Z
zhangjinchao01 已提交
4047 4048 4049 4050 4051
    """
    Connectionist Temporal Classification (CTC) is designed for temporal
    classication task. That is, for sequence labeling problems where the
    alignment between the inputs and the target labels is unknown.

4052 4053
    More details can be found by referring to `Connectionist Temporal
    Classification: Labelling Unsegmented Sequence Data with Recurrent
4054 4055
    Neural Networks <http://machinelearning.wustl.edu/mlpapers/paper_files/
    icml2006_GravesFGS06.pdf>`_
4056 4057 4058 4059 4060 4061 4062 4063

    Note:
        Considering the 'blank' label needed by CTC, you need to use
        (num_classes + 1) as the input size. num_classes is the category number.
        And the 'blank' is the last category index. So the size of 'input' layer, such as
        fc_layer with softmax activation, should be num_classes + 1. The size of ctc_layer
        should also be num_classes + 1.

Z
zhangjinchao01 已提交
4064 4065 4066 4067 4068 4069 4070 4071 4072
    The simple usage:

    .. code-block:: python

      ctc = ctc_layer(input=input,
                      label=label,
                      size=9055,
                      norm_by_times=True)

4073
    :param input: The input layer.
Z
zhangjinchao01 已提交
4074 4075 4076
    :type input: LayerOutput
    :param label: The data layer of label with variable length.
    :type label: LayerOutput
4077
    :param size: category numbers + 1.
Z
zhangjinchao01 已提交
4078
    :type size: int
4079 4080
    :param name: The name of this layer
    :type name: basestring|None
Z
zhangjinchao01 已提交
4081 4082
    :param norm_by_times: Whether to normalization by times. False by default.
    :type norm_by_times: bool
L
luotao1 已提交
4083 4084
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4085
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4086 4087 4088 4089
    :rtype: LayerOutput
    """
    assert isinstance(input, LayerOutput)
    assert isinstance(label, LayerOutput)
4090 4091 4092 4093 4094
    if label.size is not None:
        if size is not None:
            assert size == label.size + 1
        else:
            size = label.size + 1
Z
zhangjinchao01 已提交
4095
    Layer(
4096 4097 4098 4099
        name=name,
        type=LayerType.CTC_LAYER,
        size=size,
        norm_by_times=norm_by_times,
L
luotao1 已提交
4100
        inputs=[input.name, label.name],
Q
qijun 已提交
4101
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
4102 4103
    return LayerOutput(name, LayerType.CTC_LAYER, [input, label], size=size)

4104

4105 4106 4107 4108 4109 4110 4111 4112 4113 4114 4115 4116 4117 4118 4119 4120 4121 4122 4123 4124 4125 4126 4127 4128 4129 4130 4131 4132 4133 4134 4135 4136 4137 4138 4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167 4168 4169 4170 4171 4172 4173 4174 4175 4176 4177 4178 4179 4180 4181
@wrap_name_default()
@layer_support()
def warp_ctc_layer(input,
                   label,
                   size=None,
                   name=None,
                   blank=0,
                   norm_by_times=False,
                   layer_attr=None):
    """
    A layer intergrating the open-source `warp-ctc
    <https://github.com/baidu-research/warp-ctc>` library, which is used in
    `Deep Speech 2: End-toEnd Speech Recognition in English and Mandarin
    <https://arxiv.org/pdf/1512.02595v1.pdf>`, to compute Connectionist Temporal
    Classification (CTC) loss.

    More details of CTC can be found by referring to `Connectionist Temporal
    Classification: Labelling Unsegmented Sequence Data with Recurrent
    Neural Networks <http://machinelearning.wustl.edu/mlpapers/paper_files/
    icml2006_GravesFGS06.pdf>`_

    Note:
        - Let num_classes represent the category number. Considering the 'blank'
          label needed by CTC, you need to use (num_classes + 1) as the input size.
          Thus, the size of both warp_ctc_layer and 'input' layer should be set to
          num_classes + 1.
        - You can set 'blank' to [0, num_classes - 1], which should be consistent
          as that used in your labels.
        - As a native 'softmax' activation is interated to the warp-ctc library,
         'linear' activation is expected instead in the 'input' layer.

    The simple usage:

    .. code-block:: python

      ctc = warp_ctc_layer(input=input,
                           label=label,
                           size=1001,
                           blank=1000,
                           norm_by_times=False)

    :param input: The input layer.
    :type input: LayerOutput
    :param label: The data layer of label with variable length.
    :type label: LayerOutput
    :param size: category numbers + 1.
    :type size: int
    :param name: The name of this layer, which can not specify.
    :type name: basestring|None
    :param blank: the 'blank' label used in ctc
    :type blank: int
    :param norm_by_times: Whether to normalization by times. False by default.
    :type norm_by_times: bool
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
    :return: LayerOutput object.
    :rtype: LayerOutput
    """
    assert isinstance(input, LayerOutput)
    assert isinstance(label, LayerOutput)
    if label.size is not None:
        if size is not None:
            assert size == label.size + 1
        else:
            size = label.size + 1
    Layer(
        name=name,
        type=LayerType.WARP_CTC_LAYER,
        size=size,
        blank=blank,
        norm_by_times=norm_by_times,
        inputs=[input.name, label.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.WARP_CTC_LAYER, parents=[input, label], size=size)


Z
zhangjinchao01 已提交
4182
@wrap_name_default()
4183
@wrap_param_attr_default()
L
luotao1 已提交
4184
@layer_support()
Q
qijun 已提交
4185 4186 4187 4188 4189 4190
def crf_layer(input,
              label,
              size=None,
              weight=None,
              param_attr=None,
              name=None,
L
luotao1 已提交
4191
              layer_attr=None):
Z
zhangjinchao01 已提交
4192 4193 4194 4195 4196 4197 4198 4199 4200 4201 4202 4203 4204 4205 4206
    """
    A layer for calculating the cost of sequential conditional random
    field model.

    The simple usage:

    .. code-block:: python

      crf = crf_layer(input=input,
                      label=label,
                      size=label_dim)

    :param input: The first input layer is the feature.
    :type input: LayerOutput
    :param label: The second input layer is label.
4207
    :type label: LayerOutput
Z
zhangjinchao01 已提交
4208 4209 4210 4211 4212 4213 4214 4215 4216
    :param size: The category number.
    :type size: int
    :param weight: The third layer is "weight" of each sample, which is an
                  optional argument.
    :type weight: LayerOutput
    :param param_attr: Parameter attribute. None means default attribute
    :type param_attr: ParameterAttribute
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring
L
luotao1 已提交
4217 4218
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4219
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4220 4221 4222 4223 4224
    :rtype: LayerOutput
    """
    assert isinstance(input, LayerOutput)
    assert isinstance(label, LayerOutput)
    assert weight is None or isinstance(weight, LayerOutput)
4225 4226 4227 4228 4229 4230
    if input.size is not None and label.size is not None:
        assert input.size == label.size
        if size is None:
            size = input.size
        else:
            assert size == input.size
Z
zhangjinchao01 已提交
4231

Q
qijun 已提交
4232
    ipts = [Input(input.name, **param_attr.attr), Input(label.name)]
Z
zhangjinchao01 已提交
4233 4234 4235 4236
    if weight is not None:
        ipts.append(Input(weight.name))

    Layer(
4237 4238 4239 4240
        name=name,
        type=LayerType.CRF_LAYER,
        size=size,
        inputs=ipts,
Q
qijun 已提交
4241
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
4242 4243 4244
    parents = [input, label]
    if weight is not None:
        parents.append(weight)
X
xuwei06 已提交
4245 4246 4247 4248
    # The size for LayerOutput means the dimension of the output.
    # It's different from the meaning of crf layer, which is the number of
    # classes.
    return LayerOutput(name, LayerType.CRF_LAYER, parents, size=1)
Z
zhangjinchao01 已提交
4249

4250

Z
zhangjinchao01 已提交
4251
@wrap_name_default()
4252
@wrap_param_attr_default()
L
luotao1 已提交
4253
@layer_support()
Q
qijun 已提交
4254 4255 4256 4257 4258
def crf_decoding_layer(input,
                       size,
                       label=None,
                       param_attr=None,
                       name=None,
L
luotao1 已提交
4259
                       layer_attr=None):
Z
zhangjinchao01 已提交
4260 4261 4262 4263 4264 4265 4266 4267 4268 4269 4270 4271 4272 4273 4274 4275 4276
    """
    A layer for calculating the decoding sequence of sequential conditional
    random field model. The decoding sequence is stored in output.ids.
    If a second input is provided, it is treated as the ground-truth label, and
    this layer will also calculate error. output.value[i] is 1 for incorrect
    decoding or 0 for correct decoding.

    :param input: The first input layer.
    :type input: LayerOutput
    :param size: size of this layer.
    :type size: int
    :param label: None or ground-truth label.
    :type label: LayerOutput or None
    :param param_attr: Parameter attribute. None means default attribute
    :type param_attr: ParameterAttribute
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring
L
luotao1 已提交
4277 4278
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4279
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4280 4281 4282 4283 4284 4285
    :rtype: LayerOutput
    """

    assert isinstance(input, LayerOutput)
    assert label is None or isinstance(label, LayerOutput)

4286
    ipts = [Input(input.name, **param_attr.attr)]
Z
zhangjinchao01 已提交
4287 4288 4289 4290
    if label is not None:
        ipts.append(Input(label.name))

    Layer(
4291 4292 4293 4294
        name=name,
        type=LayerType.CRF_DECODING_LAYER,
        size=size,
        inputs=ipts,
Q
qijun 已提交
4295
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
4296 4297 4298
    parents = [input]
    if label is not None:
        parents.append(label)
X
xuwei06 已提交
4299 4300 4301 4302
    # The size for LayerOutput means the dimension of the output.
    # It's different from the meaning of crf layer, which is the number of
    # classes.
    return LayerOutput(name, LayerType.CRF_DECODING_LAYER, parents, size=1)
Z
zhangjinchao01 已提交
4303

Q
qijun 已提交
4304

4305 4306 4307
@wrap_bias_attr_default(has_bias=True)
@wrap_name_default()
@layer_support()
Q
qijun 已提交
4308 4309 4310 4311 4312 4313 4314 4315 4316
def nce_layer(input,
              label,
              num_classes,
              weight=None,
              num_neg_samples=10,
              neg_distribution=None,
              name=None,
              bias_attr=None,
              layer_attr=None):
4317 4318 4319 4320 4321 4322 4323 4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335 4336 4337
    """
    Noise-contrastive estimation.
    Implements the method in the following paper:
    A fast and simple algorithm for training neural probabilistic language models.

    The example usage is:

    .. code-block:: python

       cost = nce_layer(input=layer1, label=layer2, weight=layer3,
                        num_classes=3, neg_distribution=[0.1,0.3,0.6])

    :param name: layer name
    :type name: basestring
    :param input: input layers. It could be a LayerOutput of list/tuple of LayerOutput.
    :type input: LayerOutput|list|tuple|collections.Sequence
    :param label: label layer
    :type label: LayerOutput
    :param weight: weight layer, can be None(default)
    :type weight: LayerOutput
    :param num_classes: number of classes.
4338
    :type num_classes: int
4339
    :param num_neg_samples: number of negative samples. Default is 10.
4340
    :type num_neg_samples: int
4341 4342 4343 4344 4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360
    :param neg_distribution: The distribution for generating the random negative labels.
                             A uniform distribution will be used if not provided.
                             If not None, its length must be equal to num_classes.
    :type neg_distribution: list|tuple|collections.Sequence|None
    :param bias_attr: Bias parameter attribute. True if no bias.
    :type bias_attr: ParameterAttribute|None|False
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
    :return: layer name.
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
    assert isinstance(input, collections.Sequence)
    assert isinstance(label, LayerOutput)
    assert label.layer_type == LayerType.DATA
    if neg_distribution is not None:
        assert isinstance(neg_distribution, collections.Sequence)
        assert len(neg_distribution) == num_classes
        assert sum(neg_distribution) == 1
4361

4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376
    ipts_for_layer = []
    parents = []
    for each_input in input:
        assert isinstance(each_input, LayerOutput)
        ipts_for_layer.append(each_input.name)
        parents.append(each_input)
    ipts_for_layer.append(label.name)
    parents.append(label)

    if weight is not None:
        assert isinstance(weight, LayerOutput)
        assert weight.layer_type == LayerType.DATA
        ipts_for_layer.append(weight.name)
        parents.append(weight)

X
xuwei06 已提交
4377
    l = Layer(
4378 4379 4380 4381 4382 4383 4384
        name=name,
        type=LayerType.NCE_LAYER,
        num_classes=num_classes,
        neg_sampling_dist=neg_distribution,
        num_neg_samples=num_neg_samples,
        inputs=ipts_for_layer,
        bias=ParamAttr.to_bias(bias_attr),
Q
qijun 已提交
4385 4386 4387 4388
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.NCE_LAYER, parents=parents, size=l.config.size)

4389

Z
zhangjinchao01 已提交
4390 4391 4392
"""
following are cost Layers.
"""
4393 4394


Z
zhangjinchao01 已提交
4395
@wrap_name_default()
L
luotao1 已提交
4396
@layer_support()
Q
qijun 已提交
4397 4398 4399 4400 4401 4402 4403
def rank_cost(left,
              right,
              label,
              weight=None,
              name=None,
              coeff=1.0,
              layer_attr=None):
Z
zhangjinchao01 已提交
4404
    """
4405
    A cost Layer for learning to rank using gradient descent. Details can refer
4406 4407
    to `papers <http://research.microsoft.com/en-us/um/people/cburges/papers/
    ICML_ranking.pdf>`_.
Z
zhangjinchao01 已提交
4408 4409 4410 4411 4412
    This layer contains at least three inputs. The weight is an optional
    argument, which affects the cost.

    .. math::

L
luotao02 已提交
4413
       C_{i,j} & = -\\tilde{P_{ij}} * o_{i,j} + log(1 + e^{o_{i,j}})
Z
zhangjinchao01 已提交
4414

L
luotao02 已提交
4415
       o_{i,j} & =  o_i - o_j
Z
zhangjinchao01 已提交
4416

L
luotao02 已提交
4417
       \\tilde{P_{i,j}} & = \\{0, 0.5, 1\\} \ or \ \\{0, 1\\}
Z
zhangjinchao01 已提交
4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429 4430 4431 4432 4433 4434 4435 4436 4437 4438 4439 4440 4441 4442 4443 4444 4445 4446

    In this formula:
      - :math:`C_{i,j}` is the cross entropy cost.
      - :math:`\\tilde{P_{i,j}}` is the label. 1 means positive order
        and 0 means reverse order.
      - :math:`o_i` and :math:`o_j`: the left output and right output.
        Their dimension is one.

    The simple usage:

    .. code-block:: python

      cost = rank_cost(left=out_left,
                       right=out_right,
                       label=label)

    :param left: The first input, the size of this layer is 1.
    :type left: LayerOutput
    :param right: The right input, the size of this layer is 1.
    :type right: LayerOutput
    :param label: Label is 1 or 0, means positive order and reverse order.
    :type label: LayerOutput
    :param weight: The weight affects the cost, namely the scale of cost.
                   It is an optional argument.
    :type weight: LayerOutput
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring
    :param coeff: The coefficient affects the gradient in the backward.
    :type coeff: float
L
luotao1 已提交
4447 4448
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
4449
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4450 4451 4452 4453 4454 4455 4456 4457 4458 4459 4460 4461
    :rtype: LayerOutput
    """
    assert left.size == 1
    assert right.size == 1
    assert label.size == 1

    ipts = [left.name, right.name, label.name]
    parents = [left, right, label]
    if weight is not None:
        ipts.append(weight.name)
        parents.append(weight)

Q
qijun 已提交
4462 4463 4464 4465 4466 4467
    Layer(
        name=name,
        type=LayerType.RANK_COST,
        inputs=ipts,
        coeff=coeff,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
4468

X
xuwei06 已提交
4469
    return LayerOutput(name, LayerType.RANK_COST, parents=parents, size=1)
Z
zhangjinchao01 已提交
4470

4471

Z
zhangjinchao01 已提交
4472
@wrap_name_default()
L
luotao1 已提交
4473
@layer_support()
Q
qijun 已提交
4474 4475 4476 4477 4478 4479
def lambda_cost(input,
                score,
                name,
                NDCG_num=5,
                max_sort_size=-1,
                layer_attr=None):
Z
zhangjinchao01 已提交
4480 4481 4482 4483 4484 4485 4486 4487 4488 4489 4490 4491
    """
    lambdaCost for lambdaRank LTR approach.

    The simple usage:

    .. code-block:: python

      cost = lambda_cost(input=input,
                         score=score,
                         NDCG_num=8,
                         max_sort_size=-1)

4492
    :param input: Samples of the same query should be loaded as sequence.
Z
zhangjinchao01 已提交
4493 4494 4495 4496 4497 4498 4499 4500 4501 4502 4503
    :type input: LayerOutput
    :param score: The 2nd input. Score of each sample.
    :type input: LayerOutput
    :param NDCG_num: The size of NDCG (Normalized Discounted Cumulative Gain),
                     e.g., 5 for NDCG@5. It must be less than for equal to the
                     minimum size of lists.
    :type NDCG_num: int
    :param max_sort_size: The size of partial sorting in calculating gradient.
                          If max_sort_size = -1, then for each list, the
                          algorithm will sort the entire list to get gradient.
                          In other cases, max_sort_size must be greater than or
C
caoying03 已提交
4504 4505 4506
                          equal to NDCG_num. And if max_sort_size is greater
                          than the size of a list, the algorithm will sort the
                          entire list of get gradient.
Z
zhangjinchao01 已提交
4507 4508 4509
    :type max_sort_size: int
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring
L
luotao1 已提交
4510 4511
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
4512
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4513 4514
    :rtype: LayerOutput
    """
4515 4516 4517
    assert isinstance(input, LayerOutput) and isinstance(score, LayerOutput)
    if score.size is not None:
        assert score.size == 1
Q
qijun 已提交
4518 4519 4520 4521 4522 4523 4524
    Layer(
        name=name,
        type=LayerType.LAMBDA_COST,
        inputs=[input.name, score.name],
        NDCG_num=NDCG_num,
        max_sort_size=max_sort_size,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
4525

Q
qijun 已提交
4526 4527
    return LayerOutput(
        name, LayerType.LAMBDA_COST, parents=[input, score], size=1)
Z
zhangjinchao01 已提交
4528

4529

Z
zhangjinchao01 已提交
4530
@wrap_name_default()
L
luotao1 已提交
4531 4532
@layer_support()
def cross_entropy(input, label, name=None, coeff=1.0, layer_attr=None):
Z
zhangjinchao01 已提交
4533 4534 4535 4536 4537
    """
    A loss layer for multi class entropy.

    .. code-block:: python

X
xuwei06 已提交
4538
       cost = cross_entropy(input=input_layer,
L
Luo Tao 已提交
4539
                            label=label_layer)
Z
zhangjinchao01 已提交
4540 4541 4542 4543 4544 4545 4546 4547 4548

    :param input: The first input layer.
    :type input: LayerOutput.
    :param label: The input label.
    :type input: LayerOutput.
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring.
    :param coeff: The coefficient affects the gradient in the backward.
    :type coeff: float.
L
luotao1 已提交
4549 4550
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
4551
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4552 4553 4554
    :rtype: LayerOutput.
    """

Q
qijun 已提交
4555 4556 4557 4558 4559 4560 4561 4562
    Layer(
        name=name,
        type=LayerType.CROSS_ENTROPY,
        inputs=[input.name, label.name],
        coeff=coeff,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.CROSS_ENTROPY, parents=[input, label], size=1)
Z
zhangjinchao01 已提交
4563

4564

Z
zhangjinchao01 已提交
4565
@wrap_name_default()
L
luotao1 已提交
4566
@layer_support()
Q
qijun 已提交
4567 4568 4569 4570
def cross_entropy_with_selfnorm(input,
                                label,
                                name=None,
                                coeff=1.0,
L
luotao1 已提交
4571 4572
                                softmax_selfnorm_alpha=0.1,
                                layer_attr=None):
Z
zhangjinchao01 已提交
4573 4574 4575 4576 4577
    """
    A loss layer for multi class entropy with selfnorm.

    .. code-block:: python

X
xuwei06 已提交
4578
       cost = cross_entropy_with_selfnorm(input=input_layer,
L
Luo Tao 已提交
4579
                                          label=label_layer)
Z
zhangjinchao01 已提交
4580 4581 4582 4583 4584 4585 4586 4587 4588 4589 4590

    :param input: The first input layer.
    :type input: LayerOutput.
    :param label: The input label.
    :type input: LayerOutput.
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring.
    :param coeff: The coefficient affects the gradient in the backward.
    :type coeff: float.
    :param softmax_selfnorm_alpha: The scale factor affects the cost.
    :type softmax_selfnorm_alpha: float.
L
luotao1 已提交
4591 4592
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
4593
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4594 4595
    :rtype: LayerOutput.
    """
Q
qijun 已提交
4596 4597 4598 4599 4600 4601 4602
    Layer(
        name=name,
        type=LayerType.CROSS_ENTROPY_WITH_SELFNORM,
        inputs=[input.name, label.name],
        coeff=coeff,
        softmax_selfnorm_alpha=softmax_selfnorm_alpha,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
4603

Q
qijun 已提交
4604 4605 4606 4607 4608
    return LayerOutput(
        name,
        LayerType.CROSS_ENTROPY_WITH_SELFNORM,
        parents=[input, label],
        size=1)
Z
zhangjinchao01 已提交
4609

4610

X
xuwei06 已提交
4611 4612 4613 4614 4615 4616 4617 4618
@wrap_name_default()
@layer_support()
def sum_cost(input, name=None, layer_attr=None):
    """
    A loss layer which calculate the sum of the input as loss

    .. code-block:: python

L
Luo Tao 已提交
4619
       cost = sum_cost(input=input_layer)
X
xuwei06 已提交
4620 4621 4622 4623 4624 4625 4626 4627 4628 4629

    :param input: The first input layer.
    :type input: LayerOutput.
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring.
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
    :return: LayerOutput object.
    :rtype: LayerOutput.
    """
L
Luo Tao 已提交
4630
    assert isinstance(input, LayerOutput)
Q
qijun 已提交
4631 4632 4633 4634 4635
    Layer(
        name=name,
        type=LayerType.SUM_COST,
        inputs=[input.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
X
xuwei06 已提交
4636

Q
qijun 已提交
4637
    return LayerOutput(name, LayerType.SUM_COST, parents=[input], size=1)
X
xuwei06 已提交
4638 4639


Z
zhangjinchao01 已提交
4640
@wrap_name_default()
L
luotao1 已提交
4641 4642
@layer_support()
def huber_cost(input, label, name=None, coeff=1.0, layer_attr=None):
Z
zhangjinchao01 已提交
4643 4644 4645 4646 4647
    """
    A loss layer for huber loss.

    .. code-block:: python

X
xuwei06 已提交
4648
       cost = huber_cost(input=input_layer,
L
Luo Tao 已提交
4649
                         label=label_layer)
Z
zhangjinchao01 已提交
4650 4651 4652 4653 4654 4655 4656 4657 4658

    :param input: The first input layer.
    :type input: LayerOutput.
    :param label: The input label.
    :type input: LayerOutput.
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring.
    :param coeff: The coefficient affects the gradient in the backward.
    :type coeff: float.
L
luotao1 已提交
4659 4660
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
4661
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4662 4663
    :rtype: LayerOutput.
    """
4664 4665 4666
    assert isinstance(input, LayerOutput)
    if input.size is not None:
        assert input.size == 1
Q
qijun 已提交
4667 4668 4669 4670 4671 4672
    Layer(
        name=name,
        type=LayerType.HUBER,
        inputs=[input.name, label.name],
        coeff=coeff,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
X
xuwei06 已提交
4673
    return LayerOutput(name, LayerType.HUBER, parents=[input, label], size=1)
Z
zhangjinchao01 已提交
4674

4675

Z
zhangjinchao01 已提交
4676
@wrap_name_default()
L
luotao1 已提交
4677
@layer_support()
Q
qijun 已提交
4678 4679 4680 4681
def multi_binary_label_cross_entropy(input,
                                     label,
                                     name=None,
                                     coeff=1.0,
L
luotao1 已提交
4682
                                     layer_attr=None):
Z
zhangjinchao01 已提交
4683 4684 4685 4686 4687
    """
    A loss layer for multi binary label cross entropy.

    .. code-block:: python

X
xuwei06 已提交
4688
       cost = multi_binary_label_cross_entropy(input=input_layer,
L
Luo Tao 已提交
4689
                                               label=label_layer)
Z
zhangjinchao01 已提交
4690 4691 4692 4693 4694 4695 4696 4697 4698 4699 4700

    :param input: The first input layer.
    :type input: LayerOutput
    :param label: The input label.
    :type input: LayerOutput
    :param type: The type of cost.
    :type type: basestring
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring
    :param coeff: The coefficient affects the gradient in the backward.
    :type coeff: float
L
luotao1 已提交
4701 4702
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
4703
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4704 4705 4706
    :rtype: LayerOutput
    """

4707 4708
    if input.activation is None or \
            not isinstance(input.activation, SigmoidActivation):
Q
qijun 已提交
4709 4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722 4723 4724
        logger.log(
            logging.WARN,
            "%s is not recommend for multi_binary_label_cross_entropy's activation, "
            "maybe the sigmoid is better" % repr(input.activation))

    Layer(
        name=name,
        type=LayerType.MULTI_BIN_LABEL_CROSS_ENTROPY,
        inputs=[input.name, label.name],
        coeff=coeff,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.MULTI_BIN_LABEL_CROSS_ENTROPY,
        parents=[input, label],
        size=1)