layers.py 172.3 KB
Newer Older
1
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
Z
zhangjinchao01 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15
#
# 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
Y
Yu Yang 已提交
17
import inspect
Z
zhangjinchao01 已提交
18 19 20

from paddle.trainer.config_parser import *
from .activations import LinearActivation, SigmoidActivation, TanhActivation, \
Y
Yu Yang 已提交
21
    ReluActivation, IdentityActivation, SoftmaxActivation, BaseActivation
Z
zhangjinchao01 已提交
22 23 24 25
from .evaluators import *
from .poolings import MaxPooling, AvgPooling, BasePoolingType
from .attrs import *
from .default_decorators import *
26

Z
zhangjinchao01 已提交
27 28 29 30 31 32
try:
    import cPickle as pickle
except ImportError:
    import pickle
import copy

Q
qijun 已提交
33 34 35 36 37 38 39 40
__all__ = [
    "full_matrix_projection",
    "AggregateLevel",
    "ExpandLevel",
    "identity_projection",
    "dotmul_projection",
    "dotmul_operator",
    "repeat_layer",
41
    "seq_reshape_layer",
Q
qijun 已提交
42 43 44 45 46 47 48 49 50 51 52 53 54
    "table_projection",
    "mixed_layer",
    "data_layer",
    "embedding_layer",
    "fc_layer",
    "grumemory",
    "pooling_layer",
    "lstmemory",
    "last_seq",
    "first_seq",
    "cos_sim",
    "hsigmoid",
    "conv_projection",
L
Luo Tao 已提交
55
    "mse_cost",
Q
qijun 已提交
56 57 58 59 60 61 62 63 64
    "regression_cost",
    'classification_cost',
    "LayerOutput",
    'img_conv_layer',
    'img_pool_layer',
    'batch_norm_layer',
    'img_cmrnorm_layer',
    'addto_layer',
    'concat_layer',
65
    'seq_concat_layer',
Q
qijun 已提交
66 67 68 69 70 71
    'lstm_step_layer',
    'recurrent_group',
    'memory',
    'StaticInput',
    'expand_layer',
    'scaling_layer',
X
xuwei06 已提交
72
    'scaling_projection',
Q
qijun 已提交
73 74 75 76
    'power_layer',
    'interpolation_layer',
    'bilinear_interp_layer',
    'trans_layer',
77
    'rotate_layer',
Q
qijun 已提交
78 79 80 81 82 83 84 85 86
    'sum_to_one_norm_layer',
    'get_output_layer',
    'LayerType',
    'context_projection',
    'beam_search',
    'maxid_layer',
    'GeneratedInput',
    'SubsequenceInput',
    'gru_step_layer',
Y
Yu Yang 已提交
87
    'gru_step_naive_layer',
Q
qijun 已提交
88 89 90 91 92 93 94 95 96 97 98 99
    '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',
100
    'warp_ctc_layer',
Q
qijun 已提交
101 102 103 104 105 106 107 108 109 110 111 112 113 114
    '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',
Y
yuan 已提交
115
    'priorbox_layer',
116
    'cross_channel_norm_layer',
Q
qijun 已提交
117
    'spp_layer',
D
dangqingqing 已提交
118
    'pad_layer',
L
Luo Tao 已提交
119
    'eos_layer',
120
    'smooth_l1_cost',
121
    'layer_support',
Q
qijun 已提交
122
]
Z
zhangjinchao01 已提交
123 124 125 126 127 128 129 130 131 132 133 134 135


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

    DATA = "data"
    MIXED_LAYER = "mixed"
    LSTMEMORY = "lstmemory"
    GRUMEMORY = "gated_recurrent"
    SEQUENCE_LAST_INSTANCE = "seqlastins"
    SEQUENCE_FIRST_INSTANCE = "seqfirstins"
136
    SEQUENCE_RESHAPE = "seqreshape"
Z
zhangjinchao01 已提交
137 138 139 140
    POOLING_MAX = "max"
    POOLING_AVG = 'average'
    FC_LAYER = "fc"
    COST = 'cost'
141 142
    COSINE_SIM_VEC = 'cos_vm'
    COSINE_SIM = 'cos'
Z
zhangjinchao01 已提交
143 144
    HSIGMOID = 'hsigmoid'
    CONV_LAYER = "conv"
145
    CONVTRANS_LAYER = "convt"
146 147 148
    EXCONV_LAYER = "exconv"
    EXCONVTRANS_LAYER = "exconvt"
    CUDNNCONV_LAYER = "cudnn_conv"
Z
zhangjinchao01 已提交
149 150 151 152 153 154 155 156
    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'
157
    SEQUENCE_CONCAT_LAYER = 'seqconcat'
Z
zhangjinchao01 已提交
158 159 160 161 162 163 164

    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
165
    BILINEAR_INTERP_LAYER = 'bilinear_interp'
Z
zhangjinchao01 已提交
166 167 168
    POWER_LAYER = 'power'
    SCALING_LAYER = 'scaling'
    TRANS_LAYER = 'trans'
169
    ROTATE_LAYER = 'rotate'
H
Haonan 已提交
170
    OUT_PROD_LAYER = 'out_prod'
X
xuwei06 已提交
171
    FEATURE_MAP_EXPAND_LAYER = 'featmap_expand'
Z
zhangjinchao01 已提交
172 173 174 175 176 177 178 179 180 181 182

    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"
183
    LINEAR_COMBINATION_LAYER = "convex_comb"
Z
zhangjinchao01 已提交
184
    BLOCK_EXPAND = "blockexpand"
185
    MAXOUT = "maxout"
Q
qijun 已提交
186
    SPP_LAYER = "spp"
D
dangqingqing 已提交
187
    PAD_LAYER = "pad"
Z
zhangjinchao01 已提交
188

189
    PRINT_LAYER = "print"
Y
yuan 已提交
190
    PRIORBOX_LAYER = "priorbox"
191

Z
zhangjinchao01 已提交
192
    CTC_LAYER = "ctc"
193
    WARP_CTC_LAYER = "warp_ctc"
Z
zhangjinchao01 已提交
194 195
    CRF_LAYER = "crf"
    CRF_DECODING_LAYER = "crf_decoding"
196
    NCE_LAYER = 'nce'
Z
zhangjinchao01 已提交
197 198 199 200 201 202 203 204

    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
205
    SUM_COST = "sum_cost"
D
dangqingqing 已提交
206
    SMOOTH_L1 = "smooth_l1"
Z
zhangjinchao01 已提交
207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251

    @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.
252
    :type parents: list|tuple|collections.Sequence
Z
zhangjinchao01 已提交
253 254
    """

Q
qijun 已提交
255 256 257 258 259 260 261 262 263
    def __init__(self,
                 name,
                 layer_type,
                 parents=None,
                 activation=None,
                 num_filters=None,
                 img_norm_type=None,
                 size=None,
                 outputs=None,
264
                 reverse=None):
Z
zhangjinchao01 已提交
265 266
        assert isinstance(name, basestring)
        assert isinstance(layer_type, basestring)
X
xuwei06 已提交
267
        assert size is not None
Z
zhangjinchao01 已提交
268 269 270
        assert LayerType.is_layer_type(layer_type)
        self.name = name
        self.layer_type = layer_type
271 272
        if parents is not None and type(parents) != list:
            parents = [parents]
Z
zhangjinchao01 已提交
273 274 275 276 277 278 279 280
        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
281
        self.reverse = reverse
Z
zhangjinchao01 已提交
282 283 284 285 286 287 288 289 290 291 292 293 294

    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"

295 296 297 298 299 300 301 302
    def set_input(self, input):
        """
        Set the input for a memory layer. Can only be used for memory layer
        """
        assert isinstance(input, LayerOutput)
        assert self.layer_type == LayerType.MEMORY
        SetMemoryInput(self.name, input.name)

Z
zhangjinchao01 已提交
303 304 305

ERROR_CLIPPING = 'error_clipping_threshold'
DROPOUT = 'drop_rate'
306
DEVICE = 'device'
Z
zhangjinchao01 已提交
307 308 309


def layer_support(*attrs):
310
    attrs_list = list(attrs)
311
    attrs_list.append(DEVICE)
Q
qijun 已提交
312

Z
zhangjinchao01 已提交
313 314 315
    def decorator(method):
        @functools.wraps(method)
        def wrapper(*args, **kwargs):
316
            for attr in attrs_list:
Z
zhangjinchao01 已提交
317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332
                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)

Y
Yu Yang 已提交
333 334 335 336 337
        if hasattr(method, 'argspec'):
            wrapper.argspec = method.argspec
        else:
            wrapper.argspec = inspect.getargspec(method)

Z
zhangjinchao01 已提交
338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376
        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
377 378
    proj = FullMatrixProjection(
        input_layer_name=input.name, size=size, **param_attr.attr)
Z
zhangjinchao01 已提交
379 380 381 382
    proj.origin = input
    return proj


383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412
@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
413 414
    proj = TransposedFullMatrixProjection(
        input_layer_name=input.name, size=size, **param_attr.attr)
415 416 417 418
    proj.origin = input
    return proj


Z
zhangjinchao01 已提交
419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457
@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
458 459
    proj = TableProjection(
        input_layer_name=input.name, size=size, **param_attr.attr)
Z
zhangjinchao01 已提交
460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494
    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.
495
    :type input: LayerOutput
Z
zhangjinchao01 已提交
496 497
    :param offset: Offset, None if use default.
    :type offset: int
X
xuwei06 已提交
498
    :return: A IdentityProjection or IdentityOffsetProjection object
Z
zhangjinchao01 已提交
499 500 501 502 503 504
    :rtype: IdentityProjection or IdentityOffsetProjection
    """
    if offset is None:
        proj = IdentityProjection(input_layer_name=input.name)
        proj.origin = input
    else:
Q
qijun 已提交
505 506
        proj = IdentityOffsetProjection(
            input_layer_name=input.name, offset=offset)
Z
zhangjinchao01 已提交
507 508 509 510
        proj.origin = input
    return proj


X
xuwei06 已提交
511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532
@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
533
    proj = ScalingProjection(input_layer_name=input.name, **param_attr.attr)
X
xuwei06 已提交
534 535 536 537
    proj.origin = input
    return proj


Z
zhangjinchao01 已提交
538
@wrap_param_attr_default()
539
def dotmul_projection(input, param_attr=None):
Z
zhangjinchao01 已提交
540
    """
541
    DotMulProjection with a layer as input.
Z
zhangjinchao01 已提交
542 543 544 545 546 547 548 549 550 551 552 553 554
    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)

555 556 557 558 559 560 561
    :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
562 563
    proj = DotMulProjection(
        input_layer_name=input.name, size=input.size, **param_attr.attr)
564
    proj.origin = input
565
    return proj
Z
zhangjinchao01 已提交
566

567 568

def dotmul_operator(a=None, b=None, scale=1, **kwargs):
569 570
    """
    DotMulOperator takes two inputs and performs element-wise multiplication:
571

Z
zhangjinchao01 已提交
572
    .. math::
L
Luo Tao 已提交
573
       out.row[i] += scale * (a.row[i] .* b.row[i])
574

Z
zhangjinchao01 已提交
575 576
    where :math:`.*` means element-wise multiplication, and
    scale is a config scalar, its default value is one.
577

Z
zhangjinchao01 已提交
578
    The example usage is:
579

Z
zhangjinchao01 已提交
580
    .. code-block:: python
581

L
Luo Tao 已提交
582
       op = dotmul_operator(a=layer1, b=layer2, scale=0.5)
583

584 585 586 587
    :param a: Input layer1
    :type a: LayerOutput
    :param b: Input layer2
    :type b: LayerOutput
Z
zhangjinchao01 已提交
588 589
    :param scale: config scalar, default value is one.
    :type scale: float
590 591
    :return: A DotMulOperator Object.
    :rtype: DotMulOperator
Z
zhangjinchao01 已提交
592
    """
593 594 595
    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
596
    a = kwargs.get('x', a)  # For Backward capacity.
597 598 599 600 601 602
    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
603
    op = DotMulOperator(input_layer_names=[a.name, b.name], scale=scale)
604
    op.origin = [a, b]
605
    return op
Z
zhangjinchao01 已提交
606

607

Z
zhangjinchao01 已提交
608
@wrap_bias_attr_default(['padding_attr'])
Q
qijun 已提交
609 610 611
def context_projection(input,
                       context_len,
                       context_start=None,
Z
zhangjinchao01 已提交
612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647
                       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
648 649 650 651 652 653
    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
654 655 656 657 658 659 660 661 662 663 664 665 666
    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
667
    def __init__(self, name, size, act, bias_attr, layer_attr, parents=None):
Z
zhangjinchao01 已提交
668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683
        """
        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
684 685 686 687 688 689 690
        LayerOutput.__init__(
            self,
            name,
            LayerType.MIXED_LAYER,
            parents,
            size=size,
            activation=act)
Z
zhangjinchao01 已提交
691 692 693 694 695
        self.bias_attr = bias_attr
        self.layer_attr = layer_attr
        self.inputs = []
        self.finalized = False

696
    def __iadd__(self, other):
Z
zhangjinchao01 已提交
697 698 699 700 701 702 703 704
        """
        + += operator
        :param other: Other projection.
        :type other: Projection
        :return: self.
        :rtype: MixedLayerType
        """
        if not self.finalized:
705
            assert isinstance(other, Projection) or isinstance(other, Operator)
Z
zhangjinchao01 已提交
706
            self.inputs.append(other)
707 708 709 710
            if isinstance(other, Projection):
                self.parents.append(other.origin)
            else:
                self.parents.extend(other.origin)
Z
zhangjinchao01 已提交
711 712 713 714 715 716 717 718
            return self
        else:
            raise MixedLayerType.AddToSealedMixedLayerException()

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

719
    def __exit__(self, exc_type, exc_value, tb):
W
wangyang59 已提交
720 721
        if exc_value is not None:
            raise exc_value
Z
zhangjinchao01 已提交
722
        assert len(self.inputs) != 0
723
        ml = MixedLayer(
Z
zhangjinchao01 已提交
724 725 726 727 728
            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
729
            **ExtraLayerAttribute.to_kwargs(self.layer_attr))
730 731 732
        # update the size which might be computed inside MixedLayer
        # according to the operator's output size
        self.size = ml.config.size
733
        self.finalized = True
Z
zhangjinchao01 已提交
734 735 736 737 738 739


@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
740 741 742 743 744
def mixed_layer(size=0,
                input=None,
                name=None,
                act=None,
                bias_attr=False,
Z
zhangjinchao01 已提交
745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788
                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
789 790 791 792 793 794
        with mixed_layer(
                name=name,
                size=size,
                act=act,
                bias_attr=bias_attr,
                layer_attr=layer_attr) as m:
795
            if isinstance(input, collections.Sequence):
Z
zhangjinchao01 已提交
796 797 798 799 800 801 802 803
                for each in input:
                    m += each
            else:
                m += input
        return m


@layer_support()
L
Luo Tao 已提交
804
def data_layer(name, size, height=None, width=None, layer_attr=None):
Z
zhangjinchao01 已提交
805 806 807 808 809 810 811
    """
    Define DataLayer For NeuralNetwork.

    The example usage is:

    ..  code-block:: python

Y
Yu Yang 已提交
812
        data = data_layer(name="input", size=1000)
Z
zhangjinchao01 已提交
813 814 815 816 817

    :param name: Name of this data layer.
    :type name: basestring
    :param size: Size of this data layer.
    :type size: int
L
Luo Tao 已提交
818
    :param height: Height of this data layer, used for image
Y
Yu Yang 已提交
819
    :type height: int|None
L
Luo Tao 已提交
820
    :param width: Width of this data layer, used for image
Y
Yu Yang 已提交
821
    :type width: int|None
Z
zhangjinchao01 已提交
822 823
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
824
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
825 826
    :rtype: LayerOutput
    """
Q
qijun 已提交
827 828 829 830
    Layer(
        type=LayerType.DATA,
        name=name,
        size=size,
L
Luo Tao 已提交
831 832
        height=height,
        width=width,
Q
qijun 已提交
833
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855

    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
856
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
857 858
    :rtype: LayerOutput
    """
Q
qijun 已提交
859 860 861 862 863 864
    with mixed_layer(
            name=name,
            size=size,
            act=LinearActivation(),
            bias_attr=False,
            layer_attr=layer_attr) as mix:
Z
zhangjinchao01 已提交
865 866 867 868 869 870 871 872 873
        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
874 875 876 877 878 879 880
def fc_layer(input,
             size,
             act=None,
             name=None,
             param_attr=None,
             bias_attr=None,
             layer_attr=None):
Z
zhangjinchao01 已提交
881 882 883 884 885 886 887 888 889 890 891 892
    """
    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
893
    which is equal to:
Z
zhangjinchao01 已提交
894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915

    .. 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
916
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
917 918 919 920
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
921
        assert not isinstance(param_attr, collections.Sequence)
Z
zhangjinchao01 已提交
922 923
        param_attr = [param_attr]
    else:
924
        if isinstance(param_attr, collections.Sequence):
Z
zhangjinchao01 已提交
925 926 927 928
            assert len(input) == len(param_attr)
        else:
            param_attr = [copy.deepcopy(param_attr) for _ in range(len(input))]

929
    assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
930 931

    Layer(
Q
qijun 已提交
932 933 934
        inputs=[
            Input(ipt.name, **attr.attr) for ipt, attr in zip(input, param_attr)
        ],
Z
zhangjinchao01 已提交
935 936 937 938 939
        name=name,
        type=LayerType.FC_LAYER,
        size=size,
        bias=ParamAttr.to_bias(bias_attr),
        active_type=act.name,
Q
qijun 已提交
940 941 942
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.FC_LAYER, input, activation=act, size=size)
Z
zhangjinchao01 已提交
943

944

945 946 947 948
@wrap_name_default("print")
def print_layer(input, name=None):
    """
    Print the output value of input layers. This layer is useful for debugging.
949 950 951 952 953

    :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
954
    :return: LayerOutput
955
    """
956 957 958 959 960
    if isinstance(input, LayerOutput):
        input = [input]
    assert isinstance(input, collections.Sequence)  # list or tuple
    for each in input:
        assert isinstance(each, LayerOutput)
961 962 963 964

    Layer(
        name=name,
        type=LayerType.PRINT_LAYER,
Q
qijun 已提交
965
        inputs=[l.name for l in input], )
966
    # this layer don't return anything, can not be input of other layer.
967

Z
zhangjinchao01 已提交
968

Y
yuan 已提交
969
@wrap_name_default("priorbox")
G
gaoyuan 已提交
970
def priorbox_layer(input,
G
gaoyuan 已提交
971
                   image,
G
gaoyuan 已提交
972 973 974 975 976
                   aspect_ratio,
                   variance,
                   min_size,
                   max_size=[],
                   name=None):
Y
yuan 已提交
977 978 979 980 981 982 983
    """
    Compute the priorbox and set the variance. This layer is necessary for ssd.

    :param name: The Layer Name.
    :type name: basestring
    :param input: The input layer.
    :type input: LayerOutput
G
gaoyuan 已提交
984 985
    :param image: The network input image.
    :type image: LayerOutput
Y
yuan 已提交
986 987 988 989 990 991 992 993 994 995 996
    :param aspect_ratio: The aspect ratio.
    :type aspect_ratio: list
    :param variance: The bounding box variance.
    :type min_size: The min size of the priorbox width/height.
    :param min_size: list
    :type max_size: The max size of the priorbox width/height. Could be NULL.
    :param max_size: list
    :return: LayerOutput
    """
    # plus one for ratio 1.
    num_filters = (len(aspect_ratio) * 2 + 1 + len(max_size)) * 4
G
gaoyuan 已提交
997
    size = (input.size / input.num_filters) * num_filters * 2
Y
yuan 已提交
998 999 1000
    Layer(
        name=name,
        type=LayerType.PRIORBOX_LAYER,
G
gaoyuan 已提交
1001
        inputs=[input.name, image.name],
Y
yuan 已提交
1002 1003 1004 1005 1006 1007
        size=size,
        min_size=min_size,
        max_size=max_size,
        aspect_ratio=aspect_ratio,
        variance=variance)
    return LayerOutput(
G
gaoyuan 已提交
1008 1009
        name,
        LayerType.PRIORBOX_LAYER,
G
gaoyuan 已提交
1010
        parents=[input, image],
G
gaoyuan 已提交
1011 1012 1013
        num_filters=num_filters,
        size=size)

Z
zhangjinchao01 已提交
1014

1015 1016
@wrap_name_default("cross_channel_norm")
def cross_channel_norm_layer(input, name=None, param_attr=None):
G
gaoyuan 已提交
1017 1018 1019 1020 1021
    """
    Normalize a layer's output. This layer is necessary for ssd.
    This layer applys normalize across the channels of each sample to
    a conv layer's output and scale the output by a group of trainable
    factors which dimensions equal to the channel's number.
G
gaoyuan 已提交
1022

G
gaoyuan 已提交
1023 1024 1025 1026 1027 1028 1029 1030
    :param name: The Layer Name.
    :type name: basestring
    :param input: The input layer.
    :type input: LayerOutput
    :param param_attr: The Parameter Attribute|list.
    :type param_attr: ParameterAttribute
    :return: LayerOutput
    """
1031
    assert input.num_filters is not None
G
gaoyuan 已提交
1032 1033
    Layer(
        name=name,
1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046
        type=LayerType.NORM_LAYER,
        inputs=[
            Input(
                input.name,
                norm=Norm(
                    norm_type="cross-channel-norm",
                    channels=input.num_filters,
                    size=input.size,
                    scale=0,
                    pow=0,
                    blocked=0),
                **param_attr.attr)
        ])
G
gaoyuan 已提交
1047 1048
    return LayerOutput(
        name,
1049
        LayerType.NORM_LAYER,
G
gaoyuan 已提交
1050 1051 1052 1053 1054
        parents=input,
        num_filters=input.num_filters,
        size=input.size)


Z
zhangjinchao01 已提交
1055 1056 1057 1058
@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
1059 1060 1061 1062
def pooling_layer(input,
                  pooling_type=None,
                  name=None,
                  bias_attr=None,
Z
zhangjinchao01 已提交
1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075
                  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
1076 1077
    :param agg_level: AggregateLevel.EACH_TIMESTEP or
                      AggregateLevel.EACH_SEQUENCE
Z
zhangjinchao01 已提交
1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089
    :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
1090
    :return: LayerOutput object.
Y
Yu Yang 已提交
1091
    :rtype: LayerOutput
Z
zhangjinchao01 已提交
1092 1093
    """
    extra_dict = dict()
1094
    # noinspection PyUnresolvedReferences
Z
zhangjinchao01 已提交
1095 1096
    if isinstance(pooling_type, AvgPooling):
        extra_dict['average_strategy'] = pooling_type.strategy
1097 1098 1099 1100
    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
1101 1102 1103 1104 1105 1106 1107 1108
    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
1109
        **extra_dict)
Z
zhangjinchao01 已提交
1110

Q
qijun 已提交
1111 1112
    return LayerOutput(
        name, pooling_type.name, parents=[input], size=input.size)
Z
zhangjinchao01 已提交
1113

Q
qijun 已提交
1114

Z
zhangjinchao01 已提交
1115 1116
@wrap_bias_attr_default()
@wrap_param_attr_default()
Q
qijun 已提交
1117
@wrap_act_default(param_names=['gate_act'], act=SigmoidActivation())
Z
zhangjinchao01 已提交
1118 1119 1120
@wrap_act_default(param_names=["act", 'state_act'], act=TanhActivation())
@wrap_name_default("lstmemory")
@layer_support(DROPOUT)
Q
qijun 已提交
1121 1122 1123 1124 1125 1126 1127 1128 1129
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
1130 1131 1132 1133 1134 1135 1136 1137
              layer_attr=None):
    """
    Long Short-term Memory Cell.

    The memory cell was implemented as follow equations.

    ..  math::

L
luotao02 已提交
1138
        i_t & = \\sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i)
Z
zhangjinchao01 已提交
1139

L
luotao02 已提交
1140
        f_t & = \\sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f)
Z
zhangjinchao01 已提交
1141

L
luotao02 已提交
1142
        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
1143

L
luotao02 已提交
1144
        o_t & = \\sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o)
Z
zhangjinchao01 已提交
1145

L
luotao02 已提交
1146
        h_t & = o_t tanh(c_t)
Z
zhangjinchao01 已提交
1147 1148


C
caoying03 已提交
1149
    NOTE: In PaddlePaddle's implementation, the multiplications
Z
zhangjinchao01 已提交
1150
    :math:`W_{xi}x_{t}` , :math:`W_{xf}x_{t}`,
C
caoying03 已提交
1151 1152 1153 1154
    :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
1155

C
caoying03 已提交
1156
    NOTE: This is a low level user interface. You can use network.simple_lstm
Z
zhangjinchao01 已提交
1157 1158
    to config a simple plain lstm layer.

C
caoying03 已提交
1159 1160 1161 1162
    Please refer to **Generating Sequences With Recurrent Neural Networks** for
    more details about LSTM.

    Link_ goes as below.
Z
zhangjinchao01 已提交
1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185

    .. _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
1186
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1187 1188 1189 1190 1191 1192
    :rtype: LayerOutput
    """

    assert gate_act.support_hppl
    assert state_act.support_hppl
    assert act.support_hppl
1193 1194 1195 1196 1197 1198 1199 1200 1201 1202
    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
1203

Q
qijun 已提交
1204 1205 1206 1207 1208 1209 1210 1211 1212 1213
    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
1214

Q
qijun 已提交
1215 1216 1217 1218 1219
    return LayerOutput(
        name,
        LayerType.LSTMEMORY, [input],
        size=input.size / 4,
        reverse=reverse)
1220

Z
zhangjinchao01 已提交
1221 1222 1223

@wrap_bias_attr_default()
@wrap_param_attr_default()
Q
qijun 已提交
1224
@wrap_act_default(param_names=['gate_act'], act=SigmoidActivation())
Z
zhangjinchao01 已提交
1225 1226 1227
@wrap_act_default(param_names=["act"], act=TanhActivation())
@wrap_name_default("gru")
@layer_support(DROPOUT)
Q
qijun 已提交
1228 1229 1230 1231 1232 1233 1234 1235
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
1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256
              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
1257 1258
    3. The candidate activation :math:`\\tilde{h_t}` is computed similarly to
    that of the traditional recurrent unit:
Z
zhangjinchao01 已提交
1259 1260 1261 1262 1263

    ..  math::

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

C
caoying03 已提交
1264 1265 1266
    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
1267 1268 1269 1270 1271

    ..  math::

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

C
caoying03 已提交
1272
    NOTE: In PaddlePaddle's implementation, the multiplication operations
Z
zhangjinchao01 已提交
1273
    :math:`W_{r}x_{t}`, :math:`W_{z}x_{t}` and :math:`W x_t` are not computed in
C
caoying03 已提交
1274 1275 1276
    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
1277

C
caoying03 已提交
1278 1279 1280
    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
1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291

    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.
1292
    :param reverse: Whether sequence process is reversed or not.
Z
zhangjinchao01 已提交
1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307
    :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
1308 1309 1310
    :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
1311
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1312 1313 1314 1315
    :rtype: LayerOutput
    """
    assert act.support_hppl
    assert gate_act.support_hppl
1316 1317 1318 1319 1320 1321 1322 1323 1324
    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
1325

Q
qijun 已提交
1326 1327 1328 1329 1330 1331 1332 1333 1334
    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
1335

Q
qijun 已提交
1336 1337 1338 1339 1340
    return LayerOutput(
        name,
        LayerType.GRUMEMORY, [input],
        size=input.size / 3,
        reverse=reverse)
1341

Z
zhangjinchao01 已提交
1342 1343 1344

@wrap_name_default()
@layer_support()
Q
qijun 已提交
1345 1346 1347
def last_seq(input,
             name=None,
             agg_level=AggregateLevel.EACH_TIMESTEP,
1348
             stride=-1,
Z
zhangjinchao01 已提交
1349 1350 1351 1352
             layer_attr=None):
    """
    Get Last Timestamp Activation of a sequence.

1353 1354 1355
    If stride > 0, this layer slides a window whose size is determined by stride,
    and return the last value of the window as the output. Thus, a long sequence
    will be shorten. Note that for sequence with sub-sequence, the default value
L
Luo Tao 已提交
1356
    of stride is -1.
1357

L
Luo Tao 已提交
1358 1359 1360 1361 1362 1363
    The simple usage is:

    .. code-block:: python

       seq = last_seq(input=layer)

Z
zhangjinchao01 已提交
1364 1365 1366 1367 1368
    :param agg_level: Aggregated level
    :param name: Layer name.
    :type name: basestring
    :param input: Input layer name.
    :type input: LayerOutput
1369
    :param stride: window size.
1370
    :type stride: Int
Z
zhangjinchao01 已提交
1371 1372
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1373
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1374 1375
    :rtype: LayerOutput
    """
1376 1377 1378 1379 1380 1381
    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.")

1382 1383 1384
    if agg_level == AggregateLevel.EACH_SEQUENCE:
        assert stride == -1

Z
zhangjinchao01 已提交
1385 1386 1387 1388 1389
    Layer(
        name=name,
        type=LayerType.SEQUENCE_LAST_INSTANCE,
        inputs=[input.name],
        trans_type=agg_level,
1390
        stride=stride,
Q
qijun 已提交
1391 1392 1393 1394 1395 1396
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.SEQUENCE_LAST_INSTANCE,
        parents=[input],
        size=input.size)
Z
zhangjinchao01 已提交
1397 1398 1399 1400


@wrap_name_default()
@layer_support()
Q
qijun 已提交
1401 1402 1403
def first_seq(input,
              name=None,
              agg_level=AggregateLevel.EACH_TIMESTEP,
1404
              stride=-1,
Z
zhangjinchao01 已提交
1405 1406 1407 1408
              layer_attr=None):
    """
    Get First Timestamp Activation of a sequence.

1409 1410 1411
    If stride > 0, this layer slides a window whose size is determined by stride,
    and return the first value of the window as the output. Thus, a long sequence
    will be shorten. Note that for sequence with sub-sequence, the default value
L
Luo Tao 已提交
1412
    of stride is -1.
1413

L
Luo Tao 已提交
1414 1415 1416 1417 1418 1419
    The simple usage is:

    .. code-block:: python

       seq = first_seq(input=layer)

Z
zhangjinchao01 已提交
1420 1421 1422 1423 1424
    :param agg_level: aggregation level
    :param name: Layer name.
    :type name: basestring
    :param input: Input layer name.
    :type input: LayerOutput
1425
    :param stride: window size.
1426
    :type stride: Int
Z
zhangjinchao01 已提交
1427 1428
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1429
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1430 1431
    :rtype: LayerOutput
    """
1432 1433 1434 1435 1436 1437 1438

    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.')

1439 1440 1441
    if agg_level == AggregateLevel.EACH_SEQUENCE:
        assert stride == -1

Z
zhangjinchao01 已提交
1442 1443 1444 1445 1446
    Layer(
        name=name,
        type=LayerType.SEQUENCE_FIRST_INSTANCE,
        inputs=[input.name],
        trans_type=agg_level,
1447
        stride=stride,
Q
qijun 已提交
1448 1449 1450 1451 1452 1453
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.SEQUENCE_FIRST_INSTANCE,
        parents=[input],
        size=input.size)
Z
zhangjinchao01 已提交
1454 1455 1456 1457 1458 1459


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

1460

Z
zhangjinchao01 已提交
1461 1462
@wrap_name_default()
@layer_support()
Q
qijun 已提交
1463 1464
def expand_layer(input,
                 expand_as,
Z
zhangjinchao01 已提交
1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493
                 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
1494
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1495 1496 1497 1498 1499 1500 1501 1502 1503
    :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
1504 1505 1506 1507 1508 1509
        **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
1510 1511


X
xuwei06 已提交
1512 1513
@wrap_name_default()
@layer_support()
Q
qijun 已提交
1514
def repeat_layer(input, num_repeats, name=None, layer_attr=None):
X
xuwei06 已提交
1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525
    """
    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

L
Luo Tao 已提交
1526
       expand = repeat_layer(input=layer, num_repeats=4)
X
xuwei06 已提交
1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544

    :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
1545 1546 1547 1548 1549 1550 1551
        **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
1552

1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564
@wrap_name_default("seqreshape")
@wrap_act_default(act=IdentityActivation())
@wrap_bias_attr_default(has_bias=False)
@layer_support()
def seq_reshape_layer(input,
                      reshape_size,
                      act=None,
                      name=None,
                      layer_attr=None,
                      bias_attr=None):
    """
    A layer for reshaping the sequence. Assume the input sequence has T instances,
1565
    the dimension of each instance is M, and the input reshape_size is N, then the
1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607
    output sequence has T*M/N instances, the dimension of each instance is N.

    Note that T*M/N must be an integer.

    The example usage is:

    .. code-block:: python

       reshape = seq_reshape_layer(input=layer, reshape_size=4)

    :param input: Input layer.
    :type input: LayerOutput
    :param reshape_size: the size of reshaped sequence.
    :type reshape_size: int
    :param name: Layer name.
    :type name: basestring
    :param act: Activation type.
    :type act: BaseActivation
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
    :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
    :return: LayerOutput object.
    :rtype: LayerOutput
    """

    Layer(
        inputs=[input.name],
        name=name,
        size=reshape_size,
        type=LayerType.SEQUENCE_RESHAPE,
        bias=ParamAttr.to_bias(bias_attr),
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name=name,
        size=reshape_size,
        layer_type=LayerType.SEQUENCE_RESHAPE,
        parents=[input])


Z
zhangjinchao01 已提交
1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635
@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
1636
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1637 1638
    :rtype: LayerOutput
    """
1639
    assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
1640
    assert len(input) == 2
1641 1642 1643 1644 1645 1646 1647
    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
1648 1649 1650 1651
    Layer(
        name=name,
        type=LayerType.INTERPOLATION_LAYER,
        inputs=[weight.name, input[0].name, input[1].name],
Q
qijun 已提交
1652 1653 1654 1655 1656 1657
        **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
1658 1659


L
liaogang 已提交
1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675
@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
1676
       bilinear = bilinear_interp_layer(input=layer1, out_size_x=64, out_size_y=64)
X
xuwei06 已提交
1677

L
liaogang 已提交
1678
    :param   input:        A input layer.
L
liaogang 已提交
1679
    :type    input:        LayerOutput.
L
liaogang 已提交
1680
    :param   out_size_x:   bilinear interpolation output width.
X
xuwei06 已提交
1681
    :type    out_size_x:   int|None
L
liaogang 已提交
1682
    :param   out_size_y:   bilinear interpolation output height.
L
liaogang 已提交
1683
    :type    out_size_y:   int|None
L
liaogang 已提交
1684
    :param   name:         The layer's name, which cna not be specified.
L
liaogang 已提交
1685
    :type    name:         None|basestring
L
liaogang 已提交
1686
    :param   layer_attr:   Extra Layer attribute.
L
liaogang 已提交
1687 1688 1689 1690 1691 1692 1693
    :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
1694
    assert input.num_filters is not None
L
liaogang 已提交
1695
    num_channels = input.num_filters
Q
qijun 已提交
1696 1697 1698 1699 1700 1701 1702
    l = Layer(
        name=name,
        inputs=Input(
            input.name,
            bilinear_interp=BilinearInterp(
                out_size_x=out_size_x,
                out_size_y=out_size_y,
L
Luo Tao 已提交
1703
                channels=num_channels)),
Q
qijun 已提交
1704 1705 1706 1707 1708 1709 1710 1711 1712
        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
1713

Z
zhangjinchao01 已提交
1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740
@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
1741
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1742 1743
    :rtype: LayerOutput
    """
1744 1745 1746
    assert isinstance(input, LayerOutput) and isinstance(weight, LayerOutput)
    if weight.size is not None:
        assert weight.size == 1
Z
zhangjinchao01 已提交
1747 1748 1749
    Layer(
        name=name,
        type=LayerType.POWER_LAYER,
1750
        inputs=[weight.name, input.name],
Q
qijun 已提交
1751 1752 1753
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.POWER_LAYER, parents=[input, weight], size=input.size)
Z
zhangjinchao01 已提交
1754 1755 1756 1757 1758 1759


@wrap_name_default()
@layer_support()
def scaling_layer(input, weight, name=None, layer_attr=None):
    """
1760
    A layer for multiplying input vector by weight scalar.
Z
zhangjinchao01 已提交
1761 1762

    .. math::
1763
       y  = w x
Z
zhangjinchao01 已提交
1764

1765 1766 1767 1768 1769
    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
1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784

    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
1785
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1786 1787
    :rtype: LayerOutput
    """
1788 1789 1790
    assert isinstance(weight, LayerOutput) and isinstance(input, LayerOutput)
    if weight.size is not None:
        assert weight.size == 1
Z
zhangjinchao01 已提交
1791 1792 1793 1794
    Layer(
        name=name,
        type=LayerType.SCALING_LAYER,
        inputs=[weight.name, input.name],
Q
qijun 已提交
1795 1796 1797
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.SCALING_LAYER, parents=[weight, input], size=input.size)
Z
zhangjinchao01 已提交
1798 1799 1800 1801 1802 1803


@wrap_name_default()
@layer_support()
def trans_layer(input, name=None, layer_attr=None):
    """
1804
    A layer for transposing a minibatch matrix.
Z
zhangjinchao01 已提交
1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822

    .. 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
1823
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1824 1825 1826 1827 1828 1829
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.TRANS_LAYER,
        inputs=[input.name],
Q
qijun 已提交
1830 1831 1832
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.TRANS_LAYER, parents=[input], size=input.size)
Z
zhangjinchao01 已提交
1833 1834


1835 1836
@wrap_name_default()
@layer_support()
H
Haonan 已提交
1837
def rotate_layer(input, height, width, name=None, layer_attr=None):
1838
    """
H
Haonan 已提交
1839 1840
    A layer for rotating 90 degrees (clock-wise) for each feature channel,
    usually used when the input sample is some image or feature map.
1841 1842

    .. math::
H
Haonan 已提交
1843
       y(j,i,:) = x(M-i-1,j,:)
1844

H
Haonan 已提交
1845
    where :math:`x` is (M x N x C) input, and :math:`y` is (N x M x C) output.
1846 1847 1848 1849 1850 1851

    The example usage is:

    .. code-block:: python

       rot = rotate_layer(input=layer,
H
Haonan 已提交
1852 1853
                          height=100,
                          width=100)
1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866

    :param input: Input layer.
    :type input: LayerOutput
    :param height: The height of the sample matrix
    :type height: int
    :param name: Layer name.
    :type name: basestring
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
    :return: LayerOutput object.
    :rtype: LayerOutput
    """
    assert isinstance(input, LayerOutput)
H
Haonan 已提交
1867 1868 1869
    l = Layer(
        name=name,
        height=height,
H
Haonan 已提交
1870
        width=width,
H
Haonan 已提交
1871 1872 1873 1874 1875 1876 1877 1878
        type=LayerType.ROTATE_LAYER,
        inputs=[input.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name=name,
        layer_type=LayerType.ROTATE_LAYER,
        parents=[input],
        size=l.config.size)
1879 1880


Z
zhangjinchao01 已提交
1881 1882
@wrap_name_default()
@layer_support()
1883
def cos_sim(a, b, scale=1, size=1, name=None, layer_attr=None):
Z
zhangjinchao01 已提交
1884 1885 1886 1887
    """
    Cosine Similarity Layer. The cosine similarity equation is here.

    ..  math::
D
dangqingqing 已提交
1888
        similarity = cos(\\theta) = {\\mathbf{a} \\cdot \\mathbf{b}
1889 1890 1891 1892 1893
        \\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
1894

1895 1896
    Note that the above computation is for one sample. Multiple samples are
    processed in one batch.
Z
zhangjinchao01 已提交
1897

L
Luo Tao 已提交
1898 1899 1900 1901 1902 1903
    The example usage is:

    .. code-block:: python

       cos = cos_sim(a=layer1, b=layer2, size=3)

Z
zhangjinchao01 已提交
1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915
    :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
1916
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1917 1918
    :rtype: LayerOutput
    """
1919
    assert isinstance(a, LayerOutput) and isinstance(b, LayerOutput)
1920 1921 1922 1923 1924 1925
    if size == 1:
        Layer(
            name=name,
            type=LayerType.COSINE_SIM,
            cos_scale=scale,
            inputs=[a.name, b.name],
Q
qijun 已提交
1926
            **ExtraLayerAttribute.to_kwargs(layer_attr))
1927
    else:
1928 1929
        if a.size is not None and b.size is not None:
            assert size == b.size / a.size
1930 1931 1932 1933 1934 1935
        Layer(
            name=name,
            type=LayerType.COSINE_SIM_VEC,
            size=size,
            cos_scale=scale,
            inputs=[a.name, b.name],
Q
qijun 已提交
1936
            **ExtraLayerAttribute.to_kwargs(layer_attr))
X
xuwei06 已提交
1937
    return LayerOutput(name, LayerType.COSINE_SIM, parents=[a, b], size=size)
Z
zhangjinchao01 已提交
1938

1939

Z
zhangjinchao01 已提交
1940 1941
@wrap_name_default()
@wrap_bias_attr_default(has_bias=True)
1942
@wrap_param_attr_default()
Z
zhangjinchao01 已提交
1943
@layer_support()
Q
qijun 已提交
1944 1945
def hsigmoid(input,
             label,
1946
             num_classes=None,
Q
qijun 已提交
1947 1948 1949 1950
             name=None,
             bias_attr=None,
             param_attr=None,
             layer_attr=None):
Z
zhangjinchao01 已提交
1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961
    """
    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],
1962
                        label=data_layer)
Z
zhangjinchao01 已提交
1963 1964 1965 1966 1967 1968 1969

    :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.
1970
    :type num_classes: int|None
L
luotao02 已提交
1971 1972
    :param name: layer name
    :type name: basestring
Z
zhangjinchao01 已提交
1973 1974 1975
    :param bias_attr: Bias attribute. None means default bias.
                      False means no bias.
    :type bias_attr: ParameterAttribute|False
1976 1977
    :param param_attr: Parameter Attribute. None means default parameter.
    :type param_attr: ParameterAttribute|None
Z
zhangjinchao01 已提交
1978 1979
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
1980
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1981 1982 1983 1984
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
1985 1986 1987 1988 1989 1990 1991 1992 1993
        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
1994 1995 1996
    assert isinstance(label, LayerOutput)
    assert label.layer_type == LayerType.DATA

1997 1998 1999 2000 2001
    if num_classes is None:
        num_classes = label.size
    if num_classes is None or num_classes <= 2:
        raise ValueError("hsigmoid label size must larger than 2.")

Z
zhangjinchao01 已提交
2002 2003
    ipts_for_layer = []
    parents = []
2004
    for each_input, each_param_attr in zip(input, param_attr):
Z
zhangjinchao01 已提交
2005
        assert isinstance(each_input, LayerOutput)
2006
        ipts_for_layer.append(Input(each_input.name, **each_param_attr.attr))
Z
zhangjinchao01 已提交
2007 2008 2009 2010
        parents.append(each_input)
    ipts_for_layer.append(label.name)
    parents.append(label)

X
xuwei06 已提交
2011
    l = Layer(
Z
zhangjinchao01 已提交
2012 2013 2014 2015 2016
        name=name,
        type=LayerType.HSIGMOID,
        num_classes=num_classes,
        bias=ParamAttr.to_bias(bias_attr),
        inputs=ipts_for_layer,
Q
qijun 已提交
2017 2018 2019
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.HSIGMOID, parents=parents, size=l.config.size)
Z
zhangjinchao01 已提交
2020

2021

Z
zhangjinchao01 已提交
2022 2023 2024 2025 2026
@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
2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042
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,
2043 2044
                   trans=False,
                   layer_type=None):
Z
zhangjinchao01 已提交
2045
    """
2046
    Convolution layer for image. Paddle can support both square and non-square
2047
    input currently.
Z
zhangjinchao01 已提交
2048 2049 2050 2051

    The details of convolution layer, please refer UFLDL's `convolution
    <http://ufldl.stanford.edu/tutorial/supervised/
    FeatureExtractionUsingConvolution/>`_ .
X
xuwei06 已提交
2052

2053
    Convolution Transpose (deconv) layer for image. Paddle can support both square
2054
    and non-square input currently.
2055

X
xuwei06 已提交
2056
    The details of convolution transpose layer,
2057 2058 2059
    please refer to the following explanation and references therein
    <http://datascience.stackexchange.com/questions/6107/
    what-are-deconvolutional-layers/>`_ .
Z
zhangjinchao01 已提交
2060 2061 2062 2063
    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
2064 2065 2066
    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
2067
    32*4 = 128 filters to process inputs. The channels will be split into 4
C
caoying03 已提交
2068 2069
    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
2070

L
Luo Tao 已提交
2071 2072 2073 2074 2075 2076 2077 2078 2079 2080
    The example usage is:

    ..  code-block:: python

        conv = img_conv_layer(input=data, filter_size=1, filter_size_y=1,
                              num_channels=8,
                              num_filters=16, stride=1,
                              bias_attr=False,
                              act=ReluActivation())

Z
zhangjinchao01 已提交
2081 2082 2083 2084
    :param name: Layer name.
    :type name: basestring
    :param input: Layer Input.
    :type input: LayerOutput
2085 2086 2087
    :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
2088 2089 2090
    :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).
2091
    :type filter_size_y: int|None
Z
zhangjinchao01 已提交
2092 2093 2094 2095 2096
    :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
2097 2098 2099
    :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
2100 2101
    :param stride_y: The y dimension of the stride.
    :type stride_y: int
2102 2103 2104
    :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
2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118
    :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
2119 2120
    :param trans: true if it is a convTransLayer, false if it is a convLayer
    :type trans: bool
2121
    :param layer_type: specify the layer_type, default is None. If trans=True,
2122 2123
                       layer_type has to be "exconvt" or "cudnn_convt",
                       otherwise layer_type has to be either "exconv" or
2124
                       "cudnn_conv"
2125
    :type layer_type: String
D
dangqingqing 已提交
2126
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2127 2128 2129 2130 2131
    :rtype: LayerOutput
    """
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters
2132

Z
zhangjinchao01 已提交
2133
    if filter_size_y is None:
2134 2135 2136 2137 2138 2139
        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
2140
    if stride_y is None:
2141 2142 2143 2144 2145 2146
        if isinstance(stride, collections.Sequence):
            assert len(stride) == 2
            stride, stride_y = stride
        else:
            stride_y = stride

Z
zhangjinchao01 已提交
2147
    if padding_y is None:
2148 2149 2150 2151 2152 2153 2154 2155
        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
2156
        init_w = (2.0 / (filter_size**2 * num_channels))**0.5
2157 2158 2159 2160
        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
2161

2162 2163
    if layer_type:
        if trans:
2164
            assert layer_type in ["exconvt", "cudnn_convt"]
2165 2166 2167 2168 2169
        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
2170

X
xuwei06 已提交
2171
    l = Layer(
Z
zhangjinchao01 已提交
2172
        name=name,
Q
qijun 已提交
2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184
        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
2185 2186 2187 2188
        active_type=act.name,
        num_filters=num_filters,
        bias=ParamAttr.to_bias(bias_attr),
        shared_biases=shared_biases,
2189
        type=lt,
Q
qijun 已提交
2190 2191 2192 2193 2194 2195 2196 2197
        **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
2198 2199 2200 2201


@wrap_name_default("pool")
@layer_support()
Q
qijun 已提交
2202 2203 2204 2205 2206 2207 2208 2209 2210 2211
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,
2212 2213
                   padding_y=None,
                   ceil_mode=True):
Z
zhangjinchao01 已提交
2214 2215 2216 2217 2218 2219 2220
    """
    Image pooling Layer.

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

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

L
Luo Tao 已提交
2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248
    - ceil_mode=True:

    ..  math::

        w = 1 + int(ceil(input\_width + 2 * padding - pool\_size) / float(stride))
        h = 1 + int(ceil(input\_height + 2 * padding\_y - pool\_size\_y) / float(stride\_y))

    - ceil_mode=False:

    ..  math::

        w = 1 + int(floor(input\_width + 2 * padding - pool\_size) / float(stride))
        h = 1 + int(floor(input\_height + 2 * padding\_y - pool\_size\_y) / float(stride\_y))

    The example usage is:

    ..  code-block:: python

        maxpool = img_pool_layer(input=conv,
                                 pool_size=3,
                                 pool_size_y=5,
                                 num_channels=8,
                                 stride=1,
                                 stride_y=2,
                                 padding=1,
                                 padding_y=2,
                                 pool_type=MaxPooling())

2249
    :param padding: pooling padding width.
Z
zhangjinchao01 已提交
2250
    :type padding: int
2251 2252
    :param padding_y: pooling padding height. It's equal to padding by default.
    :type padding_y: int|None
Z
zhangjinchao01 已提交
2253 2254 2255 2256
    :param name: name of pooling layer
    :type name: basestring.
    :param input: layer's input
    :type input: LayerOutput
2257
    :param pool_size: pooling window width
Z
zhangjinchao01 已提交
2258
    :type pool_size: int
2259 2260
    :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
2261 2262
    :param num_channels: number of input channel.
    :type num_channels: int
2263
    :param pool_type: pooling type. MaxPooling or AvgPooling. Default is
Z
zhangjinchao01 已提交
2264 2265
                      MaxPooling.
    :type pool_type: BasePoolingType
2266
    :param stride: stride width of pooling.
Z
zhangjinchao01 已提交
2267
    :type stride: int
2268 2269
    :param stride_y: stride height of pooling. It is equal to stride by default.
    :type stride_y: int|None
Z
zhangjinchao01 已提交
2270 2271
    :param layer_attr: Extra Layer attribute.
    :type layer_attr: ExtraLayerAttribute
2272 2273 2274 2275
    :param ceil_mode: Wether to use ceil mode to calculate output height and with.
                      Defalut is True. If set false, Otherwise use floor.

    :type ceil_mode: bool
D
dangqingqing 已提交
2276 2277
    :return: LayerOutput object.
    :rtype: LayerOutput
Z
zhangjinchao01 已提交
2278 2279 2280 2281 2282 2283 2284 2285 2286 2287
    """
    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'

2288
    type_name = pool_type.name + '-projection' \
Y
Yu Yang 已提交
2289
        if (
Y
Yu Yang 已提交
2290
        isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)) \
Y
Yu Yang 已提交
2291
        else pool_type.name
2292 2293 2294 2295 2296

    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
2297
    l = Layer(
Z
zhangjinchao01 已提交
2298 2299
        name=name,
        type=LayerType.POOL_LAYER,
Q
qijun 已提交
2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311
        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,
L
Luo Tao 已提交
2312
                    padding_y=padding_y))
Q
qijun 已提交
2313
        ],
2314
        ceil_mode=ceil_mode,
Q
qijun 已提交
2315 2316 2317 2318 2319 2320 2321
        **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
2322 2323


Q
qijun 已提交
2324 2325
@wrap_name_default("spp")
@layer_support()
Q
qijun 已提交
2326 2327 2328 2329 2330 2331
def spp_layer(input,
              name=None,
              num_channels=None,
              pool_type=None,
              pyramid_height=None,
              layer_attr=None):
Q
qijun 已提交
2332 2333 2334 2335 2336
    """
    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>`_.

L
Luo Tao 已提交
2337 2338 2339 2340
    The example usage is:

    ..  code-block:: python

2341 2342 2343
        spp = spp_layer(input=data,
                        pyramid_height=2,
                        num_channels=16,
L
Luo Tao 已提交
2344 2345
                        pool_type=MaxPooling())

Q
qijun 已提交
2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373
    :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 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
2374
    l = Layer(
Q
qijun 已提交
2375 2376
        name=name,
        type=LayerType.SPP_LAYER,
Q
qijun 已提交
2377 2378 2379 2380 2381
        inputs=Input(
            input.name,
            spp=SpatialPyramidPool(
                pool_type=type_name,
                channels=num_channels,
L
Luo Tao 已提交
2382
                pyramid_height=pyramid_height)),
Q
qijun 已提交
2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393
        **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
2394 2395 2396 2397
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters

X
xuwei06 已提交
2398
    l = Layer(
Q
qijun 已提交
2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417
        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
2418 2419 2420 2421


@wrap_name_default("crmnorm")
@layer_support()
Q
qijun 已提交
2422 2423 2424 2425 2426 2427
def img_cmrnorm_layer(input,
                      size,
                      scale=0.0128,
                      power=0.75,
                      name=None,
                      num_channels=None,
2428
                      layer_attr=None):
Z
zhangjinchao01 已提交
2429
    """
2430
    Response normalization across feature maps.
D
dangqingqing 已提交
2431 2432
    The details please refer to
    `Alex's paper <http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf>`_.
Z
zhangjinchao01 已提交
2433

L
Luo Tao 已提交
2434 2435 2436
    The example usage is:

    ..  code-block:: python
2437

L
Luo Tao 已提交
2438 2439
        norm = img_cmrnorm_layer(input=net, size=5)

Z
zhangjinchao01 已提交
2440
    :param name: layer name.
D
dangqingqing 已提交
2441
    :type name: None|basestring
Z
zhangjinchao01 已提交
2442 2443
    :param input: layer's input.
    :type input: LayerOutput
2444
    :param size: Normalize in number of :math:`size` feature maps.
Z
zhangjinchao01 已提交
2445
    :type size: int
D
dangqingqing 已提交
2446
    :param scale: The hyper-parameter.
Z
zhangjinchao01 已提交
2447
    :type scale: float
D
dangqingqing 已提交
2448
    :param power: The hyper-parameter.
Z
zhangjinchao01 已提交
2449 2450 2451 2452 2453
    :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
2454
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2455 2456 2457
    :rtype: LayerOutput
    """
    return __img_norm_layer__(name, input, size, "cmrnorm-projection", scale,
2458
                              power, num_channels, 0, layer_attr)
Z
zhangjinchao01 已提交
2459 2460 2461 2462 2463 2464 2465 2466


@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
2467 2468 2469 2470 2471 2472 2473
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
2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494
                     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>`_.

L
Luo Tao 已提交
2495 2496 2497
    The example usage is:

    ..  code-block:: python
2498

L
Luo Tao 已提交
2499 2500
        norm = batch_norm_layer(input=net, act=ReluActivation())

Z
zhangjinchao01 已提交
2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514
    :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.
2515
    :type batch_norm_type: None|string, None or "batch_norm" or "cudnn_batch_norm"
Z
zhangjinchao01 已提交
2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542
    :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
2543
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562
    :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
2563
    l = Layer(
Z
zhangjinchao01 已提交
2564
        name=name,
Q
qijun 已提交
2565 2566
        inputs=Input(
            input.name, image=Image(channels=num_channels), **param_attr.attr),
Z
zhangjinchao01 已提交
2567 2568 2569 2570 2571 2572
        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
2573
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
2574

Q
qijun 已提交
2575 2576 2577 2578 2579 2580 2581
    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
2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608


@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
2609
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2610 2611 2612 2613 2614 2615
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.SUM_TO_ONE_NORM_LAYER,
        inputs=[input.name],
Q
qijun 已提交
2616 2617 2618
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.SUM_TO_ONE_NORM_LAYER, parents=[input], size=input.size)
Z
zhangjinchao01 已提交
2619 2620 2621 2622 2623 2624


@wrap_name_default("addto")
@wrap_act_default(act=LinearActivation())
@wrap_bias_attr_default(has_bias=False)
@layer_support(DROPOUT)
Q
qijun 已提交
2625
def addto_layer(input, act=None, name=None, bias_attr=None, layer_attr=None):
Z
zhangjinchao01 已提交
2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647
    """
    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
2648 2649 2650
    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
2651 2652

    It is a very good way to set dropout outside the layers. Since not all
C
caoying03 已提交
2653 2654
    PaddlePaddle layer support dropout, you can add an add_to layer, set
    dropout here.
Z
zhangjinchao01 已提交
2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668
    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
2669
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2670 2671 2672 2673 2674 2675
    :rtype: LayerOutput
    """
    num_filters = None
    if isinstance(input, LayerOutput):
        input = [input]

2676
    assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
2677 2678 2679 2680 2681 2682 2683
    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
2684
    l = Layer(
Q
qijun 已提交
2685 2686 2687
        name=name,
        type=LayerType.ADDTO_LAYER,
        inputs=ipts_for_layer,
Z
zhangjinchao01 已提交
2688 2689
        bias=ParamAttr.to_bias(bias_attr),
        active_type=act.name,
Q
qijun 已提交
2690
        **ExtraLayerAttribute.to_kwargs(layer_attr))
2691

Q
qijun 已提交
2692 2693 2694 2695 2696 2697 2698
    return LayerOutput(
        name,
        LayerType.ADDTO_LAYER,
        parents=input,
        activation=act,
        num_filters=num_filters,
        size=l.config.size)
Z
zhangjinchao01 已提交
2699 2700 2701 2702 2703


@wrap_act_default(act=IdentityActivation())
@wrap_name_default("concat")
@layer_support()
2704
def concat_layer(input, act=None, name=None, layer_attr=None, bias_attr=None):
Z
zhangjinchao01 已提交
2705 2706 2707 2708
    """
    Concat all input vector into one huge vector.
    Inputs can be list of LayerOutput or list of projection.

2709 2710 2711 2712 2713 2714
    The example usage is:

    ..  code-block:: python

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

Z
zhangjinchao01 已提交
2715 2716 2717
    :param name: Layer name.
    :type name: basestring
    :param input: input layers or projections
2718
    :type input: list|tuple|collections.Sequence
Z
zhangjinchao01 已提交
2719 2720 2721 2722
    :param act: Activation type.
    :type act: BaseActivation
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
2723
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2724 2725 2726 2727 2728 2729 2730 2731
    :rtype: LayerOutput
    """

    if isinstance(input, LayerOutput):
        input = [input]
    elif isinstance(input, Projection):
        input = [input]
    else:
2732
        assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
2733 2734

    def __is_type__(o, tp):
2735
        if not isinstance(o, collections.Sequence):
Z
zhangjinchao01 已提交
2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756
            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
2757 2758
    is_concat_layer = __is_type__(
        reduce(__reduce_concat_type__, map(type, input)), LayerOutput)
Z
zhangjinchao01 已提交
2759

Q
qijun 已提交
2760 2761
    layer_type = (LayerType.CONCAT_LAYER
                  if is_concat_layer else LayerType.CONCAT_PROJ_LAYER)
Z
zhangjinchao01 已提交
2762

2763 2764
    if layer_type == LayerType.CONCAT_LAYER:
        assert not bias_attr
2765

Z
zhangjinchao01 已提交
2766
    Layer(
Q
qijun 已提交
2767 2768
        name=name,
        type=layer_type,
Z
zhangjinchao01 已提交
2769 2770
        inputs=[x.name for x in input] if is_concat_layer else input,
        active_type=act.name,
2771
        bias=ParamAttr.to_bias(bias_attr),
Q
qijun 已提交
2772
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
2773 2774 2775 2776 2777 2778 2779 2780 2781

    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
2782 2783 2784 2785 2786 2787 2788 2789
    return LayerOutput(
        name,
        layer_type=layer_type,
        parents=input if is_concat_layer else [x.origin for x in input],
        activation=act,
        size=sz)


2790 2791
@wrap_name_default("seqconcat")
@wrap_act_default(act=IdentityActivation())
2792
@wrap_bias_attr_default(has_bias=False)
2793 2794 2795 2796 2797
@layer_support()
def seq_concat_layer(a, b, act=None, name=None, layer_attr=None,
                     bias_attr=None):
    """
    Concat sequence a with sequence b.
2798

2799
    Inputs:
2800 2801 2802
      - a = [a1, a2, ..., an]
      - b = [b1, b2, ..., bn]
      - Note that the length of a and b should be the same.
2803

2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821
    Output: [a1, b1, a2, b2, ..., an, bn]

    The example usage is:

    ..  code-block:: python

        concat = seq_concat_layer(a=layer1, b=layer2)

    :param name: Layer name.
    :type name: basestring
    :param a: input sequence layer
    :type a: LayerOutput
    :param b: input sequence layer
    :type b: LayerOutput
    :param act: Activation type.
    :type act: BaseActivation
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
2822 2823 2824 2825
    :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
2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846
    :return: LayerOutput object.
    :rtype: LayerOutput
    """
    assert isinstance(a, LayerOutput) and isinstance(b, LayerOutput)
    assert a.size == b.size
    Layer(
        name=name,
        type=LayerType.SEQUENCE_CONCAT_LAYER,
        inputs=[a.name, b.name],
        active_type=act.name,
        bias=ParamAttr.to_bias(bias_attr),
        **ExtraLayerAttribute.to_kwargs(layer_attr))

    return LayerOutput(
        name,
        layer_type=LayerType.SEQUENCE_CONCAT_LAYER,
        parents=[a, b],
        activation=act,
        size=a.size)


2847
@wrap_name_default("memory", "memory_name")
Q
qijun 已提交
2848 2849
def memory(name,
           size,
2850
           memory_name=None,
Q
qijun 已提交
2851 2852 2853 2854
           is_seq=False,
           boot_layer=None,
           boot_bias=None,
           boot_bias_active_type=None,
Z
zhangjinchao01 已提交
2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874
           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.

2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891
    .. code-block:: python

       mem = memory(size=256, name='state')
       state = fc_layer(input=mem, size=256, name='state')

    If you do not want to specify the name, you can equivalently use set_input()
    to specify the layer needs to be remembered as the following:

    .. code-block:: python
       mem = memory(size=256)
       state = fc_layer(input=mem, size=256)
       mem.set_input(mem)


    :param name: the name of the layer which this memory remembers.
                 If name is None, user should call set_input() to specify the
                 name of the layer which this memory remembers.
Z
zhangjinchao01 已提交
2892 2893 2894
    :type name: basestring
    :param size: size of memory.
    :type size: int
2895 2896 2897
    :param memory_name: the name of the memory.
                        It is ignored when name is provided.
    :type memory_name: basestring
Z
zhangjinchao01 已提交
2898 2899 2900 2901 2902 2903 2904 2905 2906 2907
    :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
2908
    :return: LayerOutput object which is a memory.
Z
zhangjinchao01 已提交
2909 2910 2911 2912 2913 2914 2915 2916 2917 2918
    :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)
2919 2920
    if name is not None:
        memory_name = None
Z
zhangjinchao01 已提交
2921

2922 2923 2924 2925 2926 2927 2928 2929 2930
    memory_name = Memory(
        name,
        size,
        is_sequence=is_seq,
        boot_layer=boot_layer.name if boot_layer is not None else None,
        boot_bias=boot_bias,
        boot_bias_active_type=boot_bias_active_type.name,
        boot_with_const_id=boot_with_const_id,
        memory_name=memory_name)
Q
qijun 已提交
2931 2932

    lout = LayerOutput(
2933
        name=memory_name,
Q
qijun 已提交
2934 2935 2936
        size=size,
        layer_type=LayerType.MEMORY,
        parents=[boot_layer] if boot_layer is not None else None)
Z
zhangjinchao01 已提交
2937 2938 2939 2940
    return lout


@wrap_bias_attr_default()
Q
qijun 已提交
2941 2942
@wrap_act_default(
    param_names=['gate_act', 'state_act'], act=SigmoidActivation())
Z
zhangjinchao01 已提交
2943 2944 2945
@wrap_act_default(act=TanhActivation())
@wrap_name_default('lstm_step')
@layer_support()
Q
qijun 已提交
2946 2947 2948 2949 2950 2951 2952 2953 2954
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
2955 2956 2957 2958 2959 2960
    """
    LSTM Step Layer. It used in recurrent_group. The lstm equations are shown
    as follow.

    ..  math::

L
luotao02 已提交
2961
        i_t & = \\sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i)
Z
zhangjinchao01 已提交
2962

L
luotao02 已提交
2963
        f_t & = \\sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f)
Z
zhangjinchao01 已提交
2964

L
luotao02 已提交
2965
        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
2966

L
luotao02 已提交
2967
        o_t & = \\sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o)
Z
zhangjinchao01 已提交
2968

L
luotao02 已提交
2969
        h_t & = o_t tanh(c_t)
Z
zhangjinchao01 已提交
2970 2971


L
luotao02 已提交
2972
    The input of lstm step is :math:`Wx_t + Wh_{t-1}`, and user should use
Z
zhangjinchao01 已提交
2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010
    :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
3011
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3012 3013 3014 3015 3016 3017 3018 3019 3020
    :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
3021 3022 3023
        size=size,
        inputs=[input.name, state.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
3024

Q
qijun 已提交
3025 3026 3027 3028 3029 3030 3031
    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
3032 3033 3034


@wrap_bias_attr_default()
W
wangyang59 已提交
3035
@wrap_param_attr_default()
Q
qijun 已提交
3036
@wrap_act_default(param_names=['gate_act'], act=SigmoidActivation())
Z
zhangjinchao01 已提交
3037 3038 3039
@wrap_act_default(act=TanhActivation())
@wrap_name_default('gru_step')
@layer_support()
Q
qijun 已提交
3040 3041 3042 3043 3044 3045 3046
def gru_step_layer(input,
                   output_mem,
                   size=None,
                   act=None,
                   name=None,
                   gate_act=None,
                   bias_attr=None,
W
wangyang59 已提交
3047
                   param_attr=None,
Q
qijun 已提交
3048
                   layer_attr=None):
Z
zhangjinchao01 已提交
3049 3050 3051 3052 3053 3054 3055 3056 3057 3058
    """

    :param input:
    :type input: LayerOutput
    :param output_mem:
    :param size:
    :param act:
    :param name:
    :param gate_act:
    :param bias_attr:
3059 3060
    :param param_attr: the parameter_attribute for transforming the output_mem
                       from previous step.
Z
zhangjinchao01 已提交
3061
    :param layer_attr:
D
dangqingqing 已提交
3062
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3063 3064 3065 3066 3067 3068 3069 3070
    :rtype: LayerOutput
    """
    assert input.size % 3 == 0
    if size is None:
        size = input.size / 3
    Layer(
        name=name,
        type=LayerType.GRU_STEP_LAYER,
3071 3072 3073 3074
        # The parameter here is for transforming the output_mem. The input has
        # already been transformed outside this module so it does not need
        # parameter associated with it.
        # The parameter here is instead grouped with input is due to
3075
        # backward model compatibility.
3076
        inputs=[Input(input.name, **param_attr.attr), output_mem.name],
Z
zhangjinchao01 已提交
3077 3078 3079 3080
        bias=ParamAttr.to_bias(bias_attr),
        size=size,
        active_type=act.name,
        active_gate_type=gate_act.name,
Q
qijun 已提交
3081
        **ExtraAttr.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
3082
    return LayerOutput(
Q
qijun 已提交
3083 3084
        name=name,
        layer_type=LayerType.GRU_STEP_LAYER,
Z
zhangjinchao01 已提交
3085
        parents=[input, output_mem],
Q
qijun 已提交
3086 3087
        size=size,
        activation=act)
Z
zhangjinchao01 已提交
3088 3089


Y
Yu Yang 已提交
3090 3091 3092 3093
@wrap_bias_attr_default()
@wrap_param_attr_default()
@wrap_act_default(param_names=['gate_act'], act=SigmoidActivation())
@wrap_act_default(act=TanhActivation())
Q
qijun 已提交
3094
@wrap_name_default('gru_step_naive')
Y
Yu Yang 已提交
3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161
@layer_support(ERROR_CLIPPING, DROPOUT)
def gru_step_naive_layer(input,
                         output_mem,
                         size=None,
                         name=None,
                         act=None,
                         gate_act=None,
                         bias_attr=None,
                         param_attr=None,
                         layer_attr=None):
    """
    GRU Step Layer, but using MixedLayer to generate. It support ERROR_CLIPPING
    and DROPOUT.

    :param input:
    :param output_mem:
    :param size:
    :param name:
    :param act:
    :param gate_act:
    :param bias_attr:
    :param param_attr:
    :param layer_attr:
    :return:
    """
    if input.size % 3 != 0:
        raise ValueError("GruStep input size must be divided by 3")
    if size is None:
        size = input.size / 3

    def __gate__(gate_name, offset):
        with mixed_layer(
                name=name + "_" + gate_name,
                size=size,
                layer_attr=layer_attr,
                bias_attr=bias_attr,
                act=gate_act) as gate:
            gate += identity_projection(input=input, offset=offset)
            gate += full_matrix_projection(
                input=output_mem, param_attr=param_attr)
        return gate

    update_gate = __gate__("update", 0)
    reset_gate = __gate__("reset", size)

    with mixed_layer(
            name=name + "_reset_output", bias_attr=False) as reset_output:
        reset_output += dotmul_operator(a=output_mem, b=reset_gate)

    with mixed_layer(
            name=name + "_output_candidate",
            size=size,
            layer_attr=layer_attr,
            bias_attr=bias_attr,
            act=act) as output_candidate:
        output_candidate += identity_projection(input=input, offset=2 * size)
        output_candidate += full_matrix_projection(
            input=reset_output, param_attr=param_attr)

    with mixed_layer(name=name) as output:
        output += identity_projection(output_mem)
        output += dotmul_operator(a=output_mem, b=update_gate, scale=-1.0)
        output += dotmul_operator(a=output_candidate, b=update_gate)

    return output


Z
zhangjinchao01 已提交
3162 3163 3164 3165
@wrap_name_default()
@layer_support()
def get_output_layer(input, arg_name, name=None, layer_attr=None):
    """
C
caoying03 已提交
3166 3167 3168 3169
    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
3170 3171 3172 3173 3174 3175 3176 3177 3178

    :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
3179
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3180 3181 3182 3183 3184 3185 3186
    :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
3187 3188 3189 3190 3191 3192 3193
    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
3194

Q
qijun 已提交
3195 3196 3197 3198 3199
    return LayerOutput(
        name=name,
        layer_type=LayerType.GET_OUTPUT_LAYER,
        parents=[input],
        size=input.size)
Z
zhangjinchao01 已提交
3200 3201 3202 3203 3204 3205 3206


@wrap_name_default()
@wrap_act_default()
@wrap_bias_attr_default()
@wrap_param_attr_default()
@layer_support()
Q
qijun 已提交
3207 3208 3209 3210 3211 3212 3213
def recurrent_layer(input,
                    act=None,
                    bias_attr=None,
                    param_attr=None,
                    name=None,
                    reverse=False,
                    layer_attr=None):
Z
zhangjinchao01 已提交
3214
    """
3215 3216
    Simple recurrent unit layer. It is just a fully connect layer through both
    time and neural network.
Z
zhangjinchao01 已提交
3217

3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244
    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
3245
    :return: LayerOutput object.
3246
    :rtype: LayerOutput
Z
zhangjinchao01 已提交
3247
    """
Q
qijun 已提交
3248 3249 3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262
    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
3263 3264 3265 3266 3267 3268 3269


class StaticInput(object):
    """
    StaticInput is only used in recurrent_group which defines a read-only memory
    that can be a sequence or non-sequence.
    """
3270

Z
zhangjinchao01 已提交
3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289
    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)
    """
3290

Z
zhangjinchao01 已提交
3291 3292 3293 3294 3295 3296 3297
    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
3298 3299 3300 3301 3302
def recurrent_group(step,
                    input,
                    reverse=False,
                    name=None,
                    targetInlink=None,
L
Luo Tao 已提交
3303
                    is_generating=False):
Z
zhangjinchao01 已提交
3304
    """
C
caoying03 已提交
3305 3306 3307 3308 3309
    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
3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353

    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

3354 3355
    :param reverse: If reverse is set true, the recurrent unit will process the
                    input sequence in a reverse order.
Z
zhangjinchao01 已提交
3356
    :type reverse: bool
3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367

    :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
3368
    :param is_generating: If is generating, none of input type should be LayerOutput;
3369
                          else, for training or testing, one of the input type must
L
Luo Tao 已提交
3370
                          be LayerOutput.
L
Luo Tao 已提交
3371

L
Luo Tao 已提交
3372
    : type is_generating: bool
3373

D
dangqingqing 已提交
3374
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3375 3376 3377 3378 3379 3380 3381 3382 3383 3384
    :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]
3385
    assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
3386 3387 3388 3389 3390 3391

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

    in_links = filter(is_in_links, input)

3392 3393 3394 3395 3396 3397 3398 3399 3400
    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
3401
    assert (targetInlink == None or targetInlink_in_inlinks())
3402
    targetInlinkName = None if targetInlink == None \
Y
Yu Yang 已提交
3403 3404
        else targetInlink.name if isinstance(targetInlink, LayerOutput) \
        else targetInlink.input.name
3405

Z
zhangjinchao01 已提交
3406 3407 3408 3409 3410 3411 3412 3413 3414 3415
    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
3416 3417
        name=name,
        in_links=map(map_in_links, in_links),
3418 3419
        seq_reversed=reverse,
        target_inlinkname=targetInlinkName)
Z
zhangjinchao01 已提交
3420
    in_args = []
3421
    has_LayerOutput = False
Z
zhangjinchao01 已提交
3422 3423 3424 3425
    for each_input in input:
        assert is_single_input(each_input)
        if isinstance(each_input, LayerOutput):
            in_args.append(each_input)
3426
            has_LayerOutput = True
Z
zhangjinchao01 已提交
3427 3428
        elif isinstance(each_input, SubsequenceInput):
            in_args.append(each_input.input)
3429
            has_LayerOutput = True
Z
zhangjinchao01 已提交
3430 3431
        else:
            mem_name = "__%s_memory__" % each_input.input.name
Q
qijun 已提交
3432 3433 3434 3435 3436 3437 3438 3439 3440
            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
3441 3442 3443
                mix += identity_projection(mem)
            in_args.append(mem)

L
Luo Tao 已提交
3444
    assert (is_generating != has_LayerOutput)
L
Luo Tao 已提交
3445

Z
zhangjinchao01 已提交
3446 3447 3448 3449 3450 3451 3452
    layer_outs = step(*in_args)

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

    for ot in layer_outs:
        assert isinstance(ot, LayerOutput)
3453
        ot.reverse = reverse
Z
zhangjinchao01 已提交
3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465
        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

3466

Z
zhangjinchao01 已提交
3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483
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
3484 3485 3486 3487 3488 3489 3490 3491 3492
        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
3493 3494 3495
        return trg_emb

    def __init__(self, size, embedding_name, embedding_size):
3496
        super(GeneratedInput, self).__init__()
Z
zhangjinchao01 已提交
3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519
        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
3520
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3521 3522 3523 3524
    :rtype: LayerOutput
    """

    assert isinstance(input, LayerOutput)
Q
qijun 已提交
3525 3526 3527 3528 3529 3530 3531 3532 3533 3534
    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
3535

3536

H
Haonan 已提交
3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557 3558 3559 3560 3561 3562
@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
3563 3564 3565 3566 3567 3568 3569 3570 3571 3572
    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)
3573

Z
zhangjinchao01 已提交
3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589

@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
3590 3591
    :param name: Layer name.
    :type name: basestring
Z
zhangjinchao01 已提交
3592 3593 3594 3595 3596 3597
    :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
3598
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3599 3600
    :rtype: LayerOutput
    """
Q
qijun 已提交
3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611
    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
3612 3613 3614


@wrap_name_default()
Q
qijun 已提交
3615 3616 3617 3618 3619 3620 3621
def beam_search(step,
                input,
                bos_id,
                eos_id,
                beam_size,
                max_length=500,
                name=None,
Z
zhangjinchao01 已提交
3622
                num_results_per_sample=None):
3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633
    """
    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)
3634
            with mixed_layer(size=512, name='rnn') as simple_rnn:
3635 3636 3637 3638
                simple_rnn += full_matrix_projection(input)
                simple_rnn += last_time_step_output
            return simple_rnn

3639 3640 3641 3642 3643
        generated_word_embedding = GeneratedInput(
                               size=target_dictionary_dim,
                               embedding_name="target_language_embedding",
                               embedding_size=word_vector_dim)

3644 3645
        beam_gen = beam_search(name="decoder",
                               step=rnn_step,
3646 3647
                               input=[StaticInput(encoder_last),
                                      generated_word_embedding],
3648 3649
                               bos_id=0,
                               eos_id=1,
3650
                               beam_size=5)
3651 3652 3653 3654 3655 3656 3657 3658 3659

    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
3660
                 step, and it is applied to sequences with arbitrary length by
3661 3662 3663 3664 3665
                 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
3666 3667
    :param input: Input data for the recurrent unit, which should include the
                  previously generated words as a GeneratedInput object.
3668
    :type input: list
3669 3670 3671
    :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
3672
                   symbol is essential, since it is used to initialize the RNN
3673 3674 3675 3676 3677 3678 3679 3680
                   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
3681 3682
    :param max_length: Max generated sequence length.
    :type max_length: int
3683 3684 3685 3686 3687 3688 3689 3690 3691 3692
    :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
3693 3694
    :return: The generated word index.
    :rtype: LayerOutput
3695 3696
    """

Z
zhangjinchao01 已提交
3697 3698 3699 3700 3701
    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
3702
    if isinstance(input, StaticInput) or isinstance(input, BaseGeneratedInput):
Z
zhangjinchao01 已提交
3703 3704 3705 3706 3707 3708
        input = [input]

    generated_input_index = -1

    real_input = []
    for i, each_input in enumerate(input):
3709 3710
        assert isinstance(each_input, StaticInput) or isinstance(
            each_input, BaseGeneratedInput)
Z
zhangjinchao01 已提交
3711 3712 3713 3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726
        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
3727 3728 3729 3730 3731 3732
        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
3733 3734 3735 3736 3737 3738 3739 3740 3741 3742

        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
3743
    tmp = recurrent_group(
L
Luo Tao 已提交
3744 3745 3746 3747
        step=__real_step__,
        input=real_input,
        reverse=False,
        name=name,
L
Luo Tao 已提交
3748
        is_generating=True)
3749

Z
zhangjinchao01 已提交
3750 3751
    return tmp

Q
qijun 已提交
3752

3753 3754
def __cost_input__(input, label, weight=None):
    """
3755
    inputs and parents for cost layers.
3756 3757 3758 3759
    """
    ipts = [Input(input.name), Input(label.name)]
    parents = [input, label]
    if weight is not None:
3760
        assert weight.size == 1
3761 3762 3763
        ipts.append(Input(weight.name))
        parents.append(weight)
    return ipts, parents
3764

Z
zhangjinchao01 已提交
3765 3766

@wrap_name_default()
L
luotao1 已提交
3767
@layer_support()
3768
def mse_cost(input, label, weight=None, name=None, coeff=1.0, layer_attr=None):
Z
zhangjinchao01 已提交
3769
    """
L
Luo Tao 已提交
3770 3771 3772 3773
    mean squared error cost:

    ..  math::

L
Luo Tao 已提交
3774
        \frac{1}{N}\sum_{i=1}^N(t_i-y_i)^2
Z
zhangjinchao01 已提交
3775 3776

    :param name: layer name.
3777
    :type name: basestring
Z
zhangjinchao01 已提交
3778
    :param input: Network prediction.
3779
    :type input: LayerOutput
Z
zhangjinchao01 已提交
3780
    :param label: Data label.
3781 3782 3783 3784
    :type label: LayerOutput
    :param weight: The weight affects the cost, namely the scale of cost.
                   It is an optional argument.
    :type weight: LayerOutput
3785 3786
    :param coeff: The coefficient affects the gradient in the backward.
    :type coeff: float
L
luotao1 已提交
3787 3788
    :param layer_attr: layer's extra attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
3789
    :return: LayerOutput object.
3790
    :rtype: LayerOutput
Z
zhangjinchao01 已提交
3791
    """
3792 3793
    ipts, parents = __cost_input__(input, label, weight)

Q
qijun 已提交
3794 3795 3796 3797
    Layer(
        inputs=ipts,
        type="square_error",
        name=name,
3798
        coeff=coeff,
Q
qijun 已提交
3799
        **ExtraLayerAttribute.to_kwargs(layer_attr))
X
xuwei06 已提交
3800
    return LayerOutput(name, LayerType.COST, parents=parents, size=1)
Z
zhangjinchao01 已提交
3801 3802


L
Luo Tao 已提交
3803 3804 3805
regression_cost = mse_cost


Z
zhangjinchao01 已提交
3806
@wrap_name_default("cost")
3807
@layer_support()
Q
qijun 已提交
3808 3809 3810 3811
def classification_cost(input,
                        label,
                        weight=None,
                        name=None,
L
Liang Zhao 已提交
3812
                        top_k=None,
3813 3814
                        evaluator=classification_error_evaluator,
                        layer_attr=None):
Z
zhangjinchao01 已提交
3815 3816 3817 3818 3819 3820 3821 3822 3823
    """
    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
3824 3825 3826
    :param weight: The weight affects the cost, namely the scale of cost.
                   It is an optional argument.
    :type weight: LayerOutput
L
Liang Zhao 已提交
3827 3828
    :param top_k: number k in top-k error rate
    :type top_k: int
Z
zhangjinchao01 已提交
3829
    :param evaluator: Evaluator method.
3830 3831
    :param layer_attr: layer's extra attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
3832
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3833 3834 3835 3836 3837
    :rtype: LayerOutput
    """
    assert input.layer_type != LayerType.DATA
    assert isinstance(input.activation, SoftmaxActivation)
    assert label.layer_type == LayerType.DATA
3838 3839 3840

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

Q
qijun 已提交
3841 3842 3843 3844 3845
    Layer(
        name=name,
        type="multi-class-cross-entropy",
        inputs=ipts,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
3846 3847 3848 3849 3850 3851 3852 3853 3854 3855

    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

L
Liang Zhao 已提交
3856
        e(name=e.__name__, input=input, label=label, weight=weight, top_k=top_k)
Z
zhangjinchao01 已提交
3857

3858
    if not isinstance(evaluator, collections.Sequence):
Z
zhangjinchao01 已提交
3859 3860 3861 3862 3863
        evaluator = [evaluator]

    for each_evaluator in evaluator:
        __add_evaluator__(each_evaluator)

X
xuwei06 已提交
3864
    return LayerOutput(name, LayerType.COST, parents=parents, size=1)
Z
zhangjinchao01 已提交
3865

3866

Q
qijun 已提交
3867 3868 3869 3870 3871 3872 3873 3874 3875
def conv_operator(img,
                  filter,
                  filter_size,
                  num_filters,
                  num_channels=None,
                  stride=1,
                  padding=0,
                  filter_size_y=None,
                  stride_y=None,
3876 3877
                  padding_y=None,
                  trans=False):
Z
zhangjinchao01 已提交
3878 3879 3880 3881 3882 3883 3884 3885 3886 3887
    """
    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

3888 3889
       op = conv_operator(img=input1,
                          filter=input2,
3890
                          filter_size=3,
Z
zhangjinchao01 已提交
3891 3892 3893
                          num_filters=64,
                          num_channels=64)

3894 3895 3896 3897
    :param img: input image
    :type img: LayerOutput
    :param filter: input filter
    :type filter: LayerOutput
Z
zhangjinchao01 已提交
3898 3899
    :param filter_size: The x dimension of a filter kernel.
    :type filter_size: int
C
caoying03 已提交
3900 3901 3902
    :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
3903
    :type filter_size_y: int
3904 3905
    :param num_filters: channel of output data.
    :type num_filters: int
3906 3907
    :param num_channels: channel of input data.
    :type num_channels: int
Z
zhangjinchao01 已提交
3908
    :param stride: The x dimension of the stride.
L
luotao02 已提交
3909
    :type stride: int
Z
zhangjinchao01 已提交
3910
    :param stride_y: The y dimension of the stride.
L
luotao02 已提交
3911
    :type stride_y: int
Z
zhangjinchao01 已提交
3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924
    :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
3925

3926 3927
    if num_channels is None:
        num_channels = img.num_filters
3928 3929 3930

    assert isinstance(filter, LayerOutput)
    if filter.size is not None:
3931
        filter.size = filter_size * filter_size_y * num_filters * num_channels
3932

3933 3934 3935
    opCls = ConvTransOperator if trans else ConvOperator

    op = opCls(
Q
qijun 已提交
3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946
        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))
3947

3948
    op.origin = [img, filter]
Z
zhangjinchao01 已提交
3949 3950
    return op

Q
qijun 已提交
3951

3952
@wrap_param_attr_default()
Q
qijun 已提交
3953 3954 3955 3956 3957 3958 3959 3960 3961 3962
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,
3963 3964
                    param_attr=None,
                    trans=False):
3965 3966 3967 3968 3969 3970 3971 3972 3973
    """
    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

D
dangqingqing 已提交
3974
       proj = conv_projection(input=input1,
3975 3976 3977 3978 3979 3980 3981 3982 3983 3984 3985 3986 3987 3988
                              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
3989 3990
    :param num_channels: channel of input data.
    :type num_channels: int
3991 3992 3993 3994 3995 3996 3997 3998 3999 4000 4001 4002
    :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
4003 4004
    :param trans: whether it is convTrans or conv
    :type trans: boolean
4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031 4032 4033 4034
    :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
4035
        init_w = (2.0 / (filter_size**2 * num_channels))**0.5
4036 4037 4038 4039 4040
        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

4041 4042 4043
    projCls = ConvTransProjection if trans else ConvProjection

    proj = projCls(
Q
qijun 已提交
4044 4045 4046 4047 4048 4049 4050 4051 4052 4053 4054 4055
        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)
4056 4057 4058 4059

    proj.origin = input
    return proj

Z
zhangjinchao01 已提交
4060

D
dangqingqing 已提交
4061 4062 4063 4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077
@wrap_name_default("pad")
@layer_support()
def pad_layer(input,
              pad_c=None,
              pad_h=None,
              pad_w=None,
              name=None,
              layer_attr=None):
    """
    This operation pads zeros to the input data according to pad_c,pad_h
    and pad_w. pad_c, pad_h, pad_w specifies the which dimension and size
    of padding. And the input data shape is NCHW.

    For example, pad_c=[2,3] means padding 2 zeros before the
    input data and 3 zeros after the input data in channel dimension.
    pad_h means padding zeros in height dimension. pad_w means padding zeros
    in width dimension.
4078

D
dangqingqing 已提交
4079
    For example,
4080

4081 4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095 4096 4097 4098 4099 4100 4101
    .. code-block:: python

       input(2,2,2,3)  = [
                           [ [[1,2,3], [3,4,5]],
                             [[2,3,5], [1,6,7]] ],
                           [ [[4,3,1], [1,8,7]],
                             [[3,8,9], [2,3,5]] ]
                         ]

       pad_c=[1,1], pad_h=[0,0], pad_w=[0,0]

       output(2,4,2,3) = [
                           [ [[0,0,0], [0,0,0]],
                             [[1,2,3], [3,4,5]],
                             [[2,3,5], [1,6,7]],
                             [[0,0,0], [0,0,0]] ],
                           [ [[0,0,0], [0,0,0]],
                             [[4,3,1], [1,8,7]],
                             [[3,8,9], [2,3,5]],
                             [[0,0,0], [0,0,0]] ]
                         ]
D
dangqingqing 已提交
4102 4103

    The simply usage is:
D
dangqingqing 已提交
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

    .. code-block:: python

       pad = pad_layer(input=ipt,
                       pad_c=[4,4],
                       pad_h=[0,0],
                       pad_w=[2,2])

    :param input: layer's input.
    :type input: LayerOutput
    :param pad_c: padding size in channel dimension.
    :type pad_c: list|None
    :param pad_h: padding size in height dimension.
    :type pad_h: list|None
    :param pad_w: padding size in width dimension.
    :type pad_w: list|None
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
    :param name: layer name.
    :type name: basestring
    :return: LayerOutput object.
    :rtype: LayerOutput
    """
    if pad_c is not None:
        assert isinstance(pad_c, collections.Sequence) and len(pad_c) == 2
    else:
        pad_c = [0, 0]

    if pad_h is not None:
        assert isinstance(pad_h, collections.Sequence) and len(pad_h) == 2
    else:
        pad_h = [0, 0]

    if pad_w is not None:
        assert isinstance(pad_w, collections.Sequence) and len(pad_w) == 2
    else:
        pad_w = [0, 0]

    assert input.num_filters is not None
    in_ch = input.num_filters
    out_ch = in_ch + pad_c[0] + pad_c[1]

    l = Layer(
        name=name,
        type=LayerType.PAD_LAYER,
        inputs=Input(
            input.name,
            pad=Pad(
                channels=in_ch,
                pad_c=pad_c,
                pad_h=pad_h,
                pad_w=pad_w, )),
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        layer_type=LayerType.PAD_LAYER,
        parents=[input],
        num_filters=out_ch,
        size=l.config.size)


Z
zhangjinchao01 已提交
4165
@wrap_name_default()
L
luotao1 已提交
4166 4167
@layer_support()
def conv_shift_layer(a, b, name=None, layer_attr=None):
Z
zhangjinchao01 已提交
4168 4169 4170 4171 4172 4173 4174 4175 4176 4177 4178
    """
    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:
4179 4180 4181 4182
     - 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
4183 4184 4185 4186 4187

    The example usage is:

    .. code-block:: python

L
Luo Tao 已提交
4188
       conv_shift = conv_shift_layer(a=layer1, b=layer2)
Z
zhangjinchao01 已提交
4189 4190 4191

    :param name: layer name
    :type name: basestring
4192 4193
    :param a: Input layer a.
    :type a: LayerOutput
L
Luo Tao 已提交
4194
    :param b: input layer b.
4195
    :type b: LayerOutput
L
luotao1 已提交
4196 4197
    :param layer_attr: layer's extra attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
4198
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4199 4200
    :rtype: LayerOutput
    """
4201 4202
    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
4203 4204 4205
    Layer(
        name=name,
        type=LayerType.CONV_SHIFT_LAYER,
4206
        inputs=[a.name, b.name],
Q
qijun 已提交
4207
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
4208

Q
qijun 已提交
4209 4210
    return LayerOutput(
        name, LayerType.CONV_SHIFT_LAYER, parents=[a, b], size=a.size)
Z
zhangjinchao01 已提交
4211 4212 4213 4214 4215


@wrap_name_default()
@wrap_param_attr_default()
@wrap_bias_attr_default()
4216
@wrap_act_default(act=LinearActivation())
Z
zhangjinchao01 已提交
4217
@layer_support(ERROR_CLIPPING, DROPOUT)
Q
qijun 已提交
4218 4219 4220 4221 4222 4223 4224 4225
def tensor_layer(a,
                 b,
                 size,
                 act=None,
                 name=None,
                 param_attr=None,
                 bias_attr=None,
                 layer_attr=None):
Z
zhangjinchao01 已提交
4226 4227 4228 4229 4230
    """
    This layer performs tensor operation for two input.
    For example, each sample:

    .. math::
4231
       y_{i} = a * W_{i} * {b^\mathrm{T}}, i=0,1,...,K-1
Z
zhangjinchao01 已提交
4232 4233

    In this formular:
4234 4235
      - :math:`a`: the first input contains M elements.
      - :math:`b`: the second input contains N elements.
Z
zhangjinchao01 已提交
4236 4237
      - :math:`y_{i}`: the i-th element of y.
      - :math:`W_{i}`: the i-th learned weight, shape if [M, N]
4238
      - :math:`b^\mathrm{T}`: the transpose of :math:`b_{2}`.
Z
zhangjinchao01 已提交
4239 4240 4241 4242 4243

    The simple usage is:

    .. code-block:: python

4244
       tensor = tensor_layer(a=layer1, b=layer2, size=1000)
Z
zhangjinchao01 已提交
4245 4246 4247

    :param name: layer name
    :type name: basestring
4248 4249 4250 4251
    :param a: Input layer a.
    :type a: LayerOutput
    :param b: input layer b.
    :type b: LayerOutput
Z
zhangjinchao01 已提交
4252
    :param size: the layer dimension.
L
luotao02 已提交
4253
    :type size: int.
Z
zhangjinchao01 已提交
4254 4255 4256
    :param act: Activation Type. Default is tanh.
    :type act: BaseActivation
    :param param_attr: The Parameter Attribute.
4257
    :type param_attr: ParameterAttribute
Z
zhangjinchao01 已提交
4258 4259 4260 4261 4262 4263
    :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
4264
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4265 4266
    :rtype: LayerOutput
    """
4267
    assert isinstance(a, LayerOutput) and isinstance(b, LayerOutput)
Z
zhangjinchao01 已提交
4268 4269 4270 4271 4272 4273
    Layer(
        name=name,
        size=size,
        type=LayerType.TENSOR_LAYER,
        active_type=act.name,
        bias=ParamAttr.to_bias(bias_attr),
Q
qijun 已提交
4274 4275 4276 4277
        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
4278 4279 4280 4281 4282 4283


@wrap_name_default()
@wrap_param_attr_default()
@wrap_bias_attr_default()
@wrap_act_default()
L
luotao1 已提交
4284
@layer_support()
Q
qijun 已提交
4285 4286
def selective_fc_layer(input,
                       size,
L
Luo Tao 已提交
4287
                       select=None,
Q
qijun 已提交
4288 4289
                       act=None,
                       name=None,
Z
zhangjinchao01 已提交
4290 4291 4292
                       pass_generation=False,
                       has_selected_colums=True,
                       mul_ratio=0.02,
Q
qijun 已提交
4293 4294 4295
                       param_attr=None,
                       bias_attr=None,
                       layer_attr=None):
Z
zhangjinchao01 已提交
4296 4297 4298 4299 4300 4301 4302 4303 4304 4305
    """
    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

4306
       sel_fc = selective_fc_layer(input=input, size=128, act=TanhActivation())
Z
zhangjinchao01 已提交
4307 4308 4309 4310 4311

    :param name: The Layer Name.
    :type name: basestring
    :param input: The input layer.
    :type input: LayerOutput|list|tuple
4312 4313
    :param select: The select layer. The output of select layer should be a
                   sparse binary matrix, and treat as the mask of selective fc.
L
Luo Tao 已提交
4314
                   If is None, acts exactly like fc_layer.
4315
    :type select: LayerOutput
Z
zhangjinchao01 已提交
4316 4317 4318 4319 4320 4321 4322 4323 4324 4325 4326 4327
    :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
4328
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4329 4330 4331 4332
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
4333
        assert not isinstance(param_attr, collections.Sequence)
Z
zhangjinchao01 已提交
4334 4335
        param_attr = [param_attr]
    else:
4336
        if isinstance(param_attr, collections.Sequence):
Z
zhangjinchao01 已提交
4337 4338 4339 4340
            assert len(input) == len(param_attr)
        else:
            param_attr = [copy.deepcopy(param_attr) for _ in range(len(input))]

4341 4342 4343 4344
    assert isinstance(input, collections.Sequence)
    assert isinstance(select, LayerOutput)
    if select.size is not None:
        assert select.size == size
Z
zhangjinchao01 已提交
4345
    Layer(
Q
qijun 已提交
4346 4347 4348
        inputs=[
            Input(ipt.name, **attr.attr) for ipt, attr in zip(input, param_attr)
        ] + [select.name],
Z
zhangjinchao01 已提交
4349 4350 4351
        name=name,
        type=LayerType.SEL_FC_LAYER,
        size=size,
4352
        bias=ParameterAttribute.to_bias(bias_attr),
Z
zhangjinchao01 已提交
4353 4354 4355 4356
        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
4357 4358 4359 4360 4361 4362 4363
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.SEL_FC_LAYER,
        list(input) + [select],
        activation=act,
        size=size)
Z
zhangjinchao01 已提交
4364 4365 4366


@wrap_name_default()
L
luotao1 已提交
4367 4368
@layer_support()
def sampling_id_layer(input, name=None, layer_attr=None):
Z
zhangjinchao01 已提交
4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382
    """
    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
4383 4384
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4385
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4386 4387
    :rtype: LayerOutput
    """
X
xuwei06 已提交
4388
    l = Layer(
Z
zhangjinchao01 已提交
4389 4390 4391
        name=name,
        type=LayerType.SAMPLING_ID_LAYER,
        inputs=[Input(input.name)],
Q
qijun 已提交
4392 4393 4394
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.SAMPLING_ID_LAYER, input, size=l.config.size)
Z
zhangjinchao01 已提交
4395 4396 4397


@wrap_name_default()
L
luotao1 已提交
4398
@layer_support()
Q
qijun 已提交
4399 4400 4401 4402
def slope_intercept_layer(input,
                          name=None,
                          slope=1.0,
                          intercept=0.0,
L
luotao1 已提交
4403
                          layer_attr=None):
Z
zhangjinchao01 已提交
4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424
    """
    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
4425 4426
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4427
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4428 4429 4430 4431 4432 4433 4434 4435
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.SLOPE_INTERCEPT_LAYER,
        slope=slope,
        intercept=intercept,
        inputs=[Input(input.name)],
Q
qijun 已提交
4436 4437 4438
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.SLOPE_INTERCEPT_LAYER, input, size=input.size)
Z
zhangjinchao01 已提交
4439 4440 4441


@wrap_name_default()
L
luotao1 已提交
4442
@layer_support()
Q
qijun 已提交
4443
def linear_comb_layer(weights, vectors, size=None, name=None, layer_attr=None):
Z
zhangjinchao01 已提交
4444
    """
4445 4446 4447 4448
    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
4449 4450 4451

    .. math::

4452
       z(i) = \sum_{j=0}^{M-1} x(j) y(i+Nj)
4453

4454 4455 4456 4457 4458
    where :math:`0 \le i \le N-1`

    Or in the matrix notation:

    .. math::
Z
zhangjinchao01 已提交
4459

4460
       z = x^\mathrm{T} Y
Z
zhangjinchao01 已提交
4461 4462

    In this formular:
4463 4464 4465 4466 4467 4468
      - :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
4469 4470 4471 4472 4473

    The simple usage is:

    .. code-block:: python

4474
       linear_comb = linear_comb_layer(weights=weight, vectors=vectors,
Z
zhangjinchao01 已提交
4475 4476
                                       size=elem_dim)

4477 4478 4479 4480
    :param weights: The weight layer.
    :type weights: LayerOutput
    :param vectors: The vector layer.
    :type vectors: LayerOutput
Z
zhangjinchao01 已提交
4481 4482 4483 4484
    :param size: the dimension of this layer.
    :type size: int
    :param name: The Layer Name.
    :type name: basestring
L
luotao1 已提交
4485 4486
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4487
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4488 4489
    :rtype: LayerOutput
    """
4490 4491 4492 4493
    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
4494
            size = vectors.size / weights.size
4495 4496
        else:
            assert size == vectors.size / weights.size
Z
zhangjinchao01 已提交
4497 4498
    Layer(
        name=name,
4499
        type=LayerType.LINEAR_COMBINATION_LAYER,
Z
zhangjinchao01 已提交
4500
        size=size,
4501
        inputs=[Input(weights.name), Input(vectors.name)],
Q
qijun 已提交
4502 4503 4504
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.LINEAR_COMBINATION_LAYER, [weights, vectors], size=size)
4505

4506

4507
convex_comb_layer = linear_comb_layer
Z
zhangjinchao01 已提交
4508

4509

Z
zhangjinchao01 已提交
4510
@wrap_name_default()
L
luotao1 已提交
4511
@layer_support()
Z
zhangjinchao01 已提交
4512 4513 4514 4515 4516 4517 4518
def block_expand_layer(input,
                       block_x=0,
                       block_y=0,
                       stride_x=0,
                       stride_y=0,
                       padding_x=0,
                       padding_y=0,
4519
                       num_channels=None,
L
luotao1 已提交
4520 4521
                       name=None,
                       layer_attr=None):
Z
zhangjinchao01 已提交
4522 4523
    """
    Expand feature map to minibatch matrix.
4524
       - matrix width is: block_y * block_x * num_channels
L
luotao02 已提交
4525
       - matirx height is: outputH * outputW
Z
zhangjinchao01 已提交
4526 4527 4528 4529 4530 4531 4532 4533 4534 4535

    .. 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
4536
    time step is block_y * block_x * num_channels. This layer can be used after
Z
zhangjinchao01 已提交
4537 4538
    convolution neural network, and before recurrent neural network.

4539 4540 4541 4542
    The simple usage is:

    .. code-block:: python

L
Luo Tao 已提交
4543
       block_expand = block_expand_layer(input=layer,
4544
                                         num_channels=128,
4545 4546 4547 4548 4549
                                         stride_x=1,
                                         stride_y=1,
                                         block_x=1,
                                         block_x=3)

Z
zhangjinchao01 已提交
4550 4551
    :param input: The input layer.
    :type input: LayerOutput
4552 4553
    :param num_channels: The channel number of input layer.
    :type num_channels: int|None
Z
zhangjinchao01 已提交
4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567
    :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
4568 4569
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4570
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4571 4572
    :rtype: LayerOutput
    """
4573 4574 4575
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters
Q
qijun 已提交
4576 4577 4578 4579 4580 4581 4582 4583 4584 4585 4586 4587 4588 4589 4590 4591 4592
    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
4593 4594


4595 4596
@wrap_name_default()
@layer_support()
4597
def maxout_layer(input, groups, num_channels=None, name=None, layer_attr=None):
4598 4599 4600 4601 4602
    """
    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.

4603
    So groups should be larger than 1, and the num of channels should be able
4604 4605
    to devided by groups.

4606
    Please refer to Paper:
4607 4608 4609 4610
      - 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
4611

4612 4613 4614 4615 4616 4617 4618 4619 4620 4621 4622 4623 4624 4625 4626 4627 4628 4629 4630 4631 4632 4633 4634 4635 4636 4637 4638 4639 4640
    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 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
4641 4642 4643 4644 4645 4646 4647 4648 4649
    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)
4650 4651


Z
zhangjinchao01 已提交
4652
@wrap_name_default()
L
luotao1 已提交
4653
@layer_support()
Q
qijun 已提交
4654 4655 4656 4657 4658
def ctc_layer(input,
              label,
              size=None,
              name=None,
              norm_by_times=False,
L
luotao1 已提交
4659
              layer_attr=None):
Z
zhangjinchao01 已提交
4660 4661 4662 4663 4664
    """
    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.

4665 4666
    More details can be found by referring to `Connectionist Temporal
    Classification: Labelling Unsegmented Sequence Data with Recurrent
4667 4668
    Neural Networks <http://machinelearning.wustl.edu/mlpapers/paper_files/
    icml2006_GravesFGS06.pdf>`_
4669 4670 4671 4672 4673 4674 4675 4676

    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
4677 4678 4679 4680 4681 4682 4683 4684 4685
    The simple usage:

    .. code-block:: python

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

4686
    :param input: The input layer.
Z
zhangjinchao01 已提交
4687 4688 4689
    :type input: LayerOutput
    :param label: The data layer of label with variable length.
    :type label: LayerOutput
4690
    :param size: category numbers + 1.
Z
zhangjinchao01 已提交
4691
    :type size: int
4692 4693
    :param name: The name of this layer
    :type name: basestring|None
Z
zhangjinchao01 已提交
4694 4695
    :param norm_by_times: Whether to normalization by times. False by default.
    :type norm_by_times: bool
L
luotao1 已提交
4696 4697
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4698
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4699 4700 4701 4702
    :rtype: LayerOutput
    """
    assert isinstance(input, LayerOutput)
    assert isinstance(label, LayerOutput)
4703 4704 4705 4706 4707
    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
4708
    Layer(
4709 4710 4711 4712
        name=name,
        type=LayerType.CTC_LAYER,
        size=size,
        norm_by_times=norm_by_times,
L
luotao1 已提交
4713
        inputs=[input.name, label.name],
Q
qijun 已提交
4714
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
4715 4716
    return LayerOutput(name, LayerType.CTC_LAYER, [input, label], size=size)

4717

4718 4719 4720 4721 4722 4723 4724 4725 4726 4727 4728 4729 4730 4731 4732 4733 4734 4735 4736 4737 4738 4739 4740
@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'
4741 4742 4743 4744 4745
          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 any value ranged in [0, num_classes], which
          should be consistent as that used in your labels.
4746
        - As a native 'softmax' activation is interated to the warp-ctc library,
L
Luo Tao 已提交
4747
          'linear' activation is expected instead in the 'input' layer.
4748 4749 4750 4751 4752 4753 4754 4755 4756 4757 4758 4759 4760 4761 4762 4763 4764 4765 4766 4767 4768 4769 4770 4771 4772 4773 4774 4775 4776 4777 4778 4779 4780 4781 4782 4783 4784 4785 4786 4787 4788 4789 4790 4791 4792 4793 4794

    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
4795
@wrap_name_default()
4796
@wrap_param_attr_default()
L
luotao1 已提交
4797
@layer_support()
Q
qijun 已提交
4798 4799 4800 4801 4802 4803
def crf_layer(input,
              label,
              size=None,
              weight=None,
              param_attr=None,
              name=None,
4804
              coeff=1.0,
L
luotao1 已提交
4805
              layer_attr=None):
Z
zhangjinchao01 已提交
4806 4807 4808 4809 4810 4811 4812 4813 4814 4815 4816 4817 4818 4819 4820
    """
    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.
4821
    :type label: LayerOutput
Z
zhangjinchao01 已提交
4822 4823 4824 4825 4826 4827 4828 4829 4830
    :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
4831 4832
    :param coeff: The coefficient affects the gradient in the backward.
    :type coeff: float
L
luotao1 已提交
4833 4834
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4835
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4836 4837 4838 4839 4840
    :rtype: LayerOutput
    """
    assert isinstance(input, LayerOutput)
    assert isinstance(label, LayerOutput)
    assert weight is None or isinstance(weight, LayerOutput)
4841 4842 4843 4844 4845 4846
    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
4847

Q
qijun 已提交
4848
    ipts = [Input(input.name, **param_attr.attr), Input(label.name)]
Z
zhangjinchao01 已提交
4849 4850 4851 4852
    if weight is not None:
        ipts.append(Input(weight.name))

    Layer(
4853 4854 4855 4856
        name=name,
        type=LayerType.CRF_LAYER,
        size=size,
        inputs=ipts,
4857
        coeff=coeff,
Q
qijun 已提交
4858
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
4859 4860 4861
    parents = [input, label]
    if weight is not None:
        parents.append(weight)
X
xuwei06 已提交
4862 4863 4864 4865
    # 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
4866

4867

Z
zhangjinchao01 已提交
4868
@wrap_name_default()
4869
@wrap_param_attr_default()
L
luotao1 已提交
4870
@layer_support()
Q
qijun 已提交
4871 4872 4873 4874 4875
def crf_decoding_layer(input,
                       size,
                       label=None,
                       param_attr=None,
                       name=None,
L
luotao1 已提交
4876
                       layer_attr=None):
Z
zhangjinchao01 已提交
4877 4878 4879 4880 4881 4882 4883
    """
    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.

L
Luo Tao 已提交
4884 4885 4886 4887 4888 4889 4890
    The simple usage:

    .. code-block:: python

      crf_decoding = crf_decoding_layer(input=input,
                                        size=label_dim)

Z
zhangjinchao01 已提交
4891 4892 4893 4894 4895 4896 4897 4898 4899 4900
    :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
4901 4902
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4903
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4904 4905 4906 4907 4908 4909
    :rtype: LayerOutput
    """

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

4910
    ipts = [Input(input.name, **param_attr.attr)]
Z
zhangjinchao01 已提交
4911 4912 4913 4914
    if label is not None:
        ipts.append(Input(label.name))

    Layer(
4915 4916 4917 4918
        name=name,
        type=LayerType.CRF_DECODING_LAYER,
        size=size,
        inputs=ipts,
Q
qijun 已提交
4919
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
4920 4921 4922
    parents = [input]
    if label is not None:
        parents.append(label)
X
xuwei06 已提交
4923 4924 4925 4926
    # 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
4927

Q
qijun 已提交
4928

Y
Yu Yang 已提交
4929
@wrap_act_default(act=SigmoidActivation())
4930
@wrap_bias_attr_default(has_bias=True)
4931
@wrap_param_attr_default()
4932 4933
@wrap_name_default()
@layer_support()
Q
qijun 已提交
4934 4935
def nce_layer(input,
              label,
C
caoying03 已提交
4936
              num_classes=None,
Y
Yu Yang 已提交
4937
              act=None,
4938
              param_attr=None,
Q
qijun 已提交
4939 4940 4941 4942 4943 4944
              weight=None,
              num_neg_samples=10,
              neg_distribution=None,
              name=None,
              bias_attr=None,
              layer_attr=None):
4945 4946 4947 4948 4949 4950 4951 4952 4953
    """
    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

C
caoying03 已提交
4954 4955
       cost = nce_layer(input=[layer1, layer2], label=layer2,
                        param_attr=[attr1, attr2], weight=layer3,
4956 4957 4958 4959 4960 4961 4962 4963 4964 4965 4966
                        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.
4967
    :type num_classes: int
Y
Yu Yang 已提交
4968 4969
    :param act: Activation, default is Sigmoid.
    :type act: BaseActivation
4970 4971
    :param param_attr: The Parameter Attribute|list.
    :type param_attr: ParameterAttribute
4972
    :param num_neg_samples: number of negative samples. Default is 10.
4973
    :type num_neg_samples: int
4974 4975 4976 4977 4978 4979 4980 4981 4982 4983 4984 4985 4986
    :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]
4987 4988 4989 4990 4991 4992 4993 4994
        assert not isinstance(param_attr, collections.Sequence)
        param_attr = [param_attr]
    else:
        if isinstance(param_attr, collections.Sequence):
            assert len(input) == len(param_attr)
        else:
            param_attr = [copy.deepcopy(param_attr) for _ in range(len(input))]

4995
    assert isinstance(input, collections.Sequence)
4996

4997 4998
    assert isinstance(label, LayerOutput)
    assert label.layer_type == LayerType.DATA
C
caoying03 已提交
4999 5000
    if num_classes is None:
        num_classes = label.size
5001 5002 5003
    if neg_distribution is not None:
        assert isinstance(neg_distribution, collections.Sequence)
        assert len(neg_distribution) == num_classes
5004
        assert abs(sum(neg_distribution) - 1.0) < 1e-5
Y
Yu Yang 已提交
5005 5006
    if not isinstance(act, BaseActivation):
        raise TypeError()
5007

5008 5009
    ipts_for_layer = []
    parents = []
5010
    for each_input, attr in zip(input, param_attr):
5011
        assert isinstance(each_input, LayerOutput)
5012
        ipts_for_layer.append(Input(each_input.name, **attr.attr))
5013 5014 5015 5016 5017 5018 5019 5020 5021 5022
        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
5023
    l = Layer(
5024 5025 5026 5027
        name=name,
        type=LayerType.NCE_LAYER,
        num_classes=num_classes,
        neg_sampling_dist=neg_distribution,
Y
Yu Yang 已提交
5028
        active_type=act.name,
5029 5030 5031
        num_neg_samples=num_neg_samples,
        inputs=ipts_for_layer,
        bias=ParamAttr.to_bias(bias_attr),
Q
qijun 已提交
5032 5033
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
Y
Yu Yang 已提交
5034 5035 5036 5037 5038
        name,
        LayerType.NCE_LAYER,
        parents=parents,
        size=l.config.size,
        activation=act)
Q
qijun 已提交
5039

5040

Z
zhangjinchao01 已提交
5041 5042 5043
"""
following are cost Layers.
"""
5044 5045


Z
zhangjinchao01 已提交
5046
@wrap_name_default()
L
luotao1 已提交
5047
@layer_support()
Q
qijun 已提交
5048 5049 5050 5051 5052 5053 5054
def rank_cost(left,
              right,
              label,
              weight=None,
              name=None,
              coeff=1.0,
              layer_attr=None):
Z
zhangjinchao01 已提交
5055
    """
5056
    A cost Layer for learning to rank using gradient descent. Details can refer
5057 5058
    to `papers <http://research.microsoft.com/en-us/um/people/cburges/papers/
    ICML_ranking.pdf>`_.
Z
zhangjinchao01 已提交
5059 5060 5061 5062 5063
    This layer contains at least three inputs. The weight is an optional
    argument, which affects the cost.

    .. math::

L
luotao02 已提交
5064
       C_{i,j} & = -\\tilde{P_{ij}} * o_{i,j} + log(1 + e^{o_{i,j}})
Z
zhangjinchao01 已提交
5065

L
luotao02 已提交
5066
       o_{i,j} & =  o_i - o_j
Z
zhangjinchao01 已提交
5067

L
luotao02 已提交
5068
       \\tilde{P_{i,j}} & = \\{0, 0.5, 1\\} \ or \ \\{0, 1\\}
Z
zhangjinchao01 已提交
5069 5070 5071 5072 5073 5074 5075 5076 5077 5078 5079 5080 5081 5082 5083 5084 5085 5086 5087 5088 5089 5090 5091 5092 5093 5094 5095 5096 5097

    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
5098 5099
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
5100
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
5101 5102 5103 5104 5105 5106 5107 5108 5109 5110 5111 5112
    :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
5113 5114 5115 5116 5117 5118
    Layer(
        name=name,
        type=LayerType.RANK_COST,
        inputs=ipts,
        coeff=coeff,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
5119

X
xuwei06 已提交
5120
    return LayerOutput(name, LayerType.RANK_COST, parents=parents, size=1)
Z
zhangjinchao01 已提交
5121

5122

Z
zhangjinchao01 已提交
5123
@wrap_name_default()
L
luotao1 已提交
5124
@layer_support()
Q
qijun 已提交
5125 5126 5127 5128 5129 5130
def lambda_cost(input,
                score,
                name,
                NDCG_num=5,
                max_sort_size=-1,
                layer_attr=None):
Z
zhangjinchao01 已提交
5131 5132 5133 5134 5135 5136 5137 5138 5139 5140 5141 5142
    """
    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)

5143
    :param input: Samples of the same query should be loaded as sequence.
Z
zhangjinchao01 已提交
5144 5145 5146 5147 5148 5149 5150 5151 5152 5153 5154
    :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
5155 5156 5157
                          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
5158 5159 5160
    :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
5161 5162
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
5163
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
5164 5165
    :rtype: LayerOutput
    """
5166 5167 5168
    assert isinstance(input, LayerOutput) and isinstance(score, LayerOutput)
    if score.size is not None:
        assert score.size == 1
Q
qijun 已提交
5169 5170 5171 5172 5173 5174 5175
    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
5176

Q
qijun 已提交
5177 5178
    return LayerOutput(
        name, LayerType.LAMBDA_COST, parents=[input, score], size=1)
Z
zhangjinchao01 已提交
5179

5180

Z
zhangjinchao01 已提交
5181
@wrap_name_default()
L
luotao1 已提交
5182
@layer_support()
5183 5184 5185 5186 5187 5188
def cross_entropy(input,
                  label,
                  name=None,
                  coeff=1.0,
                  weight=None,
                  layer_attr=None):
Z
zhangjinchao01 已提交
5189 5190 5191 5192 5193
    """
    A loss layer for multi class entropy.

    .. code-block:: python

X
xuwei06 已提交
5194
       cost = cross_entropy(input=input_layer,
L
Luo Tao 已提交
5195
                            label=label_layer)
Z
zhangjinchao01 已提交
5196 5197 5198 5199 5200 5201 5202

    :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.
5203 5204
    :param coeff: The cost is multiplied with coeff.
                  The coefficient affects the gradient in the backward.
Z
zhangjinchao01 已提交
5205
    :type coeff: float.
5206 5207 5208 5209
    :param weight: The cost of each sample is multiplied with each weight.
                   The weight should be a layer with size=1. Note that gradient
                   will not be calculated for weight.
    :type weight: LayerOutout
L
luotao1 已提交
5210 5211
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
5212
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
5213 5214 5215
    :rtype: LayerOutput.
    """

5216
    ipts, parents = __cost_input__(input, label, weight)
Q
qijun 已提交
5217 5218 5219
    Layer(
        name=name,
        type=LayerType.CROSS_ENTROPY,
5220
        inputs=ipts,
Q
qijun 已提交
5221 5222
        coeff=coeff,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
5223
    return LayerOutput(name, LayerType.CROSS_ENTROPY, parents=parents, size=1)
Z
zhangjinchao01 已提交
5224

5225

Z
zhangjinchao01 已提交
5226
@wrap_name_default()
L
luotao1 已提交
5227
@layer_support()
Q
qijun 已提交
5228 5229 5230 5231
def cross_entropy_with_selfnorm(input,
                                label,
                                name=None,
                                coeff=1.0,
L
luotao1 已提交
5232 5233
                                softmax_selfnorm_alpha=0.1,
                                layer_attr=None):
Z
zhangjinchao01 已提交
5234 5235
    """
    A loss layer for multi class entropy with selfnorm.
5236
    Input should be a vector of positive numbers, without normalization.
Z
zhangjinchao01 已提交
5237 5238 5239

    .. code-block:: python

X
xuwei06 已提交
5240
       cost = cross_entropy_with_selfnorm(input=input_layer,
L
Luo Tao 已提交
5241
                                          label=label_layer)
Z
zhangjinchao01 已提交
5242 5243 5244 5245 5246 5247 5248 5249 5250 5251 5252

    :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
5253 5254
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
5255
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
5256 5257
    :rtype: LayerOutput.
    """
Q
qijun 已提交
5258 5259 5260 5261 5262 5263 5264
    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
5265

Q
qijun 已提交
5266 5267 5268 5269 5270
    return LayerOutput(
        name,
        LayerType.CROSS_ENTROPY_WITH_SELFNORM,
        parents=[input, label],
        size=1)
Z
zhangjinchao01 已提交
5271

5272

X
xuwei06 已提交
5273 5274 5275 5276 5277 5278 5279 5280
@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
5281
       cost = sum_cost(input=input_layer)
X
xuwei06 已提交
5282 5283 5284 5285 5286 5287 5288 5289 5290 5291

    :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
5292
    assert isinstance(input, LayerOutput)
Q
qijun 已提交
5293 5294 5295 5296 5297
    Layer(
        name=name,
        type=LayerType.SUM_COST,
        inputs=[input.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
X
xuwei06 已提交
5298

Q
qijun 已提交
5299
    return LayerOutput(name, LayerType.SUM_COST, parents=[input], size=1)
X
xuwei06 已提交
5300 5301


Z
zhangjinchao01 已提交
5302
@wrap_name_default()
L
luotao1 已提交
5303 5304
@layer_support()
def huber_cost(input, label, name=None, coeff=1.0, layer_attr=None):
Z
zhangjinchao01 已提交
5305 5306 5307 5308 5309
    """
    A loss layer for huber loss.

    .. code-block:: python

X
xuwei06 已提交
5310
       cost = huber_cost(input=input_layer,
L
Luo Tao 已提交
5311
                         label=label_layer)
Z
zhangjinchao01 已提交
5312 5313 5314 5315 5316 5317 5318 5319 5320

    :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
5321 5322
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
5323
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
5324 5325
    :rtype: LayerOutput.
    """
5326 5327 5328
    assert isinstance(input, LayerOutput)
    if input.size is not None:
        assert input.size == 1
Q
qijun 已提交
5329 5330 5331 5332 5333 5334
    Layer(
        name=name,
        type=LayerType.HUBER,
        inputs=[input.name, label.name],
        coeff=coeff,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
X
xuwei06 已提交
5335
    return LayerOutput(name, LayerType.HUBER, parents=[input, label], size=1)
Z
zhangjinchao01 已提交
5336

5337

Z
zhangjinchao01 已提交
5338
@wrap_name_default()
L
luotao1 已提交
5339
@layer_support()
Q
qijun 已提交
5340 5341 5342 5343
def multi_binary_label_cross_entropy(input,
                                     label,
                                     name=None,
                                     coeff=1.0,
L
luotao1 已提交
5344
                                     layer_attr=None):
Z
zhangjinchao01 已提交
5345 5346 5347 5348 5349
    """
    A loss layer for multi binary label cross entropy.

    .. code-block:: python

X
xuwei06 已提交
5350
       cost = multi_binary_label_cross_entropy(input=input_layer,
L
Luo Tao 已提交
5351
                                               label=label_layer)
Z
zhangjinchao01 已提交
5352 5353 5354 5355 5356 5357 5358 5359 5360

    :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
5361 5362
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
5363
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
5364 5365 5366
    :rtype: LayerOutput
    """

5367 5368
    if input.activation is None or \
            not isinstance(input.activation, SigmoidActivation):
Q
qijun 已提交
5369 5370 5371 5372 5373 5374 5375 5376 5377 5378 5379 5380 5381 5382 5383 5384
        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)
D
dangqingqing 已提交
5385 5386 5387 5388


@wrap_name_default()
@layer_support()
5389
def smooth_l1_cost(input, label, name=None, coeff=1.0, layer_attr=None):
D
dangqingqing 已提交
5390 5391
    """
    This is a L1 loss but more smooth. It requires that the
D
dangqingqing 已提交
5392
    size of input and label are equal. The formula is as follows,
D
dangqingqing 已提交
5393 5394 5395 5396 5397 5398 5399 5400 5401

    .. math::

        L = \sum_{i} smooth_{L1}(input_i - label_i)

    in which

    .. math::

5402
        smooth_{L1}(x) = \\begin{cases} 0.5x^2& \\text{if}  \\ |x| < 1 \\\\ |x|-0.5& \\text{otherwise} \end{cases}
D
dangqingqing 已提交
5403

D
dangqingqing 已提交
5404 5405 5406
    More details can be found by referring to `Fast R-CNN
    <https://arxiv.org/pdf/1504.08083v2.pdf>`_

D
dangqingqing 已提交
5407 5408
    .. code-block:: python

5409 5410
       cost = smooth_l1_cost(input=input_layer,
                             label=label_layer)
D
dangqingqing 已提交
5411 5412 5413 5414 5415 5416 5417

    :param input: The 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
5418 5419
    :param coeff: The coefficient affects the gradient in the backward.
    :type coeff: float
D
dangqingqing 已提交
5420 5421 5422 5423 5424 5425 5426 5427 5428 5429 5430 5431 5432
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
    :return: LayerOutput object.
    :rtype: LayerOutput
    """
    assert isinstance(input, LayerOutput)
    assert isinstance(label, LayerOutput)
    assert input.size == label.size

    Layer(
        name=name,
        type=LayerType.SMOOTH_L1,
        inputs=[input.name, label.name],
5433
        coeff=coeff,
D
dangqingqing 已提交
5434 5435 5436
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.SMOOTH_L1, parents=[input, label], size=1)