layers.py 162.4 KB
Newer Older
1
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
2 3 4 5 6 7 8 9 10 11 12 13 14 15
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import functools
16
import collections
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

26 27 28 29 30 31
try:
    import cPickle as pickle
except ImportError:
    import pickle
import copy

32 33 34 35 36 37 38 39
__all__ = [
    "full_matrix_projection",
    "AggregateLevel",
    "ExpandLevel",
    "identity_projection",
    "dotmul_projection",
    "dotmul_operator",
    "repeat_layer",
40
    "seq_reshape_layer",
41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62
    "table_projection",
    "mixed_layer",
    "data_layer",
    "embedding_layer",
    "fc_layer",
    "grumemory",
    "pooling_layer",
    "lstmemory",
    "last_seq",
    "first_seq",
    "cos_sim",
    "hsigmoid",
    "conv_projection",
    "regression_cost",
    'classification_cost',
    "LayerOutput",
    'img_conv_layer',
    'img_pool_layer',
    'batch_norm_layer',
    'img_cmrnorm_layer',
    'addto_layer',
    'concat_layer',
63
    'seq_concat_layer',
64 65 66 67 68 69
    'lstm_step_layer',
    'recurrent_group',
    'memory',
    'StaticInput',
    'expand_layer',
    'scaling_layer',
X
xuwei06 已提交
70
    'scaling_projection',
71 72 73 74
    'power_layer',
    'interpolation_layer',
    'bilinear_interp_layer',
    'trans_layer',
75
    'rotate_layer',
76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96
    'sum_to_one_norm_layer',
    'get_output_layer',
    'LayerType',
    'context_projection',
    'beam_search',
    'maxid_layer',
    'GeneratedInput',
    'SubsequenceInput',
    'gru_step_layer',
    'recurrent_layer',
    'BaseGeneratedInput',
    'conv_operator',
    'conv_shift_layer',
    'tensor_layer',
    'selective_fc_layer',
    'sampling_id_layer',
    'slope_intercept_layer',
    'trans_full_matrix_projection',
    'linear_comb_layer',
    'convex_comb_layer',
    'ctc_layer',
97
    'warp_ctc_layer',
98 99 100 101 102 103 104 105 106 107 108 109 110 111
    'crf_layer',
    'crf_decoding_layer',
    'nce_layer',
    'cross_entropy_with_selfnorm',
    'cross_entropy',
    'multi_binary_label_cross_entropy',
    'sum_cost',
    'rank_cost',
    'lambda_cost',
    'huber_cost',
    'block_expand_layer',
    'maxout_layer',
    'out_prod_layer',
    'print_layer',
Y
yuan 已提交
112
    'priorbox_layer',
113
    'spp_layer',
D
dangqingqing 已提交
114
    'pad_layer',
115
    'eos_layer',
116
    'layer_support',
117
]
118 119 120 121 122 123 124 125 126 127 128 129 130


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

    DATA = "data"
    MIXED_LAYER = "mixed"
    LSTMEMORY = "lstmemory"
    GRUMEMORY = "gated_recurrent"
    SEQUENCE_LAST_INSTANCE = "seqlastins"
    SEQUENCE_FIRST_INSTANCE = "seqfirstins"
131
    SEQUENCE_RESHAPE = "seqreshape"
132 133 134 135
    POOLING_MAX = "max"
    POOLING_AVG = 'average'
    FC_LAYER = "fc"
    COST = 'cost'
136 137
    COSINE_SIM_VEC = 'cos_vm'
    COSINE_SIM = 'cos'
138 139
    HSIGMOID = 'hsigmoid'
    CONV_LAYER = "conv"
140
    CONVTRANS_LAYER = "convt"
141 142 143
    EXCONV_LAYER = "exconv"
    EXCONVTRANS_LAYER = "exconvt"
    CUDNNCONV_LAYER = "cudnn_conv"
144 145 146 147 148 149 150 151
    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'
152
    SEQUENCE_CONCAT_LAYER = 'seqconcat'
153 154 155 156 157 158 159

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

    EXPAND_LAYER = 'expand'
    INTERPOLATION_LAYER = 'interpolation'
160
    BILINEAR_INTERP_LAYER = 'bilinear_interp'
161 162 163
    POWER_LAYER = 'power'
    SCALING_LAYER = 'scaling'
    TRANS_LAYER = 'trans'
164
    ROTATE_LAYER = 'rotate'
165
    OUT_PROD_LAYER = 'out_prod'
166
    FEATURE_MAP_EXPAND_LAYER = 'featmap_expand'
167 168 169 170 171 172 173 174 175 176 177

    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"
178
    LINEAR_COMBINATION_LAYER = "convex_comb"
179
    BLOCK_EXPAND = "blockexpand"
180
    MAXOUT = "maxout"
Q
qijun 已提交
181
    SPP_LAYER = "spp"
D
dangqingqing 已提交
182
    PAD_LAYER = "pad"
183

184
    PRINT_LAYER = "print"
Y
yuan 已提交
185
    PRIORBOX_LAYER = "priorbox"
186

187
    CTC_LAYER = "ctc"
188
    WARP_CTC_LAYER = "warp_ctc"
189 190
    CRF_LAYER = "crf"
    CRF_DECODING_LAYER = "crf_decoding"
191
    NCE_LAYER = 'nce'
192 193 194 195 196 197 198 199

    RANK_COST = "rank-cost"
    LAMBDA_COST = "lambda_cost"
    HUBER = "huber"
    CROSS_ENTROPY = "multi-class-cross-entropy"
    CROSS_ENTROPY_WITH_SELFNORM = "multi_class_cross_entropy_with_selfnorm"
    SOFT_BIN_CLASS_CROSS_ENTROPY = "soft_binary_class_cross_entropy"
    MULTI_BIN_LABEL_CROSS_ENTROPY = "multi_binary_label_cross_entropy"
X
xuwei06 已提交
200
    SUM_COST = "sum_cost"
201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245

    @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.
246
    :type parents: list|tuple|collections.Sequence
247 248
    """

249 250 251 252 253 254 255 256 257
    def __init__(self,
                 name,
                 layer_type,
                 parents=None,
                 activation=None,
                 num_filters=None,
                 img_norm_type=None,
                 size=None,
                 outputs=None,
258
                 reverse=None):
259 260
        assert isinstance(name, basestring)
        assert isinstance(layer_type, basestring)
261
        assert size is not None
262 263 264
        assert LayerType.is_layer_type(layer_type)
        self.name = name
        self.layer_type = layer_type
265 266
        if parents is not None and type(parents) != list:
            parents = [parents]
267 268 269 270 271 272 273 274
        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
275
        self.reverse = reverse
276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291

    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'
292
DEVICE = 'device'
293 294 295


def layer_support(*attrs):
296
    attrs_list = list(attrs)
297
    attrs_list.append(DEVICE)
298

299 300 301
    def decorator(method):
        @functools.wraps(method)
        def wrapper(*args, **kwargs):
302
            for attr in attrs_list:
303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357
                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
    """
358 359
    proj = FullMatrixProjection(
        input_layer_name=input.name, size=size, **param_attr.attr)
360 361 362 363
    proj.origin = input
    return proj


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
@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
    """
394 395
    proj = TransposedFullMatrixProjection(
        input_layer_name=input.name, size=size, **param_attr.attr)
396 397 398 399
    proj.origin = input
    return proj


400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438
@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
    """
439 440
    proj = TableProjection(
        input_layer_name=input.name, size=size, **param_attr.attr)
441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475
    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.
476
    :type input: LayerOutput
477 478
    :param offset: Offset, None if use default.
    :type offset: int
X
xuwei06 已提交
479
    :return: A IdentityProjection or IdentityOffsetProjection object
480 481 482 483 484 485
    :rtype: IdentityProjection or IdentityOffsetProjection
    """
    if offset is None:
        proj = IdentityProjection(input_layer_name=input.name)
        proj.origin = input
    else:
486 487
        proj = IdentityOffsetProjection(
            input_layer_name=input.name, offset=offset)
488 489 490 491
        proj.origin = input
    return proj


X
xuwei06 已提交
492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513
@wrap_param_attr_default()
def scaling_projection(input, param_attr=None):
    """
    scaling_projection multiplies the input with a scalar parameter and add to
    the output.

    .. math::
       out += w * in

    The example usage is:

    .. code-block:: python

       proj = scaling_projection(input=layer)

    :param input: Input Layer.
    :type input: LayerOutput
    :param param_attr: Parameter config, None if use default.
    :type param_attr: ParameterAttribute
    :return: A ScalingProjection object
    :rtype: ScalingProjection
    """
514
    proj = ScalingProjection(input_layer_name=input.name, **param_attr.attr)
X
xuwei06 已提交
515 516 517 518
    proj.origin = input
    return proj


519
@wrap_param_attr_default()
520
def dotmul_projection(input, param_attr=None):
521
    """
522
    DotMulProjection with a layer as input.
523 524 525 526 527 528 529 530 531 532 533 534 535
    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)

536 537 538 539 540 541 542
    :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
    """
543 544
    proj = DotMulProjection(
        input_layer_name=input.name, size=input.size, **param_attr.attr)
545
    proj.origin = input
546
    return proj
547

548 549

def dotmul_operator(a=None, b=None, scale=1, **kwargs):
550 551
    """
    DotMulOperator takes two inputs and performs element-wise multiplication:
552

553
    .. math::
554 555
       out.row[i] += scale * (x.row[i] .* y.row[i])

556 557
    where :math:`.*` means element-wise multiplication, and
    scale is a config scalar, its default value is one.
558

559
    The example usage is:
560

561
    .. code-block:: python
562 563 564

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

565 566 567 568
    :param a: Input layer1
    :type a: LayerOutput
    :param b: Input layer2
    :type b: LayerOutput
569 570
    :param scale: config scalar, default value is one.
    :type scale: float
571 572
    :return: A DotMulOperator Object.
    :rtype: DotMulOperator
573
    """
574 575 576
    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.')
577
    a = kwargs.get('x', a)  # For Backward capacity.
578 579 580 581 582 583
    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

584
    op = DotMulOperator(input_layer_names=[a.name, b.name], scale=scale)
585
    op.origin = [a, b]
586
    return op
587

588

589
@wrap_bias_attr_default(['padding_attr'])
590 591 592
def context_projection(input,
                       context_len,
                       context_start=None,
593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628
                       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

629 630 631 632 633 634
    proj = ContextProjection(
        input_layer_name=input.name,
        context_length=context_len,
        context_start=context_start,
        trainable_padding=trainable,
        **extra_dict)
635 636 637 638 639 640 641 642 643 644 645 646 647
    proj.origin = input
    return proj


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

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

648
    def __init__(self, name, size, act, bias_attr, layer_attr, parents=None):
649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664
        """
        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
        """
665 666 667 668 669 670 671
        LayerOutput.__init__(
            self,
            name,
            LayerType.MIXED_LAYER,
            parents,
            size=size,
            activation=act)
672 673 674 675 676
        self.bias_attr = bias_attr
        self.layer_attr = layer_attr
        self.inputs = []
        self.finalized = False

677
    def __iadd__(self, other):
678 679 680 681 682 683 684 685
        """
        + += operator
        :param other: Other projection.
        :type other: Projection
        :return: self.
        :rtype: MixedLayerType
        """
        if not self.finalized:
686
            assert isinstance(other, Projection) or isinstance(other, Operator)
687
            self.inputs.append(other)
688 689 690 691
            if isinstance(other, Projection):
                self.parents.append(other.origin)
            else:
                self.parents.extend(other.origin)
692 693 694 695 696 697 698 699 700 701 702
            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
703
        ml = MixedLayer(
704 705 706 707 708
            name=self.name,
            size=self.size,
            active_type=self.activation.name,
            bias=ParamAttr.to_bias(self.bias_attr),
            inputs=self.inputs,
709
            **ExtraLayerAttribute.to_kwargs(self.layer_attr))
710 711 712
        # update the size which might be computed inside MixedLayer
        # according to the operator's output size
        self.size = ml.config.size
713
        self.finalized = True
714 715 716 717 718 719


@wrap_name_default("mixed")
@wrap_act_default(act=LinearActivation())
@wrap_bias_attr_default(has_bias=False)
@layer_support(ERROR_CLIPPING, DROPOUT)
720 721 722 723 724
def mixed_layer(size=0,
                input=None,
                name=None,
                act=None,
                bias_attr=False,
725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768
                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:
769 770 771 772 773 774
        with mixed_layer(
                name=name,
                size=size,
                act=act,
                bias_attr=bias_attr,
                layer_attr=layer_attr) as m:
775
            if isinstance(input, collections.Sequence):
776 777 778 779 780 781 782 783
                for each in input:
                    m += each
            else:
                m += input
        return m


@layer_support()
784
def data_layer(name, size, height=None, width=None, layer_attr=None):
785 786 787 788 789 790 791 792 793 794 795 796 797 798
    """
    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
799 800 801 802
    :param height: Height of this data layer, used for image
    :type size: int|None
    :param width: Width of this data layer, used for image
    :type size: int|None
803 804
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
805
    :return: LayerOutput object.
806 807
    :rtype: LayerOutput
    """
808 809 810 811
    Layer(
        type=LayerType.DATA,
        name=name,
        size=size,
812 813
        height=height,
        width=width,
814
        **ExtraLayerAttribute.to_kwargs(layer_attr))
815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836

    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
837
    :return: LayerOutput object.
838 839
    :rtype: LayerOutput
    """
840 841 842 843 844 845
    with mixed_layer(
            name=name,
            size=size,
            act=LinearActivation(),
            bias_attr=False,
            layer_attr=layer_attr) as mix:
846 847 848 849 850 851 852 853 854
        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)
855 856 857 858 859 860 861
def fc_layer(input,
             size,
             act=None,
             name=None,
             param_attr=None,
             bias_attr=None,
             layer_attr=None):
862 863 864 865 866 867 868 869 870 871 872 873
    """
    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)

874
    which is equal to:
875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896

    .. 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
897
    :return: LayerOutput object.
898 899 900 901
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
902
        assert not isinstance(param_attr, collections.Sequence)
903 904
        param_attr = [param_attr]
    else:
905
        if isinstance(param_attr, collections.Sequence):
906 907 908 909
            assert len(input) == len(param_attr)
        else:
            param_attr = [copy.deepcopy(param_attr) for _ in range(len(input))]

910
    assert isinstance(input, collections.Sequence)
911 912

    Layer(
913 914 915
        inputs=[
            Input(ipt.name, **attr.attr) for ipt, attr in zip(input, param_attr)
        ],
916 917 918 919 920
        name=name,
        type=LayerType.FC_LAYER,
        size=size,
        bias=ParamAttr.to_bias(bias_attr),
        active_type=act.name,
921 922 923
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.FC_LAYER, input, activation=act, size=size)
924

925

926 927 928 929
@wrap_name_default("print")
def print_layer(input, name=None):
    """
    Print the output value of input layers. This layer is useful for debugging.
930 931 932 933 934

    :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
935
    :return: LayerOutput
936
    """
937 938 939 940 941
    if isinstance(input, LayerOutput):
        input = [input]
    assert isinstance(input, collections.Sequence)  # list or tuple
    for each in input:
        assert isinstance(each, LayerOutput)
942 943 944 945

    Layer(
        name=name,
        type=LayerType.PRINT_LAYER,
946
        inputs=[l.name for l in input], )
947
    # this layer don't return anything, can not be input of other layer.
948

949

Y
yuan 已提交
950
@wrap_name_default("priorbox")
G
gaoyuan 已提交
951
def priorbox_layer(input,
952
                   image,
G
gaoyuan 已提交
953 954 955 956 957
                   aspect_ratio,
                   variance,
                   min_size,
                   max_size=[],
                   name=None):
Y
yuan 已提交
958 959 960 961 962 963 964
    """
    Compute the priorbox and set the variance. This layer is necessary for ssd.

    :param name: The Layer Name.
    :type name: basestring
    :param input: The input layer.
    :type input: LayerOutput
965 966
    :param image: The network input image.
    :type image: LayerOutput
Y
yuan 已提交
967 968 969 970 971 972 973 974 975 976 977
    :param aspect_ratio: The aspect ratio.
    :type aspect_ratio: list
    :param variance: The bounding box variance.
    :type min_size: The min size of the priorbox width/height.
    :param min_size: list
    :type max_size: The max size of the priorbox width/height. Could be NULL.
    :param max_size: list
    :return: LayerOutput
    """
    # plus one for ratio 1.
    num_filters = (len(aspect_ratio) * 2 + 1 + len(max_size)) * 4
G
gaoyuan 已提交
978
    size = (input.size / input.num_filters) * num_filters * 2
Y
yuan 已提交
979 980 981
    Layer(
        name=name,
        type=LayerType.PRIORBOX_LAYER,
982
        inputs=[input.name, image.name],
Y
yuan 已提交
983 984 985 986 987 988
        size=size,
        min_size=min_size,
        max_size=max_size,
        aspect_ratio=aspect_ratio,
        variance=variance)
    return LayerOutput(
G
gaoyuan 已提交
989 990
        name,
        LayerType.PRIORBOX_LAYER,
991
        parents=[input, image],
G
gaoyuan 已提交
992 993 994
        num_filters=num_filters,
        size=size)

995 996 997 998 999

@wrap_name_default("seq_pooling")
@wrap_bias_attr_default(has_bias=False)
@wrap_param_default(['pooling_type'], default_factory=lambda _: MaxPooling())
@layer_support()
1000 1001 1002 1003
def pooling_layer(input,
                  pooling_type=None,
                  name=None,
                  bias_attr=None,
1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016
                  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)

1017 1018
    :param agg_level: AggregateLevel.EACH_TIMESTEP or
                      AggregateLevel.EACH_SEQUENCE
1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030
    :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
1031
    :return: LayerOutput object.
1032
    :rtype: LayerOutput
1033 1034
    """
    extra_dict = dict()
1035
    # noinspection PyUnresolvedReferences
1036 1037
    if isinstance(pooling_type, AvgPooling):
        extra_dict['average_strategy'] = pooling_type.strategy
1038 1039 1040 1041
    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
1042 1043 1044 1045 1046 1047 1048 1049
    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,
1050
        **extra_dict)
1051

1052 1053
    return LayerOutput(
        name, pooling_type.name, parents=[input], size=input.size)
1054

Q
qijun 已提交
1055

1056 1057
@wrap_bias_attr_default()
@wrap_param_attr_default()
1058
@wrap_act_default(param_names=['gate_act'], act=SigmoidActivation())
1059 1060 1061
@wrap_act_default(param_names=["act", 'state_act'], act=TanhActivation())
@wrap_name_default("lstmemory")
@layer_support(DROPOUT)
1062 1063 1064 1065 1066 1067 1068 1069 1070
def lstmemory(input,
              name=None,
              reverse=False,
              act=None,
              gate_act=None,
              size=None,
              state_act=None,
              bias_attr=None,
              param_attr=None,
1071 1072 1073 1074 1075 1076 1077 1078
              layer_attr=None):
    """
    Long Short-term Memory Cell.

    The memory cell was implemented as follow equations.

    ..  math::

1079
        i_t & = \\sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i)
1080

1081
        f_t & = \\sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f)
1082

1083
        c_t & = f_tc_{t-1} + i_t tanh (W_{xc}x_t+W_{hc}h_{t-1} + b_c)
1084

1085
        o_t & = \\sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o)
1086

1087
        h_t & = o_t tanh(c_t)
1088 1089


1090
    NOTE: In PaddlePaddle's implementation, the multiplications
1091
    :math:`W_{xi}x_{t}` , :math:`W_{xf}x_{t}`,
1092 1093 1094 1095
    :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.
1096

1097
    NOTE: This is a low level user interface. You can use network.simple_lstm
1098 1099
    to config a simple plain lstm layer.

1100 1101 1102 1103
    Please refer to **Generating Sequences With Recurrent Neural Networks** for
    more details about LSTM.

    Link_ goes as below.
1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126

    .. _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
1127
    :return: LayerOutput object.
1128 1129 1130 1131 1132 1133
    :rtype: LayerOutput
    """

    assert gate_act.support_hppl
    assert state_act.support_hppl
    assert act.support_hppl
1134 1135 1136 1137 1138 1139 1140 1141 1142 1143
    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)
1144

1145 1146 1147 1148 1149 1150 1151 1152 1153 1154
    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))
1155

1156 1157 1158 1159 1160
    return LayerOutput(
        name,
        LayerType.LSTMEMORY, [input],
        size=input.size / 4,
        reverse=reverse)
1161

1162 1163 1164

@wrap_bias_attr_default()
@wrap_param_attr_default()
1165
@wrap_act_default(param_names=['gate_act'], act=SigmoidActivation())
1166 1167 1168
@wrap_act_default(param_names=["act"], act=TanhActivation())
@wrap_name_default("gru")
@layer_support(DROPOUT)
1169 1170 1171 1172 1173 1174 1175 1176
def grumemory(input,
              name=None,
              reverse=False,
              act=None,
              gate_act=None,
              size=None,
              bias_attr=None,
              param_attr=None,
1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197
              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)

1198 1199
    3. The candidate activation :math:`\\tilde{h_t}` is computed similarly to
    that of the traditional recurrent unit:
1200 1201 1202 1203 1204

    ..  math::

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

1205 1206 1207
    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}`:
1208 1209 1210 1211 1212

    ..  math::

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

1213
    NOTE: In PaddlePaddle's implementation, the multiplication operations
1214
    :math:`W_{r}x_{t}`, :math:`W_{z}x_{t}` and :math:`W x_t` are not computed in
1215 1216 1217
    gate_recurrent layer. Consequently, an additional mixed_layer with
    full_matrix_projection or a fc_layer must be included before grumemory
    is called.
1218

1219 1220 1221
    More details can be found by referring to `Empirical Evaluation of Gated
    Recurrent Neural Networks on Sequence Modeling.
    <https://arxiv.org/abs/1412.3555>`_
1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232

    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.
1233
    :param reverse: Whether sequence process is reversed or not.
1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248
    :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
1249 1250 1251
    :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
1252
    :return: LayerOutput object.
1253 1254 1255 1256
    :rtype: LayerOutput
    """
    assert act.support_hppl
    assert gate_act.support_hppl
1257 1258 1259 1260 1261 1262 1263 1264 1265
    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.")
1266

1267 1268 1269 1270 1271 1272 1273 1274 1275
    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))
1276

1277 1278 1279 1280 1281
    return LayerOutput(
        name,
        LayerType.GRUMEMORY, [input],
        size=input.size / 3,
        reverse=reverse)
1282

1283 1284 1285

@wrap_name_default()
@layer_support()
1286 1287 1288
def last_seq(input,
             name=None,
             agg_level=AggregateLevel.EACH_TIMESTEP,
1289 1290 1291 1292
             layer_attr=None):
    """
    Get Last Timestamp Activation of a sequence.

L
Luo Tao 已提交
1293 1294 1295 1296 1297 1298
    The simple usage is:

    .. code-block:: python

       seq = last_seq(input=layer)

1299 1300 1301 1302 1303 1304 1305
    :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
1306
    :return: LayerOutput object.
1307 1308
    :rtype: LayerOutput
    """
1309 1310 1311 1312 1313 1314
    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.")

1315 1316 1317 1318 1319
    Layer(
        name=name,
        type=LayerType.SEQUENCE_LAST_INSTANCE,
        inputs=[input.name],
        trans_type=agg_level,
1320 1321 1322 1323 1324 1325
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.SEQUENCE_LAST_INSTANCE,
        parents=[input],
        size=input.size)
1326 1327 1328 1329


@wrap_name_default()
@layer_support()
1330 1331 1332
def first_seq(input,
              name=None,
              agg_level=AggregateLevel.EACH_TIMESTEP,
1333 1334 1335 1336
              layer_attr=None):
    """
    Get First Timestamp Activation of a sequence.

L
Luo Tao 已提交
1337 1338 1339 1340 1341 1342
    The simple usage is:

    .. code-block:: python

       seq = first_seq(input=layer)

1343 1344 1345 1346 1347 1348 1349
    :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
1350
    :return: LayerOutput object.
1351 1352
    :rtype: LayerOutput
    """
1353 1354 1355 1356 1357 1358 1359

    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.')

1360 1361 1362 1363 1364
    Layer(
        name=name,
        type=LayerType.SEQUENCE_FIRST_INSTANCE,
        inputs=[input.name],
        trans_type=agg_level,
1365 1366 1367 1368 1369 1370
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.SEQUENCE_FIRST_INSTANCE,
        parents=[input],
        size=input.size)
1371 1372 1373 1374 1375 1376


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

1377

1378 1379
@wrap_name_default()
@layer_support()
1380 1381
def expand_layer(input,
                 expand_as,
1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410
                 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
1411
    :return: LayerOutput object.
1412 1413 1414 1415 1416 1417 1418 1419 1420
    :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,
1421 1422 1423 1424 1425 1426
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name=name,
        size=input.size,
        layer_type=LayerType.EXPAND_LAYER,
        parents=[input, expand_as])
1427 1428


1429 1430
@wrap_name_default()
@layer_support()
1431
def repeat_layer(input, num_repeats, name=None, layer_attr=None):
1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442
    """
    A layer for repeating the input for num_repeats times. This is equivalent
    to apply concat_layer() with num_repeats same input.

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

    The example usage is:

    .. code-block:: python

L
Luo Tao 已提交
1443
       expand = repeat_layer(input=layer, num_repeats=4)
1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461

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

    l = Layer(
        inputs=[input.name],
        name=name,
        num_filters=num_repeats,
        type=LayerType.FEATURE_MAP_EXPAND_LAYER,
1462 1463 1464 1465 1466 1467 1468
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name=name,
        size=l.config.size,
        layer_type=LayerType.FEATURE_MAP_EXPAND_LAYER,
        parents=[input])

1469

1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524
@wrap_name_default("seqreshape")
@wrap_act_default(act=IdentityActivation())
@wrap_bias_attr_default(has_bias=False)
@layer_support()
def seq_reshape_layer(input,
                      reshape_size,
                      act=None,
                      name=None,
                      layer_attr=None,
                      bias_attr=None):
    """
    A layer for reshaping the sequence. Assume the input sequence has T instances,
    the dimension of each instance is M, and the input reshape_size is N, then the 
    output sequence has T*M/N instances, the dimension of each instance is N.

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

    The example usage is:

    .. code-block:: python

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

    :param input: Input layer.
    :type input: LayerOutput
    :param reshape_size: the size of reshaped sequence.
    :type reshape_size: int
    :param name: Layer name.
    :type name: basestring
    :param act: Activation type.
    :type act: BaseActivation
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
    :param bias_attr: The Bias Attribute. If no bias, then pass False or
                      something not type of ParameterAttribute. None will get a
                      default Bias.
    :type bias_attr: ParameterAttribute or None or bool
    :return: LayerOutput object.
    :rtype: LayerOutput
    """

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


1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552
@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
1553
    :return: LayerOutput object.
1554 1555
    :rtype: LayerOutput
    """
1556
    assert isinstance(input, collections.Sequence)
1557
    assert len(input) == 2
1558 1559 1560 1561 1562 1563 1564
    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
1565 1566 1567 1568
    Layer(
        name=name,
        type=LayerType.INTERPOLATION_LAYER,
        inputs=[weight.name, input[0].name, input[1].name],
1569 1570 1571 1572 1573 1574
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.INTERPOLATION_LAYER,
        parents=[weight, input[0], input[1]],
        size=input[0].size)
1575 1576


1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592
@wrap_name_default()
@layer_support()
def bilinear_interp_layer(input,
                          out_size_x=None,
                          out_size_y=None,
                          name=None,
                          layer_attr=None):
    """
    This layer is to implement bilinear interpolation on conv layer output.

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

    The simple usage is:

    .. code-block:: python

L
liaogang 已提交
1593
       bilinear = bilinear_interp_layer(input=layer1, out_size_x=64, out_size_y=64)
X
xuwei06 已提交
1594

L
liaogang 已提交
1595
    :param   input:        A input layer.
1596
    :type    input:        LayerOutput.
L
liaogang 已提交
1597
    :param   out_size_x:   bilinear interpolation output width.
X
xuwei06 已提交
1598
    :type    out_size_x:   int|None
L
liaogang 已提交
1599
    :param   out_size_y:   bilinear interpolation output height.
1600
    :type    out_size_y:   int|None
L
liaogang 已提交
1601
    :param   name:         The layer's name, which cna not be specified.
1602
    :type    name:         None|basestring
L
liaogang 已提交
1603
    :param   layer_attr:   Extra Layer attribute.
1604 1605 1606 1607 1608 1609 1610
    :type    layer_attr:   ExtraLayerAttribute
    :return: LayerOutput object.
    :rtype:  LayerOutput
    """
    assert input.layer_type == LayerType.CONV_LAYER
    assert isinstance(input.activation, LinearActivation)
    assert out_size_x > 0 and out_size_y > 0
L
liaogang 已提交
1611
    assert input.num_filters is not None
L
liaogang 已提交
1612
    num_channels = input.num_filters
1613 1614 1615 1616 1617 1618 1619
    l = Layer(
        name=name,
        inputs=Input(
            input.name,
            bilinear_interp=BilinearInterp(
                out_size_x=out_size_x,
                out_size_y=out_size_y,
1620
                channels=num_channels)),
1621 1622 1623 1624 1625 1626 1627 1628 1629
        type=LayerType.BILINEAR_INTERP_LAYER,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.BILINEAR_INTERP_LAYER,
        parents=[input],
        num_filters=num_channels,
        size=l.config.size)

1630

1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657
@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
1658
    :return: LayerOutput object.
1659 1660
    :rtype: LayerOutput
    """
1661 1662 1663
    assert isinstance(input, LayerOutput) and isinstance(weight, LayerOutput)
    if weight.size is not None:
        assert weight.size == 1
1664 1665 1666
    Layer(
        name=name,
        type=LayerType.POWER_LAYER,
1667
        inputs=[weight.name, input.name],
1668 1669 1670
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.POWER_LAYER, parents=[input, weight], size=input.size)
1671 1672 1673 1674 1675 1676


@wrap_name_default()
@layer_support()
def scaling_layer(input, weight, name=None, layer_attr=None):
    """
1677
    A layer for multiplying input vector by weight scalar.
1678 1679

    .. math::
1680
       y  = w x
1681

1682 1683 1684 1685 1686
    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.
1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701

    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
1702
    :return: LayerOutput object.
1703 1704
    :rtype: LayerOutput
    """
1705 1706 1707
    assert isinstance(weight, LayerOutput) and isinstance(input, LayerOutput)
    if weight.size is not None:
        assert weight.size == 1
1708 1709 1710 1711
    Layer(
        name=name,
        type=LayerType.SCALING_LAYER,
        inputs=[weight.name, input.name],
1712 1713 1714
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.SCALING_LAYER, parents=[weight, input], size=input.size)
1715 1716 1717 1718 1719 1720


@wrap_name_default()
@layer_support()
def trans_layer(input, name=None, layer_attr=None):
    """
1721
    A layer for transposing a minibatch matrix.
1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739

    .. 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
1740
    :return: LayerOutput object.
1741 1742 1743 1744 1745 1746
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.TRANS_LAYER,
        inputs=[input.name],
1747 1748 1749
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.TRANS_LAYER, parents=[input], size=input.size)
1750 1751


1752 1753
@wrap_name_default()
@layer_support()
1754
def rotate_layer(input, height, width, name=None, layer_attr=None):
1755
    """
H
Haonan 已提交
1756 1757
    A layer for rotating 90 degrees (clock-wise) for each feature channel,
    usually used when the input sample is some image or feature map.
1758 1759

    .. math::
1760
       y(j,i,:) = x(M-i-1,j,:)
1761

1762
    where :math:`x` is (M x N x C) input, and :math:`y` is (N x M x C) output.
1763 1764 1765 1766 1767 1768

    The example usage is:

    .. code-block:: python

       rot = rotate_layer(input=layer,
1769 1770
                          height=100,
                          width=100)
1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783

    :param input: Input layer.
    :type input: LayerOutput
    :param height: The height of the sample matrix
    :type height: int
    :param name: Layer name.
    :type name: basestring
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
    :return: LayerOutput object.
    :rtype: LayerOutput
    """
    assert isinstance(input, LayerOutput)
H
Haonan 已提交
1784 1785 1786
    l = Layer(
        name=name,
        height=height,
1787
        width=width,
H
Haonan 已提交
1788 1789 1790 1791 1792 1793 1794 1795
        type=LayerType.ROTATE_LAYER,
        inputs=[input.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name=name,
        layer_type=LayerType.ROTATE_LAYER,
        parents=[input],
        size=l.config.size)
1796 1797


1798 1799
@wrap_name_default()
@layer_support()
1800
def cos_sim(a, b, scale=1, size=1, name=None, layer_attr=None):
1801 1802 1803 1804
    """
    Cosine Similarity Layer. The cosine similarity equation is here.

    ..  math::
D
dangqingqing 已提交
1805
        similarity = cos(\\theta) = {\\mathbf{a} \\cdot \\mathbf{b}
1806 1807 1808 1809 1810
        \\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.
1811

1812 1813
    Note that the above computation is for one sample. Multiple samples are
    processed in one batch.
1814

L
Luo Tao 已提交
1815 1816 1817 1818 1819 1820
    The example usage is:

    .. code-block:: python

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

1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832
    :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
1833
    :return: LayerOutput object.
1834 1835
    :rtype: LayerOutput
    """
1836
    assert isinstance(a, LayerOutput) and isinstance(b, LayerOutput)
1837 1838 1839 1840 1841 1842
    if size == 1:
        Layer(
            name=name,
            type=LayerType.COSINE_SIM,
            cos_scale=scale,
            inputs=[a.name, b.name],
1843
            **ExtraLayerAttribute.to_kwargs(layer_attr))
1844
    else:
1845 1846
        if a.size is not None and b.size is not None:
            assert size == b.size / a.size
1847 1848 1849 1850 1851 1852
        Layer(
            name=name,
            type=LayerType.COSINE_SIM_VEC,
            size=size,
            cos_scale=scale,
            inputs=[a.name, b.name],
1853
            **ExtraLayerAttribute.to_kwargs(layer_attr))
1854
    return LayerOutput(name, LayerType.COSINE_SIM, parents=[a, b], size=size)
1855

1856

1857 1858
@wrap_name_default()
@wrap_bias_attr_default(has_bias=True)
1859
@wrap_param_attr_default()
1860
@layer_support()
1861 1862 1863 1864 1865 1866 1867
def hsigmoid(input,
             label,
             num_classes,
             name=None,
             bias_attr=None,
             param_attr=None,
             layer_attr=None):
1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888
    """
    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
1889 1890
    :param name: layer name
    :type name: basestring
1891 1892 1893 1894 1895
    :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
1896
    :return: LayerOutput object.
1897 1898 1899 1900
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
1901 1902 1903 1904 1905 1906 1907 1908 1909
        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)
1910 1911 1912 1913 1914
    assert isinstance(label, LayerOutput)
    assert label.layer_type == LayerType.DATA

    ipts_for_layer = []
    parents = []
1915
    for each_input, each_param_attr in zip(input, param_attr):
1916
        assert isinstance(each_input, LayerOutput)
1917
        ipts_for_layer.append(Input(each_input.name, **each_param_attr.attr))
1918 1919 1920 1921
        parents.append(each_input)
    ipts_for_layer.append(label.name)
    parents.append(label)

1922
    l = Layer(
1923 1924 1925 1926 1927
        name=name,
        type=LayerType.HSIGMOID,
        num_classes=num_classes,
        bias=ParamAttr.to_bias(bias_attr),
        inputs=ipts_for_layer,
1928 1929 1930
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.HSIGMOID, parents=parents, size=l.config.size)
1931

1932

1933 1934 1935 1936 1937
@wrap_name_default("conv")
@wrap_param_attr_default()
@wrap_bias_attr_default()
@wrap_act_default(act=ReluActivation())
@layer_support(DROPOUT)
1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953
def img_conv_layer(input,
                   filter_size,
                   num_filters,
                   name=None,
                   num_channels=None,
                   act=None,
                   groups=1,
                   stride=1,
                   padding=0,
                   bias_attr=None,
                   param_attr=None,
                   shared_biases=True,
                   layer_attr=None,
                   filter_size_y=None,
                   stride_y=None,
                   padding_y=None,
1954 1955
                   trans=False,
                   layer_type=None):
1956
    """
1957
    Convolution layer for image. Paddle can support both square and non-square
1958
    input currently.
1959 1960 1961 1962

    The details of convolution layer, please refer UFLDL's `convolution
    <http://ufldl.stanford.edu/tutorial/supervised/
    FeatureExtractionUsingConvolution/>`_ .
X
xuwei06 已提交
1963

1964
    Convolution Transpose (deconv) layer for image. Paddle can support both square
1965
    and non-square input currently.
1966

X
xuwei06 已提交
1967
    The details of convolution transpose layer,
1968 1969 1970
    please refer to the following explanation and references therein
    <http://datascience.stackexchange.com/questions/6107/
    what-are-deconvolutional-layers/>`_ .
1971 1972 1973 1974
    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.

1975 1976 1977
    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
1978
    32*4 = 128 filters to process inputs. The channels will be split into 4
1979 1980
    pieces. First 256/4 = 64 channels will process by first 32 filters. The
    rest channels will be processed by rest group of filters.
1981

L
Luo Tao 已提交
1982 1983 1984 1985 1986 1987 1988 1989 1990 1991
    The example usage is:

    ..  code-block:: python

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

1992 1993 1994 1995
    :param name: Layer name.
    :type name: basestring
    :param input: Layer Input.
    :type input: LayerOutput
1996 1997 1998
    :param filter_size: The x dimension of a filter kernel. Or input a tuple for
                        two image dimension.
    :type filter_size: int|tuple|list
1999 2000 2001
    :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).
2002
    :type filter_size_y: int|None
2003 2004 2005 2006 2007
    :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
2008 2009 2010
    :param stride: The x dimension of the stride. Or input a tuple for two image
                   dimension.
    :type stride: int|tuple|list
2011 2012
    :param stride_y: The y dimension of the stride.
    :type stride_y: int
2013 2014 2015
    :param padding: The x dimension of the padding. Or input a tuple for two
                    image dimension
    :type padding: int|tuple|list
2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029
    :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
2030 2031
    :param trans: true if it is a convTransLayer, false if it is a convLayer
    :type trans: bool
2032
    :param layer_type: specify the layer_type, default is None. If trans=True,
2033
                       layer_type has to be "exconvt", otherwise layer_type
2034 2035
                       has to be either "exconv" or "cudnn_conv"
    :type layer_type: String
D
dangqingqing 已提交
2036
    :return: LayerOutput object.
2037 2038 2039 2040 2041
    :rtype: LayerOutput
    """
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters
2042

2043
    if filter_size_y is None:
2044 2045 2046 2047 2048 2049
        if isinstance(filter_size, collections.Sequence):
            assert len(filter_size) == 2
            filter_size, filter_size_y = filter_size
        else:
            filter_size_y = filter_size

2050
    if stride_y is None:
2051 2052 2053 2054 2055 2056
        if isinstance(stride, collections.Sequence):
            assert len(stride) == 2
            stride, stride_y = stride
        else:
            stride_y = stride

2057
    if padding_y is None:
2058 2059 2060 2061 2062 2063 2064 2065
        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.
2066
        init_w = (2.0 / (filter_size**2 * num_channels))**0.5
2067 2068 2069 2070
        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
2071

2072 2073 2074 2075 2076 2077 2078 2079
    if layer_type:
        if trans:
            assert layer_type in ["exconvt"]
        else:
            assert layer_type in ["exconv", "cudnn_conv"]
        lt = layer_type
    else:
        lt = LayerType.CONVTRANS_LAYER if trans else LayerType.CONV_LAYER
2080

2081
    l = Layer(
2082
        name=name,
2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094
        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),
2095 2096 2097 2098
        active_type=act.name,
        num_filters=num_filters,
        bias=ParamAttr.to_bias(bias_attr),
        shared_biases=shared_biases,
2099
        type=lt,
2100 2101 2102 2103 2104 2105 2106 2107
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        lt,
        parents=[input],
        activation=act,
        num_filters=num_filters,
        size=l.config.size)
2108 2109 2110 2111


@wrap_name_default("pool")
@layer_support()
2112 2113 2114 2115 2116 2117 2118 2119 2120 2121
def img_pool_layer(input,
                   pool_size,
                   name=None,
                   num_channels=None,
                   pool_type=None,
                   stride=1,
                   padding=0,
                   layer_attr=None,
                   pool_size_y=None,
                   stride_y=None,
2122 2123
                   padding_y=None,
                   ceil_mode=True):
2124 2125 2126 2127 2128 2129 2130
    """
    Image pooling Layer.

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

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

L
Luo Tao 已提交
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
    - ceil_mode=True:

    ..  math::

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

    - ceil_mode=False:

    ..  math::

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

    The example usage is:

    ..  code-block:: python

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

2159
    :param padding: pooling padding width.
2160
    :type padding: int
2161 2162
    :param padding_y: pooling padding height. It's equal to padding by default.
    :type padding_y: int|None
2163 2164 2165 2166
    :param name: name of pooling layer
    :type name: basestring.
    :param input: layer's input
    :type input: LayerOutput
2167
    :param pool_size: pooling window width
2168
    :type pool_size: int
2169 2170
    :param pool_size_y: pooling window height. It's eaqual to pool_size by default.
    :type pool_size_y: int|None
2171 2172
    :param num_channels: number of input channel.
    :type num_channels: int
2173
    :param pool_type: pooling type. MaxPooling or AvgPooling. Default is
2174 2175
                      MaxPooling.
    :type pool_type: BasePoolingType
2176
    :param stride: stride width of pooling.
2177
    :type stride: int
2178 2179
    :param stride_y: stride height of pooling. It is equal to stride by default.
    :type stride_y: int|None
2180 2181
    :param layer_attr: Extra Layer attribute.
    :type layer_attr: ExtraLayerAttribute
2182 2183 2184 2185
    :param ceil_mode: Wether to use ceil mode to calculate output height and with.
                      Defalut is True. If set false, Otherwise use floor.

    :type ceil_mode: bool
D
dangqingqing 已提交
2186 2187
    :return: LayerOutput object.
    :rtype: LayerOutput
2188 2189 2190 2191 2192 2193 2194 2195 2196 2197
    """
    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'

2198 2199 2200 2201 2202 2203 2204 2205
    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

2206
    l = Layer(
2207 2208
        name=name,
        type=LayerType.POOL_LAYER,
2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220
        inputs=[
            Input(
                input.name,
                pool=Pool(
                    pool_type=type_name,
                    channels=num_channels,
                    size_x=pool_size,
                    start=None,
                    stride=stride,
                    padding=padding,
                    size_y=pool_size_y,
                    stride_y=stride_y,
2221
                    padding_y=padding_y))
2222
        ],
2223
        ceil_mode=ceil_mode,
2224 2225 2226 2227 2228 2229 2230
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.POOL_LAYER,
        parents=[input],
        num_filters=num_channels,
        size=l.config.size)
2231 2232


Q
qijun 已提交
2233 2234
@wrap_name_default("spp")
@layer_support()
2235 2236 2237 2238 2239 2240
def spp_layer(input,
              name=None,
              num_channels=None,
              pool_type=None,
              pyramid_height=None,
              layer_attr=None):
Q
qijun 已提交
2241 2242 2243 2244 2245
    """
    Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition.
    The details please refer to
    `Kaiming He's paper <https://arxiv.org/abs/1406.4729>`_.

L
Luo Tao 已提交
2246 2247 2248 2249 2250 2251 2252 2253 2254
    The example usage is:

    ..  code-block:: python

        spp = spp_layer(input=data, 
                        pyramid_height=2, 
                        num_channels=16, 
                        pool_type=MaxPooling())

Q
qijun 已提交
2255 2256 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
    :param name: layer name.
    :type name: basestring
    :param input: layer's input.
    :type input: LayerOutput
    :param num_channels: number of input channel.
    :type num_channels: int
    :param pool_type: Pooling type. MaxPooling or AveragePooling. Default is MaxPooling.
    :type scale: BasePoolingType
    :param pyramid_height: pyramid height.
    :type pyramid_height: int
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
    :return: LayerOutput object.
    :rtype: LayerOutput
    """
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters

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

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

2283
    l = Layer(
Q
qijun 已提交
2284 2285
        name=name,
        type=LayerType.SPP_LAYER,
2286 2287 2288 2289 2290
        inputs=Input(
            input.name,
            spp=SpatialPyramidPool(
                pool_type=type_name,
                channels=num_channels,
2291
                pyramid_height=pyramid_height)),
2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        layer_type=LayerType.SPP_LAYER,
        parents=[input],
        num_filters=num_channels,
        size=l.config.size)


def __img_norm_layer__(name, input, size, norm_type, scale, power, num_channels,
                       blocked, layer_attr):
2303 2304 2305 2306
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters

2307
    l = Layer(
2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326
        name=name,
        type=LayerType.NORM_LAYER,
        inputs=Input(
            input.name,
            norm=Norm(
                norm_type=norm_type,
                channels=num_channels,
                size=size,
                scale=scale,
                pow=power,
                blocked=blocked)),
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        layer_type=LayerType.NORM_LAYER,
        parents=[input],
        num_filters=num_channels,
        img_norm_type=norm_type,
        size=l.config.size)
2327 2328 2329 2330


@wrap_name_default("crmnorm")
@layer_support()
2331 2332 2333 2334 2335 2336
def img_cmrnorm_layer(input,
                      size,
                      scale=0.0128,
                      power=0.75,
                      name=None,
                      num_channels=None,
2337
                      layer_attr=None):
2338
    """
2339
    Response normalization across feature maps.
D
dangqingqing 已提交
2340 2341
    The details please refer to
    `Alex's paper <http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf>`_.
2342

L
Luo Tao 已提交
2343 2344 2345 2346 2347 2348
    The example usage is:

    ..  code-block:: python
    
        norm = img_cmrnorm_layer(input=net, size=5)

2349
    :param name: layer name.
D
dangqingqing 已提交
2350
    :type name: None|basestring
2351 2352
    :param input: layer's input.
    :type input: LayerOutput
2353
    :param size: Normalize in number of :math:`size` feature maps.
2354
    :type size: int
D
dangqingqing 已提交
2355
    :param scale: The hyper-parameter.
2356
    :type scale: float
D
dangqingqing 已提交
2357
    :param power: The hyper-parameter.
2358 2359 2360 2361 2362
    :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
2363
    :return: LayerOutput object.
2364 2365 2366
    :rtype: LayerOutput
    """
    return __img_norm_layer__(name, input, size, "cmrnorm-projection", scale,
2367
                              power, num_channels, 0, layer_attr)
2368 2369 2370 2371 2372 2373 2374 2375


@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)
2376 2377 2378 2379 2380 2381 2382
def batch_norm_layer(input,
                     act=None,
                     name=None,
                     num_channels=None,
                     bias_attr=None,
                     param_attr=None,
                     layer_attr=None,
2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403
                     batch_norm_type=None,
                     moving_average_fraction=0.9,
                     use_global_stats=None):
    """
    Batch Normalization Layer. The notation of this layer as follow.

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

    ..  math::

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

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

L
Luo Tao 已提交
2404 2405 2406 2407 2408 2409
    The example usage is:

    ..  code-block:: python
    
        norm = batch_norm_layer(input=net, act=ReluActivation())

2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423
    :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.
2424
    :type batch_norm_type: None|string, None or "batch_norm" or "cudnn_batch_norm"
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
    :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
2452
    :return: LayerOutput object.
2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471
    :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")
2472
    l = Layer(
2473
        name=name,
2474 2475
        inputs=Input(
            input.name, image=Image(channels=num_channels), **param_attr.attr),
2476 2477 2478 2479 2480 2481
        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,
2482
        **ExtraLayerAttribute.to_kwargs(layer_attr))
2483

2484 2485 2486 2487 2488 2489 2490
    return LayerOutput(
        name=name,
        layer_type=LayerType.BATCH_NORM_LAYER,
        parents=[input],
        activation=act,
        num_filters=num_channels,
        size=l.config.size)
2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517


@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
2518
    :return: LayerOutput object.
2519 2520 2521 2522 2523 2524
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.SUM_TO_ONE_NORM_LAYER,
        inputs=[input.name],
2525 2526 2527
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.SUM_TO_ONE_NORM_LAYER, parents=[input], size=input.size)
2528 2529 2530 2531 2532 2533


@wrap_name_default("addto")
@wrap_act_default(act=LinearActivation())
@wrap_bias_attr_default(has_bias=False)
@layer_support(DROPOUT)
2534
def addto_layer(input, act=None, name=None, bias_attr=None, layer_attr=None):
2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556
    """
    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.

2557 2558 2559
    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.
2560 2561

    It is a very good way to set dropout outside the layers. Since not all
2562 2563
    PaddlePaddle layer support dropout, you can add an add_to layer, set
    dropout here.
2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577
    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
2578
    :return: LayerOutput object.
2579 2580 2581 2582 2583 2584
    :rtype: LayerOutput
    """
    num_filters = None
    if isinstance(input, LayerOutput):
        input = [input]

2585
    assert isinstance(input, collections.Sequence)
2586 2587 2588 2589 2590 2591 2592
    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

2593
    l = Layer(
2594 2595 2596
        name=name,
        type=LayerType.ADDTO_LAYER,
        inputs=ipts_for_layer,
2597 2598
        bias=ParamAttr.to_bias(bias_attr),
        active_type=act.name,
2599
        **ExtraLayerAttribute.to_kwargs(layer_attr))
2600

2601 2602 2603 2604 2605 2606 2607
    return LayerOutput(
        name,
        LayerType.ADDTO_LAYER,
        parents=input,
        activation=act,
        num_filters=num_filters,
        size=l.config.size)
2608 2609 2610 2611 2612


@wrap_act_default(act=IdentityActivation())
@wrap_name_default("concat")
@layer_support()
2613
def concat_layer(input, act=None, name=None, layer_attr=None, bias_attr=None):
2614 2615 2616 2617
    """
    Concat all input vector into one huge vector.
    Inputs can be list of LayerOutput or list of projection.

2618 2619 2620 2621 2622 2623
    The example usage is:

    ..  code-block:: python

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

2624 2625 2626
    :param name: Layer name.
    :type name: basestring
    :param input: input layers or projections
2627
    :type input: list|tuple|collections.Sequence
2628 2629 2630 2631
    :param act: Activation type.
    :type act: BaseActivation
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
2632
    :return: LayerOutput object.
2633 2634 2635 2636 2637 2638 2639 2640
    :rtype: LayerOutput
    """

    if isinstance(input, LayerOutput):
        input = [input]
    elif isinstance(input, Projection):
        input = [input]
    else:
2641
        assert isinstance(input, collections.Sequence)
2642 2643

    def __is_type__(o, tp):
2644
        if not isinstance(o, collections.Sequence):
2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665
            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

2666 2667
    is_concat_layer = __is_type__(
        reduce(__reduce_concat_type__, map(type, input)), LayerOutput)
2668

2669 2670
    layer_type = (LayerType.CONCAT_LAYER
                  if is_concat_layer else LayerType.CONCAT_PROJ_LAYER)
2671

2672 2673
    if layer_type == LayerType.CONCAT_LAYER:
        assert not bias_attr
2674

2675
    Layer(
2676 2677
        name=name,
        type=layer_type,
2678 2679
        inputs=[x.name for x in input] if is_concat_layer else input,
        active_type=act.name,
2680
        bias=ParamAttr.to_bias(bias_attr),
2681
        **ExtraLayerAttribute.to_kwargs(layer_attr))
2682 2683 2684 2685 2686 2687 2688 2689 2690

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

2691 2692 2693 2694 2695 2696 2697 2698
    return LayerOutput(
        name,
        layer_type=layer_type,
        parents=input if is_concat_layer else [x.origin for x in input],
        activation=act,
        size=sz)


2699 2700
@wrap_name_default("seqconcat")
@wrap_act_default(act=IdentityActivation())
2701
@wrap_bias_attr_default(has_bias=False)
2702 2703 2704 2705 2706
@layer_support()
def seq_concat_layer(a, b, act=None, name=None, layer_attr=None,
                     bias_attr=None):
    """
    Concat sequence a with sequence b.
2707 2708 2709 2710 2711 2712

    Inputs: 
      - a = [a1, a2, ..., an]
      - b = [b1, b2, ..., bn]
      - Note that the length of a and b should be the same.
        
2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730
    Output: [a1, b1, a2, b2, ..., an, bn]

    The example usage is:

    ..  code-block:: python

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

    :param name: Layer name.
    :type name: basestring
    :param a: input sequence layer
    :type a: LayerOutput
    :param b: input sequence layer
    :type b: LayerOutput
    :param act: Activation type.
    :type act: BaseActivation
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
2731 2732 2733 2734
    :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
2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755
    :return: LayerOutput object.
    :rtype: LayerOutput
    """
    assert isinstance(a, LayerOutput) and isinstance(b, LayerOutput)
    assert a.size == b.size
    Layer(
        name=name,
        type=LayerType.SEQUENCE_CONCAT_LAYER,
        inputs=[a.name, b.name],
        active_type=act.name,
        bias=ParamAttr.to_bias(bias_attr),
        **ExtraLayerAttribute.to_kwargs(layer_attr))

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


2756 2757 2758 2759 2760 2761
def memory(name,
           size,
           is_seq=False,
           boot_layer=None,
           boot_bias=None,
           boot_bias_active_type=None,
2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 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
           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
2797
    :return: LayerOutput object which is a memory.
2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808
    :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)

2809 2810 2811 2812 2813 2814 2815 2816 2817
    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)
2818 2819 2820 2821
    return lout


@wrap_bias_attr_default()
2822 2823
@wrap_act_default(
    param_names=['gate_act', 'state_act'], act=SigmoidActivation())
2824 2825 2826
@wrap_act_default(act=TanhActivation())
@wrap_name_default('lstm_step')
@layer_support()
2827 2828 2829 2830 2831 2832 2833 2834 2835
def lstm_step_layer(input,
                    state,
                    size,
                    act=None,
                    name=None,
                    gate_act=None,
                    state_act=None,
                    bias_attr=None,
                    layer_attr=None):
2836 2837 2838 2839 2840 2841
    """
    LSTM Step Layer. It used in recurrent_group. The lstm equations are shown
    as follow.

    ..  math::

2842
        i_t & = \\sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i)
2843

2844
        f_t & = \\sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f)
2845

2846
        c_t & = f_tc_{t-1} + i_t tanh (W_{xc}x_t+W_{hc}h_{t-1} + b_c)
2847

2848
        o_t & = \\sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o)
2849

2850
        h_t & = o_t tanh(c_t)
2851 2852


2853
    The input of lstm step is :math:`Wx_t + Wh_{t-1}`, and user should use
2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891
    :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
2892
    :return: LayerOutput object.
2893 2894 2895 2896 2897 2898 2899 2900 2901
    :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),
2902 2903 2904
        size=size,
        inputs=[input.name, state.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
2905

2906 2907 2908 2909 2910 2911 2912
    return LayerOutput(
        name=name,
        layer_type=LayerType.LSTM_STEP_LAYER,
        parents=[input, state],
        activation=act,
        size=size,
        outputs=['default', 'state'])
2913 2914 2915


@wrap_bias_attr_default()
2916
@wrap_param_attr_default()
2917
@wrap_act_default(param_names=['gate_act'], act=SigmoidActivation())
2918 2919 2920
@wrap_act_default(act=TanhActivation())
@wrap_name_default('gru_step')
@layer_support()
2921 2922 2923 2924 2925 2926 2927
def gru_step_layer(input,
                   output_mem,
                   size=None,
                   act=None,
                   name=None,
                   gate_act=None,
                   bias_attr=None,
2928
                   param_attr=None,
2929
                   layer_attr=None):
2930 2931 2932 2933 2934 2935 2936 2937 2938 2939
    """

    :param input:
    :type input: LayerOutput
    :param output_mem:
    :param size:
    :param act:
    :param name:
    :param gate_act:
    :param bias_attr:
2940 2941
    :param param_attr: the parameter_attribute for transforming the output_mem
                       from previous step.
2942
    :param layer_attr:
D
dangqingqing 已提交
2943
    :return: LayerOutput object.
2944 2945 2946 2947 2948 2949 2950 2951
    :rtype: LayerOutput
    """
    assert input.size % 3 == 0
    if size is None:
        size = input.size / 3
    Layer(
        name=name,
        type=LayerType.GRU_STEP_LAYER,
2952 2953 2954 2955
        # The parameter here is for transforming the output_mem. The input has
        # already been transformed outside this module so it does not need
        # parameter associated with it.
        # The parameter here is instead grouped with input is due to
2956
        # backward model compatibility.
2957
        inputs=[Input(input.name, **param_attr.attr), output_mem.name],
2958 2959 2960 2961
        bias=ParamAttr.to_bias(bias_attr),
        size=size,
        active_type=act.name,
        active_gate_type=gate_act.name,
2962
        **ExtraAttr.to_kwargs(layer_attr))
2963
    return LayerOutput(
2964 2965
        name=name,
        layer_type=LayerType.GRU_STEP_LAYER,
2966
        parents=[input, output_mem],
2967 2968
        size=size,
        activation=act)
2969 2970 2971 2972 2973 2974


@wrap_name_default()
@layer_support()
def get_output_layer(input, arg_name, name=None, layer_attr=None):
    """
2975 2976 2977 2978
    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.
2979 2980 2981 2982 2983 2984 2985 2986 2987

    :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
2988
    :return: LayerOutput object.
2989 2990 2991 2992 2993 2994 2995
    :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))
2996 2997 2998 2999 3000 3001 3002
    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))
3003

3004 3005 3006 3007 3008
    return LayerOutput(
        name=name,
        layer_type=LayerType.GET_OUTPUT_LAYER,
        parents=[input],
        size=input.size)
3009 3010 3011 3012 3013 3014 3015


@wrap_name_default()
@wrap_act_default()
@wrap_bias_attr_default()
@wrap_param_attr_default()
@layer_support()
3016 3017 3018 3019 3020 3021 3022
def recurrent_layer(input,
                    act=None,
                    bias_attr=None,
                    param_attr=None,
                    name=None,
                    reverse=False,
                    layer_attr=None):
3023
    """
3024 3025
    Simple recurrent unit layer. It is just a fully connect layer through both
    time and neural network.
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
    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
3054
    :return: LayerOutput object.
3055
    :rtype: LayerOutput
3056
    """
3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071
    Layer(
        name=name,
        type=LayerType.RECURRENT_LAYER,
        inputs=Input(input.name, **param_attr.attr),
        active_type=act.name,
        bias=ParamAttr.to_bias(bias_attr),
        reversed=reverse,
        **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(
        name=name,
        layer_type=LayerType.RECURRENT_LAYER,
        parents=[input],
        size=input.size,
        activation=act,
        reverse=reverse)
3072 3073 3074 3075 3076 3077 3078


class StaticInput(object):
    """
    StaticInput is only used in recurrent_group which defines a read-only memory
    that can be a sequence or non-sequence.
    """
3079

3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098
    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)
    """
3099

3100 3101 3102 3103 3104 3105 3106
    def __init__(self, input):
        assert isinstance(input, LayerOutput)
        assert input.size is not None
        self.input = input


@wrap_name_default("recurrent_group")
3107 3108 3109 3110 3111
def recurrent_group(step,
                    input,
                    reverse=False,
                    name=None,
                    targetInlink=None,
3112
                    is_generating=False):
3113
    """
3114 3115 3116 3117 3118
    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.
3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162

    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

3163 3164
    :param reverse: If reverse is set true, the recurrent unit will process the
                    input sequence in a reverse order.
3165
    :type reverse: bool
3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176

    :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

3177
    :param is_generating: If is generating, none of input type should be LayerOutput;
3178
                          else, for training or testing, one of the input type must
3179
                          be LayerOutput.
3180

3181
    : type is_generating: bool
3182

D
dangqingqing 已提交
3183
    :return: LayerOutput object.
3184 3185 3186 3187 3188 3189 3190 3191 3192 3193
    :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]
3194
    assert isinstance(input, collections.Sequence)
3195 3196 3197 3198 3199 3200

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

    in_links = filter(is_in_links, input)

3201 3202 3203 3204 3205 3206 3207 3208 3209
    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

3210
    assert (targetInlink == None or targetInlink_in_inlinks())
3211 3212 3213 3214
    targetInlinkName = None if targetInlink == None \
                            else targetInlink.name if isinstance(targetInlink, LayerOutput) \
                                                   else targetInlink.input.name

3215 3216 3217 3218 3219 3220 3221 3222 3223 3224
    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(
3225 3226
        name=name,
        in_links=map(map_in_links, in_links),
3227 3228
        seq_reversed=reverse,
        target_inlinkname=targetInlinkName)
3229
    in_args = []
3230
    has_LayerOutput = False
3231 3232 3233 3234
    for each_input in input:
        assert is_single_input(each_input)
        if isinstance(each_input, LayerOutput):
            in_args.append(each_input)
3235
            has_LayerOutput = True
3236 3237
        elif isinstance(each_input, SubsequenceInput):
            in_args.append(each_input.input)
3238
            has_LayerOutput = True
3239 3240
        else:
            mem_name = "__%s_memory__" % each_input.input.name
3241 3242 3243 3244 3245 3246 3247 3248 3249
            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:
3250 3251 3252
                mix += identity_projection(mem)
            in_args.append(mem)

3253
    assert (is_generating != has_LayerOutput)
3254

3255 3256 3257 3258 3259 3260 3261
    layer_outs = step(*in_args)

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

    for ot in layer_outs:
        assert isinstance(ot, LayerOutput)
3262
        ot.reverse = reverse
3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274
        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

3275

3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292
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):
3293 3294 3295 3296 3297 3298 3299 3300 3301
        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))
3302 3303 3304
        return trg_emb

    def __init__(self, size, embedding_name, embedding_size):
3305
        super(GeneratedInput, self).__init__()
3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328
        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
3329
    :return: LayerOutput object.
3330 3331 3332 3333
    :rtype: LayerOutput
    """

    assert isinstance(input, LayerOutput)
3334 3335 3336 3337 3338 3339 3340 3341 3342 3343
    l = Layer(
        name=name,
        type='maxid',
        inputs=[input.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name=name,
        layer_type=LayerType.MAXID_LAYER,
        parents=[input],
        size=l.config.size)
3344

3345

3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371
@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)
3372 3373 3374 3375 3376 3377 3378 3379 3380 3381
    l = Layer(
        name=name,
        type=LayerType.OUT_PROD_LAYER,
        inputs=[input1.name, input2.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name=name,
        layer_type=LayerType.OUT_PROD_LAYER,
        parents=[input1, input2],
        size=l.config.size)
3382

3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398

@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)

3399 3400
    :param name: Layer name.
    :type name: basestring
3401 3402 3403 3404 3405 3406
    :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
3407
    :return: LayerOutput object.
3408 3409
    :rtype: LayerOutput
    """
3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420
    l = Layer(
        name=name,
        type=LayerType.EOSID_LAYER,
        eos_id=eos_id,
        inputs=[input.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name=name,
        layer_type=LayerType.EOSID_LAYER,
        parents=[input],
        size=l.config.size)
3421 3422 3423


@wrap_name_default()
3424 3425 3426 3427 3428 3429 3430
def beam_search(step,
                input,
                bos_id,
                eos_id,
                beam_size,
                max_length=500,
                name=None,
3431
                num_results_per_sample=None):
3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442
    """
    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)
3443
            with mixed_layer(size=512, name='rnn') as simple_rnn:
3444 3445 3446 3447 3448 3449
                simple_rnn += full_matrix_projection(input)
                simple_rnn += last_time_step_output
            return simple_rnn

        beam_gen = beam_search(name="decoder",
                               step=rnn_step,
3450
                               input=[StaticInput(encoder_last)],
3451 3452
                               bos_id=0,
                               eos_id=1,
3453
                               beam_size=5)
3454 3455 3456 3457 3458 3459 3460 3461 3462

    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
3463
                 step, and it is applied to sequences with arbitrary length by
3464 3465 3466 3467 3468 3469
                 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
3470
    :type input: list
3471 3472 3473
    :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
3474
                   symbol is essential, since it is used to initialize the RNN
3475 3476 3477 3478 3479 3480 3481 3482
                   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
3483 3484
    :param max_length: Max generated sequence length.
    :type max_length: int
3485 3486 3487 3488 3489 3490 3491 3492 3493 3494
    :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
3495 3496
    :return: The generated word index.
    :rtype: LayerOutput
3497 3498
    """

3499 3500 3501 3502 3503
    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")

3504
    if isinstance(input, StaticInput) or isinstance(input, BaseGeneratedInput):
3505 3506 3507 3508 3509 3510
        input = [input]

    generated_input_index = -1

    real_input = []
    for i, each_input in enumerate(input):
3511 3512
        assert isinstance(each_input, StaticInput) or isinstance(
            each_input, BaseGeneratedInput)
3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528
        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
3529 3530 3531 3532 3533 3534
        RecurrentLayerGroupSetGenerator(
            Generator(
                eos_layer_name=eos_name,
                max_num_frames=max_length,
                beam_size=beam_size,
                num_results_per_sample=num_results_per_sample))
3535 3536 3537 3538 3539 3540 3541 3542 3543 3544

        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

3545
    tmp = recurrent_group(
3546 3547 3548 3549
        step=__real_step__,
        input=real_input,
        reverse=False,
        name=name,
3550
        is_generating=True)
3551

3552 3553
    return tmp

3554

3555 3556
def __cost_input__(input, label, weight=None):
    """
3557
    inputs and parents for cost layers.
3558 3559 3560 3561 3562 3563 3564 3565
    """
    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
3566

3567 3568

@wrap_name_default()
L
luotao1 已提交
3569
@layer_support()
3570
def regression_cost(input, label, weight=None, name=None, layer_attr=None):
3571 3572 3573 3574 3575 3576
    """
    Regression Layer.

    TODO(yuyang18): Complete this method.

    :param name: layer name.
3577
    :type name: basestring
3578
    :param input: Network prediction.
3579
    :type input: LayerOutput
3580
    :param label: Data label.
3581 3582 3583 3584
    :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
3585 3586
    :param layer_attr: layer's extra attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
3587
    :return: LayerOutput object.
3588
    :rtype: LayerOutput
3589
    """
3590 3591
    ipts, parents = __cost_input__(input, label, weight)

3592 3593 3594 3595 3596
    Layer(
        inputs=ipts,
        type="square_error",
        name=name,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
3597
    return LayerOutput(name, LayerType.COST, parents=parents, size=1)
3598 3599 3600


@wrap_name_default("cost")
3601
@layer_support()
3602 3603 3604 3605
def classification_cost(input,
                        label,
                        weight=None,
                        name=None,
3606
                        top_k=None,
3607 3608
                        evaluator=classification_error_evaluator,
                        layer_attr=None):
3609 3610 3611 3612 3613 3614 3615 3616 3617
    """
    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
3618 3619 3620
    :param weight: The weight affects the cost, namely the scale of cost.
                   It is an optional argument.
    :type weight: LayerOutput
3621 3622
    :param top_k: number k in top-k error rate
    :type top_k: int
3623
    :param evaluator: Evaluator method.
3624 3625
    :param layer_attr: layer's extra attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
3626
    :return: LayerOutput object.
3627 3628 3629 3630 3631
    :rtype: LayerOutput
    """
    assert input.layer_type != LayerType.DATA
    assert isinstance(input.activation, SoftmaxActivation)
    assert label.layer_type == LayerType.DATA
3632 3633 3634

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

3635 3636 3637 3638 3639
    Layer(
        name=name,
        type="multi-class-cross-entropy",
        inputs=ipts,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
3640 3641 3642 3643 3644 3645 3646 3647 3648 3649

    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

3650
        e(name=e.__name__, input=input, label=label, weight=weight, top_k=top_k)
3651

3652
    if not isinstance(evaluator, collections.Sequence):
3653 3654 3655 3656 3657
        evaluator = [evaluator]

    for each_evaluator in evaluator:
        __add_evaluator__(each_evaluator)

3658
    return LayerOutput(name, LayerType.COST, parents=parents, size=1)
3659

3660

3661 3662 3663 3664 3665 3666 3667 3668 3669 3670
def conv_operator(img,
                  filter,
                  filter_size,
                  num_filters,
                  num_channels=None,
                  stride=1,
                  padding=0,
                  filter_size_y=None,
                  stride_y=None,
                  padding_y=None):
3671 3672 3673 3674 3675 3676 3677 3678 3679 3680
    """
    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

3681 3682
       op = conv_operator(img=input1,
                          filter=input2,
3683
                          filter_size=3,
3684 3685 3686
                          num_filters=64,
                          num_channels=64)

3687 3688 3689 3690
    :param img: input image
    :type img: LayerOutput
    :param filter: input filter
    :type filter: LayerOutput
3691 3692
    :param filter_size: The x dimension of a filter kernel.
    :type filter_size: int
3693 3694 3695
    :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).
3696
    :type filter_size_y: int
3697 3698
    :param num_filters: channel of output data.
    :type num_filters: int
3699 3700
    :param num_channels: channel of input data.
    :type num_channels: int
3701
    :param stride: The x dimension of the stride.
3702
    :type stride: int
3703
    :param stride_y: The y dimension of the stride.
3704
    :type stride_y: int
3705 3706 3707 3708 3709 3710 3711 3712 3713 3714 3715 3716 3717
    :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
3718

3719 3720
    if num_channels is None:
        num_channels = img.num_filters
3721 3722 3723

    assert isinstance(filter, LayerOutput)
    if filter.size is not None:
3724
        filter.size = filter_size * filter_size_y * num_filters * num_channels
3725

3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737
    op = ConvOperator(
        input_layer_names=[img.name, filter.name],
        num_filters=num_filters,
        conv_conf=Conv(
            filter_size=filter_size,
            padding=padding,
            stride=stride,
            channels=num_channels,
            filter_size_y=filter_size_y,
            padding_y=padding_y,
            stride_y=stride_y,
            groups=1))
3738
    op.origin = [img, filter]
3739 3740
    return op

3741

3742
@wrap_param_attr_default()
3743 3744 3745 3746 3747 3748 3749 3750 3751 3752 3753
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):
3754 3755 3756 3757 3758 3759 3760 3761 3762
    """
    Different from img_conv_layer and conv_op, conv_projection is an Projection,
    which can be used in mixed_layer and conat_layer. It use cudnn to implement
    conv and only support GPU mode.

    The example usage is:

    .. code-block:: python

D
dangqingqing 已提交
3763
       proj = conv_projection(input=input1,
3764 3765 3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777
                              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
3778 3779
    :param num_channels: channel of input data.
    :type num_channels: int
3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821
    :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.
3822
        init_w = (2.0 / (filter_size**2 * num_channels))**0.5
3823 3824 3825 3826 3827
        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

3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840
    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)
3841 3842 3843 3844

    proj.origin = input
    return proj

3845

D
dangqingqing 已提交
3846 3847 3848 3849 3850 3851 3852 3853 3854 3855 3856 3857 3858 3859 3860 3861 3862
@wrap_name_default("pad")
@layer_support()
def pad_layer(input,
              pad_c=None,
              pad_h=None,
              pad_w=None,
              name=None,
              layer_attr=None):
    """
    This operation pads zeros to the input data according to pad_c,pad_h
    and pad_w. pad_c, pad_h, pad_w specifies the which dimension and size
    of padding. And the input data shape is NCHW.

    For example, pad_c=[2,3] means padding 2 zeros before the
    input data and 3 zeros after the input data in channel dimension.
    pad_h means padding zeros in height dimension. pad_w means padding zeros
    in width dimension.
3863

D
dangqingqing 已提交
3864
    For example,
3865

3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883 3884 3885 3886
    .. code-block:: python

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

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

       output(2,4,2,3) = [
                           [ [[0,0,0], [0,0,0]],
                             [[1,2,3], [3,4,5]],
                             [[2,3,5], [1,6,7]],
                             [[0,0,0], [0,0,0]] ],
                           [ [[0,0,0], [0,0,0]],
                             [[4,3,1], [1,8,7]],
                             [[3,8,9], [2,3,5]],
                             [[0,0,0], [0,0,0]] ]
                         ]
D
dangqingqing 已提交
3887 3888

    The simply usage is:
D
dangqingqing 已提交
3889 3890 3891 3892 3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949

    .. code-block:: python

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

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

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

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

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

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


3950
@wrap_name_default()
L
luotao1 已提交
3951 3952
@layer_support()
def conv_shift_layer(a, b, name=None, layer_attr=None):
3953 3954 3955 3956 3957 3958 3959 3960 3961 3962 3963
    """
    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:
3964 3965 3966 3967
     - 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.
3968 3969 3970 3971 3972

    The example usage is:

    .. code-block:: python

L
Luo Tao 已提交
3973
       conv_shift = conv_shift_layer(a=layer1, b=layer2)
3974 3975 3976

    :param name: layer name
    :type name: basestring
3977 3978
    :param a: Input layer a.
    :type a: LayerOutput
L
Luo Tao 已提交
3979
    :param b: input layer b.
3980
    :type b: LayerOutput
L
luotao1 已提交
3981 3982
    :param layer_attr: layer's extra attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
3983
    :return: LayerOutput object.
3984 3985
    :rtype: LayerOutput
    """
3986 3987
    assert isinstance(a, LayerOutput) and isinstance(b, LayerOutput)
    assert b.size is None or b.size % 2 == 1  # size of b must be odd.
3988 3989 3990
    Layer(
        name=name,
        type=LayerType.CONV_SHIFT_LAYER,
3991
        inputs=[a.name, b.name],
3992
        **ExtraLayerAttribute.to_kwargs(layer_attr))
3993

3994 3995
    return LayerOutput(
        name, LayerType.CONV_SHIFT_LAYER, parents=[a, b], size=a.size)
3996 3997 3998 3999 4000


@wrap_name_default()
@wrap_param_attr_default()
@wrap_bias_attr_default()
4001
@wrap_act_default(act=LinearActivation())
4002
@layer_support(ERROR_CLIPPING, DROPOUT)
4003 4004 4005 4006 4007 4008 4009 4010
def tensor_layer(a,
                 b,
                 size,
                 act=None,
                 name=None,
                 param_attr=None,
                 bias_attr=None,
                 layer_attr=None):
4011 4012 4013 4014 4015
    """
    This layer performs tensor operation for two input.
    For example, each sample:

    .. math::
4016
       y_{i} = a * W_{i} * {b^\mathrm{T}}, i=0,1,...,K-1
4017 4018

    In this formular:
4019 4020
      - :math:`a`: the first input contains M elements.
      - :math:`b`: the second input contains N elements.
4021 4022
      - :math:`y_{i}`: the i-th element of y.
      - :math:`W_{i}`: the i-th learned weight, shape if [M, N]
4023
      - :math:`b^\mathrm{T}`: the transpose of :math:`b_{2}`.
4024 4025 4026 4027 4028

    The simple usage is:

    .. code-block:: python

4029
       tensor = tensor_layer(a=layer1, b=layer2, size=1000)
4030 4031 4032

    :param name: layer name
    :type name: basestring
4033 4034 4035 4036
    :param a: Input layer a.
    :type a: LayerOutput
    :param b: input layer b.
    :type b: LayerOutput
4037
    :param size: the layer dimension.
4038
    :type size: int.
4039 4040 4041
    :param act: Activation Type. Default is tanh.
    :type act: BaseActivation
    :param param_attr: The Parameter Attribute.
4042
    :type param_attr: ParameterAttribute
4043 4044 4045 4046 4047 4048
    :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
4049
    :return: LayerOutput object.
4050 4051
    :rtype: LayerOutput
    """
4052
    assert isinstance(a, LayerOutput) and isinstance(b, LayerOutput)
4053 4054 4055 4056 4057 4058
    Layer(
        name=name,
        size=size,
        type=LayerType.TENSOR_LAYER,
        active_type=act.name,
        bias=ParamAttr.to_bias(bias_attr),
4059 4060 4061 4062
        inputs=[Input(a.name, **param_attr.attr), Input(b.name)],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.TENSOR_LAYER, parents=[a, b], activation=act, size=size)
4063 4064 4065 4066 4067 4068


@wrap_name_default()
@wrap_param_attr_default()
@wrap_bias_attr_default()
@wrap_act_default()
L
luotao1 已提交
4069
@layer_support()
4070 4071
def selective_fc_layer(input,
                       size,
L
Luo Tao 已提交
4072
                       select=None,
4073 4074
                       act=None,
                       name=None,
4075 4076 4077
                       pass_generation=False,
                       has_selected_colums=True,
                       mul_ratio=0.02,
4078 4079 4080
                       param_attr=None,
                       bias_attr=None,
                       layer_attr=None):
4081 4082 4083 4084 4085 4086 4087 4088 4089 4090
    """
    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

4091
       sel_fc = selective_fc_layer(input=input, size=128, act=TanhActivation())
4092 4093 4094 4095 4096

    :param name: The Layer Name.
    :type name: basestring
    :param input: The input layer.
    :type input: LayerOutput|list|tuple
4097 4098
    :param select: The select layer. The output of select layer should be a
                   sparse binary matrix, and treat as the mask of selective fc.
L
Luo Tao 已提交
4099
                   If is None, acts exactly like fc_layer.
4100
    :type select: LayerOutput
4101 4102 4103 4104 4105 4106 4107 4108 4109 4110 4111 4112
    :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
4113
    :return: LayerOutput object.
4114 4115 4116 4117
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
4118
        assert not isinstance(param_attr, collections.Sequence)
4119 4120
        param_attr = [param_attr]
    else:
4121
        if isinstance(param_attr, collections.Sequence):
4122 4123 4124 4125
            assert len(input) == len(param_attr)
        else:
            param_attr = [copy.deepcopy(param_attr) for _ in range(len(input))]

4126 4127 4128 4129
    assert isinstance(input, collections.Sequence)
    assert isinstance(select, LayerOutput)
    if select.size is not None:
        assert select.size == size
4130
    Layer(
4131 4132 4133
        inputs=[
            Input(ipt.name, **attr.attr) for ipt, attr in zip(input, param_attr)
        ] + [select.name],
4134 4135 4136
        name=name,
        type=LayerType.SEL_FC_LAYER,
        size=size,
4137
        bias=ParameterAttribute.to_bias(bias_attr),
4138 4139 4140 4141
        active_type=act.name,
        selective_fc_pass_generation=pass_generation,
        has_selected_colums=has_selected_colums,
        selective_fc_full_mul_ratio=mul_ratio,
4142 4143 4144 4145 4146 4147 4148
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name,
        LayerType.SEL_FC_LAYER,
        list(input) + [select],
        activation=act,
        size=size)
4149 4150 4151


@wrap_name_default()
L
luotao1 已提交
4152 4153
@layer_support()
def sampling_id_layer(input, name=None, layer_attr=None):
4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167
    """
    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
4168 4169
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4170
    :return: LayerOutput object.
4171 4172
    :rtype: LayerOutput
    """
4173
    l = Layer(
4174 4175 4176
        name=name,
        type=LayerType.SAMPLING_ID_LAYER,
        inputs=[Input(input.name)],
4177 4178 4179
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.SAMPLING_ID_LAYER, input, size=l.config.size)
4180 4181 4182


@wrap_name_default()
L
luotao1 已提交
4183
@layer_support()
4184 4185 4186 4187
def slope_intercept_layer(input,
                          name=None,
                          slope=1.0,
                          intercept=0.0,
L
luotao1 已提交
4188
                          layer_attr=None):
4189 4190 4191 4192 4193 4194 4195 4196 4197 4198 4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209
    """
    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
4210 4211
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4212
    :return: LayerOutput object.
4213 4214 4215 4216 4217 4218 4219 4220
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.SLOPE_INTERCEPT_LAYER,
        slope=slope,
        intercept=intercept,
        inputs=[Input(input.name)],
4221 4222 4223
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.SLOPE_INTERCEPT_LAYER, input, size=input.size)
4224 4225 4226


@wrap_name_default()
L
luotao1 已提交
4227
@layer_support()
4228
def linear_comb_layer(weights, vectors, size=None, name=None, layer_attr=None):
4229
    """
4230 4231 4232 4233
    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
4234 4235 4236

    .. math::

4237
       z(i) = \sum_{j=0}^{M-1} x(j) y(i+Nj)
4238

4239 4240 4241 4242 4243
    where :math:`0 \le i \le N-1`

    Or in the matrix notation:

    .. math::
4244

4245
       z = x^\mathrm{T} Y
4246 4247

    In this formular:
4248 4249 4250 4251 4252 4253
      - :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.
4254 4255 4256 4257 4258

    The simple usage is:

    .. code-block:: python

4259
       linear_comb = linear_comb_layer(weights=weight, vectors=vectors,
4260 4261
                                       size=elem_dim)

4262 4263 4264 4265
    :param weights: The weight layer.
    :type weights: LayerOutput
    :param vectors: The vector layer.
    :type vectors: LayerOutput
4266 4267 4268 4269
    :param size: the dimension of this layer.
    :type size: int
    :param name: The Layer Name.
    :type name: basestring
L
luotao1 已提交
4270 4271
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4272
    :return: LayerOutput object.
4273 4274
    :rtype: LayerOutput
    """
4275 4276 4277 4278
    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:
4279
            size = vectors.size / weights.size
4280 4281
        else:
            assert size == vectors.size / weights.size
4282 4283
    Layer(
        name=name,
4284
        type=LayerType.LINEAR_COMBINATION_LAYER,
4285
        size=size,
4286
        inputs=[Input(weights.name), Input(vectors.name)],
4287 4288 4289
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.LINEAR_COMBINATION_LAYER, [weights, vectors], size=size)
4290

4291

4292
convex_comb_layer = linear_comb_layer
4293

4294

4295
@wrap_name_default()
L
luotao1 已提交
4296
@layer_support()
4297 4298 4299 4300 4301 4302 4303
def block_expand_layer(input,
                       block_x=0,
                       block_y=0,
                       stride_x=0,
                       stride_y=0,
                       padding_x=0,
                       padding_y=0,
4304
                       num_channels=None,
L
luotao1 已提交
4305 4306
                       name=None,
                       layer_attr=None):
4307 4308
    """
    Expand feature map to minibatch matrix.
4309
       - matrix width is: block_y * block_x * num_channels
4310
       - matirx height is: outputH * outputW
4311 4312 4313 4314 4315 4316 4317 4318 4319 4320

    .. 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
4321
    time step is block_y * block_x * num_channels. This layer can be used after
4322 4323
    convolution neural network, and before recurrent neural network.

4324 4325 4326 4327
    The simple usage is:

    .. code-block:: python

L
Luo Tao 已提交
4328
       block_expand = block_expand_layer(input=layer,
4329
                                         num_channels=128,
4330 4331 4332 4333 4334
                                         stride_x=1,
                                         stride_y=1,
                                         block_x=1,
                                         block_x=3)

4335 4336
    :param input: The input layer.
    :type input: LayerOutput
4337 4338
    :param num_channels: The channel number of input layer.
    :type num_channels: int|None
4339 4340 4341 4342 4343 4344 4345 4346 4347 4348 4349 4350 4351 4352
    :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
4353 4354
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4355
    :return: LayerOutput object.
4356 4357
    :rtype: LayerOutput
    """
4358 4359 4360
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters
4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377
    l = Layer(
        name=name,
        inputs=Input(
            input.name,
            block_expand=BlockExpand(
                channels=num_channels,
                block_x=block_x,
                block_y=block_y,
                stride_x=stride_x,
                stride_y=stride_y,
                padding_x=padding_x,
                padding_y=padding_y)),
        type=LayerType.BLOCK_EXPAND,
        **ExtraLayerAttribute.to_kwargs(layer_attr))

    return LayerOutput(
        name, LayerType.BLOCK_EXPAND, parents=[input], size=l.config.size)
4378 4379


4380 4381
@wrap_name_default()
@layer_support()
4382
def maxout_layer(input, groups, num_channels=None, name=None, layer_attr=None):
4383 4384 4385 4386 4387
    """
    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.

4388
    So groups should be larger than 1, and the num of channels should be able
4389 4390
    to devided by groups.

4391
    Please refer to Paper:
4392 4393 4394 4395
      - 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
4396

4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425
    The simple usage is:

    .. code-block:: python

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

    :param input: The input layer.
    :type input: LayerOutput
    :param num_channels: The channel number of input layer. If None will be set
                     automatically from previous output.
    :type num_channels: int|None
    :param groups: The group number of input layer.
    :type groups: int
    :param name: The name of this layer, which can not specify.
    :type name: None|basestring.
    :param layer_attr: Extra Layer attribute.
    :type layer_attr: ExtraLayerAttribute
    :return: LayerOutput object.
    :rtype: LayerOutput
    """
    assert input.layer_type == LayerType.CONV_LAYER
    assert isinstance(input.activation, LinearActivation)
    assert groups > 1
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters
    assert num_channels % groups == 0
4426 4427 4428 4429 4430 4431 4432 4433 4434
    l = Layer(
        name=name,
        inputs=Input(
            input.name, maxout=MaxOut(
                channels=num_channels, groups=groups)),
        type=LayerType.MAXOUT,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.MAXOUT, parents=[input], size=l.config.size)
4435 4436


4437
@wrap_name_default()
L
luotao1 已提交
4438
@layer_support()
4439 4440 4441 4442 4443
def ctc_layer(input,
              label,
              size=None,
              name=None,
              norm_by_times=False,
L
luotao1 已提交
4444
              layer_attr=None):
4445 4446 4447 4448 4449
    """
    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.

4450 4451
    More details can be found by referring to `Connectionist Temporal
    Classification: Labelling Unsegmented Sequence Data with Recurrent
4452 4453
    Neural Networks <http://machinelearning.wustl.edu/mlpapers/paper_files/
    icml2006_GravesFGS06.pdf>`_
4454 4455 4456 4457 4458 4459 4460 4461

    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.

4462 4463 4464 4465 4466 4467 4468 4469 4470
    The simple usage:

    .. code-block:: python

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

4471
    :param input: The input layer.
4472 4473 4474
    :type input: LayerOutput
    :param label: The data layer of label with variable length.
    :type label: LayerOutput
4475
    :param size: category numbers + 1.
4476
    :type size: int
4477 4478
    :param name: The name of this layer
    :type name: basestring|None
4479 4480
    :param norm_by_times: Whether to normalization by times. False by default.
    :type norm_by_times: bool
L
luotao1 已提交
4481 4482
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4483
    :return: LayerOutput object.
4484 4485 4486 4487
    :rtype: LayerOutput
    """
    assert isinstance(input, LayerOutput)
    assert isinstance(label, LayerOutput)
4488 4489 4490 4491 4492
    if label.size is not None:
        if size is not None:
            assert size == label.size + 1
        else:
            size = label.size + 1
4493
    Layer(
4494 4495 4496 4497
        name=name,
        type=LayerType.CTC_LAYER,
        size=size,
        norm_by_times=norm_by_times,
L
luotao1 已提交
4498
        inputs=[input.name, label.name],
4499
        **ExtraLayerAttribute.to_kwargs(layer_attr))
4500 4501
    return LayerOutput(name, LayerType.CTC_LAYER, [input, label], size=size)

4502

4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525
@wrap_name_default()
@layer_support()
def warp_ctc_layer(input,
                   label,
                   size=None,
                   name=None,
                   blank=0,
                   norm_by_times=False,
                   layer_attr=None):
    """
    A layer intergrating the open-source `warp-ctc
    <https://github.com/baidu-research/warp-ctc>` library, which is used in
    `Deep Speech 2: End-toEnd Speech Recognition in English and Mandarin
    <https://arxiv.org/pdf/1512.02595v1.pdf>`, to compute Connectionist Temporal
    Classification (CTC) loss.

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

    Note:
        - Let num_classes represent the category number. Considering the 'blank'
4526 4527 4528 4529 4530
          label needed by CTC, you need to use (num_classes + 1) as the input
          size. Thus, the size of both warp_ctc_layer and 'input' layer should
          be set to num_classes + 1.
        - You can set 'blank' to any value ranged in [0, num_classes], which
          should be consistent as that used in your labels.
4531
        - As a native 'softmax' activation is interated to the warp-ctc library,
L
Luo Tao 已提交
4532
          'linear' activation is expected instead in the 'input' layer.
4533 4534 4535 4536 4537 4538 4539 4540 4541 4542 4543 4544 4545 4546 4547 4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576 4577 4578 4579

    The simple usage:

    .. code-block:: python

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

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


4580
@wrap_name_default()
4581
@wrap_param_attr_default()
L
luotao1 已提交
4582
@layer_support()
4583 4584 4585 4586 4587 4588
def crf_layer(input,
              label,
              size=None,
              weight=None,
              param_attr=None,
              name=None,
L
luotao1 已提交
4589
              layer_attr=None):
4590 4591 4592 4593 4594 4595 4596 4597 4598 4599 4600 4601 4602 4603 4604
    """
    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.
4605
    :type label: LayerOutput
4606 4607 4608 4609 4610 4611 4612 4613 4614
    :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
4615 4616
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4617
    :return: LayerOutput object.
4618 4619 4620 4621 4622
    :rtype: LayerOutput
    """
    assert isinstance(input, LayerOutput)
    assert isinstance(label, LayerOutput)
    assert weight is None or isinstance(weight, LayerOutput)
4623 4624 4625 4626 4627 4628
    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
4629

4630
    ipts = [Input(input.name, **param_attr.attr), Input(label.name)]
4631 4632 4633 4634
    if weight is not None:
        ipts.append(Input(weight.name))

    Layer(
4635 4636 4637 4638
        name=name,
        type=LayerType.CRF_LAYER,
        size=size,
        inputs=ipts,
4639
        **ExtraLayerAttribute.to_kwargs(layer_attr))
4640 4641 4642
    parents = [input, label]
    if weight is not None:
        parents.append(weight)
4643 4644 4645 4646
    # The size for LayerOutput means the dimension of the output.
    # It's different from the meaning of crf layer, which is the number of
    # classes.
    return LayerOutput(name, LayerType.CRF_LAYER, parents, size=1)
4647

4648

4649
@wrap_name_default()
4650
@wrap_param_attr_default()
L
luotao1 已提交
4651
@layer_support()
4652 4653 4654 4655 4656
def crf_decoding_layer(input,
                       size,
                       label=None,
                       param_attr=None,
                       name=None,
L
luotao1 已提交
4657
                       layer_attr=None):
4658 4659 4660 4661 4662 4663 4664
    """
    A layer for calculating the decoding sequence of sequential conditional
    random field model. The decoding sequence is stored in output.ids.
    If a second input is provided, it is treated as the ground-truth label, and
    this layer will also calculate error. output.value[i] is 1 for incorrect
    decoding or 0 for correct decoding.

L
Luo Tao 已提交
4665 4666 4667 4668 4669 4670 4671
    The simple usage:

    .. code-block:: python

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

4672 4673 4674 4675 4676 4677 4678 4679 4680 4681
    :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
4682 4683
    :param layer_attr: Extra Layer config.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
4684
    :return: LayerOutput object.
4685 4686 4687 4688 4689 4690
    :rtype: LayerOutput
    """

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

4691
    ipts = [Input(input.name, **param_attr.attr)]
4692 4693 4694 4695
    if label is not None:
        ipts.append(Input(label.name))

    Layer(
4696 4697 4698 4699
        name=name,
        type=LayerType.CRF_DECODING_LAYER,
        size=size,
        inputs=ipts,
4700
        **ExtraLayerAttribute.to_kwargs(layer_attr))
4701 4702 4703
    parents = [input]
    if label is not None:
        parents.append(label)
4704 4705 4706 4707
    # The size for LayerOutput means the dimension of the output.
    # It's different from the meaning of crf layer, which is the number of
    # classes.
    return LayerOutput(name, LayerType.CRF_DECODING_LAYER, parents, size=1)
4708

4709

4710 4711 4712
@wrap_bias_attr_default(has_bias=True)
@wrap_name_default()
@layer_support()
4713 4714 4715 4716 4717 4718 4719 4720 4721
def nce_layer(input,
              label,
              num_classes,
              weight=None,
              num_neg_samples=10,
              neg_distribution=None,
              name=None,
              bias_attr=None,
              layer_attr=None):
4722 4723 4724 4725 4726 4727 4728 4729 4730 4731 4732 4733 4734 4735 4736 4737 4738 4739 4740 4741 4742
    """
    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.
4743
    :type num_classes: int
4744
    :param num_neg_samples: number of negative samples. Default is 10.
4745
    :type num_neg_samples: int
4746 4747 4748 4749 4750 4751 4752 4753 4754 4755 4756 4757 4758 4759 4760 4761 4762 4763 4764 4765
    :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
4766

4767 4768 4769 4770 4771 4772 4773 4774 4775 4776 4777 4778 4779 4780 4781
    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)

4782
    l = Layer(
4783 4784 4785 4786 4787 4788 4789
        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),
4790 4791 4792 4793
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.NCE_LAYER, parents=parents, size=l.config.size)

4794

4795 4796 4797
"""
following are cost Layers.
"""
4798 4799


4800
@wrap_name_default()
L
luotao1 已提交
4801
@layer_support()
4802 4803 4804 4805 4806 4807 4808
def rank_cost(left,
              right,
              label,
              weight=None,
              name=None,
              coeff=1.0,
              layer_attr=None):
4809
    """
4810
    A cost Layer for learning to rank using gradient descent. Details can refer
4811 4812
    to `papers <http://research.microsoft.com/en-us/um/people/cburges/papers/
    ICML_ranking.pdf>`_.
4813 4814 4815 4816 4817
    This layer contains at least three inputs. The weight is an optional
    argument, which affects the cost.

    .. math::

4818
       C_{i,j} & = -\\tilde{P_{ij}} * o_{i,j} + log(1 + e^{o_{i,j}})
4819

4820
       o_{i,j} & =  o_i - o_j
4821

4822
       \\tilde{P_{i,j}} & = \\{0, 0.5, 1\\} \ or \ \\{0, 1\\}
4823 4824 4825 4826 4827 4828 4829 4830 4831 4832 4833 4834 4835 4836 4837 4838 4839 4840 4841 4842 4843 4844 4845 4846 4847 4848 4849 4850 4851

    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
4852 4853
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
4854
    :return: LayerOutput object.
4855 4856 4857 4858 4859 4860 4861 4862 4863 4864 4865 4866
    :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)

4867 4868 4869 4870 4871 4872
    Layer(
        name=name,
        type=LayerType.RANK_COST,
        inputs=ipts,
        coeff=coeff,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
4873

4874
    return LayerOutput(name, LayerType.RANK_COST, parents=parents, size=1)
4875

4876

4877
@wrap_name_default()
L
luotao1 已提交
4878
@layer_support()
4879 4880 4881 4882 4883 4884
def lambda_cost(input,
                score,
                name,
                NDCG_num=5,
                max_sort_size=-1,
                layer_attr=None):
4885 4886 4887 4888 4889 4890 4891 4892 4893 4894 4895 4896
    """
    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)

4897
    :param input: Samples of the same query should be loaded as sequence.
4898 4899 4900 4901 4902 4903 4904 4905 4906 4907 4908
    :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
4909 4910 4911
                          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.
4912 4913 4914
    :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
4915 4916
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
4917
    :return: LayerOutput object.
4918 4919
    :rtype: LayerOutput
    """
4920 4921 4922
    assert isinstance(input, LayerOutput) and isinstance(score, LayerOutput)
    if score.size is not None:
        assert score.size == 1
4923 4924 4925 4926 4927 4928 4929
    Layer(
        name=name,
        type=LayerType.LAMBDA_COST,
        inputs=[input.name, score.name],
        NDCG_num=NDCG_num,
        max_sort_size=max_sort_size,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
4930

4931 4932
    return LayerOutput(
        name, LayerType.LAMBDA_COST, parents=[input, score], size=1)
4933

4934

4935
@wrap_name_default()
L
luotao1 已提交
4936 4937
@layer_support()
def cross_entropy(input, label, name=None, coeff=1.0, layer_attr=None):
4938 4939 4940 4941 4942
    """
    A loss layer for multi class entropy.

    .. code-block:: python

X
xuwei06 已提交
4943
       cost = cross_entropy(input=input_layer,
L
Luo Tao 已提交
4944
                            label=label_layer)
4945 4946 4947 4948 4949 4950 4951 4952 4953

    :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
4954 4955
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
4956
    :return: LayerOutput object.
4957 4958 4959
    :rtype: LayerOutput.
    """

4960 4961 4962 4963 4964 4965 4966 4967
    Layer(
        name=name,
        type=LayerType.CROSS_ENTROPY,
        inputs=[input.name, label.name],
        coeff=coeff,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(
        name, LayerType.CROSS_ENTROPY, parents=[input, label], size=1)
4968

4969

4970
@wrap_name_default()
L
luotao1 已提交
4971
@layer_support()
4972 4973 4974 4975
def cross_entropy_with_selfnorm(input,
                                label,
                                name=None,
                                coeff=1.0,
L
luotao1 已提交
4976 4977
                                softmax_selfnorm_alpha=0.1,
                                layer_attr=None):
4978 4979
    """
    A loss layer for multi class entropy with selfnorm.
4980
    Input should be a vector of positive numbers, without normalization.
4981 4982 4983

    .. code-block:: python

X
xuwei06 已提交
4984
       cost = cross_entropy_with_selfnorm(input=input_layer,
L
Luo Tao 已提交
4985
                                          label=label_layer)
4986 4987 4988 4989 4990 4991 4992 4993 4994 4995 4996

    :param input: The first input layer.
    :type input: LayerOutput.
    :param label: The input label.
    :type input: LayerOutput.
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring.
    :param coeff: The coefficient affects the gradient in the backward.
    :type coeff: float.
    :param softmax_selfnorm_alpha: The scale factor affects the cost.
    :type softmax_selfnorm_alpha: float.
L
luotao1 已提交
4997 4998
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
4999
    :return: LayerOutput object.
5000 5001
    :rtype: LayerOutput.
    """
5002 5003 5004 5005 5006 5007 5008
    Layer(
        name=name,
        type=LayerType.CROSS_ENTROPY_WITH_SELFNORM,
        inputs=[input.name, label.name],
        coeff=coeff,
        softmax_selfnorm_alpha=softmax_selfnorm_alpha,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
5009

5010 5011 5012 5013 5014
    return LayerOutput(
        name,
        LayerType.CROSS_ENTROPY_WITH_SELFNORM,
        parents=[input, label],
        size=1)
5015

5016

X
xuwei06 已提交
5017 5018 5019 5020 5021 5022 5023 5024
@wrap_name_default()
@layer_support()
def sum_cost(input, name=None, layer_attr=None):
    """
    A loss layer which calculate the sum of the input as loss

    .. code-block:: python

L
Luo Tao 已提交
5025
       cost = sum_cost(input=input_layer)
X
xuwei06 已提交
5026 5027 5028 5029 5030 5031 5032 5033 5034 5035

    :param input: The first input layer.
    :type input: LayerOutput.
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring.
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
    :return: LayerOutput object.
    :rtype: LayerOutput.
    """
L
Luo Tao 已提交
5036
    assert isinstance(input, LayerOutput)
5037 5038 5039 5040 5041
    Layer(
        name=name,
        type=LayerType.SUM_COST,
        inputs=[input.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr))
X
xuwei06 已提交
5042

5043
    return LayerOutput(name, LayerType.SUM_COST, parents=[input], size=1)
X
xuwei06 已提交
5044 5045


5046
@wrap_name_default()
L
luotao1 已提交
5047 5048
@layer_support()
def huber_cost(input, label, name=None, coeff=1.0, layer_attr=None):
5049 5050 5051 5052 5053
    """
    A loss layer for huber loss.

    .. code-block:: python

X
xuwei06 已提交
5054
       cost = huber_cost(input=input_layer,
L
Luo Tao 已提交
5055
                         label=label_layer)
5056 5057 5058 5059 5060 5061 5062 5063 5064

    :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
5065 5066
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
5067
    :return: LayerOutput object.
5068 5069
    :rtype: LayerOutput.
    """
5070 5071 5072
    assert isinstance(input, LayerOutput)
    if input.size is not None:
        assert input.size == 1
5073 5074 5075 5076 5077 5078
    Layer(
        name=name,
        type=LayerType.HUBER,
        inputs=[input.name, label.name],
        coeff=coeff,
        **ExtraLayerAttribute.to_kwargs(layer_attr))
5079
    return LayerOutput(name, LayerType.HUBER, parents=[input, label], size=1)
5080

5081

5082
@wrap_name_default()
L
luotao1 已提交
5083
@layer_support()
5084 5085 5086 5087
def multi_binary_label_cross_entropy(input,
                                     label,
                                     name=None,
                                     coeff=1.0,
L
luotao1 已提交
5088
                                     layer_attr=None):
5089 5090 5091 5092 5093
    """
    A loss layer for multi binary label cross entropy.

    .. code-block:: python

X
xuwei06 已提交
5094
       cost = multi_binary_label_cross_entropy(input=input_layer,
L
Luo Tao 已提交
5095
                                               label=label_layer)
5096 5097 5098 5099 5100 5101 5102 5103 5104 5105 5106

    :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
5107 5108
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
5109
    :return: LayerOutput object.
5110 5111 5112
    :rtype: LayerOutput
    """

5113 5114
    if input.activation is None or \
            not isinstance(input.activation, SigmoidActivation):
5115 5116 5117 5118 5119 5120 5121 5122 5123 5124 5125 5126 5127 5128 5129 5130
        logger.log(
            logging.WARN,
            "%s is not recommend for multi_binary_label_cross_entropy's activation, "
            "maybe the sigmoid is better" % repr(input.activation))

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