layers.py 114.9 KB
Newer Older
Z
zhangjinchao01 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38
# Copyright (c) 2016 Baidu, Inc. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import functools

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 *

try:
    import cPickle as pickle
except ImportError:
    import pickle
import copy

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


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

    DATA = "data"
    MIXED_LAYER = "mixed"
    LSTMEMORY = "lstmemory"
    GRUMEMORY = "gated_recurrent"
    SEQUENCE_LAST_INSTANCE = "seqlastins"
    SEQUENCE_FIRST_INSTANCE = "seqfirstins"
    POOLING_MAX = "max"
    POOLING_AVG = 'average'
    FC_LAYER = "fc"
    COST = 'cost'
74 75
    COSINE_SIM_VEC = 'cos_vm'
    COSINE_SIM = 'cos'
Z
zhangjinchao01 已提交
76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106
    HSIGMOID = 'hsigmoid'
    CONV_LAYER = "conv"
    POOL_LAYER = "pool"
    BATCH_NORM_LAYER = 'batch_norm'
    NORM_LAYER = 'norm'
    SUM_TO_ONE_NORM_LAYER = 'sum_to_one_norm'
    ADDTO_LAYER = 'addto'

    CONCAT_LAYER = 'concat'
    CONCAT_PROJ_LAYER = 'concat2'

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

    EXPAND_LAYER = 'expand'
    INTERPOLATION_LAYER = 'interpolation'
    POWER_LAYER = 'power'
    SCALING_LAYER = 'scaling'
    TRANS_LAYER = 'trans'

    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"
107
    LINEAR_COMBINATION_LAYER = "convex_comb"
Z
zhangjinchao01 已提交
108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175
    BLOCK_EXPAND = "blockexpand"

    CTC_LAYER = "ctc"
    CRF_LAYER = "crf"
    CRF_DECODING_LAYER = "crf_decoding"

    RANK_COST = "rank-cost"
    LAMBDA_COST = "lambda_cost"
    HUBER = "huber"
    CROSS_ENTROPY = "multi-class-cross-entropy"
    CROSS_ENTROPY_WITH_SELFNORM = "multi_class_cross_entropy_with_selfnorm"
    SOFT_BIN_CLASS_CROSS_ENTROPY = "soft_binary_class_cross_entropy"
    MULTI_BIN_LABEL_CROSS_ENTROPY = "multi_binary_label_cross_entropy"

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

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


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


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

    - Check layer connection make sense.

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

    - Tracking layer connection.

    - Pass to layer methods as input.

    :param name: Layer output name.
    :type name: basestring
    :param layer_type: Current Layer Type. One of LayerType enumeration.
    :type layer_type: basestring
    :param activation: Layer Activation.
    :type activation: BaseActivation.
    :param parents: Layer's parents.
    :type parents: list|tuple
    """

    def __init__(self, name, layer_type, parents=None, activation=None,
                 num_filters=None, img_norm_type=None, size=None, outputs=None):
        assert isinstance(name, basestring)
        assert isinstance(layer_type, basestring)
        assert LayerType.is_layer_type(layer_type)
        self.name = name
        self.layer_type = layer_type
176 177
        if parents is not None and type(parents) != list:
            parents = [parents]
Z
zhangjinchao01 已提交
178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518
        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

    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'


def layer_support(*attrs):
    def decorator(method):
        @functools.wraps(method)
        def wrapper(*args, **kwargs):
            for attr in attrs:
                for each in args:
                    if isinstance(each, ExtraLayerAttribute):
                        setattr(each, '_'.join(['can', attr]), True)
                for key in kwargs:
                    val = kwargs[key]
                    if isinstance(val, ExtraLayerAttribute):
                        setattr(val, '_'.join(['can', attr]), True)
            for each in args:
                if isinstance(each, ExtraLayerAttribute):
                    each.check(method.__name__)
            for key in kwargs:
                val = kwargs[key]
                if isinstance(val, ExtraLayerAttribute):
                    val.check(method.__name__)
            return method(*args, **kwargs)

        return wrapper

    return decorator


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

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

    There are two styles of usage.

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

    .. code-block:: python

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

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

    .. code-block:: python

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

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


@wrap_param_attr_default()
def table_projection(input, size=0, param_attr=None):
    """
    Table Projection. It selects rows from parameter where row\_id
    is in input\_ids.

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

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

    There are two styles of usage.

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

    .. code-block:: python

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

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

    .. code-block:: python

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


    :param input: Input layer, which must contains id fields.
    :type input: LayerOutput
    :param size: The parameter size. Means the width of parameter.
    :type size: int
    :param param_attr: Parameter config, None if use default.
    :type param_attr: ParameterAttribute
    :return: A TableProjection Object.
    :rtype: TableProjection
    """
    proj = TableProjection(input_layer_name=input.name,
                           size=size,
                           **param_attr.attr)
    proj.origin = input
    proj.origin.projection = "table"
    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.
    :type input: LayerOutput.
    :param offset: Offset, None if use default.
    :type offset: int
    :return: A IdentityProjection or IdentityOffsetProjection Object
    :rtype: IdentityProjection or IdentityOffsetProjection
    """
    if offset is None:
        proj = IdentityProjection(input_layer_name=input.name)
        proj.origin = input
        proj.origin.projection = 'identity'
    else:
        proj = IdentityOffsetProjection(input_layer_name=input.name,
                                        offset=offset)
        proj.origin = input
        proj.origin.projection = 'identity_offset'
    return proj


@wrap_param_attr_default()
def dotmul_projection(input, param_attr=None, scale=1):
    """
    1. DotMulProjection if input is a layer.
    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)

    2. DotMulOperator if input is a list or tuple.
    It takes two inputs, performs element-wise multiplication:

    .. math::
       out.row[i] += scale * (in1.row[i] .* in2.row[i])

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

    The example usage is:

    .. code-block:: python

       op = dotmul_projection(input=[layer1, layer2],
                              scale=2.0)

    :param input: Input layer.
    :type input: LayerOutput|list|tuple
    :param param_attr: Parameter config, None if use default.
    :type param_attr: ParameterAttribute
    :param scale: config scalar, default value is one.
    :type scale: float
    :return: A DotMulProjection or DotMulOperator Object.
    :rtype: DotMulProjection or DotMulOperator
    """
    if isinstance(input, LayerOutput):
        proj = DotMulProjection(input_layer_name=input.name,
                                size=input.size,
                                **param_attr.attr)
        proj.origin = input
        proj.origin.projection = "dot_mul"
        return proj
    else:
        assert isinstance(input, list) or isinstance(input, tuple)
        assert len(input) == 2
        assert param_attr is None
        op = DotMulOperator(input_layer_name=[x.name for x in input],
                            scale=scale)
        op.origin = input
        op.origin.operator = "dot_mul"
        return op


@wrap_bias_attr_default(['padding_attr'])
def context_projection(input, context_len, context_start=None,
                       padding_attr=False):
    """
    Context Projection.

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

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

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

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

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


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

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

    def __init__(self, name, size, act, bias_attr, layer_attr,
                 parents=None):
        """
        Ctor.
        :param name: layer name.
        :type name: basestring
        :param size: layer size.
        :type size: int
        :param act: activation type.
        :type act: BaseActivation
        :param bias_attr: The Bias Attribute. If no bias, then pass False or
                          something not type of ParameterAttribute. None will
                          get a default Bias.
        :type bias_attr: ParameterAttribute or None means has bias. Any other
                         type means no bias.
        :param layer_attr: Extra Layer Attribute.
        :type layer_attr: ExtraLayerAttribute or None
        """
        LayerOutput.__init__(self, name, LayerType.MIXED_LAYER, parents,
                             size=size, activation=act)
        self.bias_attr = bias_attr
        self.layer_attr = layer_attr
        self.inputs = []
        self.finalized = False

    def __add__(self, other):
        """
        + += operator
        :param other: Other projection.
        :type other: Projection
        :return: self.
        :rtype: MixedLayerType
        """
        if not self.finalized:
519
            assert isinstance(other, Projection) or isinstance(other, Operator)
Z
zhangjinchao01 已提交
520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619
            self.inputs.append(other)
            self.parents.append(other.origin)
            return self
        else:
            raise MixedLayerType.AddToSealedMixedLayerException()

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

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


@wrap_name_default("mixed")
@wrap_act_default(act=LinearActivation())
@wrap_bias_attr_default(has_bias=False)
@layer_support(ERROR_CLIPPING, DROPOUT)
def mixed_layer(size, input=None, name=None, act=None, bias_attr=False,
                layer_attr=None):
    """
    Mixed Layer. A mixed layer will add all inputs together, then activate.
    Each inputs is a projection or operator.

    There are two styles of usages.

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

    .. code-block:: python

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

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

    .. code-block:: python

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

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

    if input is None:
        return MixedLayerType(name, size, act, bias_attr, layer_attr)
    else:
        with mixed_layer(name=name, size=size, act=act, bias_attr=bias_attr,
                         layer_attr=layer_attr) as m:
            if isinstance(input, list) or isinstance(input, tuple):
                for each in input:
                    m += each
            else:
                m += input
        return m


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

    The example usage is:

    ..  code-block:: python

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

    :param name: Name of this data layer.
    :type name: basestring
    :param size: Size of this data layer.
    :type size: int
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
620
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646
    :rtype: LayerOutput
    """
    Layer(type=LayerType.DATA, name=name, size=size,
          **ExtraLayerAttribute.to_kwargs(layer_attr))

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


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

    :param name: Name of this embedding layer.
    :type name: basestring
    :param input: The input layer for this embedding. NOTE: must be Index Data.
    :type input: LayerOutput
    :param size: The embedding dimension.
    :type size: int
    :param param_attr: The embedding parameter attribute. See ParameterAttribute
                      for details.
    :type param_attr: ParameterAttribute|None
    :param layer_attr: Extra layer Config. Default is None.
    :type layer_attr: ExtraLayerAttribute|None
D
dangqingqing 已提交
647
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675
    :rtype: LayerOutput
    """
    with mixed_layer(name=name, size=size, act=LinearActivation(),
                     bias_attr=False,
                     layer_attr=layer_attr) as mix:
        mix += table_projection(input=input, size=size, param_attr=param_attr)
    return mix


@wrap_name_default()
@wrap_param_attr_default()
@wrap_bias_attr_default()
@wrap_act_default()
@layer_support(ERROR_CLIPPING, DROPOUT)
def fc_layer(input, size, act=None, name=None,
             param_attr=None, bias_attr=None, layer_attr=None):
    """
    Helper for declare fully connected layer.

    The example usage is:

    .. code-block:: python

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

L
luotao02 已提交
676
    which is equal to:
Z
zhangjinchao01 已提交
677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698

    .. 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
699
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
        assert not isinstance(param_attr, list)
        param_attr = [param_attr]
    else:
        if isinstance(param_attr, list) or isinstance(param_attr, tuple):
            assert len(input) == len(param_attr)
        else:
            param_attr = [copy.deepcopy(param_attr) for _ in range(len(input))]

    assert isinstance(input, list)

    def __idx_to_input__(i):
        attr = param_attr[i]
        assert isinstance(attr, ParameterAttribute)
        return Input(input[i].name, **attr.attr)

    Layer(
        inputs=map(__idx_to_input__, range(len(input))),
        name=name,
        type=LayerType.FC_LAYER,
        size=size,
        bias=ParamAttr.to_bias(bias_attr),
        active_type=act.name,
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.FC_LAYER, input, activation=act,
                       size=size)


@wrap_name_default("seq_pooling")
@wrap_bias_attr_default(has_bias=False)
@wrap_param_default(['pooling_type'], default_factory=lambda _: MaxPooling())
@layer_support()
def pooling_layer(input, pooling_type=None, name=None, bias_attr=None,
                  agg_level=AggregateLevel.EACH_TIMESTEP,
                  layer_attr=None):
    """
    Pooling layer for sequence inputs, not used for Image.

    The example usage is:

    .. code-block:: python

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

C
caoying03 已提交
750 751
    :param agg_level: AggregateLevel.EACH_TIMESTEP or
                      AggregateLevel.EACH_SEQUENCE
Z
zhangjinchao01 已提交
752 753 754 755 756 757 758 759 760 761 762 763
    :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
764
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802
    :rtype: LayerType
    """
    extra_dict = dict()
    if isinstance(pooling_type, AvgPooling):
        extra_dict['average_strategy'] = pooling_type.strategy
    extra_dict.update(ExtraLayerAttribute.to_kwargs(layer_attr))

    Layer(
        name=name,
        type=pooling_type.name,
        inputs=[Input(input.name)],
        bias=ParamAttr.to_bias(bias_attr),
        trans_type=agg_level,
        **extra_dict
    )

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


@wrap_bias_attr_default()
@wrap_param_attr_default()
@wrap_act_default(param_names=['gate_act'],
                  act=SigmoidActivation())
@wrap_act_default(param_names=["act", 'state_act'], act=TanhActivation())
@wrap_name_default("lstmemory")
@layer_support(DROPOUT)
def lstmemory(input, name=None, reverse=False, act=None,
              gate_act=None,
              state_act=None, bias_attr=None, param_attr=None,
              layer_attr=None):
    """
    Long Short-term Memory Cell.

    The memory cell was implemented as follow equations.

    ..  math::

L
luotao02 已提交
803
        i_t & = \\sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i)
Z
zhangjinchao01 已提交
804

L
luotao02 已提交
805
        f_t & = \\sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f)
Z
zhangjinchao01 已提交
806

L
luotao02 已提交
807
        c_t & = f_tc_{t-1} + i_t tanh (W_{xc}x_t+W_{hc}h_{t-1} + b_c)
Z
zhangjinchao01 已提交
808

L
luotao02 已提交
809
        o_t & = \\sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o)
Z
zhangjinchao01 已提交
810

L
luotao02 已提交
811
        h_t & = o_t tanh(c_t)
Z
zhangjinchao01 已提交
812 813


C
caoying03 已提交
814
    NOTE: In PaddlePaddle's implementation, the multiplications
Z
zhangjinchao01 已提交
815
    :math:`W_{xi}x_{t}` , :math:`W_{xf}x_{t}`,
C
caoying03 已提交
816 817 818 819
    :math:`W_{xc}x_t`, :math:`W_{xo}x_{t}` are not done in the lstmemory layer,
    so an additional mixed_layer with full_matrix_projection or a fc_layer must
    be included in the configuration file to complete the input-to-hidden
    mappings before lstmemory is called.
Z
zhangjinchao01 已提交
820

C
caoying03 已提交
821
    NOTE: This is a low level user interface. You can use network.simple_lstm
Z
zhangjinchao01 已提交
822 823
    to config a simple plain lstm layer.

C
caoying03 已提交
824 825 826 827
    Please refer to **Generating Sequences With Recurrent Neural Networks** for
    more details about LSTM.

    Link_ goes as below.
Z
zhangjinchao01 已提交
828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850

    .. _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
851
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902
    :rtype: LayerOutput
    """

    assert gate_act.support_hppl
    assert state_act.support_hppl
    assert act.support_hppl

    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))

    return LayerOutput(name, LayerType.LSTMEMORY, [input],
                       size=input.size / 4 if input.size is not None else None)

@wrap_bias_attr_default()
@wrap_param_attr_default()
@wrap_act_default(param_names=['gate_act'],
                  act=SigmoidActivation())
@wrap_act_default(param_names=["act"], act=TanhActivation())
@wrap_name_default("gru")
@layer_support(DROPOUT)
def grumemory(input, name=None, reverse=False, act=None,
              gate_act=None,
              bias_attr=None, param_attr=None,
              layer_attr=None):
    """
    Gate Recurrent Unit Layer.

    The memory cell was implemented as follow equations.

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

    ..  math::

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

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

    ..  math::

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

C
caoying03 已提交
903 904
    3. The candidate activation :math:`\\tilde{h_t}` is computed similarly to
    that of the traditional recurrent unit:
Z
zhangjinchao01 已提交
905 906 907 908 909

    ..  math::

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

C
caoying03 已提交
910 911 912
    4. The hidden activation :math:`h_t` of the GRU at time t is a linear
    interpolation between the previous activation :math:`h_{t-1}` and the
    candidate activation :math:`\\tilde{h_t}`:
Z
zhangjinchao01 已提交
913 914 915 916 917

    ..  math::

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

C
caoying03 已提交
918
    NOTE: In PaddlePaddle's implementation, the multiplication operations
Z
zhangjinchao01 已提交
919
    :math:`W_{r}x_{t}`, :math:`W_{z}x_{t}` and :math:`W x_t` are not computed in
C
caoying03 已提交
920 921 922
    gate_recurrent layer. Consequently, an additional mixed_layer with
    full_matrix_projection or a fc_layer must be included before grumemory
    is called.
Z
zhangjinchao01 已提交
923

C
caoying03 已提交
924 925 926
    More details can be found by referring to `Empirical Evaluation of Gated
    Recurrent Neural Networks on Sequence Modeling.
    <https://arxiv.org/abs/1412.3555>`_
Z
zhangjinchao01 已提交
927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953

    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.
    :param reverse: Wether sequence process is reversed or not.
    :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
D
dangqingqing 已提交
954
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987
    :rtype: LayerOutput
    """

    assert act.support_hppl
    assert gate_act.support_hppl

    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)
          )

    return LayerOutput(name, LayerType.GRUMEMORY, [input],
                       size=input.size / 3 if input.size is not None else None)

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

    :param agg_level: Aggregated level
    :param name: Layer name.
    :type name: basestring
    :param input: Input layer name.
    :type input: LayerOutput
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
988
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.SEQUENCE_LAST_INSTANCE,
        inputs=[input.name],
        trans_type=agg_level,
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.SEQUENCE_LAST_INSTANCE, parents=[input],
                       size=input.size)


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

    :param agg_level: aggregation level
    :param name: Layer name.
    :type name: basestring
    :param input: Input layer name.
    :type input: LayerOutput
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1016
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.SEQUENCE_FIRST_INSTANCE,
        inputs=[input.name],
        trans_type=agg_level,
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.SEQUENCE_FIRST_INSTANCE,
                       parents=[input], size=input.size)


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

@wrap_name_default()
@layer_support()
def expand_layer(input, expand_as,
                 name=None,
                 bias_attr=False,
                 expand_level=ExpandLevel.FROM_TIMESTEP,
                 layer_attr=None):
    """
    A layer for "Expand Dense data or (sequence data where the length of each
    sequence is one) to sequence data."

    The example usage is:

    .. code-block:: python

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

    :param input: Input layer
    :type input: LayerOutput
    :param expand_as: Expand as this layer's sequence info.
    :type expand_as: LayerOutput
    :param name: Layer name.
    :type name: basestring
    :param bias_attr: Bias attribute. None means default bias. False means no
                      bias.
    :type bias_attr: ParameterAttribute|None|False
    :param expand_level: whether input layer is timestep(default) or sequence.
    :type expand_level: ExpandLevel
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1066
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112
    :rtype: LayerOutput
    """

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



@wrap_name_default()
@layer_support()
def interpolation_layer(input, weight, name=None, layer_attr=None):
    """
    This layer is for linear interpolation with two inputs,
    which is used in NEURAL TURING MACHINE.

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

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

    The example usage is:

    .. code-block:: python

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

    :param input: Input layer.
    :type input: list|tuple
    :param weight: Weight layer.
    :type weight: LayerOutput
    :param name: Layer name.
    :type name: basestring
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1113
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157
    :rtype: LayerOutput
    """
    assert isinstance(input, list) or isinstance(input, tuple)
    assert len(input) == 2
    assert input[0].size == input[1].size
    assert weight.size == 1
    Layer(
        name=name,
        type=LayerType.INTERPOLATION_LAYER,
        inputs=[weight.name, input[0].name, input[1].name],
        **ExtraAttr.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.INTERPOLATION_LAYER,
                       parents=[weight, input[0], input[1]],
                       size=input[0].size)


@wrap_name_default()
@layer_support()
def power_layer(input, weight, name=None, layer_attr=None):
    """
    This layer applies a power function to a vector element-wise,
    which is used in NEURAL TURING MACHINE.

    .. math::
       y = x^w

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

    The example usage is:

    .. code-block:: python

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

    :param input: Input layer.
    :type input: LayerOutput
    :param weight: Weight layer.
    :type weight: LayerOutput
    :param name: Layer name.
    :type name: basestring
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1158
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175
    :rtype: LayerOutput
    """
    assert weight.size == 1
    Layer(
        name=name,
        type=LayerType.POWER_LAYER,
        inputs=[input.name, weight.name],
        **ExtraAttr.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.POWER_LAYER,
                       parents=[input, weight], size=input.size)


@wrap_name_default()
@layer_support()
def scaling_layer(input, weight, name=None, layer_attr=None):
    """
1176
    A layer for multiplying input vector by weight scalar.
Z
zhangjinchao01 已提交
1177 1178

    .. math::
1179
       y  = w x
Z
zhangjinchao01 已提交
1180

1181 1182 1183 1184 1185
    where :math:`x` is size=dataDim input, :math:`w` is size=1 weight,
    and :math:`y` is size=dataDim output.

    Note that the above computation is for one sample. Multiple samples are
    processed in one batch.
Z
zhangjinchao01 已提交
1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200

    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
1201
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237
    :rtype: LayerOutput
    """
    assert weight.size == 1
    Layer(
        name=name,
        type=LayerType.SCALING_LAYER,
        inputs=[weight.name, input.name],
        **ExtraAttr.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.SCALING_LAYER, parents=[weight, input],
                       size=input.size)


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

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

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

    The example usage is:

    .. code-block:: python

       trans = trans_layer(input=layer)

    :param input: Input layer.
    :type input: LayerOutput
    :param name: Layer name.
    :type name: basestring
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1238
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.TRANS_LAYER,
        inputs=[input.name],
        **ExtraAttr.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.TRANS_LAYER, parents=[input],
                       size=input.size)


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

    ..  math::
D
dangqingqing 已提交
1258
        similarity = cos(\\theta) = {\\mathbf{a} \\cdot \\mathbf{b}
1259 1260 1261 1262 1263
        \\over \\|\\mathbf{a}\\| \\|\\mathbf{b}\\|}

    The size of a is M, size of b is M*N,
    Similarity will be calculated N times by step M. The output size is
    N. The scale will be multiplied to similarity.
Z
zhangjinchao01 已提交
1264

1265 1266
    Note that the above computation is for one sample. Multiple samples are
    processed in one batch.
Z
zhangjinchao01 已提交
1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279

    :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
1280
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1281 1282
    :rtype: LayerOutput
    """
1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299
    if size == 1:
        Layer(
            name=name,
            type=LayerType.COSINE_SIM,
            cos_scale=scale,
            inputs=[a.name, b.name],
            **ExtraLayerAttribute.to_kwargs(layer_attr)
        )
    else:
        Layer(
            name=name,
            type=LayerType.COSINE_SIM_VEC,
            size=size,
            cos_scale=scale,
            inputs=[a.name, b.name],
            **ExtraLayerAttribute.to_kwargs(layer_attr)
        )
Z
zhangjinchao01 已提交
1300 1301 1302 1303 1304
    return LayerOutput(name, LayerType.COSINE_SIM, parents=[a, b])

@wrap_name_default()
@wrap_bias_attr_default(has_bias=True)
@layer_support()
C
caoying03 已提交
1305 1306
def hsigmoid(input, label, num_classes, name=None, bias_attr=None,
             layer_attr=None):
Z
zhangjinchao01 已提交
1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327
    """
    Organize the classes into a binary tree. At each node, a sigmoid function
    is used to calculate the probability of belonging to the right branch.
    This idea is from "F. Morin, Y. Bengio (AISTATS 05):
    Hierarchical Probabilistic Neural Network Language Model."

    The example usage is:

    ..  code-block:: python

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

    :param input: Input layers. It could be a LayerOutput or list/tuple of
                 LayerOutput.
    :type input: LayerOutput|list|tuple
    :param label: Label layer.
    :type label: LayerOutput
    :param num_classes: number of classes.
    :type num_classes: int
L
luotao02 已提交
1328 1329
    :param name: layer name
    :type name: basestring
Z
zhangjinchao01 已提交
1330 1331 1332 1333 1334
    :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
1335
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
    assert isinstance(input, list) or isinstance(input, tuple)
    assert isinstance(label, LayerOutput)
    assert label.layer_type == LayerType.DATA

    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)

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

@wrap_name_default("conv")
@wrap_param_attr_default()
@wrap_bias_attr_default()
@wrap_act_default(act=ReluActivation())
@layer_support(DROPOUT)
def img_conv_layer(input, filter_size, num_filters,
                   name=None, num_channels=None,
                   act=None, groups=1, stride=1, padding=0, bias_attr=None,
                   param_attr=None, shared_biases=True, layer_attr=None,
                   filter_size_y=None, stride_y=None, padding_y=None):
    """
    Convolution layer for image. Paddle only support square input currently and
    thus input image's width equals height.

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

    The num_channel means input image's channel number. It may be 1 or 3 when
    input is raw pixels of image(mono or RGB), or it may be the previous layer's
    num_filters * num_group.

C
caoying03 已提交
1385 1386 1387
    There are several group of filter in PaddlePaddle implementation.
    Each group will process some channel of the inputs. For example, if an input
    num_channel = 256, group = 4, num_filter=32, the PaddlePaddle will create
Z
zhangjinchao01 已提交
1388
    32*4 = 128 filters to process inputs. The channels will be split into 4
C
caoying03 已提交
1389 1390
    pieces. First 256/4 = 64 channels will process by first 32 filters. The
    rest channels will be processed by rest group of filters.
Z
zhangjinchao01 已提交
1391 1392 1393 1394 1395 1396 1397

    :param name: Layer name.
    :type name: basestring
    :param input: Layer Input.
    :type input: LayerOutput
    :param filter_size: The x dimension of a filter kernel.
    :type filter_size: int
C
caoying03 已提交
1398 1399 1400
    :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).
Z
zhangjinchao01 已提交
1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426
    :type filter_size_y: int
    :param num_filters: Each filter group's number of filter
    :param act: Activation type. Default is tanh
    :type act: BaseActivation
    :param groups: Group size of filters.
    :type groups: int
    :param stride: The x dimension of the stride.
    :type stride: int
    :param stride_y: The y dimension of the stride.
    :type stride_y: int
    :param padding: The x dimension of the padding.
    :type padding: int
    :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
D
dangqingqing 已提交
1427
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440
    :rtype: LayerOutput
    """
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters
    if filter_size_y is None:
        filter_size_y = filter_size
    if stride_y is None:
        stride_y = stride
    if padding_y is None:
        padding_y = padding
    if param_attr.attr.get('initial_smart') == True: # special initial for conv layers.
        init_w = (2.0 / (filter_size ** 2 * num_channels)) ** 0.5
1441 1442 1443 1444
        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
Z
zhangjinchao01 已提交
1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493
    Layer(
        name=name,
        inputs=Input(input.name, conv=Conv(
            filter_size=filter_size, padding=padding, stride=stride,
            channels=num_channels, groups=groups,
            filter_size_y=filter_size_y, padding_y=padding_y, stride_y=stride_y),
            **param_attr.attr),
        active_type=act.name,
        num_filters=num_filters,
        bias=ParamAttr.to_bias(bias_attr),
        shared_biases=shared_biases,
        type=LayerType.CONV_LAYER,
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.CONV_LAYER, parents=[input],
                       activation=act, num_filters=num_filters)


@wrap_name_default("pool")
@layer_support()
def img_pool_layer(input, pool_size, name=None,
                   num_channels=None, pool_type=None,
                   stride=1, start=None, padding=0, layer_attr=None):
    """
    Image pooling Layer.

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

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

    :param padding: pooling padding
    :type padding: int
    :param name: name of pooling layer
    :type name: basestring.
    :param input: layer's input
    :type input: LayerOutput
    :param pool_size: pooling size
    :type pool_size: int
    :param num_channels: number of input channel.
    :type num_channels: int
    :param pool_type: pooling type. MaxPooling or AveragePooling. Default is
                      MaxPooling.
    :type pool_type: BasePoolingType
    :param stride: stride of pooling.
    :type stride: int
    :param start: start position of pooling operation.
    :type start: int
    :param layer_attr: Extra Layer attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
1494 1495
    :return: LayerOutput object.
    :rtype: LayerOutput
Z
zhangjinchao01 已提交
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
    """
    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'

    Layer(
        name=name,
        type=LayerType.POOL_LAYER,
        inputs=[Input(input.name,
                      pool=Pool(
                          pool_type=pool_type.name + '-projection',
                          channels=num_channels,
                          size_x=pool_size,
                          start=start,
                          stride=stride,
                          padding=padding
                      ))],
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.POOL_LAYER, parents=[input],
                       num_filters=num_channels)


def __img_norm_layer__(name, input, size, norm_type, scale, power,
1525
                       num_channels, blocked, layer_attr):
Z
zhangjinchao01 已提交
1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544
    if num_channels is None:
        assert input.num_filters is not None
        num_channels = input.num_filters

    Layer(
        name=name, type=LayerType.NORM_LAYER, inputs=Input(
            input.name, norm=Norm(norm_type=norm_type,
                                  channels=num_channels, size=size,
                                  scale=scale,
                                  pow=power, blocked=blocked)
        ),
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
    return LayerOutput(name, layer_type=LayerType.NORM_LAYER, parents=[input],
                       num_filters=num_channels, img_norm_type=norm_type)


@wrap_name_default("crmnorm")
@layer_support()
D
dangqingqing 已提交
1545 1546
def img_cmrnorm_layer(input, size, scale=0.0128, power=0.75,
                      name=None, num_channels=None,
1547
                      layer_attr=None):
Z
zhangjinchao01 已提交
1548
    """
1549
    Response normalization across feature maps.
D
dangqingqing 已提交
1550 1551
    The details please refer to
    `Alex's paper <http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf>`_.
Z
zhangjinchao01 已提交
1552 1553

    :param name: layer name.
D
dangqingqing 已提交
1554
    :type name: None|basestring
Z
zhangjinchao01 已提交
1555 1556
    :param input: layer's input.
    :type input: LayerOutput
1557
    :param size: Normalize in number of :math:`size` feature maps.
Z
zhangjinchao01 已提交
1558
    :type size: int
D
dangqingqing 已提交
1559
    :param scale: The hyper-parameter.
Z
zhangjinchao01 已提交
1560
    :type scale: float
D
dangqingqing 已提交
1561
    :param power: The hyper-parameter.
Z
zhangjinchao01 已提交
1562 1563 1564
    :type power: float
    :param num_channels: input layer's filers number or channels. If
                         num_channels is None, it will be set automatically.
D
dangqingqing 已提交
1565
    :param blocked: namely normalize in number of blocked feature maps.
Z
zhangjinchao01 已提交
1566 1567
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
1568
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1569 1570 1571
    :rtype: LayerOutput
    """
    return __img_norm_layer__(name, input, size, "cmrnorm-projection", scale,
1572
                              power, num_channels, 0, layer_attr)
Z
zhangjinchao01 已提交
1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645


@wrap_bias_attr_default()
@wrap_param_attr_default(default_factory=lambda _: ParamAttr(initial_mean=1.0,
                                                             initial_std=0.))
@wrap_act_default(act=ReluActivation())
@wrap_name_default("batch_norm")
@layer_support(DROPOUT)
def batch_norm_layer(input, act=None, name=None, num_channels=None,
                     bias_attr=None, param_attr=None, layer_attr=None,
                     batch_norm_type=None,
                     moving_average_fraction=0.9,
                     use_global_stats=None):
    """
    Batch Normalization Layer. The notation of this layer as follow.

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

    ..  math::

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

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

    :param name: layer name.
    :type name: basestring
    :param input: batch normalization input. Better be linear activation.
                Because there is an activation inside batch_normalization.
    :type input: LayerOutput
    :param batch_norm_type: We have batch_norm and cudnn_batch_norm. batch_norm
                            supports both CPU and GPU. cudnn_batch_norm requires
                            cuDNN version greater or equal to v4 (>=v4). But
                            cudnn_batch_norm is faster and needs less memory
                            than batch_norm. By default (None), we will
                            automaticly select cudnn_batch_norm for GPU and
                            batch_norm for CPU. Otherwise, select batch norm
                            type based on the specified type. If you use cudnn_batch_norm,
                            we suggested you use latest version, such as v5.1.
    :type type: None|string, None or "batch_norm" or "cudnn_batch_norm"
    :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
1646
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709
    :rtype: LayerOutput
    """
    if not isinstance(act, ReluActivation):
        logger.log(logging.WARN,
                   "%s is not recommend for batch normalization's activation, "
                   "maybe the relu is better" % act.name)

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

    if num_channels is None:
        if input.num_filters is not None:
            num_channels = input.num_filters
        else:
            num_channels = input.size
    assert (batch_norm_type is None) or (batch_norm_type == "batch_norm") or \
           (batch_norm_type == "cudnn_batch_norm")
    Layer(
        name=name,
        inputs=Input(input.name,
                     image=Image(channels=num_channels),
                     **param_attr.attr),
        active_type=act.name,
        type=LayerType.BATCH_NORM_LAYER,
        batch_norm_type=batch_norm_type,
        bias=ParamAttr.to_bias(bias_attr),
        moving_average_fraction=moving_average_fraction,
        use_global_stats=use_global_stats,
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )

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


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

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

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

    The example usage is:

    .. code-block:: python

       sum_to_one_norm = sum_to_one_norm_layer(input=layer)

    :param input: Input layer.
    :type input: LayerOutput
    :param name: Layer name.
    :type name: basestring
    :param layer_attr: extra layer attributes.
    :type layer_attr: ExtraLayerAttribute.
D
dangqingqing 已提交
1710
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.SUM_TO_ONE_NORM_LAYER,
        inputs=[input.name],
        **ExtraAttr.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.SUM_TO_ONE_NORM_LAYER, parents=[input],
                       size=input.size)


@wrap_name_default("addto")
@wrap_act_default(act=LinearActivation())
@wrap_bias_attr_default(has_bias=False)
@layer_support(DROPOUT)
def addto_layer(input, act=None, name=None, bias_attr=None,
                layer_attr=None):
    """
    AddtoLayer.

    ..  math::

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

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

    The example usage is:

    ..  code-block:: python

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

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

C
caoying03 已提交
1751 1752 1753
    There is no weight matrix for each input, because it just a simple add
    operation. If you want a complicated operation before add, please use
    mixed_layer.
Z
zhangjinchao01 已提交
1754 1755

    It is a very good way to set dropout outside the layers. Since not all
C
caoying03 已提交
1756 1757
    PaddlePaddle layer support dropout, you can add an add_to layer, set
    dropout here.
Z
zhangjinchao01 已提交
1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771
    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
1772
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813
    :rtype: LayerOutput
    """
    num_filters = None
    if isinstance(input, LayerOutput):
        input = [input]

    assert isinstance(input, list) or isinstance(input, tuple)
    ipts_for_layer = []
    for each_input in input:
        assert isinstance(each_input, LayerOutput)
        ipts_for_layer.append(Input(each_input.name))
        if each_input.num_filters is not None:
            num_filters = each_input.num_filters

    Layer(
        name=name, type=LayerType.ADDTO_LAYER, inputs=ipts_for_layer,
        bias=ParamAttr.to_bias(bias_attr),
        active_type=act.name,
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
    assert isinstance(input, list) or isinstance(input, tuple)
    return LayerOutput(name, LayerType.ADDTO_LAYER, parents=input,
                       activation=act, num_filters=num_filters)


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

    :param name: Layer name.
    :type name: basestring
    :param input: input layers or projections
    :type input: list|tuple
    :param act: Activation type.
    :type act: BaseActivation
    :param layer_attr: Extra Layer Attribute.
    :type layer_attr: ExtraLayerAttribute
D
dangqingqing 已提交
1814
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911
    :rtype: LayerOutput
    """

    if isinstance(input, LayerOutput):
        input = [input]
    elif isinstance(input, Projection):
        input = [input]
    else:
        assert isinstance(input, list) or isinstance(input, tuple)

    def __is_type__(o, tp):
        if not isinstance(o, list) and not isinstance(o, tuple):
            if o == tp:
                return True
            elif len(o.__bases__) == 0:
                return False
            else:
                for bs in o.__bases__:
                    if __is_type__(bs, tp):
                        return True
                return False
        else:
            tmp = map(lambda _x: __is_type__(_x, tp), o)
            a = tmp[0]
            for b in tmp[1:]:
                assert a == b
            return a

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

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

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

    Layer(
        name=name, type=layer_type,
        inputs=[x.name for x in input] if is_concat_layer else input,
        active_type=act.name,
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )

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

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


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

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

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

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

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


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

    :param name: memory's name.
    :type name: basestring
    :param size: size of memory.
    :type size: int
    :param is_seq: is sequence for boot_layer
    :type is_seq: bool
    :param boot_layer: boot layer of memory.
    :type boot_layer: LayerOutput|None
    :param boot_bias: boot layer's bias
    :type boot_bias: ParameterAttribute|None
    :param boot_bias_active_type: boot layer's active type.
    :type boot_bias_active_type: BaseActivation
    :param boot_with_const_id: boot layer's id.
    :type boot_with_const_id: int
D
dangqingqing 已提交
1912
    :return: LayerOutput object which is a memory.
Z
zhangjinchao01 已提交
1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953
    :rtype: LayerOutput
    """
    if boot_bias_active_type is None:
        boot_bias_active_type = LinearActivation()

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

    assert boot_layer is None or isinstance(boot_layer, LayerOutput)

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

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


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

    ..  math::

L
luotao02 已提交
1954
        i_t & = \\sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i)
Z
zhangjinchao01 已提交
1955

L
luotao02 已提交
1956
        f_t & = \\sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f)
Z
zhangjinchao01 已提交
1957

L
luotao02 已提交
1958
        c_t & = f_tc_{t-1} + i_t tanh (W_{xc}x_t+W_{hc}h_{t-1} + b_c)
Z
zhangjinchao01 已提交
1959

L
luotao02 已提交
1960
        o_t & = \\sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o)
Z
zhangjinchao01 已提交
1961

L
luotao02 已提交
1962
        h_t & = o_t tanh(c_t)
Z
zhangjinchao01 已提交
1963 1964


L
luotao02 已提交
1965
    The input of lstm step is :math:`Wx_t + Wh_{t-1}`, and user should use
Z
zhangjinchao01 已提交
1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003
    :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
2004
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.LSTM_STEP_LAYER,
        active_type=act.name,
        active_gate_type=gate_act.name,
        active_state_type=state_act.name,
        bias=ParamAttr.to_bias(bias_attr),
        size=size, inputs=[input.name, state.name],
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )

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


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

    :param input:
    :type input: LayerOutput
    :param output_mem:
    :param size:
    :param act:
    :param name:
    :param gate_act:
    :param bias_attr:
    :param layer_attr:
D
dangqingqing 已提交
2043
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071
    :rtype: LayerOutput
    """
    assert input.size % 3 == 0
    if size is None:
        size = input.size / 3
    Layer(
        name=name,
        type=LayerType.GRU_STEP_LAYER,
        inputs=[
            input.name,
            output_mem.name
        ],
        bias=ParamAttr.to_bias(bias_attr),
        size=size,
        active_type=act.name,
        active_gate_type=gate_act.name,
        **ExtraAttr.to_kwargs(layer_attr)
    )
    return LayerOutput(
        name=name, layer_type=LayerType.GRU_STEP_LAYER,
        parents=[input, output_mem],
        size=size, activation=act)


@wrap_name_default()
@layer_support()
def get_output_layer(input, arg_name, name=None, layer_attr=None):
    """
C
caoying03 已提交
2072 2073 2074 2075
    Get layer's output by name. In PaddlePaddle, a layer might return multiple
    values, but returns one layer's output. If the user wants to use another
    output besides the default one, please use get_output_layer first to get
    the output from input.
Z
zhangjinchao01 已提交
2076 2077 2078 2079 2080 2081 2082 2083 2084

    :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
2085
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118
    :rtype: LayerOutput
    """
    # GetOutputLayer
    assert arg_name in input.outputs, 'Get Output From an not existed input.' \
                                      ' The get output name is %s, which not' \
                                      ' in %s' % (
                                          arg_name, ",".join(input.outputs))
    Layer(name=name, type=LayerType.GET_OUTPUT_LAYER,
          inputs=[Input(input.name, input_layer_argument=arg_name)],
          size=input.size,
          **ExtraLayerAttribute.to_kwargs(layer_attr))

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


@wrap_name_default()
@wrap_act_default()
@wrap_bias_attr_default()
@wrap_param_attr_default()
@layer_support()
def recurrent_layer(input, act=None, bias_attr=None,
                    param_attr=None, name=None, layer_attr=None):
    """
    TODO(yuyang18): Add docs

    :param input:
    :param size:
    :param act:
    :param bias_attr:
    :param param_attr:
    :param name:
    :param layer_attr:
D
dangqingqing 已提交
2119
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164
    """
    Layer(name=name,
          type=LayerType.RECURRENT_LAYER,
          inputs=Input(input.name, **param_attr.attr),
          active_type=act.name,
          size=input.size,
          bias=ParamAttr.to_bias(bias_attr),
          **ExtraAttr.to_kwargs(layer_attr))
    return LayerOutput(name=name, layer_type=LayerType.RECURRENT_LAYER,
                       parents=[input], size=input.size, activation=act)


class StaticInput(object):
    """
    StaticInput is only used in recurrent_group which defines a read-only memory
    that can be a sequence or non-sequence.
    """
    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)
    """
    def __init__(self, input):
        assert isinstance(input, LayerOutput)
        assert input.size is not None
        self.input = input


@wrap_name_default("recurrent_group")
def recurrent_group(step, input, reverse=False, name=None):
    """
C
caoying03 已提交
2165 2166 2167 2168 2169
    Recurrent layer group is an extremely flexible recurrent unit in
    PaddlePaddle. As long as the user defines the calculation done within a
    time step, PaddlePaddle will iterate such a recurrent calculation over
    sequence input. This is extremely usefull for attention based model, or
    Neural Turning Machine like models.
Z
zhangjinchao01 已提交
2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213

    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

2214 2215
    :param reverse: If reverse is set true, the recurrent unit will process the
                    input sequence in a reverse order.
Z
zhangjinchao01 已提交
2216
    :type reverse: bool
D
dangqingqing 已提交
2217
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 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 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334
    :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]
    assert isinstance(input, list) or isinstance(input, tuple)

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

    in_links = filter(is_in_links, input)

    contains_sub_seq = [False]

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

    RecurrentLayerGroupWithoutOutLinksBegin(
        name=name, in_links=map(map_in_links, in_links),
        seq_reversed=reverse)
    in_args = []
    for each_input in input:
        assert is_single_input(each_input)
        if isinstance(each_input, LayerOutput):
            in_args.append(each_input)
        elif isinstance(each_input, SubsequenceInput):
            in_args.append(each_input.input)
        else:
            mem_name = "__%s_memory__" % each_input.input.name
            mem = memory(name=mem_name,
                         is_seq=each_input.is_seq,
                         size=each_input.input.size,
                         boot_layer=each_input.input)
            with mixed_layer(name=mem_name, size=each_input.input.size,
                             act=IdentityActivation()) as mix:
                mix += identity_projection(mem)
            in_args.append(mem)

    layer_outs = step(*in_args)

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

    for ot in layer_outs:
        assert isinstance(ot, LayerOutput)
        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

class BaseGeneratedInput(object):
    def __init__(self):
        self.bos_id = None
        self.eos_id = None

    def before_real_step(self):
        raise NotImplementedError()

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


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

    def before_real_step(self):
        predict_id = memory(name='__beam_search_predict__',
                            size=self.size,
                            boot_with_const_id=self.bos_id)

        trg_emb = embedding_layer(input=predict_id,
                                  size=self.embedding_size,
                                  param_attr=ParamAttr(
                                      name=self.embedding_name))
        return trg_emb

    def __init__(self, size, embedding_name, embedding_size):
        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
2335
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363
    :rtype: LayerOutput
    """

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


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

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

    The example usage is:

    .. code-block:: python

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

L
luotao02 已提交
2364 2365
    :param name: Layer name.
    :type name: basestring
Z
zhangjinchao01 已提交
2366 2367 2368 2369 2370 2371
    :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
2372
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388
    :rtype: LayerOutput
    """
    Layer(name=name,
          type=LayerType.EOSID_LAYER,
          eos_id=eos_id,
          inputs=[input.name],
          **ExtraLayerAttribute.to_kwargs(layer_attr))
    return LayerOutput(name=name, layer_type=LayerType.EOSID_LAYER,
                       parents=[input])


@wrap_name_default()
def beam_search(step, input, bos_id, eos_id, beam_size,
                result_file, dict_file="", id_input=None,
                max_length=500, name=None,
                num_results_per_sample=None):
2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466
    """
    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)
            with mixed_layer(size=512) as simple_rnn:
                simple_rnn += full_matrix_projection(input)
                simple_rnn += last_time_step_output
            return simple_rnn

        beam_gen = beam_search(name="decoder",
                               step=rnn_step,
                               input=[StaticInput("encoder_last")],
                               bos_id=0,
                               eos_id=1,
                               beam_size=5,
                               result_file="./generated_sequences.txt")

    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
                 step, and it is appled to sequences with arbitrary length by
                 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
    :type input: StaticInput|GeneratedInput
    :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
                   symbol is ensential, since it is used to initialize the RNN
                   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
    :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 result_file: Path of the file to store the generated results.
    :type result_file: basestring
    :param dict_file: Path of dictionary. This is an optional parameter.
                      Every line is a word in the dictionary with
                      (line number - 1) as the word index.
                      If this parameter is set to None, or to an empty string,
                      only word index are printed in the generated results.
    :type dict_file: basestring
    :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
    :param id_input: Index of the input sequence, and the specified index will
                     be prited in the gereated results. This an optional
                     parameter.
    :type id_input: LayerOutput
    :return: The seq_text_printer that prints the generated sequence to a file.
    :rtype: evaluator
    """

Z
zhangjinchao01 已提交
2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543
    if num_results_per_sample is None:
        num_results_per_sample = beam_size
    if num_results_per_sample > beam_size:
        logger.warning("num_results_per_sample should be less than beam_size")

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

    generated_input_index = -1

    real_input = []
    for i, each_input in enumerate(input):
        # print type(each_input)
        assert isinstance(each_input, StaticInput) or isinstance(each_input,
                                                          BaseGeneratedInput)
        if isinstance(each_input, BaseGeneratedInput):
            assert generated_input_index == -1
            generated_input_index = i
        else:
            real_input.append(each_input)

    assert generated_input_index != -1

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

    gipt.bos_id = bos_id
    gipt.eos_id = eos_id

    def __real_step__(*args):
        eos_name = "__%s_eos_layer__" % name
        RecurrentLayerGroupSetGenerator(Generator(
            eos_layer_name=eos_name,
            max_num_frames=max_length,
            beam_size=beam_size,
            num_results_per_sample=num_results_per_sample))

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

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

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

        return predict

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

    if id_input is None:
        inputs = [tmp.name]
    else:
        assert isinstance(id_input, LayerOutput)
        inputs = [id_input.name, tmp.name]
        tmp.parents.append(id_input)

    Evaluator(name='target_printer',
              type='seq_text_printer',
              dict_file=dict_file,
              result_file=result_file,
              inputs=inputs
              )
    return tmp


@wrap_name_default()
def regression_cost(input, label, cost='square_error', name=None):
    """
    Regression Layer.

    TODO(yuyang18): Complete this method.

    :param name: layer name.
    :param input: Network prediction.
    :param label: Data label.
    :param cost: Cost method.
D
dangqingqing 已提交
2544
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567
    """
    Layer(inputs=[Input(input.name), Input(label.name)], type=cost, name=name)
    return LayerOutput(
        name, LayerType.COST, parents=[input, label]
    )


@wrap_name_default("cost")
def classification_cost(input, label, name=None,
                        cost="multi-class-cross-entropy",
                        evaluator=classification_error_evaluator):
    """
    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
    :param cost: cost method.
    :type cost: basestring
    :param evaluator: Evaluator method.
D
dangqingqing 已提交
2568
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616
    :rtype: LayerOutput
    """
    assert input.layer_type != LayerType.DATA
    assert isinstance(input.activation, SoftmaxActivation)
    assert label.layer_type == LayerType.DATA
    Layer(name=name, type=cost, inputs=[Input(input.name), Input(label.name)])

    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

        e(name=e.__name__, input=input, label=label)

    if not isinstance(evaluator, list) and not isinstance(evaluator, tuple):
        evaluator = [evaluator]

    for each_evaluator in evaluator:
        __add_evaluator__(each_evaluator)

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

def conv_operator(input, filter_size, num_filters,
                  num_channel=None, stride=1, padding=0,
                  filter_size_y=None, stride_y=None, padding_y=None):
    """
    Different from img_conv_layer, conv_op is an Operator, which can be used
    in mixed_layer. And conv_op takes two inputs to perform convolution.
    The first input is the image and the second is filter kernel. It only
    support GPU mode.

    The example usage is:

    .. code-block:: python

       op = conv_operator(input=[layer1, layer2],
                          filter_size=3.0,
                          num_filters=64,
                          num_channels=64)

    :param input: Input layer.
    :type input: LayerOutput|list|tuple
    :param filter_size: The x dimension of a filter kernel.
    :type filter_size: int
C
caoying03 已提交
2617 2618 2619
    :param filter_size_y: The y dimension of a filter kernel. Since
                        PaddlePaddle now supports rectangular filters,
                        the filter's shape can be (filter_size, filter_size_y).
Z
zhangjinchao01 已提交
2620 2621 2622 2623
    :type filter_size_y: int
    :param num_filter: channel of output data.
    :type num_filter: int
    :param num_channel: channel of input data.
L
luotao02 已提交
2624
    :type num_channel: int
Z
zhangjinchao01 已提交
2625
    :param stride: The x dimension of the stride.
L
luotao02 已提交
2626
    :type stride: int
Z
zhangjinchao01 已提交
2627
    :param stride_y: The y dimension of the stride.
L
luotao02 已提交
2628
    :type stride_y: int
Z
zhangjinchao01 已提交
2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682
    :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
    """
    assert isinstance(input, list) or isinstance(input, tuple)
    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
    op = ConvOperator(input_layer_name=[x.name for x in input],
                      num_filters = num_filter,
                      conv_conf=Conv(filter_size=filter_size,
                                     padding=padding,
                                     stride=stride,
                                     channels=num_channel,
                                     filter_size_y=filter_size_y,
                                     padding_y=padding_y,
                                     stride_y=stride_y))
    op.origin = input
    op.origin.operator = "conv_op"
    return op


@wrap_name_default()
def conv_shift_layer(input, name=None):
    """
    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:
     - a's index is computed modulo M.
     - b's index is computed modulo N.

    The example usage is:

    .. code-block:: python

       conv_shift = conv_shif_layer(input=[layer1, layer2])

    :param name: layer name
    :type name: basestring
    :param input: Input layer.
    :type input: LayerOutput|list|tuple.
D
dangqingqing 已提交
2683
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726
    :rtype: LayerOutput
    """
    assert isinstance(input, list) or isinstance(input, tuple)
    Layer(
        name=name,
        type=LayerType.CONV_SHIFT_LAYER,
        inputs=[x.name for x in input],
    )

    return LayerOutput(name, LayerType.CONV_SHIFT_LAYER, parents=input)


@wrap_name_default()
@wrap_param_attr_default()
@wrap_bias_attr_default()
@layer_support(ERROR_CLIPPING, DROPOUT)
def tensor_layer(input, size, act=None, name=None,
                 param_attr=None, bias_attr=None, layer_attr=None):
    """
    This layer performs tensor operation for two input.
    For example, each sample:

    .. math::
       y_{i} = x_{1} * W_{i} * {x_{2}^\mathrm{T}}, i=0,1,...,K-1

    In this formular:
      - :math:`x_{1}`: the first input contains M elements.
      - :math:`x_{2}`: the second input contains N elements.
      - :math:`y_{i}`: the i-th element of y.
      - :math:`W_{i}`: the i-th learned weight, shape if [M, N]
      - :math:`{x_{2}}^\mathrm{T}`: the transpose of :math:`x_{2}`.

    The simple usage is:

    .. code-block:: python

       tensor = tensor_layer(input=[layer1, layer2])

    :param name: layer name
    :type name: basestring
    :param input: Input layer.
    :type input: LayerOutput|list|tuple.
    :param size: the layer dimension.
L
luotao02 已提交
2727
    :type size: int.
Z
zhangjinchao01 已提交
2728 2729 2730 2731 2732 2733 2734 2735 2736 2737
    :param act: Activation Type. Default is tanh.
    :type act: BaseActivation
    :param param_attr: The Parameter Attribute.
    :type param_attr: ParameterAttribute|list
    :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
2738
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2739 2740 2741 2742 2743 2744 2745 2746 2747 2748
    :rtype: LayerOutput
    """
    assert isinstance(input, list) or isinstance(input, tuple)
    assert len(input) == 2
    Layer(
        name=name,
        size=size,
        type=LayerType.TENSOR_LAYER,
        active_type=act.name,
        bias=ParamAttr.to_bias(bias_attr),
2749
        inputs=[Input(input[0].name, **param_attr.attr),
Z
zhangjinchao01 已提交
2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 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 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831
                Input(input[1].name)],
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.TENSOR_LAYER, parents=input,
                       activation=act, size=size)


@wrap_param_attr_default()
def trans_full_matrix_projection(input, size=0, param_attr=None):
    """
    Different from full_matrix_projection, this projection performs matrix
    multiplication, using transpose of weight.

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

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

    .. code-block:: python

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

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


@wrap_name_default()
@wrap_param_attr_default()
@wrap_bias_attr_default()
@wrap_act_default()
def selective_fc_layer(input, size, act=None, name=None,
                       pass_generation=False,
                       has_selected_colums=True,
                       mul_ratio=0.02,
                       param_attr=None, bias_attr=None, layer_attr=None):
    """
    Selectived fully connected layer. Different from fc_layer, the output
    of this layer maybe sparse. It requires an additional input to indicate
    several selected columns for output. If the selected columns is not
    specified, selective_fc_layer acts exactly like fc_layer.

    The simple usage is:

    .. code-block:: python

       sel_fc = selective_fc_layer(input=input, 128, act=TanhActivation())

    :param name: The Layer Name.
    :type name: basestring
    :param input: The 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.
    :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
2832
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 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
    :rtype: LayerOutput
    """
    if isinstance(input, LayerOutput):
        input = [input]
        assert not isinstance(param_attr, list)
        param_attr = [param_attr]
    else:
        if isinstance(param_attr, list) or isinstance(param_attr, tuple):
            assert len(input) == len(param_attr)
        else:
            param_attr = [copy.deepcopy(param_attr) for _ in range(len(input))]

    assert isinstance(input, list)

    def __idx_to_input__(i):
        attr = param_attr[i]
        assert isinstance(attr, ParameterAttribute)
        return Input(input[i].name, **attr.attr)

    Layer(
        inputs=map(__idx_to_input__, range(len(input))),
        name=name,
        type=LayerType.SEL_FC_LAYER,
        size=size,
        active_type=act.name,
        selective_fc_pass_generation=pass_generation,
        has_selected_colums=has_selected_colums,
        selective_fc_full_mul_ratio=mul_ratio,
        **ExtraLayerAttribute.to_kwargs(layer_attr)
    )
    return LayerOutput(name, LayerType.SEL_FC_LAYER, input, activation=act,
                       size=size)


@wrap_name_default()
def sampling_id_layer(input, name=None):
    """
    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
D
dangqingqing 已提交
2883
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.SAMPLING_ID_LAYER,
        inputs=[Input(input.name)],
    )
    return LayerOutput(name, LayerType.SAMPLING_ID_LAYER, input)


@wrap_name_default()
def slope_intercept_layer(input, name=None, slope=1.0, intercept=0.0):
    """
    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.
D
dangqingqing 已提交
2917
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930
    :rtype: LayerOutput
    """
    Layer(
        name=name,
        type=LayerType.SLOPE_INTERCEPT_LAYER,
        slope=slope,
        intercept=intercept,
        inputs=[Input(input.name)],
    )
    return LayerOutput(name, LayerType.SLOPE_INTERCEPT_LAYER, input)


@wrap_name_default()
2931
def linear_comb_layer(weights, vectors, size, name=None):
Z
zhangjinchao01 已提交
2932
    """
2933 2934 2935 2936
    A layer for weighted sum of vectors takes two inputs.
      - Input: size of weights is M
               size of vectors is M*N
      - Output: a vector of size=N
Z
zhangjinchao01 已提交
2937 2938 2939

    .. math::

2940 2941 2942 2943 2944 2945
       z(i) = \sum_{j=0}^{M-1} x(j) y(i+Nj)
    where :math:`0 \le i \le N-1`

    Or in the matrix notation:

    .. math::
Z
zhangjinchao01 已提交
2946

2947
       z = x^\mathrm{T} Y
Z
zhangjinchao01 已提交
2948 2949

    In this formular:
2950 2951 2952 2953 2954 2955
      - :math:`x`: weights
      - :math:`y`: vectors.
      - :math:`z`: the output.

    Note that the above computation is for one sample. Multiple samples are
    processed in one batch.
Z
zhangjinchao01 已提交
2956 2957 2958 2959 2960

    The simple usage is:

    .. code-block:: python

2961
       linear_comb = linear_comb_layer(weighs=weight, vectors=vectors,
Z
zhangjinchao01 已提交
2962 2963 2964 2965 2966 2967 2968 2969
                                       size=elem_dim)

    :param input: The input layers.
    :type input: LayerOutput
    :param size: the dimension of this layer.
    :type size: int
    :param name: The Layer Name.
    :type name: basestring
D
dangqingqing 已提交
2970
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
2971 2972 2973 2974 2975
    :rtype: LayerOutput
    """

    Layer(
        name=name,
2976
        type=LayerType.LINEAR_COMBINATION_LAYER,
Z
zhangjinchao01 已提交
2977
        size=size,
2978
        inputs=[Input(weights.name), Input(vectors.name)],
Z
zhangjinchao01 已提交
2979
    )
2980 2981 2982 2983
    return LayerOutput(name, LayerType.LINEAR_COMBINATION_LAYER,
                       [weights, vectors], size=size)

convex_comb_layer = linear_comb_layer
Z
zhangjinchao01 已提交
2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996

@wrap_name_default()
def block_expand_layer(input,
                       channel=0,
                       block_x=0,
                       block_y=0,
                       stride_x=0,
                       stride_y=0,
                       padding_x=0,
                       padding_y=0,
                       name=None):
    """
    Expand feature map to minibatch matrix.
L
luotao02 已提交
2997 2998
       - matrix width is: block_y * block_x * channel
       - matirx height is: outputH * outputW
Z
zhangjinchao01 已提交
2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011

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

3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022
    The simple usage is:

    .. code-block:: python

       block_expand = block_expand_layer(input,
                                         channel=128,
                                         stride_x=1,
                                         stride_y=1,
                                         block_x=1,
                                         block_x=3)

Z
zhangjinchao01 已提交
3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040
    :param input: The input layer.
    :type input: LayerOutput
    :param channel: The channel number of input layer.
    :type channel: int
    :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.
D
dangqingqing 已提交
3041
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066
    :rtype: LayerOutput
    """
    Layer(name=name,
          input=Input(input.name,
                      block_expand=BlockExpand(channel=channel,
                                               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,
         )

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

@wrap_name_default()
def ctc_layer(input, label, size, name=None, norm_by_times=False):
    """
    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.

3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077
    More details 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:
        Considering the 'blank' label needed by CTC, you need to use
        (num_classes + 1) as the input size. num_classes is the category number.
        And the 'blank' is the last category index. So the size of 'input' layer, such as
        fc_layer with softmax activation, should be num_classes + 1. The size of ctc_layer
        should also be num_classes + 1.

Z
zhangjinchao01 已提交
3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090
    The simple usage:

    .. code-block:: python

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

    :param input: The input layers.
    :type input: LayerOutput
    :param label: The data layer of label with variable length.
    :type label: LayerOutput
3091
    :param size: category numbers + 1.
Z
zhangjinchao01 已提交
3092 3093 3094 3095 3096
    :type size: int
    :param name: The name of this layer, which can not specify.
    :type name: string|None
    :param norm_by_times: Whether to normalization by times. False by default.
    :type norm_by_times: bool
D
dangqingqing 已提交
3097
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111
    :rtype: LayerOutput
    """
    assert isinstance(input, LayerOutput)
    assert isinstance(label, LayerOutput)
    Layer(
        name = name,
        type = LayerType.CTC_LAYER,
        size = size,
        norm_by_times = norm_by_times,
        inputs = [input.name, label.name]
    )
    return LayerOutput(name, LayerType.CTC_LAYER, [input, label], size=size)

@wrap_name_default()
3112
@wrap_param_attr_default()
Z
zhangjinchao01 已提交
3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138
def crf_layer(input, label, size, weight=None, param_attr=None, name=None):
    """
    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.
    :type input: LayerOutput
    :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
D
dangqingqing 已提交
3139
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3140 3141 3142 3143 3144 3145
    :rtype: LayerOutput
    """
    assert isinstance(input, LayerOutput)
    assert isinstance(label, LayerOutput)
    assert weight is None or isinstance(weight, LayerOutput)

3146
    ipts = [Input(input.name, **param_attr.attr),
Z
zhangjinchao01 已提交
3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162
            Input(label.name)]
    if weight is not None:
        ipts.append(Input(weight.name))

    Layer(
        name = name,
        type = LayerType.CRF_LAYER,
        size = size,
        inputs = ipts,
    )
    parents = [input, label]
    if weight is not None:
        parents.append(weight)
    return LayerOutput(name, LayerType.CRF_LAYER, parents, size=size)

@wrap_name_default()
3163
@wrap_param_attr_default()
Z
zhangjinchao01 已提交
3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181
def crf_decoding_layer(input, size, label=None, param_attr=None, name=None):
    """
    A layer for calculating the decoding sequence of sequential conditional
    random field model. The decoding sequence is stored in output.ids.
    If a second input is provided, it is treated as the ground-truth label, and
    this layer will also calculate error. output.value[i] is 1 for incorrect
    decoding or 0 for correct decoding.

    :param input: The first input layer.
    :type input: LayerOutput
    :param size: size of this layer.
    :type size: int
    :param label: None or ground-truth label.
    :type label: LayerOutput or None
    :param param_attr: Parameter attribute. None means default attribute
    :type param_attr: ParameterAttribute
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring
D
dangqingqing 已提交
3182
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3183 3184 3185 3186 3187 3188
    :rtype: LayerOutput
    """

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

3189
    ipts = [Input(input.name, **param_attr.attr)]
Z
zhangjinchao01 已提交
3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209
    if label is not None:
        ipts.append(Input(label.name))

    Layer(
        name = name,
        type = LayerType.CRF_DECODING_LAYER,
        size = size,
        inputs = ipts,
    )
    parents = [input]
    if label is not None:
        parents.append(label)
    return LayerOutput(name, LayerType.CRF_DECODING_LAYER, parents, size=size)

"""
following are cost Layers.
"""
@wrap_name_default()
def rank_cost(left, right, lable, weight=None, name=None, coeff=1.0):
    """
3210
    A cost Layer for learning to rank using gradient descent. Details can refer
Z
zhangjinchao01 已提交
3211 3212 3213 3214 3215 3216
    to `papers <http://research.microsoft.com/en-us/um/people/cburges/papers/ICML_ranking.pdf>`_.
    This layer contains at least three inputs. The weight is an optional
    argument, which affects the cost.

    .. math::

L
luotao02 已提交
3217
       C_{i,j} & = -\\tilde{P_{ij}} * o_{i,j} + log(1 + e^{o_{i,j}})
Z
zhangjinchao01 已提交
3218

L
luotao02 已提交
3219
       o_{i,j} & =  o_i - o_j
Z
zhangjinchao01 已提交
3220

L
luotao02 已提交
3221
       \\tilde{P_{i,j}} & = \\{0, 0.5, 1\\} \ or \ \\{0, 1\\}
Z
zhangjinchao01 已提交
3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250

    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
D
dangqingqing 已提交
3251
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299
    :rtype: LayerOutput
    """
    assert left.size == 1
    assert right.size == 1
    assert label.size == 1

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

    Layer(name=name,
          type=LayerType.RANK_COST,
          inputs=ipts,
          coeff=coeff,
         )

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

@wrap_name_default()
def lambda_cost(input, score, NDCG_num=5, max_sort_size=-1, coeff=1.0):
    """
    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)

    :param input: The 1st input. Samples of the same query should be loaded
                  as sequence. User should provided socres for each sample.
                  The score should be the 2nd input of this layer.
    :type input: LayerOutput
    :param score: The 2nd input. Score of each sample.
    :type input: LayerOutput
    :param NDCG_num: The size of NDCG (Normalized Discounted Cumulative Gain),
                     e.g., 5 for NDCG@5. It must be less than for equal to the
                     minimum size of lists.
    :type NDCG_num: int
    :param max_sort_size: The size of partial sorting in calculating gradient.
                          If max_sort_size = -1, then for each list, the
                          algorithm will sort the entire list to get gradient.
                          In other cases, max_sort_size must be greater than or
C
caoying03 已提交
3300 3301 3302
                          equal to NDCG_num. And if max_sort_size is greater
                          than the size of a list, the algorithm will sort the
                          entire list of get gradient.
Z
zhangjinchao01 已提交
3303 3304 3305 3306 3307
    :type max_sort_size: int
    :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
D
dangqingqing 已提交
3308
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339
    :rtype: LayerOutput
    """
    Layer(name=name,
          type=LayerType.LAMBDA_COST,
          inputs=[input.name, score.name],
          NDCG_num=NDCG_num,
          max_sort_size=max_sort_size,
          coeff=coeff,
         )

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

@wrap_name_default()
def cross_entropy(input, label, name=None, coeff=1.0):
    """
    A loss layer for multi class entropy.

    .. code-block:: python

       cost = cross_entropy(input, label)

    :param input: The first input layer.
    :type input: LayerOutput.
    :param label: The input label.
    :type input: LayerOutput.
    :param type: The type of cost.
    :type type: basestring.
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring.
    :param coeff: The coefficient affects the gradient in the backward.
    :type coeff: float.
D
dangqingqing 已提交
3340
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3341 3342 3343 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 3372
    :rtype: LayerOutput.
    """

    Layer(name=name,
          type=LayerType.CROSS_ENTROPY,
          inputs=[input.name, label.name],
          coeff=coeff,
         )
    return LayerOutput(name, LayerType.CROSS_ENTROPY, parents=[input, label])

@wrap_name_default()
def cross_entropy_with_selfnorm(input, label, name=None, coeff=1.0,
                                softmax_selfnorm_alpha=0.1):
    """
    A loss layer for multi class entropy with selfnorm.

    .. code-block:: python

       cost = cross_entropy_with_selfnorm(input, label)

    :param input: The first input layer.
    :type input: LayerOutput.
    :param label: The input label.
    :type input: LayerOutput.
    :param type: The type of cost.
    :type type: basestring.
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring.
    :param coeff: The coefficient affects the gradient in the backward.
    :type coeff: float.
    :param softmax_selfnorm_alpha: The scale factor affects the cost.
    :type softmax_selfnorm_alpha: float.
D
dangqingqing 已提交
3373
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405
    :rtype: LayerOutput.
    """
    Layer(name=name,
          type=LayerType.CROSS_ENTROPY_WITH_SELFNORM,
          inputs=[input.name, label.name],
          coeff=coeff,
          softmax_selfnorm_alpha=softmax_selfnorm_alpha,
         )

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

@wrap_name_default()
def huber_cost(input, label, name=None, coeff=1.0):
    """
    A loss layer for huber loss.

    .. code-block:: python

       cost = huber_cost(input, label)

    :param input: The first input layer.
    :type input: LayerOutput.
    :param label: The input label.
    :type input: LayerOutput.
    :param 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.
D
dangqingqing 已提交
3406
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435
    :rtype: LayerOutput.
    """

    Layer(name=name,
          type=LayerType.HUBER,
          inputs=[input.name, label.name],
          coeff=coeff,
         )
    return LayerOutput(name, LayerType.HUBER, parents=[input, label])

@wrap_name_default()
def multi_binary_label_cross_entropy(input, label, name=None, coeff=1.0):
    """
    A loss layer for multi binary label cross entropy.

    .. code-block:: python

       cost = multi_binary_label_cross_entropy(input, label)

    :param input: The first input layer.
    :type input: LayerOutput
    :param label: The input label.
    :type input: LayerOutput
    :param type: The type of cost.
    :type type: basestring
    :param name: The name of this layers. It is not necessary.
    :type name: None|basestring
    :param coeff: The coefficient affects the gradient in the backward.
    :type coeff: float
D
dangqingqing 已提交
3436
    :return: LayerOutput object.
Z
zhangjinchao01 已提交
3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451
    :rtype: LayerOutput
    """

    if not isinstance(input.act, SigmoidActivation):
        logger.log(logging.WARN,
                   "%s is not recommend for batch normalization's activation, "
                   "maybe the relu is better" % act.name)

    Layer(name=name,
          type=LayerType.MULTI_BIN_LABEL_CROSS_ENTROPY,
          inputs=[input.name, label.name],
          coeff=coeff,
         )
    return LayerOutput(name, LayerType.MULTI_BIN_LABEL_CROSS_ENTROPY,
                       parents=[input, label])