layers.py 137.6 KB
Newer Older
Z
zhangjinchao01 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
# Copyright (c) 2016 Baidu, Inc. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

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

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

Z
zhangjinchao01 已提交
26 27 28 29 30 31 32
try:
    import cPickle as pickle
except ImportError:
    import pickle
import copy

__all__ = ["full_matrix_projection", "AggregateLevel", "ExpandLevel",
33
           "identity_projection", "dotmul_projection", "dotmul_operator",
Z
zhangjinchao01 已提交
34 35 36
           "table_projection", "mixed_layer", "data_layer",
           "embedding_layer", "fc_layer", "grumemory",
           "pooling_layer", "lstmemory", "last_seq", "first_seq",
37
           "cos_sim", "hsigmoid", "conv_projection",
Z
zhangjinchao01 已提交
38
           "regression_cost", 'classification_cost', "LayerOutput",
39
           'img_conv_layer', 'img_convTrans_layer', 'img_pool_layer', 'batch_norm_layer',
40
           'img_cmrnorm_layer', 'addto_layer',
Z
zhangjinchao01 已提交
41 42 43 44 45 46 47 48 49 50
           'concat_layer', 'lstm_step_layer', 'recurrent_group',
           'memory', 'StaticInput', 'expand_layer', 'scaling_layer',
           'power_layer', 'interpolation_layer', 'trans_layer',
           'sum_to_one_norm_layer',
           'get_output_layer', 'LayerType', 'context_projection',
           'beam_search', 'maxid_layer', 'GeneratedInput', 'SubsequenceInput',
           'gru_step_layer', 'recurrent_layer',
           'BaseGeneratedInput', 'conv_operator', 'conv_shift_layer',
           'tensor_layer', 'selective_fc_layer', 'sampling_id_layer',
           'slope_intercept_layer', 'trans_full_matrix_projection',
51
           'linear_comb_layer',
Z
zhangjinchao01 已提交
52
           'convex_comb_layer', 'ctc_layer', 'crf_layer', 'crf_decoding_layer',
53
           'nce_layer',
Z
zhangjinchao01 已提交
54 55 56
           'cross_entropy_with_selfnorm', 'cross_entropy',
           'multi_binary_label_cross_entropy',
           'rank_cost', 'lambda_cost', 'huber_cost',
57
           'block_expand_layer',
58
           'maxout_layer', 'out_prod_layer', 'print_layer'
Z
zhangjinchao01 已提交
59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76
           ]


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

    DATA = "data"
    MIXED_LAYER = "mixed"
    LSTMEMORY = "lstmemory"
    GRUMEMORY = "gated_recurrent"
    SEQUENCE_LAST_INSTANCE = "seqlastins"
    SEQUENCE_FIRST_INSTANCE = "seqfirstins"
    POOLING_MAX = "max"
    POOLING_AVG = 'average'
    FC_LAYER = "fc"
    COST = 'cost'
77 78
    COSINE_SIM_VEC = 'cos_vm'
    COSINE_SIM = 'cos'
Z
zhangjinchao01 已提交
79 80
    HSIGMOID = 'hsigmoid'
    CONV_LAYER = "conv"
81
    CONVTRANS_LAYER = "convt"
Z
zhangjinchao01 已提交
82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99
    POOL_LAYER = "pool"
    BATCH_NORM_LAYER = 'batch_norm'
    NORM_LAYER = 'norm'
    SUM_TO_ONE_NORM_LAYER = 'sum_to_one_norm'
    ADDTO_LAYER = 'addto'

    CONCAT_LAYER = 'concat'
    CONCAT_PROJ_LAYER = 'concat2'

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

    EXPAND_LAYER = 'expand'
    INTERPOLATION_LAYER = 'interpolation'
    POWER_LAYER = 'power'
    SCALING_LAYER = 'scaling'
    TRANS_LAYER = 'trans'
H
Haonan 已提交
100
    OUT_PROD_LAYER = 'out_prod'
Z
zhangjinchao01 已提交
101 102 103 104 105 106 107 108 109 110 111

    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"
112
    LINEAR_COMBINATION_LAYER = "convex_comb"
Z
zhangjinchao01 已提交
113
    BLOCK_EXPAND = "blockexpand"
114
    MAXOUT = "maxout"
Z
zhangjinchao01 已提交
115

116 117
    PRINT_LAYER = "print"

Z
zhangjinchao01 已提交
118 119 120
    CTC_LAYER = "ctc"
    CRF_LAYER = "crf"
    CRF_DECODING_LAYER = "crf_decoding"
121
    NCE_LAYER = 'nce'
Z
zhangjinchao01 已提交
122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174

    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"

    @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.
175
    :type parents: list|tuple|collections.Sequence
Z
zhangjinchao01 已提交
176 177 178
    """

    def __init__(self, name, layer_type, parents=None, activation=None,
179 180
                 num_filters=None, img_norm_type=None, size=None, outputs=None,
                 reverse=None):
Z
zhangjinchao01 已提交
181 182 183 184 185
        assert isinstance(name, basestring)
        assert isinstance(layer_type, basestring)
        assert LayerType.is_layer_type(layer_type)
        self.name = name
        self.layer_type = layer_type
186 187
        if parents is not None and type(parents) != list:
            parents = [parents]
Z
zhangjinchao01 已提交
188 189 190 191 192 193 194 195
        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
196
        self.reverse = reverse
Z
zhangjinchao01 已提交
197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212

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

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


ERROR_CLIPPING = 'error_clipping_threshold'
DROPOUT = 'drop_rate'
213
DEVICE = 'device'
Z
zhangjinchao01 已提交
214 215 216


def layer_support(*attrs):
217
    attrs_list = list(attrs)
218
    attrs_list.append(DEVICE)
Z
zhangjinchao01 已提交
219 220 221
    def decorator(method):
        @functools.wraps(method)
        def wrapper(*args, **kwargs):
222
            for attr in attrs_list:
Z
zhangjinchao01 已提交
223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284
                for each in args:
                    if isinstance(each, ExtraLayerAttribute):
                        setattr(each, '_'.join(['can', attr]), True)
                for key in kwargs:
                    val = kwargs[key]
                    if isinstance(val, ExtraLayerAttribute):
                        setattr(val, '_'.join(['can', attr]), True)
            for each in args:
                if isinstance(each, ExtraLayerAttribute):
                    each.check(method.__name__)
            for key in kwargs:
                val = kwargs[key]
                if isinstance(val, ExtraLayerAttribute):
                    val.check(method.__name__)
            return method(*args, **kwargs)

        return wrapper

    return decorator


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

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

    There are two styles of usage.

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

    .. code-block:: python

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

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

    .. code-block:: python

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

    :param input: input layer
    :type input: LayerOutput
    :param size: The parameter size. Means the width of parameter.
    :type size: int
    :param param_attr: Parameter config, None if use default.
    :type param_attr: ParameterAttribute
    :return: A FullMatrixProjection Object.
    :rtype: FullMatrixProjection
    """
    proj = FullMatrixProjection(input_layer_name=input.name,
                                size=size,
                                **param_attr.attr)
    proj.origin = input
    return proj


285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321
@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
    """
    proj = TransposedFullMatrixProjection(input_layer_name=input.name,
                                          size=size,
                                          **param_attr.attr)
    proj.origin = input
    return proj


Z
zhangjinchao01 已提交
322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398
@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
    """
    proj = TableProjection(input_layer_name=input.name,
                           size=size,
                           **param_attr.attr)
    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.
399
    :type input: LayerOutput
Z
zhangjinchao01 已提交
400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415
    :param offset: Offset, None if use default.
    :type offset: int
    :return: A IdentityProjection or IdentityOffsetProjection Object
    :rtype: IdentityProjection or IdentityOffsetProjection
    """
    if offset is None:
        proj = IdentityProjection(input_layer_name=input.name)
        proj.origin = input
    else:
        proj = IdentityOffsetProjection(input_layer_name=input.name,
                                        offset=offset)
        proj.origin = input
    return proj


@wrap_param_attr_default()
416
def dotmul_projection(input, param_attr=None):
Z
zhangjinchao01 已提交
417
    """
418
    DotMulProjection with a layer as input.
Z
zhangjinchao01 已提交
419 420 421 422 423 424 425 426 427 428 429 430 431
    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)

432 433 434 435 436 437 438 439
    :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
    """
    proj = DotMulProjection(input_layer_name=input.name,
440 441
                            size=input.size,
                            **param_attr.attr)
442
    proj.origin = input
443
    return proj
Z
zhangjinchao01 已提交
444

445 446

def dotmul_operator(a=None, b=None, scale=1, **kwargs):
447 448
    """
    DotMulOperator takes two inputs and performs element-wise multiplication:
449

Z
zhangjinchao01 已提交
450
    .. math::
451 452
       out.row[i] += scale * (x.row[i] .* y.row[i])

Z
zhangjinchao01 已提交
453 454
    where :math:`.*` means element-wise multiplication, and
    scale is a config scalar, its default value is one.
455

Z
zhangjinchao01 已提交
456
    The example usage is:
457

Z
zhangjinchao01 已提交
458
    .. code-block:: python
459 460 461

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

462 463 464 465
    :param a: Input layer1
    :type a: LayerOutput
    :param b: Input layer2
    :type b: LayerOutput
Z
zhangjinchao01 已提交
466 467
    :param scale: config scalar, default value is one.
    :type scale: float
468 469
    :return: A DotMulOperator Object.
    :rtype: DotMulOperator
Z
zhangjinchao01 已提交
470
    """
471 472 473 474 475 476 477 478 479 480 481
    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.')
    a = kwargs.get('x', a)    # For Backward capacity.
    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

    op = DotMulOperator(input_layer_names=[a.name, b.name],
482
                        scale=scale)
483
    op.origin = [a, b]
484
    return op
Z
zhangjinchao01 已提交
485

486

Z
zhangjinchao01 已提交
487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567
@wrap_bias_attr_default(['padding_attr'])
def context_projection(input, context_len, context_start=None,
                       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

    proj = ContextProjection(input_layer_name=input.name,
                             context_length=context_len,
                             context_start=context_start,
                             trainable_padding=trainable,
                             **extra_dict)
    proj.origin = input
    return proj


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

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

    def __init__(self, name, size, act, bias_attr, layer_attr,
                 parents=None):
        """
        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
        """
        LayerOutput.__init__(self, name, LayerType.MIXED_LAYER, parents,
                             size=size, activation=act)
        self.bias_attr = bias_attr
        self.layer_attr = layer_attr
        self.inputs = []
        self.finalized = False

568
    def __iadd__(self, other):
Z
zhangjinchao01 已提交
569 570 571 572 573 574 575 576
        """
        + += operator
        :param other: Other projection.
        :type other: Projection
        :return: self.
        :rtype: MixedLayerType
        """
        if not self.finalized:
577
            assert isinstance(other, Projection) or isinstance(other, Operator)
Z
zhangjinchao01 已提交
578
            self.inputs.append(other)
579 580 581 582
            if isinstance(other, Projection):
                self.parents.append(other.origin)
            else:
                self.parents.extend(other.origin)
Z
zhangjinchao01 已提交
583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607
            return self
        else:
            raise MixedLayerType.AddToSealedMixedLayerException()

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

    def __exit__(self, *args, **kwargs):
        del args, kwargs  # unused parameter to suppress warning
        assert len(self.inputs) != 0
        MixedLayer(
            name=self.name,
            size=self.size,
            active_type=self.activation.name,
            bias=ParamAttr.to_bias(self.bias_attr),
            inputs=self.inputs,
            **ExtraLayerAttribute.to_kwargs(self.layer_attr)
        )


@wrap_name_default("mixed")
@wrap_act_default(act=LinearActivation())
@wrap_bias_attr_default(has_bias=False)
@layer_support(ERROR_CLIPPING, DROPOUT)
608
def mixed_layer(size=0, input=None, name=None, act=None, bias_attr=False,
Z
zhangjinchao01 已提交
609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654
                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:
        with mixed_layer(name=name, size=size, act=act, bias_attr=bias_attr,
                         layer_attr=layer_attr) as m:
655
            if isinstance(input, collections.Sequence):
Z
zhangjinchao01 已提交
656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680
                for each in input:
                    m += each
            else:
                m += input
        return m


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

    The example usage is:

    ..  code-block:: python

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

    :param name: Name of this data layer.
    :type name: basestring
    :param size: Size of this data layer.
    :type size: int
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
681
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707
    :rtype: LayerOutput
    """
    Layer(type=LayerType.DATA, name=name, size=size,
          **ExtraLayerAttribute.to_kwargs(layer_attr))

    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
708
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736
    :rtype: LayerOutput
    """
    with mixed_layer(name=name, size=size, act=LinearActivation(),
                     bias_attr=False,
                     layer_attr=layer_attr) as mix:
        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)
def fc_layer(input, size, act=None, name=None,
             param_attr=None, bias_attr=None, layer_attr=None):
    """
    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
737
    which is equal to:
Z
zhangjinchao01 已提交
738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759

    .. 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
760
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
761 762 763 764
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
765
        assert not isinstance(param_attr, collections.Sequence)
Z
zhangjinchao01 已提交
766 767
        param_attr = [param_attr]
    else:
768
        if isinstance(param_attr, collections.Sequence):
Z
zhangjinchao01 已提交
769 770 771 772
            assert len(input) == len(param_attr)
        else:
            param_attr = [copy.deepcopy(param_attr) for _ in range(len(input))]

773
    assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
774 775

    Layer(
776 777
        inputs=[Input(ipt.name, **attr.attr) for ipt, attr in zip(
            input, param_attr)],
Z
zhangjinchao01 已提交
778 779 780 781 782 783 784 785 786 787
        name=name,
        type=LayerType.FC_LAYER,
        size=size,
        bias=ParamAttr.to_bias(bias_attr),
        active_type=act.name,
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.FC_LAYER, input, activation=act,
                       size=size)

788

789 790 791 792
@wrap_name_default("print")
def print_layer(input, name=None):
    """
    Print the output value of input layers. This layer is useful for debugging.
793 794 795 796 797

    :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
798
    :return: LayerOutput
799
    """
800 801 802 803 804
    if isinstance(input, LayerOutput):
        input = [input]
    assert isinstance(input, collections.Sequence)  # list or tuple
    for each in input:
        assert isinstance(each, LayerOutput)
805 806 807 808 809 810

    Layer(
        name=name,
        type=LayerType.PRINT_LAYER,
        inputs=[l.name for l in input],
    )
811
    # this layer don't return anything, can not be input of other layer.
812

Z
zhangjinchao01 已提交
813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831

@wrap_name_default("seq_pooling")
@wrap_bias_attr_default(has_bias=False)
@wrap_param_default(['pooling_type'], default_factory=lambda _: MaxPooling())
@layer_support()
def pooling_layer(input, pooling_type=None, name=None, bias_attr=None,
                  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
832 833
    :param agg_level: AggregateLevel.EACH_TIMESTEP or
                      AggregateLevel.EACH_SEQUENCE
Z
zhangjinchao01 已提交
834 835 836 837 838 839 840 841 842 843 844 845
    :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
846
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
847 848 849
    :rtype: LayerType
    """
    extra_dict = dict()
850
    # noinspection PyUnresolvedReferences
Z
zhangjinchao01 已提交
851 852
    if isinstance(pooling_type, AvgPooling):
        extra_dict['average_strategy'] = pooling_type.strategy
853 854 855 856
    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
857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879
    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,
        **extra_dict
    )

    return LayerOutput(name, pooling_type.name, parents=[input],
                       size=input.size)


@wrap_bias_attr_default()
@wrap_param_attr_default()
@wrap_act_default(param_names=['gate_act'],
                  act=SigmoidActivation())
@wrap_act_default(param_names=["act", 'state_act'], act=TanhActivation())
@wrap_name_default("lstmemory")
@layer_support(DROPOUT)
def lstmemory(input, name=None, reverse=False, act=None,
880
              gate_act=None, size=None,
Z
zhangjinchao01 已提交
881 882 883 884 885 886 887 888 889
              state_act=None, bias_attr=None, param_attr=None,
              layer_attr=None):
    """
    Long Short-term Memory Cell.

    The memory cell was implemented as follow equations.

    ..  math::

L
luotao02 已提交
890
        i_t & = \\sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i)
Z
zhangjinchao01 已提交
891

L
luotao02 已提交
892
        f_t & = \\sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f)
Z
zhangjinchao01 已提交
893

L
luotao02 已提交
894
        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
895

L
luotao02 已提交
896
        o_t & = \\sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o)
Z
zhangjinchao01 已提交
897

L
luotao02 已提交
898
        h_t & = o_t tanh(c_t)
Z
zhangjinchao01 已提交
899 900


C
caoying03 已提交
901
    NOTE: In PaddlePaddle's implementation, the multiplications
Z
zhangjinchao01 已提交
902
    :math:`W_{xi}x_{t}` , :math:`W_{xf}x_{t}`,
C
caoying03 已提交
903 904 905 906
    :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
907

C
caoying03 已提交
908
    NOTE: This is a low level user interface. You can use network.simple_lstm
Z
zhangjinchao01 已提交
909 910
    to config a simple plain lstm layer.

C
caoying03 已提交
911 912 913 914
    Please refer to **Generating Sequences With Recurrent Neural Networks** for
    more details about LSTM.

    Link_ goes as below.
Z
zhangjinchao01 已提交
915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937

    .. _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
938
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
939 940 941 942 943 944
    :rtype: LayerOutput
    """

    assert gate_act.support_hppl
    assert state_act.support_hppl
    assert act.support_hppl
945 946 947 948 949 950 951 952 953 954
    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
955 956 957 958 959 960 961 962 963 964 965

    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))

966 967 968
    return LayerOutput(name, LayerType.LSTMEMORY, [input], size=input.size / 4,
                       reverse=reverse)

Z
zhangjinchao01 已提交
969 970 971 972 973 974 975 976 977

@wrap_bias_attr_default()
@wrap_param_attr_default()
@wrap_act_default(param_names=['gate_act'],
                  act=SigmoidActivation())
@wrap_act_default(param_names=["act"], act=TanhActivation())
@wrap_name_default("gru")
@layer_support(DROPOUT)
def grumemory(input, name=None, reverse=False, act=None,
978
              gate_act=None, size=None,
Z
zhangjinchao01 已提交
979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000
              bias_attr=None, param_attr=None,
              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
1001 1002
    3. The candidate activation :math:`\\tilde{h_t}` is computed similarly to
    that of the traditional recurrent unit:
Z
zhangjinchao01 已提交
1003 1004 1005 1006 1007

    ..  math::

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

C
caoying03 已提交
1008 1009 1010
    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
1011 1012 1013 1014 1015

    ..  math::

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

C
caoying03 已提交
1016
    NOTE: In PaddlePaddle's implementation, the multiplication operations
Z
zhangjinchao01 已提交
1017
    :math:`W_{r}x_{t}`, :math:`W_{z}x_{t}` and :math:`W x_t` are not computed in
C
caoying03 已提交
1018 1019 1020
    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
1021

C
caoying03 已提交
1022 1023 1024
    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
1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035

    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.
1036
    :param reverse: Whether sequence process is reversed or not.
Z
zhangjinchao01 已提交
1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051
    :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
1052 1053 1054
    :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
1055
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1056 1057 1058 1059
    :rtype: LayerOutput
    """
    assert act.support_hppl
    assert gate_act.support_hppl
1060 1061 1062 1063 1064 1065 1066 1067 1068
    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
1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079

    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)
          )

1080 1081 1082
    return LayerOutput(name, LayerType.GRUMEMORY, [input], size=input.size / 3,
                       reverse=reverse)

Z
zhangjinchao01 已提交
1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097

@wrap_name_default()
@layer_support()
def last_seq(input, name=None, agg_level=AggregateLevel.EACH_TIMESTEP,
             layer_attr=None):
    """
    Get Last Timestamp Activation of a sequence.

    :param agg_level: Aggregated level
    :param name: Layer name.
    :type name: basestring
    :param input: Input layer name.
    :type input: LayerOutput
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1098
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1099 1100
    :rtype: LayerOutput
    """
1101 1102 1103 1104 1105 1106
    if input.reverse is not None and input.reverse:
        logger.warning("You are getting the last instance of a sequence that"
                       " is a output of a REVERSED layer. There is no time"
                       " series information at all. Maybe you want to use"
                       " first_seq instead.")

Z
zhangjinchao01 已提交
1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131
    Layer(
        name=name,
        type=LayerType.SEQUENCE_LAST_INSTANCE,
        inputs=[input.name],
        trans_type=agg_level,
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.SEQUENCE_LAST_INSTANCE, parents=[input],
                       size=input.size)


@wrap_name_default()
@layer_support()
def first_seq(input, name=None, agg_level=AggregateLevel.EACH_TIMESTEP,
              layer_attr=None):
    """
    Get First Timestamp Activation of a sequence.

    :param agg_level: aggregation level
    :param name: Layer name.
    :type name: basestring
    :param input: Input layer name.
    :type input: LayerOutput
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1132
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1133 1134
    :rtype: LayerOutput
    """
1135 1136 1137 1138 1139 1140 1141

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

Z
zhangjinchao01 已提交
1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156
    Layer(
        name=name,
        type=LayerType.SEQUENCE_FIRST_INSTANCE,
        inputs=[input.name],
        trans_type=agg_level,
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.SEQUENCE_FIRST_INSTANCE,
                       parents=[input], size=input.size)


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

1157

Z
zhangjinchao01 已提交
1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189
@wrap_name_default()
@layer_support()
def expand_layer(input, expand_as,
                 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
1190
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235
    :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,
        **ExtraAttr.to_kwargs(layer_attr)
    )
    return LayerOutput(name=name,
                       size=input.size,
                       layer_type=LayerType.EXPAND_LAYER,
                       parents=[input, expand_as])


@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
1236
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1237 1238
    :rtype: LayerOutput
    """
1239
    assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
1240
    assert len(input) == 2
1241 1242 1243 1244 1245 1246 1247
    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
1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285
    Layer(
        name=name,
        type=LayerType.INTERPOLATION_LAYER,
        inputs=[weight.name, input[0].name, input[1].name],
        **ExtraAttr.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.INTERPOLATION_LAYER,
                       parents=[weight, input[0], input[1]],
                       size=input[0].size)


@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
1286
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1287 1288
    :rtype: LayerOutput
    """
1289 1290 1291
    assert isinstance(input, LayerOutput) and isinstance(weight, LayerOutput)
    if weight.size is not None:
        assert weight.size == 1
Z
zhangjinchao01 已提交
1292 1293 1294
    Layer(
        name=name,
        type=LayerType.POWER_LAYER,
1295
        inputs=[weight.name, input.name],
Z
zhangjinchao01 已提交
1296 1297 1298 1299 1300 1301 1302 1303 1304 1305
        **ExtraAttr.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.POWER_LAYER,
                       parents=[input, weight], size=input.size)


@wrap_name_default()
@layer_support()
def scaling_layer(input, weight, name=None, layer_attr=None):
    """
1306
    A layer for multiplying input vector by weight scalar.
Z
zhangjinchao01 已提交
1307 1308

    .. math::
1309
       y  = w x
Z
zhangjinchao01 已提交
1310

1311 1312 1313 1314 1315
    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
1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330

    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
1331
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1332 1333
    :rtype: LayerOutput
    """
1334 1335 1336
    assert isinstance(weight, LayerOutput) and isinstance(input, LayerOutput)
    if weight.size is not None:
        assert weight.size == 1
Z
zhangjinchao01 已提交
1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369
    Layer(
        name=name,
        type=LayerType.SCALING_LAYER,
        inputs=[weight.name, input.name],
        **ExtraAttr.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.SCALING_LAYER, parents=[weight, input],
                       size=input.size)


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

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

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

    The example usage is:

    .. code-block:: python

       trans = trans_layer(input=layer)

    :param input: Input layer.
    :type input: LayerOutput
    :param name: Layer name.
    :type name: basestring
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1370
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.TRANS_LAYER,
        inputs=[input.name],
        **ExtraAttr.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.TRANS_LAYER, parents=[input],
                       size=input.size)


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

    ..  math::
D
dangqingqing 已提交
1390
        similarity = cos(\\theta) = {\\mathbf{a} \\cdot \\mathbf{b}
1391 1392 1393 1394 1395
        \\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
1396

1397 1398
    Note that the above computation is for one sample. Multiple samples are
    processed in one batch.
Z
zhangjinchao01 已提交
1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411

    :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
1412
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1413 1414
    :rtype: LayerOutput
    """
1415
    assert isinstance(a, LayerOutput) and isinstance(b, LayerOutput)
1416 1417 1418 1419 1420 1421 1422 1423 1424
    if size == 1:
        Layer(
            name=name,
            type=LayerType.COSINE_SIM,
            cos_scale=scale,
            inputs=[a.name, b.name],
            **ExtraLayerAttribute.to_kwargs(layer_attr)
        )
    else:
1425 1426
        if a.size is not None and b.size is not None:
            assert size == b.size / a.size
1427 1428 1429 1430 1431 1432 1433 1434
        Layer(
            name=name,
            type=LayerType.COSINE_SIM_VEC,
            size=size,
            cos_scale=scale,
            inputs=[a.name, b.name],
            **ExtraLayerAttribute.to_kwargs(layer_attr)
        )
Z
zhangjinchao01 已提交
1435 1436
    return LayerOutput(name, LayerType.COSINE_SIM, parents=[a, b])

1437

Z
zhangjinchao01 已提交
1438 1439
@wrap_name_default()
@wrap_bias_attr_default(has_bias=True)
1440
@wrap_param_attr_default()
Z
zhangjinchao01 已提交
1441
@layer_support()
C
caoying03 已提交
1442
def hsigmoid(input, label, num_classes, name=None, bias_attr=None,
1443
             param_attr=None, layer_attr=None):
Z
zhangjinchao01 已提交
1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464
    """
    Organize the classes into a binary tree. At each node, a sigmoid function
    is used to calculate the probability of belonging to the right branch.
    This idea is from "F. Morin, Y. Bengio (AISTATS 05):
    Hierarchical Probabilistic Neural Network Language Model."

    The example usage is:

    ..  code-block:: python

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

    :param input: Input layers. It could be a LayerOutput or list/tuple of
                 LayerOutput.
    :type input: LayerOutput|list|tuple
    :param label: Label layer.
    :type label: LayerOutput
    :param num_classes: number of classes.
    :type num_classes: int
L
luotao02 已提交
1465 1466
    :param name: layer name
    :type name: basestring
Z
zhangjinchao01 已提交
1467 1468 1469 1470 1471
    :param bias_attr: Bias attribute. None means default bias.
                      False means no bias.
    :type bias_attr: ParameterAttribute|False
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
1472
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1473 1474 1475 1476
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
1477 1478 1479 1480 1481 1482 1483 1484 1485
        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
1486 1487 1488 1489 1490
    assert isinstance(label, LayerOutput)
    assert label.layer_type == LayerType.DATA

    ipts_for_layer = []
    parents = []
1491
    for each_input, each_param_attr in zip(input, param_attr):
Z
zhangjinchao01 已提交
1492
        assert isinstance(each_input, LayerOutput)
1493
        ipts_for_layer.append(Input(each_input.name, **each_param_attr.attr))
Z
zhangjinchao01 已提交
1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507
        parents.append(each_input)
    ipts_for_layer.append(label.name)
    parents.append(label)

    Layer(
        name=name,
        type=LayerType.HSIGMOID,
        num_classes=num_classes,
        bias=ParamAttr.to_bias(bias_attr),
        inputs=ipts_for_layer,
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.HSIGMOID, parents=parents)

1508

Z
zhangjinchao01 已提交
1509 1510 1511 1512 1513 1514 1515 1516 1517
@wrap_name_default("conv")
@wrap_param_attr_default()
@wrap_bias_attr_default()
@wrap_act_default(act=ReluActivation())
@layer_support(DROPOUT)
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,
1518 1519
                   filter_size_y=None, stride_y=None, padding_y=None,
                   trans=False):
Z
zhangjinchao01 已提交
1520 1521 1522 1523 1524 1525 1526
    """
    Convolution layer for image. Paddle only support square input currently and
    thus input image's width equals height.

    The details of convolution layer, please refer UFLDL's `convolution
    <http://ufldl.stanford.edu/tutorial/supervised/
    FeatureExtractionUsingConvolution/>`_ .
1527
    
1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583
    Convolution Transpose (deconv) layer for image. Paddle only support square 
    input currently and thus input image's width equals height.

    The details of convolution transpose layer, 
    please refer to the following explanation and references therein
    <http://datascience.stackexchange.com/questions/6107/
    what-are-deconvolutional-layers/>`_ .
    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.

    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
    32*4 = 128 filters to process inputs. The channels will be split into 4
    pieces. First 256/4 = 64 channels will process by first 32 filters. The
    rest channels will be processed by rest group of filters.

    :param name: Layer name.
    :type name: basestring
    :param input: Layer Input.
    :type input: LayerOutput
    :param filter_size: The x dimension of a filter kernel. Or input a tuple for
                        two image dimension.
    :type filter_size: int|tuple|list
    :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).
    :type filter_size_y: int|None
    :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
    :param stride: The x dimension of the stride. Or input a tuple for two image
                   dimension.
    :type stride: int|tuple|list
    :param stride_y: The y dimension of the stride.
    :type stride_y: int
    :param padding: The x dimension of the padding. Or input a tuple for two
                    image dimension
    :type padding: int|tuple|list
    :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
1584 1585
    :param trans: true if it is a convTransLayer, false if it is a convLayer
    :type trans: bool
1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620
    :return: LayerOutput object.
    :rtype: LayerOutput
    """
    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.
        init_w = (2.0 / (filter_size ** 2 * num_channels)) ** 0.5
        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
1621 1622 1623 1624 1625 1626
    
    if trans:
        lt = LayerType.CONVTRANS_LAYER
    else:
        lt = LayerType.CONV_LAYER
    
1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638
    Layer(
        name=name,
        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),
        active_type=act.name,
        num_filters=num_filters,
        bias=ParamAttr.to_bias(bias_attr),
        shared_biases=shared_biases,
1639
        type=lt,
1640 1641
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
1642
    return LayerOutput(name, lt, parents=[input],
1643 1644 1645
                       activation=act, num_filters=num_filters)


Z
zhangjinchao01 已提交
1646 1647 1648 1649
@wrap_name_default("pool")
@layer_support()
def img_pool_layer(input, pool_size, name=None,
                   num_channels=None, pool_type=None,
1650
                   stride=1, padding=0, layer_attr=None,
1651 1652
                   pool_size_y=None, stride_y=None, padding_y=None,
                   img_width=None):
Z
zhangjinchao01 已提交
1653 1654 1655 1656 1657 1658 1659
    """
    Image pooling Layer.

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

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

1660
    :param padding: pooling padding width.
Z
zhangjinchao01 已提交
1661
    :type padding: int
1662 1663
    :param padding_y: pooling padding height. It's equal to padding by default.
    :type padding_y: int|None
Z
zhangjinchao01 已提交
1664 1665 1666 1667
    :param name: name of pooling layer
    :type name: basestring.
    :param input: layer's input
    :type input: LayerOutput
1668
    :param pool_size: pooling window width
Z
zhangjinchao01 已提交
1669
    :type pool_size: int
1670 1671
    :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
1672 1673 1674 1675 1676
    :param num_channels: number of input channel.
    :type num_channels: int
    :param pool_type: pooling type. MaxPooling or AveragePooling. Default is
                      MaxPooling.
    :type pool_type: BasePoolingType
1677
    :param stride: stride width of pooling.
Z
zhangjinchao01 已提交
1678
    :type stride: int
1679 1680
    :param stride_y: stride height of pooling. It is equal to stride by default.
    :type stride_y: int|None
Z
zhangjinchao01 已提交
1681 1682
    :param layer_attr: Extra Layer attribute.
    :type layer_attr: ExtraLayerAttribute
1683 1684 1685
    :param img_width: the width of input feature map. If it is None, the input feature
                      map should be square.
    :type img_width: int|None
D
dangqingqing 已提交
1686 1687
    :return: LayerOutput object.
    :rtype: LayerOutput
Z
zhangjinchao01 已提交
1688 1689 1690 1691 1692 1693 1694 1695 1696 1697
    """
    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'

1698 1699 1700 1701 1702 1703 1704 1705
    type_name = pool_type.name + '-projection' \
      if (isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)) \
      else pool_type.name

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

Z
zhangjinchao01 已提交
1706 1707 1708 1709 1710
    Layer(
        name=name,
        type=LayerType.POOL_LAYER,
        inputs=[Input(input.name,
                      pool=Pool(
1711
                          pool_type=type_name,
Z
zhangjinchao01 已提交
1712 1713
                          channels=num_channels,
                          size_x=pool_size,
1714
                          start=None,
Z
zhangjinchao01 已提交
1715
                          stride=stride,
1716 1717 1718 1719 1720
                          padding=padding,
                          size_y=pool_size_y,
                          stride_y=stride_y,
                          padding_y=padding_y,
                          img_width=img_width
Z
zhangjinchao01 已提交
1721 1722 1723 1724 1725 1726 1727 1728
                      ))],
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.POOL_LAYER, parents=[input],
                       num_filters=num_channels)


def __img_norm_layer__(name, input, size, norm_type, scale, power,
1729
                       num_channels, blocked, layer_attr):
Z
zhangjinchao01 已提交
1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters

    Layer(
        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)


@wrap_name_default("crmnorm")
@layer_support()
D
dangqingqing 已提交
1749 1750
def img_cmrnorm_layer(input, size, scale=0.0128, power=0.75,
                      name=None, num_channels=None,
1751
                      layer_attr=None):
Z
zhangjinchao01 已提交
1752
    """
1753
    Response normalization across feature maps.
D
dangqingqing 已提交
1754 1755
    The details please refer to
    `Alex's paper <http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf>`_.
Z
zhangjinchao01 已提交
1756 1757

    :param name: layer name.
D
dangqingqing 已提交
1758
    :type name: None|basestring
Z
zhangjinchao01 已提交
1759 1760
    :param input: layer's input.
    :type input: LayerOutput
1761
    :param size: Normalize in number of :math:`size` feature maps.
Z
zhangjinchao01 已提交
1762
    :type size: int
D
dangqingqing 已提交
1763
    :param scale: The hyper-parameter.
Z
zhangjinchao01 已提交
1764
    :type scale: float
D
dangqingqing 已提交
1765
    :param power: The hyper-parameter.
Z
zhangjinchao01 已提交
1766 1767 1768 1769 1770
    :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
1771
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1772 1773 1774
    :rtype: LayerOutput
    """
    return __img_norm_layer__(name, input, size, "cmrnorm-projection", scale,
1775
                              power, num_channels, 0, layer_attr)
Z
zhangjinchao01 已提交
1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820


@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)
def batch_norm_layer(input, act=None, name=None, num_channels=None,
                     bias_attr=None, param_attr=None, layer_attr=None,
                     batch_norm_type=None,
                     moving_average_fraction=0.9,
                     use_global_stats=None):
    """
    Batch Normalization Layer. The notation of this layer as follow.

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

    ..  math::

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

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

    :param name: layer name.
    :type name: basestring
    :param input: batch normalization input. Better be linear activation.
                Because there is an activation inside batch_normalization.
    :type input: LayerOutput
    :param batch_norm_type: We have batch_norm and cudnn_batch_norm. batch_norm
                            supports both CPU and GPU. cudnn_batch_norm requires
                            cuDNN version greater or equal to v4 (>=v4). But
                            cudnn_batch_norm is faster and needs less memory
                            than batch_norm. By default (None), we will
                            automaticly select cudnn_batch_norm for GPU and
                            batch_norm for CPU. Otherwise, select batch norm
                            type based on the specified type. If you use cudnn_batch_norm,
                            we suggested you use latest version, such as v5.1.
1821
    :type batch_norm_type: None|string, None or "batch_norm" or "cudnn_batch_norm"
Z
zhangjinchao01 已提交
1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848
    :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
1849
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912
    :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")
    Layer(
        name=name,
        inputs=Input(input.name,
                     image=Image(channels=num_channels),
                     **param_attr.attr),
        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,
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )

    return LayerOutput(name=name, layer_type=LayerType.BATCH_NORM_LAYER,
                       parents=[input], activation=act,
                       num_filters=num_channels)


@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
1913
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.SUM_TO_ONE_NORM_LAYER,
        inputs=[input.name],
        **ExtraAttr.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.SUM_TO_ONE_NORM_LAYER, parents=[input],
                       size=input.size)


@wrap_name_default("addto")
@wrap_act_default(act=LinearActivation())
@wrap_bias_attr_default(has_bias=False)
@layer_support(DROPOUT)
def addto_layer(input, act=None, name=None, bias_attr=None,
                layer_attr=None):
    """
    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
1954 1955 1956
    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
1957 1958

    It is a very good way to set dropout outside the layers. Since not all
C
caoying03 已提交
1959 1960
    PaddlePaddle layer support dropout, you can add an add_to layer, set
    dropout here.
Z
zhangjinchao01 已提交
1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974
    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
1975
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1976 1977 1978 1979 1980 1981
    :rtype: LayerOutput
    """
    num_filters = None
    if isinstance(input, LayerOutput):
        input = [input]

1982
    assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995
    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

    Layer(
        name=name, type=LayerType.ADDTO_LAYER, inputs=ipts_for_layer,
        bias=ParamAttr.to_bias(bias_attr),
        active_type=act.name,
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
1996

Z
zhangjinchao01 已提交
1997 1998 1999 2000 2001 2002 2003
    return LayerOutput(name, LayerType.ADDTO_LAYER, parents=input,
                       activation=act, num_filters=num_filters)


@wrap_act_default(act=IdentityActivation())
@wrap_name_default("concat")
@layer_support()
2004
def concat_layer(input, act=None, name=None, layer_attr=None, bias_attr=None):
Z
zhangjinchao01 已提交
2005 2006 2007 2008
    """
    Concat all input vector into one huge vector.
    Inputs can be list of LayerOutput or list of projection.

2009 2010 2011 2012 2013 2014
    The example usage is:

    ..  code-block:: python

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

Z
zhangjinchao01 已提交
2015 2016 2017
    :param name: Layer name.
    :type name: basestring
    :param input: input layers or projections
2018
    :type input: list|tuple|collections.Sequence
Z
zhangjinchao01 已提交
2019 2020 2021 2022
    :param act: Activation type.
    :type act: BaseActivation
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
2023
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2024 2025 2026 2027 2028 2029 2030 2031
    :rtype: LayerOutput
    """

    if isinstance(input, LayerOutput):
        input = [input]
    elif isinstance(input, Projection):
        input = [input]
    else:
2032
        assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
2033 2034

    def __is_type__(o, tp):
2035
        if not isinstance(o, collections.Sequence):
Z
zhangjinchao01 已提交
2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062
            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

    is_concat_layer = __is_type__(reduce(__reduce_concat_type__,
                                         map(type, input)), LayerOutput)

    layer_type = (LayerType.CONCAT_LAYER if is_concat_layer
                  else LayerType.CONCAT_PROJ_LAYER)

2063 2064 2065
    if layer_type == LayerType.CONCAT_LAYER:
        assert not bias_attr
    
Z
zhangjinchao01 已提交
2066 2067 2068 2069
    Layer(
        name=name, type=layer_type,
        inputs=[x.name for x in input] if is_concat_layer else input,
        active_type=act.name,
2070
        bias=ParamAttr.to_bias(bias_attr),
Z
zhangjinchao01 已提交
2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )

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

    return LayerOutput(name, layer_type=layer_type,
                       parents=input if is_concat_layer else [
                           x.origin for x in input],
                       activation=act, size=sz)


def memory(name, size, is_seq=False, boot_layer=None,
           boot_bias=None, boot_bias_active_type=None,
           boot_with_const_id=None):
    """
    The memory layers is a layer cross each time step. Reference this output
    as previous time step layer :code:`name` 's output.

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

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

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

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


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

    :param name: memory's name.
    :type name: basestring
    :param size: size of memory.
    :type size: int
    :param is_seq: is sequence for boot_layer
    :type is_seq: bool
    :param boot_layer: boot layer of memory.
    :type boot_layer: LayerOutput|None
    :param boot_bias: boot layer's bias
    :type boot_bias: ParameterAttribute|None
    :param boot_bias_active_type: boot layer's active type.
    :type boot_bias_active_type: BaseActivation
    :param boot_with_const_id: boot layer's id.
    :type boot_with_const_id: int
D
dangqingqing 已提交
2125
    :return: LayerOutput object which is a memory.
Z
zhangjinchao01 已提交
2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166
    :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)

    agent_name = Memory(name, size,
                        is_seq,
                        boot_layer.name if boot_layer is not None else None,
                        boot_bias,
                        boot_bias_active_type.name,
                        boot_with_const_id)

    lout = LayerOutput(name=agent_name, size=size,
                       layer_type=LayerType.MEMORY,
                       parents=[boot_layer] if boot_layer is not None
                       else None)
    return lout


@wrap_bias_attr_default()
@wrap_act_default(param_names=['gate_act',
                               'state_act'],
                  act=SigmoidActivation())
@wrap_act_default(act=TanhActivation())
@wrap_name_default('lstm_step')
@layer_support()
def lstm_step_layer(input, state, size, act=None,
                    name=None, gate_act=None, state_act=None,
                    bias_attr=None, layer_attr=None):
    """
    LSTM Step Layer. It used in recurrent_group. The lstm equations are shown
    as follow.

    ..  math::

L
luotao02 已提交
2167
        i_t & = \\sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i)
Z
zhangjinchao01 已提交
2168

L
luotao02 已提交
2169
        f_t & = \\sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f)
Z
zhangjinchao01 已提交
2170

L
luotao02 已提交
2171
        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
2172

L
luotao02 已提交
2173
        o_t & = \\sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o)
Z
zhangjinchao01 已提交
2174

L
luotao02 已提交
2175
        h_t & = o_t tanh(c_t)
Z
zhangjinchao01 已提交
2176 2177


L
luotao02 已提交
2178
    The input of lstm step is :math:`Wx_t + Wh_{t-1}`, and user should use
Z
zhangjinchao01 已提交
2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216
    :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
2217
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2218 2219 2220 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 2249 2250 2251 2252 2253 2254 2255
    :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),
        size=size, inputs=[input.name, state.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )

    return LayerOutput(name=name, layer_type=LayerType.LSTM_STEP_LAYER,
                       parents=[input, state], activation=act,
                       size=size, outputs=['default', 'state'])


@wrap_bias_attr_default()
@wrap_act_default(param_names=['gate_act'],
                  act=SigmoidActivation())
@wrap_act_default(act=TanhActivation())
@wrap_name_default('gru_step')
@layer_support()
def gru_step_layer(input, output_mem, size=None, act=None,
                   name=None, gate_act=None,
                   bias_attr=None, layer_attr=None):
    """

    :param input:
    :type input: LayerOutput
    :param output_mem:
    :param size:
    :param act:
    :param name:
    :param gate_act:
    :param bias_attr:
    :param layer_attr:
D
dangqingqing 已提交
2256
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284
    :rtype: LayerOutput
    """
    assert input.size % 3 == 0
    if size is None:
        size = input.size / 3
    Layer(
        name=name,
        type=LayerType.GRU_STEP_LAYER,
        inputs=[
            input.name,
            output_mem.name
        ],
        bias=ParamAttr.to_bias(bias_attr),
        size=size,
        active_type=act.name,
        active_gate_type=gate_act.name,
        **ExtraAttr.to_kwargs(layer_attr)
    )
    return LayerOutput(
        name=name, layer_type=LayerType.GRU_STEP_LAYER,
        parents=[input, output_mem],
        size=size, activation=act)


@wrap_name_default()
@layer_support()
def get_output_layer(input, arg_name, name=None, layer_attr=None):
    """
C
caoying03 已提交
2285 2286 2287 2288
    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
2289 2290 2291 2292 2293 2294 2295 2296 2297

    :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
2298
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320
    :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))
    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))

    return LayerOutput(name=name, layer_type=LayerType.GET_OUTPUT_LAYER,
                       parents=[input], size=input.size)


@wrap_name_default()
@wrap_act_default()
@wrap_bias_attr_default()
@wrap_param_attr_default()
@layer_support()
def recurrent_layer(input, act=None, bias_attr=None,
2321
                    param_attr=None, name=None, reverse=False, layer_attr=None):
Z
zhangjinchao01 已提交
2322
    """
2323 2324
    Simple recurrent unit layer. It is just a fully connect layer through both
    time and neural network.
Z
zhangjinchao01 已提交
2325

2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352
    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
2353
    :return: LayerOutput object.
2354
    :rtype: LayerOutput
Z
zhangjinchao01 已提交
2355 2356 2357 2358 2359 2360
    """
    Layer(name=name,
          type=LayerType.RECURRENT_LAYER,
          inputs=Input(input.name, **param_attr.attr),
          active_type=act.name,
          bias=ParamAttr.to_bias(bias_attr),
2361
          reversed=reverse,
Z
zhangjinchao01 已提交
2362 2363
          **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(name=name, layer_type=LayerType.RECURRENT_LAYER,
2364 2365
                       parents=[input], size=input.size, activation=act,
                       reverse=reverse)
Z
zhangjinchao01 已提交
2366 2367 2368 2369 2370 2371 2372


class StaticInput(object):
    """
    StaticInput is only used in recurrent_group which defines a read-only memory
    that can be a sequence or non-sequence.
    """
2373

Z
zhangjinchao01 已提交
2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392
    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)
    """
2393

Z
zhangjinchao01 已提交
2394 2395 2396 2397 2398 2399 2400
    def __init__(self, input):
        assert isinstance(input, LayerOutput)
        assert input.size is not None
        self.input = input


@wrap_name_default("recurrent_group")
2401
def recurrent_group(step, input, reverse=False, name=None, targetInlink=None):
Z
zhangjinchao01 已提交
2402
    """
C
caoying03 已提交
2403 2404 2405 2406 2407
    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
2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451

    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

2452 2453
    :param reverse: If reverse is set true, the recurrent unit will process the
                    input sequence in a reverse order.
Z
zhangjinchao01 已提交
2454
    :type reverse: bool
2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465

    :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

D
dangqingqing 已提交
2466
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2467 2468 2469 2470 2471 2472 2473 2474 2475 2476
    :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]
2477
    assert isinstance(input, collections.Sequence)
Z
zhangjinchao01 已提交
2478 2479 2480 2481 2482 2483

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

    in_links = filter(is_in_links, input)

2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497
    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

    assert(targetInlink == None or targetInlink_in_inlinks())
    targetInlinkName = None if targetInlink == None \
                            else targetInlink.name if isinstance(targetInlink, LayerOutput) \
                                                   else targetInlink.input.name

Z
zhangjinchao01 已提交
2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508
    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(
        name=name, in_links=map(map_in_links, in_links),
2509 2510
        seq_reversed=reverse,
        target_inlinkname=targetInlinkName)
Z
zhangjinchao01 已提交
2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535
    in_args = []
    for each_input in input:
        assert is_single_input(each_input)
        if isinstance(each_input, LayerOutput):
            in_args.append(each_input)
        elif isinstance(each_input, SubsequenceInput):
            in_args.append(each_input.input)
        else:
            mem_name = "__%s_memory__" % each_input.input.name
            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:
                mix += identity_projection(mem)
            in_args.append(mem)

    layer_outs = step(*in_args)

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

    for ot in layer_outs:
        assert isinstance(ot, LayerOutput)
2536
        ot.reverse = reverse
Z
zhangjinchao01 已提交
2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548
        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

2549

Z
zhangjinchao01 已提交
2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577
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):
        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))
        return trg_emb

    def __init__(self, size, embedding_name, embedding_size):
2578
        super(GeneratedInput, self).__init__()
Z
zhangjinchao01 已提交
2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601
        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
2602
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614
    :rtype: LayerOutput
    """

    assert isinstance(input, LayerOutput)
    Layer(name=name,
          type='maxid',
          inputs=[input.name],
          **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(name=name,
                       layer_type=LayerType.MAXID_LAYER,
                       parents=[input])

2615

H
Haonan 已提交
2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647
@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)
    Layer(name=name,
          type="out_prod",
          inputs=[input1.name, input2.name],
          **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(name=name,
                       layer_type=LayerType.OUT_PROD_LAYER,
2648 2649
                       parents=[input1, input2])

Z
zhangjinchao01 已提交
2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665

@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
2666 2667
    :param name: Layer name.
    :type name: basestring
Z
zhangjinchao01 已提交
2668 2669 2670 2671 2672 2673
    :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
2674
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689
    :rtype: LayerOutput
    """
    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])


@wrap_name_default()
def beam_search(step, input, bos_id, eos_id, beam_size,
                max_length=500, name=None,
                num_results_per_sample=None):
2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700
    """
    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)
2701
            with mixed_layer(size=512, name='rnn') as simple_rnn:
2702 2703 2704 2705 2706 2707
                simple_rnn += full_matrix_projection(input)
                simple_rnn += last_time_step_output
            return simple_rnn

        beam_gen = beam_search(name="decoder",
                               step=rnn_step,
2708
                               input=[StaticInput(encoder_last)],
2709 2710
                               bos_id=0,
                               eos_id=1,
2711
                               beam_size=5)
2712 2713 2714 2715 2716 2717 2718 2719 2720

    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
2721
                 step, and it is applied to sequences with arbitrary length by
2722 2723 2724 2725 2726 2727
                 sharing a same set of weights.

                 You can refer to the first parameter of recurrent_group, or
                 demo/seqToseq/seqToseq_net.py for more details.
    :type step: callable
    :param input: Input data for the recurrent unit
2728
    :type input: list
2729 2730 2731
    :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
2732
                   symbol is essential, since it is used to initialize the RNN
2733 2734 2735 2736 2737 2738 2739 2740
                   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
2741 2742
    :param max_length: Max generated sequence length.
    :type max_length: int
2743 2744 2745 2746 2747 2748 2749 2750 2751 2752
    :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
2753 2754
    :return: The generated word index.
    :rtype: LayerOutput
2755 2756
    """

Z
zhangjinchao01 已提交
2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769
    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")

    if isinstance(input, StaticInput) or isinstance(input,
                                                    BaseGeneratedInput):
        input = [input]

    generated_input_index = -1

    real_input = []
    for i, each_input in enumerate(input):
2770 2771
        assert isinstance(each_input, StaticInput) or isinstance(
            each_input, BaseGeneratedInput)
Z
zhangjinchao01 已提交
2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804
        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
        RecurrentLayerGroupSetGenerator(Generator(
            eos_layer_name=eos_name,
            max_num_frames=max_length,
            beam_size=beam_size,
            num_results_per_sample=num_results_per_sample))

        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

    tmp = recurrent_group(step=__real_step__, input=real_input, reverse=False,
                          name=name)
2805

Z
zhangjinchao01 已提交
2806 2807
    return tmp

2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819
def __cost_input__(input, label, weight=None):
    """
    inputs and parents for cost layers. 
    """
    ipts = [Input(input.name), Input(label.name)]
    parents = [input, label]
    if weight is not None:
        assert weight.layer_type == LayerType.DATA
        ipts.append(Input(weight.name))
        parents.append(weight)
    return ipts, parents
    
Z
zhangjinchao01 已提交
2820 2821

@wrap_name_default()
L
luotao1 已提交
2822 2823 2824
@layer_support()
def regression_cost(input, label, weight=None, name=None,
                    layer_attr=None):
Z
zhangjinchao01 已提交
2825 2826 2827 2828 2829 2830
    """
    Regression Layer.

    TODO(yuyang18): Complete this method.

    :param name: layer name.
2831
    :type name: basestring
Z
zhangjinchao01 已提交
2832
    :param input: Network prediction.
2833
    :type input: LayerOutput
Z
zhangjinchao01 已提交
2834
    :param label: Data label.
2835 2836 2837 2838
    :type label: LayerOutput
    :param weight: The weight affects the cost, namely the scale of cost.
                   It is an optional argument.
    :type weight: LayerOutput
L
luotao1 已提交
2839 2840
    :param layer_attr: layer's extra attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
2841
    :return: LayerOutput object.
2842
    :rtype: LayerOutput
Z
zhangjinchao01 已提交
2843
    """
2844 2845
    ipts, parents = __cost_input__(input, label, weight)

L
luotao1 已提交
2846 2847
    Layer(inputs=ipts, type="square_error", name=name,
          **ExtraLayerAttribute.to_kwargs(layer_attr))
2848
    return LayerOutput(name, LayerType.COST, parents=parents)
Z
zhangjinchao01 已提交
2849 2850 2851


@wrap_name_default("cost")
2852
@layer_support()
2853
def classification_cost(input, label, weight=None, name=None,
2854 2855
                        evaluator=classification_error_evaluator,
                        layer_attr=None):
Z
zhangjinchao01 已提交
2856 2857 2858 2859 2860 2861 2862 2863 2864
    """
    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
2865 2866 2867
    :param weight: The weight affects the cost, namely the scale of cost.
                   It is an optional argument.
    :type weight: LayerOutput
Z
zhangjinchao01 已提交
2868
    :param evaluator: Evaluator method.
2869 2870
    :param layer_attr: layer's extra attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
2871
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2872 2873 2874 2875 2876
    :rtype: LayerOutput
    """
    assert input.layer_type != LayerType.DATA
    assert isinstance(input.activation, SoftmaxActivation)
    assert label.layer_type == LayerType.DATA
2877 2878 2879

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

2880
    Layer(name=name, type="multi-class-cross-entropy", inputs=ipts,
2881
          **ExtraLayerAttribute.to_kwargs(layer_attr))
Z
zhangjinchao01 已提交
2882 2883 2884 2885 2886 2887 2888 2889 2890 2891

    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

2892
        e(name=e.__name__, input=input, label=label, weight=weight)
Z
zhangjinchao01 已提交
2893

2894
    if not isinstance(evaluator, collections.Sequence):
Z
zhangjinchao01 已提交
2895 2896 2897 2898 2899
        evaluator = [evaluator]

    for each_evaluator in evaluator:
        __add_evaluator__(each_evaluator)

2900
    return LayerOutput(name, LayerType.COST, parents=parents)
Z
zhangjinchao01 已提交
2901

2902

2903
def conv_operator(img, filter, filter_size, num_filters,
2904
                  num_channel=None, stride=1, padding=0,
Z
zhangjinchao01 已提交
2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915
                  filter_size_y=None, stride_y=None, padding_y=None):
    """
    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

2916 2917
       op = conv_operator(img=input1,
                          filter=input2,
2918
                          filter_size=3,
Z
zhangjinchao01 已提交
2919 2920 2921
                          num_filters=64,
                          num_channels=64)

2922 2923 2924 2925
    :param img: input image
    :type img: LayerOutput
    :param filter: input filter
    :type filter: LayerOutput
Z
zhangjinchao01 已提交
2926 2927
    :param filter_size: The x dimension of a filter kernel.
    :type filter_size: int
C
caoying03 已提交
2928 2929 2930
    :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
2931
    :type filter_size_y: int
2932 2933
    :param num_filters: channel of output data.
    :type num_filters: int
Z
zhangjinchao01 已提交
2934
    :param num_channel: channel of input data.
L
luotao02 已提交
2935
    :type num_channel: int
Z
zhangjinchao01 已提交
2936
    :param stride: The x dimension of the stride.
L
luotao02 已提交
2937
    :type stride: int
Z
zhangjinchao01 已提交
2938
    :param stride_y: The y dimension of the stride.
L
luotao02 已提交
2939
    :type stride_y: int
Z
zhangjinchao01 已提交
2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952
    :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
2953 2954 2955 2956 2957 2958 2959 2960

    if num_channel is None:
        num_channel = img.num_filters

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

2961
    op = ConvOperator(input_layer_names=[img.name, filter.name],
2962
                      num_filters=num_filters,
Z
zhangjinchao01 已提交
2963 2964 2965 2966 2967 2968
                      conv_conf=Conv(filter_size=filter_size,
                                     padding=padding,
                                     stride=stride,
                                     channels=num_channel,
                                     filter_size_y=filter_size_y,
                                     padding_y=padding_y,
2969
                                     stride_y=stride_y,
2970
                                     groups=1))
2971
    op.origin = [img, filter]
Z
zhangjinchao01 已提交
2972 2973
    return op

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 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070
@wrap_param_attr_default()
def conv_projection(input, filter_size, num_filters,
                    num_channels=None, stride=1, padding=0,
                    filter_size_y=None, stride_y=None, padding_y=None,
                    groups=1, param_attr=None):
    """
    ConvProjection with a layer as input.
    It performs element-wise multiplication with weight.

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

    The example usage is:

    .. code-block:: python

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

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

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

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

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

    if param_attr.attr.get('initial_smart'):
        # special initial for conv layers.
        init_w = (2.0 / (filter_size ** 2 * num_channels)) ** 0.5
        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

    proj = ConvProjection(input_layer_name=input.name,
                          num_filters=num_filters,
                          conv_conf=Conv(filter_size=filter_size,
                                         padding=padding,
                                         stride=stride,
                                         channels=num_channels,
                                         filter_size_y=filter_size_y,
                                         padding_y=padding_y,
                                         stride_y=stride_y,
                                         groups=groups),
                          **param_attr.attr)

    proj.origin = input
    return proj

Z
zhangjinchao01 已提交
3071 3072

@wrap_name_default()
L
luotao1 已提交
3073 3074
@layer_support()
def conv_shift_layer(a, b, name=None, layer_attr=None):
Z
zhangjinchao01 已提交
3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085
    """
    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:
3086 3087 3088 3089
     - 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
3090 3091 3092 3093 3094

    The example usage is:

    .. code-block:: python

3095
       conv_shift = conv_shift_layer(input=[layer1, layer2])
Z
zhangjinchao01 已提交
3096 3097 3098

    :param name: layer name
    :type name: basestring
3099 3100 3101 3102
    :param a: Input layer a.
    :type a: LayerOutput
    :param b: input layer b
    :type b: LayerOutput
L
luotao1 已提交
3103 3104
    :param layer_attr: layer's extra attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
3105
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3106 3107
    :rtype: LayerOutput
    """
3108 3109
    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
3110 3111 3112
    Layer(
        name=name,
        type=LayerType.CONV_SHIFT_LAYER,
3113
        inputs=[a.name, b.name],
L
luotao1 已提交
3114
        **ExtraLayerAttribute.to_kwargs(layer_attr)
Z
zhangjinchao01 已提交
3115 3116
    )

3117 3118
    return LayerOutput(name, LayerType.CONV_SHIFT_LAYER, parents=[a, b],
                       size=a.size)
Z
zhangjinchao01 已提交
3119 3120 3121 3122 3123


@wrap_name_default()
@wrap_param_attr_default()
@wrap_bias_attr_default()
3124
@wrap_act_default(act=LinearActivation())
Z
zhangjinchao01 已提交
3125
@layer_support(ERROR_CLIPPING, DROPOUT)
3126
def tensor_layer(a, b, size, act=None, name=None,
Z
zhangjinchao01 已提交
3127 3128 3129 3130 3131 3132
                 param_attr=None, bias_attr=None, layer_attr=None):
    """
    This layer performs tensor operation for two input.
    For example, each sample:

    .. math::
3133
       y_{i} = a * W_{i} * {b^\mathrm{T}}, i=0,1,...,K-1
Z
zhangjinchao01 已提交
3134 3135

    In this formular:
3136 3137
      - :math:`a`: the first input contains M elements.
      - :math:`b`: the second input contains N elements.
Z
zhangjinchao01 已提交
3138 3139
      - :math:`y_{i}`: the i-th element of y.
      - :math:`W_{i}`: the i-th learned weight, shape if [M, N]
3140
      - :math:`b^\mathrm{T}`: the transpose of :math:`b_{2}`.
Z
zhangjinchao01 已提交
3141 3142 3143 3144 3145

    The simple usage is:

    .. code-block:: python

3146
       tensor = tensor_layer(a=layer1, b=layer2, size=1000)
Z
zhangjinchao01 已提交
3147 3148 3149

    :param name: layer name
    :type name: basestring
3150 3151 3152 3153
    :param a: Input layer a.
    :type a: LayerOutput
    :param b: input layer b.
    :type b: LayerOutput
Z
zhangjinchao01 已提交
3154
    :param size: the layer dimension.
L
luotao02 已提交
3155
    :type size: int.
Z
zhangjinchao01 已提交
3156 3157 3158
    :param act: Activation Type. Default is tanh.
    :type act: BaseActivation
    :param param_attr: The Parameter Attribute.
3159
    :type param_attr: ParameterAttribute
Z
zhangjinchao01 已提交
3160 3161 3162 3163 3164 3165
    :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
3166
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3167 3168
    :rtype: LayerOutput
    """
3169
    assert isinstance(a, LayerOutput) and isinstance(b, LayerOutput)
Z
zhangjinchao01 已提交
3170 3171 3172 3173 3174 3175
    Layer(
        name=name,
        size=size,
        type=LayerType.TENSOR_LAYER,
        active_type=act.name,
        bias=ParamAttr.to_bias(bias_attr),
3176 3177
        inputs=[Input(a.name, **param_attr.attr),
                Input(b.name)],
Z
zhangjinchao01 已提交
3178 3179
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
3180
    return LayerOutput(name, LayerType.TENSOR_LAYER, parents=[a, b],
Z
zhangjinchao01 已提交
3181 3182 3183 3184 3185 3186 3187
                       activation=act, size=size)


@wrap_name_default()
@wrap_param_attr_default()
@wrap_bias_attr_default()
@wrap_act_default()
L
luotao1 已提交
3188
@layer_support()
3189
def selective_fc_layer(input, select, size, act=None, name=None,
Z
zhangjinchao01 已提交
3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203
                       pass_generation=False,
                       has_selected_colums=True,
                       mul_ratio=0.02,
                       param_attr=None, bias_attr=None, layer_attr=None):
    """
    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

3204
       sel_fc = selective_fc_layer(input=input, size=128, act=TanhActivation())
Z
zhangjinchao01 已提交
3205 3206 3207 3208 3209

    :param name: The Layer Name.
    :type name: basestring
    :param input: The input layer.
    :type input: LayerOutput|list|tuple
3210 3211 3212
    :param select: The select layer. The output of select layer should be a
                   sparse binary matrix, and treat as the mask of selective fc.
    :type select: LayerOutput
Z
zhangjinchao01 已提交
3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224
    :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
3225
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3226 3227 3228 3229
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
3230
        assert not isinstance(param_attr, collections.Sequence)
Z
zhangjinchao01 已提交
3231 3232
        param_attr = [param_attr]
    else:
3233
        if isinstance(param_attr, collections.Sequence):
Z
zhangjinchao01 已提交
3234 3235 3236 3237
            assert len(input) == len(param_attr)
        else:
            param_attr = [copy.deepcopy(param_attr) for _ in range(len(input))]

3238 3239 3240 3241
    assert isinstance(input, collections.Sequence)
    assert isinstance(select, LayerOutput)
    if select.size is not None:
        assert select.size == size
Z
zhangjinchao01 已提交
3242
    Layer(
3243 3244
        inputs=[Input(ipt.name, **attr.attr) for ipt, attr in zip(
            input, param_attr)] + [select.name],
Z
zhangjinchao01 已提交
3245 3246 3247
        name=name,
        type=LayerType.SEL_FC_LAYER,
        size=size,
3248
        bias=ParameterAttribute.to_bias(bias_attr),
Z
zhangjinchao01 已提交
3249 3250 3251 3252 3253 3254
        active_type=act.name,
        selective_fc_pass_generation=pass_generation,
        has_selected_colums=has_selected_colums,
        selective_fc_full_mul_ratio=mul_ratio,
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
3255 3256
    return LayerOutput(name, LayerType.SEL_FC_LAYER, list(input) + [select],
                       activation=act,
Z
zhangjinchao01 已提交
3257 3258 3259 3260
                       size=size)


@wrap_name_default()
L
luotao1 已提交
3261 3262
@layer_support()
def sampling_id_layer(input, name=None, layer_attr=None):
Z
zhangjinchao01 已提交
3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276
    """
    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
3277 3278
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
3279
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3280 3281 3282 3283 3284 3285
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.SAMPLING_ID_LAYER,
        inputs=[Input(input.name)],
L
luotao1 已提交
3286
        **ExtraLayerAttribute.to_kwargs(layer_attr)
Z
zhangjinchao01 已提交
3287 3288 3289 3290 3291
    )
    return LayerOutput(name, LayerType.SAMPLING_ID_LAYER, input)


@wrap_name_default()
L
luotao1 已提交
3292 3293 3294
@layer_support()
def slope_intercept_layer(input, name=None, slope=1.0, intercept=0.0,
                          layer_attr=None):
Z
zhangjinchao01 已提交
3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315
    """
    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
3316 3317
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
3318
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3319 3320 3321 3322 3323 3324 3325 3326
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.SLOPE_INTERCEPT_LAYER,
        slope=slope,
        intercept=intercept,
        inputs=[Input(input.name)],
L
luotao1 已提交
3327
        **ExtraLayerAttribute.to_kwargs(layer_attr)
Z
zhangjinchao01 已提交
3328 3329 3330 3331 3332
    )
    return LayerOutput(name, LayerType.SLOPE_INTERCEPT_LAYER, input)


@wrap_name_default()
L
luotao1 已提交
3333 3334 3335
@layer_support()
def linear_comb_layer(weights, vectors, size=None, name=None,
                      layer_attr=None):
Z
zhangjinchao01 已提交
3336
    """
3337 3338 3339 3340
    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
3341 3342 3343

    .. math::

3344 3345 3346 3347 3348 3349
       z(i) = \sum_{j=0}^{M-1} x(j) y(i+Nj)
    where :math:`0 \le i \le N-1`

    Or in the matrix notation:

    .. math::
Z
zhangjinchao01 已提交
3350

3351
       z = x^\mathrm{T} Y
Z
zhangjinchao01 已提交
3352 3353

    In this formular:
3354 3355 3356 3357 3358 3359
      - :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
3360 3361 3362 3363 3364

    The simple usage is:

    .. code-block:: python

3365
       linear_comb = linear_comb_layer(weights=weight, vectors=vectors,
Z
zhangjinchao01 已提交
3366 3367
                                       size=elem_dim)

3368 3369 3370 3371
    :param weights: The weight layer.
    :type weights: LayerOutput
    :param vectors: The vector layer.
    :type vectors: LayerOutput
Z
zhangjinchao01 已提交
3372 3373 3374 3375
    :param size: the dimension of this layer.
    :type size: int
    :param name: The Layer Name.
    :type name: basestring
L
luotao1 已提交
3376 3377
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
3378
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3379 3380
    :rtype: LayerOutput
    """
3381 3382 3383 3384 3385 3386 3387
    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:
                size = vectors.size / weights.size
        else:
            assert size == vectors.size / weights.size
Z
zhangjinchao01 已提交
3388 3389
    Layer(
        name=name,
3390
        type=LayerType.LINEAR_COMBINATION_LAYER,
Z
zhangjinchao01 已提交
3391
        size=size,
3392
        inputs=[Input(weights.name), Input(vectors.name)],
L
luotao1 已提交
3393
        **ExtraLayerAttribute.to_kwargs(layer_attr)
Z
zhangjinchao01 已提交
3394
    )
3395 3396 3397
    return LayerOutput(name, LayerType.LINEAR_COMBINATION_LAYER,
                       [weights, vectors], size=size)

3398

3399
convex_comb_layer = linear_comb_layer
Z
zhangjinchao01 已提交
3400

3401

Z
zhangjinchao01 已提交
3402
@wrap_name_default()
L
luotao1 已提交
3403
@layer_support()
Z
zhangjinchao01 已提交
3404 3405 3406 3407 3408 3409 3410
def block_expand_layer(input,
                       block_x=0,
                       block_y=0,
                       stride_x=0,
                       stride_y=0,
                       padding_x=0,
                       padding_y=0,
3411
                       num_channels=None,
L
luotao1 已提交
3412 3413
                       name=None,
                       layer_attr=None):
Z
zhangjinchao01 已提交
3414 3415
    """
    Expand feature map to minibatch matrix.
3416
       - matrix width is: block_y * block_x * num_channels
L
luotao02 已提交
3417
       - matirx height is: outputH * outputW
Z
zhangjinchao01 已提交
3418 3419 3420 3421 3422 3423 3424 3425 3426 3427

    .. 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
3428
    time step is block_y * block_x * num_channels. This layer can be used after
Z
zhangjinchao01 已提交
3429 3430
    convolution neural network, and before recurrent neural network.

3431 3432 3433 3434 3435
    The simple usage is:

    .. code-block:: python

       block_expand = block_expand_layer(input,
3436
                                         num_channels=128,
3437 3438 3439 3440 3441
                                         stride_x=1,
                                         stride_y=1,
                                         block_x=1,
                                         block_x=3)

Z
zhangjinchao01 已提交
3442 3443
    :param input: The input layer.
    :type input: LayerOutput
3444 3445
    :param num_channels: The channel number of input layer.
    :type num_channels: int|None
Z
zhangjinchao01 已提交
3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459
    :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
3460 3461
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
3462
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3463 3464
    :rtype: LayerOutput
    """
3465 3466 3467
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters
Z
zhangjinchao01 已提交
3468
    Layer(name=name,
3469 3470 3471 3472 3473 3474 3475 3476
          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)),
Z
zhangjinchao01 已提交
3477
          type=LayerType.BLOCK_EXPAND,
L
luotao1 已提交
3478
          **ExtraLayerAttribute.to_kwargs(layer_attr)
3479 3480 3481
          )

    return LayerOutput(name, LayerType.BLOCK_EXPAND, parents=[input])
Z
zhangjinchao01 已提交
3482 3483


3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550
@wrap_name_default()
@layer_support()
def maxout_layer(input,
                 groups,
                 num_channels=None,
                 size_x=None,
                 size_y=None,
                 name=None,
                 layer_attr=None):
    """
    A layer to do max out on conv layer output.
      - Input: output of a conv layer.
      - Output: feature map size same as input. Channel is (input channel) / groups.

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

    Please refer to Paper: 
      - 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
    
    The simple usage is:

    .. code-block:: python

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

    :param input: The input layer.
    :type input: LayerOutput
    :param num_channels: The channel number of input layer. If None will be set
                     automatically from previous output.
    :type num_channels: int|None
    :param groups: The group number of input layer.
    :type groups: int
    :param size_x: conv output width. If None will be set
                   automatically from previous output.
    :type size_x: int|None
    :param size_y: conv output height. If None will be set
                   automatically from previous output.
    :type size_y: int|None
    :param name: The name of this layer, which can not specify.
    :type name: None|basestring.
    :param layer_attr: Extra Layer attribute.
    :type layer_attr: ExtraLayerAttribute
    :return: LayerOutput object.
    :rtype: LayerOutput
    """
    assert input.layer_type == LayerType.CONV_LAYER
    assert isinstance(input.activation, LinearActivation)
    assert groups > 1
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters
    assert num_channels % groups == 0
    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])


Z
zhangjinchao01 已提交
3551
@wrap_name_default()
L
luotao1 已提交
3552 3553 3554
@layer_support()
def ctc_layer(input, label, size=None, name=None, norm_by_times=False,
              layer_attr=None):
Z
zhangjinchao01 已提交
3555 3556 3557 3558 3559
    """
    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.

3560 3561
    More details can be found by referring to `Connectionist Temporal
    Classification: Labelling Unsegmented Sequence Data with Recurrent
3562 3563
    Neural Networks <http://machinelearning.wustl.edu/mlpapers/paper_files/
    icml2006_GravesFGS06.pdf>`_
3564 3565 3566 3567 3568 3569 3570 3571

    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
3572 3573 3574 3575 3576 3577 3578 3579 3580
    The simple usage:

    .. code-block:: python

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

3581
    :param input: The input layer.
Z
zhangjinchao01 已提交
3582 3583 3584
    :type input: LayerOutput
    :param label: The data layer of label with variable length.
    :type label: LayerOutput
3585
    :param size: category numbers + 1.
Z
zhangjinchao01 已提交
3586
    :type size: int
3587 3588
    :param name: The name of this layer
    :type name: basestring|None
Z
zhangjinchao01 已提交
3589 3590
    :param norm_by_times: Whether to normalization by times. False by default.
    :type norm_by_times: bool
L
luotao1 已提交
3591 3592
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
3593
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3594 3595 3596 3597
    :rtype: LayerOutput
    """
    assert isinstance(input, LayerOutput)
    assert isinstance(label, LayerOutput)
3598 3599 3600 3601 3602
    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
3603
    Layer(
3604 3605 3606 3607
        name=name,
        type=LayerType.CTC_LAYER,
        size=size,
        norm_by_times=norm_by_times,
L
luotao1 已提交
3608 3609
        inputs=[input.name, label.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr)
Z
zhangjinchao01 已提交
3610 3611 3612
    )
    return LayerOutput(name, LayerType.CTC_LAYER, [input, label], size=size)

3613

Z
zhangjinchao01 已提交
3614
@wrap_name_default()
3615
@wrap_param_attr_default()
L
luotao1 已提交
3616 3617 3618
@layer_support()
def crf_layer(input, label, size=None, weight=None, param_attr=None, name=None,
              layer_attr=None):
Z
zhangjinchao01 已提交
3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633
    """
    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.
3634
    :type label: LayerOutput
Z
zhangjinchao01 已提交
3635 3636 3637 3638 3639 3640 3641 3642 3643
    :param size: The category number.
    :type size: int
    :param weight: The third layer is "weight" of each sample, which is an
                  optional argument.
    :type weight: LayerOutput
    :param param_attr: Parameter attribute. None means default attribute
    :type param_attr: ParameterAttribute
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring
L
luotao1 已提交
3644 3645
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
3646
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3647 3648 3649 3650 3651
    :rtype: LayerOutput
    """
    assert isinstance(input, LayerOutput)
    assert isinstance(label, LayerOutput)
    assert weight is None or isinstance(weight, LayerOutput)
3652 3653 3654 3655 3656 3657
    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
3658

3659
    ipts = [Input(input.name, **param_attr.attr),
Z
zhangjinchao01 已提交
3660 3661 3662 3663 3664
            Input(label.name)]
    if weight is not None:
        ipts.append(Input(weight.name))

    Layer(
3665 3666 3667 3668
        name=name,
        type=LayerType.CRF_LAYER,
        size=size,
        inputs=ipts,
L
luotao1 已提交
3669
        **ExtraLayerAttribute.to_kwargs(layer_attr)
Z
zhangjinchao01 已提交
3670 3671 3672 3673 3674 3675
    )
    parents = [input, label]
    if weight is not None:
        parents.append(weight)
    return LayerOutput(name, LayerType.CRF_LAYER, parents, size=size)

3676

Z
zhangjinchao01 已提交
3677
@wrap_name_default()
3678
@wrap_param_attr_default()
L
luotao1 已提交
3679 3680 3681
@layer_support()
def crf_decoding_layer(input, size, label=None, param_attr=None, name=None,
                       layer_attr=None):
Z
zhangjinchao01 已提交
3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698
    """
    A layer for calculating the decoding sequence of sequential conditional
    random field model. The decoding sequence is stored in output.ids.
    If a second input is provided, it is treated as the ground-truth label, and
    this layer will also calculate error. output.value[i] is 1 for incorrect
    decoding or 0 for correct decoding.

    :param input: The first input layer.
    :type input: LayerOutput
    :param size: size of this layer.
    :type size: int
    :param label: None or ground-truth label.
    :type label: LayerOutput or None
    :param param_attr: Parameter attribute. None means default attribute
    :type param_attr: ParameterAttribute
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring
L
luotao1 已提交
3699 3700
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
3701
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3702 3703 3704 3705 3706 3707
    :rtype: LayerOutput
    """

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

3708
    ipts = [Input(input.name, **param_attr.attr)]
Z
zhangjinchao01 已提交
3709 3710 3711 3712
    if label is not None:
        ipts.append(Input(label.name))

    Layer(
3713 3714 3715 3716
        name=name,
        type=LayerType.CRF_DECODING_LAYER,
        size=size,
        inputs=ipts,
L
luotao1 已提交
3717
        **ExtraLayerAttribute.to_kwargs(layer_attr)
Z
zhangjinchao01 已提交
3718 3719 3720 3721 3722 3723
    )
    parents = [input]
    if label is not None:
        parents.append(label)
    return LayerOutput(name, LayerType.CRF_DECODING_LAYER, parents, size=size)

3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748 3749 3750 3751 3752 3753 3754 3755 3756 3757 3758 3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798 3799 3800
@wrap_bias_attr_default(has_bias=True)
@wrap_name_default()
@layer_support()
def nce_layer(input, label, num_classes, weight=None,
              num_neg_samples=10, neg_distribution=None,
              name=None, bias_attr=None, layer_attr=None):
    """
    Noise-contrastive estimation.
    Implements the method in the following paper:
    A fast and simple algorithm for training neural probabilistic language models.

    The example usage is:

    .. code-block:: python

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

    :param name: layer name
    :type name: basestring
    :param input: input layers. It could be a LayerOutput of list/tuple of LayerOutput.
    :type input: LayerOutput|list|tuple|collections.Sequence
    :param label: label layer
    :type label: LayerOutput
    :param weight: weight layer, can be None(default)
    :type weight: LayerOutput
    :param num_classes: number of classes.
    :type num_classes: int 
    :param num_neg_samples: number of negative samples. Default is 10.
    :type num_neg_samples: int 
    :param neg_distribution: The distribution for generating the random negative labels.
                             A uniform distribution will be used if not provided.
                             If not None, its length must be equal to num_classes.
    :type neg_distribution: list|tuple|collections.Sequence|None
    :param bias_attr: Bias parameter attribute. True if no bias.
    :type bias_attr: ParameterAttribute|None|False
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
    :return: layer name.
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
    assert isinstance(input, collections.Sequence)
    assert isinstance(label, LayerOutput)
    assert label.layer_type == LayerType.DATA
    if neg_distribution is not None:
        assert isinstance(neg_distribution, collections.Sequence)
        assert len(neg_distribution) == num_classes
        assert sum(neg_distribution) == 1
    
    ipts_for_layer = []
    parents = []
    for each_input in input:
        assert isinstance(each_input, LayerOutput)
        ipts_for_layer.append(each_input.name)
        parents.append(each_input)
    ipts_for_layer.append(label.name)
    parents.append(label)

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

    Layer(
        name=name,
        type=LayerType.NCE_LAYER,
        num_classes=num_classes,
        neg_sampling_dist=neg_distribution,
        num_neg_samples=num_neg_samples,
        inputs=ipts_for_layer,
        bias=ParamAttr.to_bias(bias_attr),
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.NCE_LAYER, parents=parents)
3801

Z
zhangjinchao01 已提交
3802 3803 3804
"""
following are cost Layers.
"""
3805 3806


Z
zhangjinchao01 已提交
3807
@wrap_name_default()
L
luotao1 已提交
3808 3809
@layer_support()
def rank_cost(left, right, label, weight=None, name=None, coeff=1.0, layer_attr=None):
Z
zhangjinchao01 已提交
3810
    """
3811
    A cost Layer for learning to rank using gradient descent. Details can refer
3812 3813
    to `papers <http://research.microsoft.com/en-us/um/people/cburges/papers/
    ICML_ranking.pdf>`_.
Z
zhangjinchao01 已提交
3814 3815 3816 3817 3818
    This layer contains at least three inputs. The weight is an optional
    argument, which affects the cost.

    .. math::

L
luotao02 已提交
3819
       C_{i,j} & = -\\tilde{P_{ij}} * o_{i,j} + log(1 + e^{o_{i,j}})
Z
zhangjinchao01 已提交
3820

L
luotao02 已提交
3821
       o_{i,j} & =  o_i - o_j
Z
zhangjinchao01 已提交
3822

L
luotao02 已提交
3823
       \\tilde{P_{i,j}} & = \\{0, 0.5, 1\\} \ or \ \\{0, 1\\}
Z
zhangjinchao01 已提交
3824 3825 3826 3827 3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848 3849 3850 3851 3852

    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
3853 3854
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
3855
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3856 3857 3858 3859 3860 3861 3862 3863 3864 3865 3866 3867 3868 3869 3870 3871
    :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)

    Layer(name=name,
          type=LayerType.RANK_COST,
          inputs=ipts,
          coeff=coeff,
L
luotao1 已提交
3872
          **ExtraLayerAttribute.to_kwargs(layer_attr)
3873
          )
Z
zhangjinchao01 已提交
3874 3875 3876

    return LayerOutput(name, LayerType.RANK_COST, parents=parents)

3877

Z
zhangjinchao01 已提交
3878
@wrap_name_default()
L
luotao1 已提交
3879 3880
@layer_support()
def lambda_cost(input, score, name, NDCG_num=5, max_sort_size=-1, layer_attr=None):
Z
zhangjinchao01 已提交
3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892
    """
    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)

3893
    :param input: Samples of the same query should be loaded as sequence.
Z
zhangjinchao01 已提交
3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904
    :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
3905 3906 3907
                          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
3908 3909 3910
    :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
3911 3912
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
3913
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3914 3915
    :rtype: LayerOutput
    """
3916 3917 3918
    assert isinstance(input, LayerOutput) and isinstance(score, LayerOutput)
    if score.size is not None:
        assert score.size == 1
Z
zhangjinchao01 已提交
3919 3920 3921 3922
    Layer(name=name,
          type=LayerType.LAMBDA_COST,
          inputs=[input.name, score.name],
          NDCG_num=NDCG_num,
L
luotao1 已提交
3923 3924
          max_sort_size=max_sort_size,
          **ExtraLayerAttribute.to_kwargs(layer_attr)
3925
          )
Z
zhangjinchao01 已提交
3926 3927 3928

    return LayerOutput(name, LayerType.LAMBDA_COST, parents=[input, score])

3929

Z
zhangjinchao01 已提交
3930
@wrap_name_default()
L
luotao1 已提交
3931 3932
@layer_support()
def cross_entropy(input, label, name=None, coeff=1.0, layer_attr=None):
Z
zhangjinchao01 已提交
3933 3934 3935 3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949
    """
    A loss layer for multi class entropy.

    .. code-block:: python

       cost = cross_entropy(input, label)

    :param input: The first input layer.
    :type input: LayerOutput.
    :param label: The input label.
    :type input: LayerOutput.
    :param type: The type of cost.
    :type type: basestring.
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring.
    :param coeff: The coefficient affects the gradient in the backward.
    :type coeff: float.
L
luotao1 已提交
3950 3951
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
3952
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3953 3954 3955 3956 3957 3958 3959
    :rtype: LayerOutput.
    """

    Layer(name=name,
          type=LayerType.CROSS_ENTROPY,
          inputs=[input.name, label.name],
          coeff=coeff,
L
luotao1 已提交
3960
          **ExtraLayerAttribute.to_kwargs(layer_attr)
3961
          )
Z
zhangjinchao01 已提交
3962 3963
    return LayerOutput(name, LayerType.CROSS_ENTROPY, parents=[input, label])

3964

Z
zhangjinchao01 已提交
3965
@wrap_name_default()
L
luotao1 已提交
3966
@layer_support()
Z
zhangjinchao01 已提交
3967
def cross_entropy_with_selfnorm(input, label, name=None, coeff=1.0,
L
luotao1 已提交
3968 3969
                                softmax_selfnorm_alpha=0.1,
                                layer_attr=None):
Z
zhangjinchao01 已提交
3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982 3983 3984 3985 3986 3987 3988
    """
    A loss layer for multi class entropy with selfnorm.

    .. code-block:: python

       cost = cross_entropy_with_selfnorm(input, label)

    :param input: The first input layer.
    :type input: LayerOutput.
    :param label: The input label.
    :type input: LayerOutput.
    :param type: The type of cost.
    :type type: basestring.
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring.
    :param coeff: The coefficient affects the gradient in the backward.
    :type coeff: float.
    :param softmax_selfnorm_alpha: The scale factor affects the cost.
    :type softmax_selfnorm_alpha: float.
L
luotao1 已提交
3989 3990
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
3991
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3992 3993 3994 3995 3996 3997 3998
    :rtype: LayerOutput.
    """
    Layer(name=name,
          type=LayerType.CROSS_ENTROPY_WITH_SELFNORM,
          inputs=[input.name, label.name],
          coeff=coeff,
          softmax_selfnorm_alpha=softmax_selfnorm_alpha,
L
luotao1 已提交
3999
          **ExtraLayerAttribute.to_kwargs(layer_attr)
4000
          )
Z
zhangjinchao01 已提交
4001 4002 4003 4004 4005

    return LayerOutput(name,
                       LayerType.CROSS_ENTROPY_WITH_SELFNORM,
                       parents=[input, label])

4006

Z
zhangjinchao01 已提交
4007
@wrap_name_default()
L
luotao1 已提交
4008 4009
@layer_support()
def huber_cost(input, label, name=None, coeff=1.0, layer_attr=None):
Z
zhangjinchao01 已提交
4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024
    """
    A loss layer for huber loss.

    .. code-block:: python

       cost = huber_cost(input, label)

    :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
4025 4026
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
4027
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4028 4029
    :rtype: LayerOutput.
    """
4030 4031 4032
    assert isinstance(input, LayerOutput)
    if input.size is not None:
        assert input.size == 1
Z
zhangjinchao01 已提交
4033 4034 4035 4036
    Layer(name=name,
          type=LayerType.HUBER,
          inputs=[input.name, label.name],
          coeff=coeff,
L
luotao1 已提交
4037
          **ExtraLayerAttribute.to_kwargs(layer_attr)
4038
          )
Z
zhangjinchao01 已提交
4039 4040
    return LayerOutput(name, LayerType.HUBER, parents=[input, label])

4041

Z
zhangjinchao01 已提交
4042
@wrap_name_default()
L
luotao1 已提交
4043 4044 4045
@layer_support()
def multi_binary_label_cross_entropy(input, label, name=None, coeff=1.0,
                                     layer_attr=None):
Z
zhangjinchao01 已提交
4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057 4058 4059 4060 4061 4062
    """
    A loss layer for multi binary label cross entropy.

    .. code-block:: python

       cost = multi_binary_label_cross_entropy(input, label)

    :param input: The first input layer.
    :type input: LayerOutput
    :param label: The input label.
    :type input: LayerOutput
    :param type: The type of cost.
    :type type: basestring
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring
    :param coeff: The coefficient affects the gradient in the backward.
    :type coeff: float
L
luotao1 已提交
4063 4064
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
4065
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
4066 4067 4068
    :rtype: LayerOutput
    """

4069 4070
    if input.activation is None or \
            not isinstance(input.activation, SigmoidActivation):
Z
zhangjinchao01 已提交
4071
        logger.log(logging.WARN,
4072 4073
                   "%s is not recommend for multi_binary_label_cross_entropy's activation, "
                   "maybe the sigmoid is better" % repr(input.activation))
Z
zhangjinchao01 已提交
4074 4075 4076 4077 4078

    Layer(name=name,
          type=LayerType.MULTI_BIN_LABEL_CROSS_ENTROPY,
          inputs=[input.name, label.name],
          coeff=coeff,
L
luotao1 已提交
4079
          **ExtraLayerAttribute.to_kwargs(layer_attr)
4080
          )
Z
zhangjinchao01 已提交
4081 4082
    return LayerOutput(name, LayerType.MULTI_BIN_LABEL_CROSS_ENTROPY,
                       parents=[input, label])