config_parser.py 120.5 KB
Newer Older
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 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102
# 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.

from __future__ import print_function
'''
The following functions are available in the config file:

Bias: define bias. To be used as value of bias argument in Layer().

Data: define data provider.

Input: define input layer for a layer. To be used as element of inputs argument
       in Layer().

Conv: define a convolution operation for an input of a layer.

Norm: define a normalization operation for an input of a layer.

Pool: define a pooling operation for an input of a layer.

Layer: define a layer.

Parameter: define a parameter.

Import: import another config file. If the imported config file name is
        a relative path, then it will be searched under the directory of the
        current config file.

Inputs(layer_names...):
    Define the name of the input layers of the NeuralNetwork.
    The type of these layers must be "data".
    These layers will be provided with the DataBatch obtained
    from DataProvider. The data streams from DataProvider must
    have the same order.

Outputs(layer_names...):
    Define the name of the output layers of the NeuralNetwork.
    Usually the output is simply the cost layer.
    You can specify other layers as outputs and  calculate the
    cost (and its derivative) yourself.


default_initial_std(val)
default_initial_mean(val)
default_momentum(val):
default_decay_rate(val): Set the default value for these parameters


get_config_arg(name, type, default): Get the value for a config parameter.


*** customized extension to config_parser ***
The functionality of the config_parser can be extended.
If the config_arg_str for parse_config() contains
extension_module_name=[MODULE_NAME], then config_parser will call
MODULE_NAME.get_config_funcs(g_config)
MODULE_NAME.get_config_funcs() should return a dictionary of name to functions,
those functions will be available in the config file.
See trainer/tests/config_parser_test.py for example

To use this from paddle_trainer, paddle_trainer should be called with
--config_args=extension_module_name=[MODULE_NAME]

'''

import copy
import logging
import os
import sys
import traceback
import math
import shutil

try:
    from paddle.proto.DataConfig_pb2 import DataConfig
    from paddle.proto.ModelConfig_pb2 import ModelConfig
    from paddle.proto.ModelConfig_pb2 import LayerConfig
    from paddle.proto.ModelConfig_pb2 import LayerInputConfig
    from paddle.proto.ModelConfig_pb2 import ProjectionConfig
    from paddle.proto.ModelConfig_pb2 import OperatorConfig
    from paddle.proto.ModelConfig_pb2 import GeneratorConfig
    from paddle.proto.ModelConfig_pb2 import LinkConfig
    from paddle.proto.ParameterConfig_pb2 import ParameterConfig
    from paddle.proto.ParameterConfig_pb2 import ParameterUpdaterHookConfig
    from paddle.proto.TrainerConfig_pb2 import TrainerConfig

except Exception as e:
    traceback.print_exc()
    raise

logging.basicConfig(
103
    format='[%(levelname)s %(asctime)s %(filename)s:%(lineno)s] %(message)s', )
104 105 106
logger = logging.getLogger('paddle')
logger.setLevel(logging.INFO)
__real_print__ = print
107
print = logger.info
108 109 110 111

# from layer type name to layer class
g_layer_type_map = {}

112

113 114 115
# Initialize global variables. We use this function so that we can
# call parse_config() multiple times
def init_config_environment(
116 117 118 119 120 121 122 123 124 125 126 127 128 129 130
        g_default_momentum=None,
        g_default_decay_rate=None,
        g_default_initial_mean=0.,
        g_default_initial_std=0.01,
        g_default_num_batches_regularization=None,
        g_default_initial_strategy=0,
        g_default_initial_smart=False,
        g_default_gradient_clipping_threshold=None,
        g_default_device=None,
        g_default_update_hooks=None,
        g_default_compact_func=None,
        g_config=TrainerConfig(),
        g_layer_map={},
        g_parameter_map={},
        g_extended_config_funcs={},
131 132

        # store command args of paddle_trainer
133
        g_command_config_args={},
134 135

        # Used for PyDataProvider to avoid duplicate module name
136 137 138 139 140 141
        g_py_module_name_list=[],
        g_current_submodel=None,
        g_root_submodel=None,
        g_submodel_map={},
        g_submodel_stack=[],
        g_add_submodel_suffix=False, ):
142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157

    for k, v in locals().iteritems():
        globals()[k] = copy.deepcopy(v)


# Because type is widely used as a variable name in this code.
# we need a different function name for the builtin type()
def type_of(x):
    return type(x)


# Check a condition derived config file
def config_assert(b, msg):
    if not b:
        logger.fatal(msg)

158

159 160
g_config_funcs = {}

161

162 163 164 165 166
# decorator for indicating a function which can be used in config file
def config_func(func):
    g_config_funcs[func.func_name] = func
    return func

167

168 169 170 171 172
# decorator for indicating a class which can be used in config file
def config_class(cls):
    g_config_funcs[cls.__name__] = cls
    return cls

173

174 175 176 177 178 179
# decorator for indicating a class for a layer type
def config_layer(layer_type):
    def wrap(cls):
        g_config_funcs[cls.__name__] = cls
        g_layer_type_map[layer_type] = cls
        return cls
180

181 182
    return wrap

183

184 185 186
def gen_parameter_name(layer_name, input_index):
    return '_%s.w%d' % (layer_name, input_index)

187

188 189 190
def gen_bias_parameter_name(layer_name):
    return '_%s.wbias' % layer_name

191

192 193 194
def default(x, default_value):
    return default_value if x is None else x

195

196 197 198 199 200 201
class Cfg(object):
    def add_keys(self, locals):
        for k, v in locals.iteritems():
            if not k.startswith('_'):
                self.__setattr__(k, v)

202

203 204
# functions available in config file

205

206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223
# Define the name of the input layers of the NeuralNetwork.
# The type of these layers must be "data".
# These layers will be provided with the DataBatch obtained
# from DataProvider. The data streams from DataProvider must
# have the same order.
@config_func
def Inputs(*args):
    for name in args:
        name = MakeLayerNameInSubmodel(name)
        global g_current_submodel, g_root_submodel
        if g_current_submodel.is_recurrent_layer_group:
            config_assert(False, "Do not set Inputs in recurrent layer group")
        else:
            g_current_submodel.input_layer_names.append(name)

        if g_current_submodel is g_root_submodel:
            g_config.model_config.input_layer_names.append(name)

224

225 226
@config_func
def HasInputsSet():
227
    return len(g_current_submodel.input_layer_names) != 0
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

# Define the name of the output layers of the NeuralNetwork.
# Usually the output is simply the cost layer.
# You can specify other layers as outputs and calculate the
# cost (and its derivative) yourself.
@config_func
def Outputs(*args):
    for name in args:
        name = MakeLayerNameInSubmodel(name)
        global g_current_submodel, g_root_submodel
        if g_current_submodel.is_recurrent_layer_group:
            config_assert(False, "Do not set Outputs in recurrent layer group")
        else:
            g_current_submodel.output_layer_names.append(name)

        if g_current_submodel is g_root_submodel:
            g_config.model_config.output_layer_names.append(name)


@config_func
def SubModelBegin(name):
    global g_current_submodel, g_root_submodel, g_submodel_stack
    g_submodel_stack.append(g_current_submodel)

253
    name = MakeLayerNameInParentSubmodel(name)  #rename in nested submodel
254 255 256 257 258 259 260 261 262

    config_assert(name not in g_submodel_map,
                  'Duplicated submodel name: %s' % name)

    sub_model = g_config.model_config.sub_models.add()
    sub_model.name = name
    g_submodel_map[name] = sub_model
    g_current_submodel = sub_model

263

264
@config_func
265
def SubModelEnd(name=None):
266
    global g_current_submodel, g_root_submodel, g_submodel_stack
267 268
    config_assert(g_current_submodel is not g_root_submodel,
                  "submodel not begin")
269
    if name is not None:
270 271 272
        config_assert(
            g_current_submodel.name == MakeLayerNameInParentSubmodel(name),
            "submodel name error")
273 274 275

    g_current_submodel = g_submodel_stack.pop()

276

277 278
def MakeLayerNameInParentSubmodel(name):
    suffix = ""
279 280
    if len(g_submodel_stack) > 1:
        suffix = "@" + g_submodel_stack[-1].name
281 282
    return name + suffix

283

284 285 286
def GetLayerBaseName(name):
    return name.split('@')[0]

287 288

def MakeLayerNameInSubmodel(name, submodel_name=None):
289 290
    global g_current_submodel
    global g_add_submodel_suffix
291 292
    if (submodel_name is None and not g_add_submodel_suffix and
            not g_current_submodel.is_recurrent_layer_group):
293 294 295 296 297
        return name
    if submodel_name is None:
        submodel_name = g_current_submodel.name
    return name + "@" + submodel_name

298

299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321
# Define a recurrent layer group begin with RecurrentLayerGroupBegin
# and end with RecurrentLayerGroupEnd.
# A recurrent layer group forward/backward one frame after previous frame
# forward/backward through all layers in layer group.
# in_links are names of layer used as input layer in the layer group.
# out_links are names of layer in layer group used as outside layer's input.
#
# If generator is set, the layer group need one or more than one outlinks.
# The first outlink should always be the generated token ids.
# If generator.num_results_per_sample is not set, the output for one sample is
# a ids sequence. Else if num_results_per_sample is more than one,
# the output for one sample is up to #num_results_per_sample generated
# sequences, which are packed in one sequence in output ids vector. Each
# generated sequence has a generation probability. The probabilities for one
# sample are stored in one row of output value matrix.
# Packed generated sequences format, for each i:
#   seq_i_length: one interger, seq_i content length,
#   [seq_i content], length = seq_i_length
#   seq_i_end_mark: one interger, for format check, always -1
# You can use "seq_text_printer" to print the output of the generator.
@config_func
def RecurrentLayerGroupWithoutOutLinksBegin(name,
                                            in_links,
322 323
                                            seq_reversed=False,
                                            target_inlinkname=""):
324 325 326 327 328 329 330
    global g_current_submodel
    config_assert(g_config.model_config.type == "recurrent_nn",
                  "RecurrentLayerGroup should be used only in recurrent_nn")
    RecurrentLayerGroup(name=name)  # add to father model
    SubModelBegin(name)
    g_current_submodel.is_recurrent_layer_group = True
    g_current_submodel.reversed = seq_reversed
331
    g_current_submodel.target_inlinkid = -1
332
    in_links_count = 0
333
    for linkid, link in enumerate(in_links):
334 335 336 337 338 339
        if isinstance(link, basestring):
            name = link
            has_subseq = False
        else:
            name = link.link_name
            has_subseq = link.has_subseq
340 341 342 343
        # assign target_inlinkid according to target_inlinkname
        if target_inlinkname == name:
            g_current_submodel.target_inlinkid = linkid

344 345 346
        if in_links_count == 0:
            in_links_has_subseq = has_subseq
        else:
347 348 349 350
            config_assert(
                in_links_has_subseq == has_subseq,
                "The sequence type of in_links should be the same in RecurrentLayerGroup"
            )
351 352 353 354 355 356 357
        in_links_count += 1
        layer_name = MakeLayerNameInParentSubmodel(name)
        layer = g_layer_map[layer_name]
        if has_subseq:
            SequenceScatterAgentLayer(name=name, size=layer.size)
        else:
            ScatterAgentLayer(name=name, size=layer.size)
358

359 360 361 362 363
        pair = g_current_submodel.in_links.add()
        pair.layer_name = layer_name
        pair.link_name = MakeLayerNameInSubmodel(name)
        pair.has_subseq = has_subseq

364

365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380
@config_func
def RecurrentLayerGroupSetOutLink(link):
    if isinstance(link, basestring):
        name = link
        has_subseq = False
    else:
        name = link.link_name
        has_subseq = link.has_subseq
    layer_name = MakeLayerNameInParentSubmodel(name)
    pair = g_current_submodel.out_links.add()
    pair.layer_name = MakeLayerNameInSubmodel(name)
    pair.link_name = layer_name
    pair.has_subseq = has_subseq


def RecurrentLayerGroupSetGenerator(generator=None):
381
    generator.eos_layer_name = MakeLayerNameInSubmodel(generator.eos_layer_name)
382 383 384 385 386 387 388 389
    g_current_submodel.generator.CopyFrom(generator)


@config_func
def RecurrentLayerGroupBegin(name,
                             in_links,
                             out_links,
                             generator=None,
390
                             target_inlinkname="",
391
                             seq_reversed=False):
392
    RecurrentLayerGroupWithoutOutLinksBegin(name, in_links, seq_reversed,
393
                                            target_inlinkname)
394 395 396 397 398
    for link in out_links:
        RecurrentLayerGroupSetOutLink(link)

    if generator is not None:
        RecurrentLayerGroupSetGenerator(generator)
399 400 401 402 403
        config_assert(
            len(in_links) == 0, "no in_links should be passed to generator")
        config_assert(
            len(out_links) >= 1,
            "one or more than one out_links should be passed to generator")
404 405 406 407 408 409 410


@config_func
def RecurrentLayerGroupEnd(name):
    global g_current_submodel
    config_assert(g_current_submodel.is_recurrent_layer_group,
                  "RecurrentLayerGroup not begin")
411
    for pair in g_current_submodel.memories:  #check exist
412
        layer = g_layer_map[pair.layer_name]
Y
Yu Yang 已提交
413 414
        config_assert(layer is not None,
                      "memory declare wrong name:%s" % pair.layer_name)
415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432
        memory_link = g_layer_map[pair.link_name]
        config_assert(layer.size == memory_link.size,
                      "memory declare wrong size:%d" % memory_link.size)

    prev_submodel = g_current_submodel
    SubModelEnd(name)

    for pair in prev_submodel.out_links:
        layer = g_layer_map[pair.layer_name]
        # add out agent to father model
        agent_name = GetLayerBaseName(pair.link_name)
        if prev_submodel.HasField("generator"):
            DataLayer(name=agent_name, size=layer.size)
        elif pair.has_subseq:
            SequenceGatherAgentLayer(name=agent_name, size=layer.size)
        else:
            GatherAgentLayer(name=agent_name, size=layer.size)

433

434 435 436 437 438 439
# Define the model type
# currently, the paddle supports "nn", "recurrent_nn", "recursive_nn" and "multi_nn"
@config_func
def model_type(name):
    g_config.model_config.type = name

440

441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457
@config_class
class Bias(Cfg):
    def __init__(
            self,
            parameter_name=None,
            learning_rate=None,
            momentum=None,
            decay_rate=None,
            decay_rate_l1=None,
            initial_mean=None,
            initial_std=None,
            initial_strategy=None,
            initial_smart=None,
            num_batches_regularization=None,
            sparse_remote_update=None,
            gradient_clipping_threshold=None,
            is_static=None,
458
            is_shared=None, ):
459 460
        self.add_keys(locals())

461

462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481
# Define one input for a layer
@config_class
class Input(Cfg):
    def __init__(
            self,
            input_layer_name,
            parameter_name=None,
            learning_rate=None,
            momentum=None,
            decay_rate=None,
            decay_rate_l1=None,
            initial_mean=None,
            initial_std=None,
            initial_strategy=None,
            initial_smart=None,
            num_batches_regularization=None,
            sparse_remote_update=None,
            sparse_update=None,
            gradient_clipping_threshold=None,
            conv=None,
482
            bilinear_interp=None,
483 484 485 486
            norm=None,
            pool=None,
            image=None,
            block_expand=None,
487
            maxout=None,
Q
qijun 已提交
488
            spp=None,
489 490 491 492 493
            format=None,
            nnz=None,
            is_static=None,
            is_shared=None,
            update_hooks=None,
494
            input_layer_argument=None, ):
495 496 497
        self.add_keys(locals())
        self.input_layer_name = MakeLayerNameInSubmodel(input_layer_name)

498

499 500 501
# Define a projection for iexed layer
@config_class
class Projection(Input):
502 503
    type = None  # subclass should set it correctly

504 505 506
    def __init__(
            self,
            input_layer_name,
507
            size=0,  # projection output size
508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526
            parameter_name=None,
            learning_rate=None,
            momentum=None,
            decay_rate=None,
            decay_rate_l1=None,
            initial_mean=None,
            initial_std=None,
            initial_strategy=None,
            initial_smart=None,
            num_batches_regularization=None,
            sparse_remote_update=None,
            sparse_update=None,
            gradient_clipping_threshold=None,
            ptype=None,
            format=None,
            nnz=None,
            is_static=None,
            is_shared=None,
            update_hooks=None,
527
            input_layer_argument=None, ):
528 529 530 531 532 533 534 535 536 537 538 539 540
        self.add_keys(locals())
        self.input_layer_name = MakeLayerNameInSubmodel(input_layer_name)

        self.proj_conf = ProjectionConfig()
        if ptype is not None:
            self.proj_conf.type = ptype
        else:
            self.proj_conf.type = self.type

    # calculate the output_size given input_size. return 0
    # to indicate using the size from Layer config
    def calc_output_size(self, input_layer_config):
        return self.size
541

542 543
    def calc_parameter_size(self, input_size, output_size):
        raise NotimplementedError
544

545 546 547 548 549 550 551 552 553 554
    def calc_parameter_dims(self, input_size, output_size):
        raise NotimplementedError


@config_class
class IdentityProjection(Projection):
    type = 'identity'

    def calc_output_size(self, input_layer_config):
        return input_layer_config.size
555

556 557
    def calc_parameter_size(self, input_size, output_size):
        return 0
558

559 560 561
    def calc_parameter_dims(self, input_size, output_size):
        return []

562

563 564 565 566 567 568
# Like IdentityProjection, but layer size may smaller than input size,
# the projection select dimesions [offset, offset+layer_size) from input
@config_class
class IdentityOffsetProjection(Projection):
    type = 'identity_offset'

569 570 571
    def __init__(self, input_layer_name, offset, **xargs):
        super(IdentityOffsetProjection, self).__init__(input_layer_name,
                                                       **xargs)
572 573 574 575
        self.proj_conf.offset = offset

    def calc_parameter_size(self, input_size, output_size):
        return 0
576

577 578 579
    def calc_parameter_dims(self, input_size, output_size):
        return []

580

581 582 583 584 585 586 587
# DotMulProjection performs element-wise multiplication with weight
@config_class
class DotMulProjection(Projection):
    type = 'dot_mul'

    def calc_output_size(self, input_layer_config):
        return input_layer_config.size
588

589 590
    def calc_parameter_size(self, input_size, output_size):
        return output_size
591

592 593 594
    def calc_parameter_dims(self, input_size, output_size):
        return [1, output_size]

D
dangqingqing 已提交
595

X
xuwei06 已提交
596 597 598 599 600 601 602 603 604 605 606 607 608 609
# ScalingProjection
@config_class
class ScalingProjection(Projection):
    type = 'scaling'

    def calc_output_size(self, input_layer_config):
        return input_layer_config.size

    def calc_parameter_size(self, input_size, output_size):
        return 1

    def calc_parameter_dims(self, input_size, output_size):
        return [1, 1]

610

611 612 613 614 615 616
@config_class
class TableProjection(Projection):
    type = 'table'

    def calc_parameter_size(self, input_size, output_size):
        return input_size * output_size
617

618 619 620
    def calc_parameter_dims(self, input_size, output_size):
        return [input_size, output_size]

621

622 623 624 625 626 627
@config_class
class FullMatrixProjection(Projection):
    type = 'fc'

    def calc_parameter_size(self, input_size, output_size):
        return input_size * output_size
628

629 630 631
    def calc_parameter_dims(self, input_size, output_size):
        return [input_size, output_size]

632

633 634 635 636 637 638
@config_class
class TransposedFullMatrixProjection(Projection):
    type = 'trans_fc'

    def calc_parameter_size(self, input_size, output_size):
        return input_size * output_size
639

640 641 642
    def calc_parameter_dims(self, input_size, output_size):
        return [output_size, input_size]

643

644 645 646 647
@config_class
class ContextProjection(Projection):
    type = 'context'

648 649
    def __init__(self, input_layer_name, context_start, context_length,
                 trainable_padding, **xargs):
650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672
        super(ContextProjection, self).__init__(input_layer_name, **xargs)
        self.proj_conf.context_start = context_start
        self.proj_conf.context_length = context_length
        self.proj_conf.trainable_padding = trainable_padding
        self._total_pad = max(0, -self.proj_conf.context_start) \
                          + max(0, self.proj_conf.context_start \
                                + self.proj_conf.context_length - 1)

    def calc_output_size(self, input_layer_config):
        return input_layer_config.size * self.proj_conf.context_length

    def calc_parameter_size(self, input_size, output_size):
        if self.proj_conf.trainable_padding == False:
            return 0
        else:
            return input_size * self._total_pad

    def calc_parameter_dims(self, input_size, output_size):
        return [self._total_pad, input_size]

    _total_pad = 0


673 674 675 676
@config_class
class ConvProjection(Projection):
    type = 'conv'

677 678 679 680 681
    def __init__(self,
                 input_layer_name,
                 num_filters=None,
                 conv_conf=None,
                 **xargs):
682 683 684 685 686
        super(ConvProjection, self).__init__(input_layer_name, **xargs)

        if num_filters is not None:
            self.proj_conf.num_filters = num_filters

687
        parse_conv(conv_conf, input_layer_name, self.proj_conf.conv_conf,
688
                   num_filters)
689
        # TODO: support rectangle input
Y
Yu Yang 已提交
690 691
        self.proj_conf.output_size = (self.proj_conf.conv_conf.output_x
                                      **2) * num_filters
692 693 694 695 696 697 698 699 700

    def calc_output_size(self, input_layer_config):
        return self.proj_conf.output_size

    def calc_parameter_size(self, input_size, output_size):
        co = self.proj_conf.num_filters
        ci = self.proj_conf.conv_conf.channels
        fh = self.proj_conf.conv_conf.filter_size
        fw = self.proj_conf.conv_conf.filter_size_y
701 702
        gr = self.proj_conf.conv_conf.groups
        return co * ci * fh * fw / gr
703 704 705 706 707 708 709

    def calc_bias_size(self):
        return self.proj_conf.num_filters

    def calc_parameter_dims(self, input_size, output_size):
        return None

710

711 712 713
# Define a operator for mixed layer
@config_class
class Operator(Cfg):
714 715
    type = None  # subclass should set it correctly

716 717
    def __init__(
            self,
718
            input_layer_names, ):
719 720 721 722 723 724 725 726 727 728
        self.add_keys(locals())
        self.operator_conf = OperatorConfig()
        self.operator_conf.type = self.type

    def check_dims(self):
        pass

    def calc_output_size(self, input_sizes):
        return 0

729

730 731 732
@config_class
class DotMulOperator(Operator):
    type = 'dot_mul'
733 734 735

    def __init__(self, input_layer_names, scale=None, **xargs):
        super(DotMulOperator, self).__init__(input_layer_names, **xargs)
736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753
        if scale is not None:
            self.operator_conf.dotmul_scale = scale

        config_assert(len(input_layer_names) == 2, "DotMul is binary operator")

    def check_dims(self):
        for i in range(2):
            config_assert(self.operator_conf.input_sizes[i] ==
                          self.operator_conf.output_size,
                          "DotMul input_size != output_size")

    def calc_output_size(self, input_sizes):
        return input_sizes[0]


@config_class
class ConvOperator(Operator):
    type = 'conv'
754 755 756 757 758 759 760

    def __init__(self,
                 input_layer_names,
                 num_filters=None,
                 conv_conf=None,
                 **xargs):
        super(ConvOperator, self).__init__(input_layer_names, **xargs)
761 762 763
        if num_filters is not None:
            self.operator_conf.num_filters = num_filters

764 765
        parse_conv(conv_conf,
                   MakeLayerNameInSubmodel(input_layer_names[0]),
766 767 768
                   self.operator_conf.conv_conf, num_filters)
        self.operator_conf.output_size = (self.operator_conf.conv_conf.output_x
                                          **2) * num_filters
769 770 771

        config_assert(len(input_layer_names) == 2, "Conv is binary operator")

772 773
    def calc_output_size(self, input_sizes):
        return self.operator_conf.output_size
774 775 776 777 778


# please refer to the comments in proto/ModelConfig.proto
@config_class
class Conv(Cfg):
779 780 781 782 783 784 785 786 787 788 789 790 791
    def __init__(self,
                 filter_size,
                 channels,
                 padding=None,
                 stride=None,
                 groups=None,
                 filter_channels=None,
                 output_x=None,
                 img_size=None,
                 caffe_mode=True,
                 filter_size_y=None,
                 padding_y=None,
                 stride_y=None):
792 793
        self.add_keys(locals())
        if filter_size_y is None:
794
            self.filter_size_y = filter_size
795
        if padding_y is None:
796
            self.padding_y = padding
797
        if stride_y is None:
798
            self.stride_y = stride
799
        if output_x is not None:
800 801
            config_assert(output_x <= 0)

802

803 804 805
# please refer to the comments in proto/ModelConfig.proto
@config_class
class BilinearInterp(Cfg):
806
    def __init__(self, out_size_x=None, out_size_y=None, num_channels=None):
807 808
        self.add_keys(locals())

809

810 811 812
# please refer to the comments in proto/ModelConfig.proto
@config_class
class Pool(Cfg):
D
dangqingqing 已提交
813 814 815 816 817 818 819 820 821 822 823 824
    def __init__(
            self,
            pool_type,
            channels,
            size_x,
            size_y=None,
            img_width=None,
            start=None,
            stride=None,  # 1 by defalut in protobuf
            stride_y=None,
            padding=None,  # 0 by defalut in protobuf
            padding_y=None):
825
        self.add_keys(locals())
826 827


Q
qijun 已提交
828 829
# please refer to the comments in proto/ModelConfig.proto
@config_class
Q
qijun 已提交
830
class SpatialPyramidPool(Cfg):
831
    def __init__(self, pool_type, pyramid_height, channels, img_width=None):
Q
qijun 已提交
832
        self.add_keys(locals())
833

834

835 836 837
# please refer to the comments in proto/ModelConfig.proto
@config_class
class Norm(Cfg):
838 839 840 841 842 843 844 845 846
    def __init__(self,
                 norm_type,
                 channels,
                 size,
                 scale,
                 pow,
                 output_x=None,
                 img_size=None,
                 blocked=None):
847 848
        self.add_keys(locals())

849

850 851 852
# please refer to the comments in proto/ModelConfig.proto
@config_class
class Image(Cfg):
853
    def __init__(self, channels, img_size=None):
854 855
        self.add_keys(locals())

856

857 858
@config_class
class BlockExpand(Cfg):
859 860 861 862 863 864 865 866 867 868 869 870
    def __init__(self,
                 channels,
                 padding_x=0,
                 padding_y=0,
                 stride_x=0,
                 stride_y=0,
                 block_x=0,
                 block_y=0,
                 img_size_x=0,
                 img_size_y=0,
                 output_x=0,
                 output_y=0):
871 872
        self.add_keys(locals())

873

874 875
@config_class
class MaxOut(Cfg):
876
    def __init__(self, channels, groups, img_size_x=0, img_size_y=0):
877 878
        self.add_keys(locals())

879

880 881 882 883 884 885 886 887 888 889 890 891 892
def DataBase(async_load_data=False,
             constant_slots=None,
             data_ratio=1,
             is_main_data=True,
             usage_ratio=None):
    # default: all sub dataproviders are treat as "main data".
    # see proto/DataConfig.proto for is_main_data
    data_config = DataConfig()

    data_config.async_load_data = async_load_data

    if constant_slots:
        data_config.constant_slots.extend(constant_slots)
893 894
    data_config.data_ratio = data_ratio
    data_config.is_main_data = is_main_data
895

896
    usage_ratio = default(usage_ratio, settings_deprecated["usage_ratio"])
897 898 899 900 901 902
    config_assert(usage_ratio >= 0 and usage_ratio <= 1,
                  "The range of usage_ratio is [0, 1]")
    data_config.usage_ratio = usage_ratio

    return data_config

903

904
@config_func
905 906 907 908 909
def SimpleData(files=None,
               feat_dim=None,
               context_len=None,
               buffer_capacity=None,
               **xargs):
910 911 912 913 914 915 916 917 918 919
    data_config = DataBase(**xargs)
    data_config.type = 'simple'
    data_config.files = files
    data_config.feat_dim = feat_dim
    if context_len is not None:
        data_config.context_len = context_len
    if buffer_capacity:
        data_config.buffer_capacity = buffer_capacity
    return data_config

920

921
@config_func
922 923 924 925 926 927 928 929 930 931
def PyData(files=None,
           type=None,
           file_group_queue_capacity=None,
           load_data_module=None,
           load_data_object=None,
           load_data_args="",
           load_file_count=None,
           constant_slots=None,
           load_thread_num=None,
           **xargs):
932 933 934
    data_config = DataBase(**xargs)
    data_config.type = 'py'
    if load_data_module in g_py_module_name_list:
935

936 937 938
        def get_path(module):
            m = __import__(load_data_module)
            return os.path.split(os.path.realpath(m.__file__))[0]
939

940 941 942
        # python C-api is not thread safe, one module can only be import once,
        # so here we nedd to copy the module with different names if it has to be
        # imported several times.
943 944
        module_new_name = "%s_copy_%d" % (load_data_module,
                                          len(g_py_module_name_list))
945
        g_py_module_name_list.append(module_new_name)
946 947 948 949
        module_path = "%s/%s.py" % (get_path(load_data_module),
                                    load_data_module)
        new_module_path = "%s/%s.py" % (get_path(load_data_module),
                                        module_new_name)
950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973
        if os.path.isfile(module_path) == False:
            raise Exception("File %s is not exist." % module_path)
        shutil.copy2(module_path, new_module_path)
        load_data_module = module_new_name
    else:
        g_py_module_name_list.append(load_data_module)
    if load_data_module is not None and load_data_object is not None:
        data_config.load_data_module = load_data_module
        data_config.load_data_object = load_data_object
    else:
        raise ValueError('load_data_module, load_data_object is not defined.')
    data_config.load_data_args = load_data_args

    data_config.files = files or ''
    if file_group_queue_capacity is not None:
        data_config.file_group_conf.queue_capacity = file_group_queue_capacity
    if load_file_count is not None:
        data_config.file_group_conf.load_file_count = load_file_count
    if load_thread_num is not None:
        data_config.file_group_conf.load_thread_num = load_thread_num
    if constant_slots:
        data_config.constant_slots.extend(constant_slots)
    return data_config

974

975
@config_func
976 977 978 979 980 981 982
def ProtoData(files=None,
              type=None,
              file_group_queue_capacity=None,
              load_file_count=None,
              constant_slots=None,
              load_thread_num=None,
              **xargs):
983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002
    data_config = DataBase(**xargs)
    if type is None:
        data_config.type = 'proto'
    else:
        data_config.type = type
    data_config.files = files

    # When type="proto_group", one data provider contains at most
    # load_file_count files, and there are at most
    # (queue_capacity + load_thread_num + 1) data providers in memory
    if file_group_queue_capacity is not None:
        data_config.file_group_conf.queue_capacity = file_group_queue_capacity
    if load_file_count is not None:
        data_config.file_group_conf.load_file_count = load_file_count
    if load_thread_num is not None:
        data_config.file_group_conf.load_thread_num = load_thread_num
    if constant_slots:
        data_config.constant_slots.extend(constant_slots)
    return data_config

1003

1004 1005
#real data for training is actually provided by "sub_data" data providers.
@config_func
1006
def MultiData(sub_data=[]):
1007 1008 1009 1010 1011
    data_config = DataConfig()
    data_config.type = 'multi'
    data_config.sub_data_configs.extend(sub_data)
    return data_config

1012

1013
@config_func
1014 1015 1016 1017 1018 1019 1020
def Data(type,
         files=None,
         feat_dim=None,
         slot_dims=None,
         context_len=None,
         buffer_capacity=None,
         **xargs):
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

    data_config = DataBase(**xargs)
    data_config.type = type
    data_config.files = files
    data_config.feat_dim = feat_dim
    data_config.slot_dims.extend(slot_dims)
    if context_len is not None:
        data_config.context_len = context_len
    data_config.buffer_capacity = buffer_capacity
    return data_config


@config_func
def TrainData(data_config, async_load_data=None):
    config_assert(not g_config.HasField('data_config'),
                  'Only one TrainData definition is allowed')
    g_config.data_config.CopyFrom(data_config)
    g_config.data_config.for_test = False
    if async_load_data is not None:
        logger.warning("Deprecated: async_load_data should be used inside"
                       " Data definition")
        g_config.data_config.async_load_data = async_load_data


@config_func
def TestData(data_config, async_load_data=None):
    config_assert(not g_config.HasField('test_data_config'),
                  'Only one TestData definition is allowed')
    g_config.test_data_config.CopyFrom(data_config)
    g_config.test_data_config.for_test = True
    if async_load_data is not None:
        logger.warning("Deprecated: async_load_data should be used inside"
                       " Data definition")
        g_config.test_data_config.async_load_data = async_load_data

1056

1057
def parse_bilinear(bilinear, input_layer_name, bilinear_conf):
1058 1059 1060 1061
    bilinear_conf.out_size_x = bilinear.out_size_x
    bilinear_conf.out_size_y = bilinear.out_size_y
    bilinear_conf.num_channels = bilinear.num_channels

1062

1063 1064 1065 1066
'''
caffe_mode: compute the output size using floor instead of ceil,
            which is consistent of caffe and CuDNN's convention.
'''
1067 1068


1069 1070 1071 1072 1073 1074 1075
def cnn_output_size(img_size, filter_size, padding, stride, caffe_mode):
    output = (2 * padding + img_size - filter_size) / float(stride)
    if caffe_mode:
        return 1 + int(math.floor(output))
    else:
        return 1 + int(math.ceil(output))

1076

1077 1078 1079 1080
'''
calcualte image_size based on output_size for convolution. 
It is the reverse function of cnn_output_size
'''
1081 1082


1083 1084 1085 1086
def cnn_image_size(output_size, filter_size, padding, stride, caffe_mode):
    if caffe_mode:
        img_size = (output_size - 1) * stride + filter_size - 2 * padding
    else:
1087
        img_size = (output_size - 2) * stride + filter_size - 2 * padding + 1
1088 1089
    return img_size

1090

1091 1092
def parse_pool(pool, input_layer_name, pool_conf):
    pool_conf.pool_type = pool.pool_type
1093 1094 1095
    config_assert(pool.pool_type in [
        'max-projection', 'avg-projection', 'cudnn-max-pool', 'cudnn-avg-pool'
    ], "pool-type %s is not in "
1096
                  "['max-projection', 'avg-projection', "
1097
                  "'cudnn-max-pool', 'cudnn-avg-pool']" % pool.pool_type)
1098 1099 1100 1101 1102 1103

    pool_conf.channels = pool.channels
    pool_conf.size_x = pool.size_x
    pool_conf.stride = pool.stride

    pool_conf.size_y = default(pool.size_y, pool_conf.size_x)
1104
    pool_conf.stride_y = default(pool.stride_y, pool_conf.stride)
1105 1106

    img_pixels = g_layer_map[input_layer_name].size / pool.channels
1107 1108
    # the img_width may be removed,
    # and it can be calculated automatically later.
1109
    pool_conf.img_size = default(pool.img_width, int(img_pixels**0.5))
1110 1111
    pool_conf.img_size_y = img_pixels / pool_conf.img_size
    config_assert(pool_conf.img_size * pool_conf.img_size_y == img_pixels,
1112 1113
                  "Incorrect input image size %d for input image pixels %d" %
                  (pool_conf.img_size, img_pixels))
1114

1115
    config_assert(not pool.start, "start is deprecated in pooling.")
1116

1117
    if pool.padding is not None:
D
dangqingqing 已提交
1118
        pool_conf.padding = pool.padding
1119
    pool_conf.padding_y = default(pool.padding_y, pool_conf.padding)
D
dangqingqing 已提交
1120 1121 1122 1123 1124 1125
    pool_conf.output_x = cnn_output_size(pool_conf.img_size, pool_conf.size_x,
                                         pool_conf.padding, pool_conf.stride,
                                         False)
    pool_conf.output_y = cnn_output_size(pool_conf.img_size_y, pool_conf.size_y,
                                         pool_conf.padding_y,
                                         pool_conf.stride_y, False)
1126

1127

Q
qijun 已提交
1128 1129 1130
def parse_spp(spp, input_layer_name, spp_conf):
    spp_conf.pool_type = spp.pool_type
    config_assert(spp.pool_type in ['max-projection', 'avg-projection'],
1131 1132
                  "pool-type %s is not in "
                  "['max-projection', 'avg-projection']" % spp.pool_type)
Q
qijun 已提交
1133 1134 1135 1136 1137
    spp_conf.pyramid_height = spp.pyramid_height
    spp_conf.channels = spp.channels

    img_pixels = g_layer_map[input_layer_name].size / spp_conf.channels

1138
    spp_conf.img_size = default(spp.img_width, int(img_pixels**0.5))
Q
qijun 已提交
1139 1140
    spp_conf.img_size_y = img_pixels / spp_conf.img_size
    config_assert(spp_conf.img_size * spp_conf.img_size_y == img_pixels,
1141 1142 1143
                  "Incorrect input image size %d for input image pixels %d" %
                  (spp_conf.img_size, img_pixels))

Q
qijun 已提交
1144

1145 1146 1147
def parse_image(image, input_layer_name, image_conf):
    image_conf.channels = image.channels
    image_pixels = g_layer_map[input_layer_name].size / image_conf.channels
1148 1149 1150 1151 1152
    image_conf.img_size = int(image_pixels**0.5)
    config_assert((image_conf.img_size**2) == image_pixels,
                  "Incorrect input image size %d for input image pixels %d" %
                  (image_conf.img_size, image_pixels))

1153 1154 1155 1156

def parse_norm(norm, input_layer_name, norm_conf):
    norm_conf.norm_type = norm.norm_type
    config_assert(norm.norm_type in ['rnorm', 'cmrnorm-projection'],
1157 1158
                  "norm-type %s is not in [rnorm, 'cmrnorm-projection']" %
                  norm.norm_type)
1159 1160 1161 1162 1163 1164 1165
    norm_conf.channels = norm.channels
    norm_conf.size = norm.size
    norm_conf.scale = norm.scale
    norm_conf.pow = norm.pow
    norm_conf.blocked = norm.blocked

    img_pixels = g_layer_map[input_layer_name].size / norm.channels
1166 1167 1168 1169
    norm_conf.img_size = int(img_pixels**0.5)
    config_assert((norm_conf.img_size**2) == img_pixels,
                  "Incorrect input image size %d for input image pixels %d" %
                  (norm_conf.img_size, img_pixels))
1170 1171 1172 1173
    norm_conf.output_x = norm_conf.img_size
    if norm.norm_type in ['cmrnorm-projection']:
        norm_conf.scale /= norm.size
    else:
1174 1175
        norm_conf.scale /= norm.size**2

1176

1177 1178 1179 1180
'''
caffe_mode: compute the output size using floor instead of ceil,
            which is consistent of caffe and CuDNN's convention.
'''
1181 1182


1183
def parse_conv(conv, input_layer_name, conv_conf, num_filters, trans=False):
1184 1185 1186 1187 1188 1189 1190 1191 1192
    conv_conf.filter_size = conv.filter_size
    conv_conf.filter_size_y = conv.filter_size_y
    conv_conf.channels = conv.channels
    conv_conf.padding = conv.padding
    conv_conf.padding_y = conv.padding_y
    conv_conf.stride = conv.stride
    conv_conf.stride_y = conv.stride_y
    conv_conf.groups = conv.groups
    conv_conf.caffe_mode = conv.caffe_mode
1193

1194
    if not trans:
1195 1196
        conv_conf.filter_channels = conv.channels / conv.groups

1197
        img_pixels = g_layer_map[input_layer_name].size / conv.channels
1198 1199 1200 1201 1202 1203 1204 1205
        print('channels=%d size=%d' % (conv.channels,
                                       g_layer_map[input_layer_name].size))
        conv_conf.img_size = int(img_pixels**0.5)
        config_assert((conv_conf.img_size**2) == img_pixels, (
            "Input layer %s: Incorrect input image size %d for input " +
            "image pixels %d") %
                      (input_layer_name, conv_conf.img_size, img_pixels))

1206
        conv_conf.output_x = cnn_output_size(
1207 1208
            conv_conf.img_size, conv_conf.filter_size, conv_conf.padding,
            conv_conf.stride, conv_conf.caffe_mode)
1209
    else:
1210
        conv_conf.filter_channels = num_filters / conv.groups
1211

1212
        outputSize = g_layer_map[input_layer_name].size / conv.channels
1213 1214 1215 1216 1217 1218 1219
        print('channels=%d size=%d' % (conv.channels,
                                       g_layer_map[input_layer_name].size))
        conv_conf.output_x = int(outputSize**0.5)
        config_assert((conv_conf.output_x**2) == outputSize, (
            "Input layer %s: Incorrect input image size %d for input " +
            "image pixels %d") %
                      (input_layer_name, conv_conf.output_x, outputSize))
1220
        conv_conf.img_size = cnn_image_size(
1221 1222 1223
            conv_conf.output_x, conv_conf.filter_size, conv_conf.padding,
            conv_conf.stride, conv_conf.caffe_mode)

1224

1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237
def parse_block_expand(block_expand, input_layer_name, block_expand_conf):
    block_expand_conf.channels = block_expand.channels
    block_expand_conf.stride_x = block_expand.stride_x
    block_expand_conf.stride_y = block_expand.stride_y
    block_expand_conf.padding_x = block_expand.padding_x
    block_expand_conf.padding_y = block_expand.padding_y
    block_expand_conf.block_x = block_expand.block_x
    block_expand_conf.block_y = block_expand.block_y
    block_expand_conf.img_size_x = block_expand.img_size_x
    block_expand_conf.img_size_y = block_expand.img_size_y
    if block_expand_conf.img_size_x == 0:
        block_expand_conf.output_x = 0
    else:
1238
        block_expand_conf.output_x = cnn_output_size(
1239
            block_expand.img_size_x, block_expand.block_x,
1240
            block_expand.padding_x, block_expand.stride_x, False)
1241 1242

    if block_expand_conf.img_size_y == 0:
1243
        block_expand_conf.output_y = 0
1244
    else:
1245
        block_expand_conf.output_y = cnn_output_size(
1246
            block_expand.img_size_y, block_expand.block_y,
1247
            block_expand.padding_y, block_expand.stride_y, False)
1248

1249

1250 1251 1252 1253 1254
def parse_maxout(maxout, input_layer_name, maxout_conf):
    maxout_conf.channels = maxout.channels
    maxout_conf.groups = maxout.groups
    maxout_conf.img_size_x = maxout.img_size_x
    maxout_conf.img_size_y = maxout.img_size_y
1255

1256

1257 1258 1259 1260 1261 1262
# Define an evaluator
@config_func
def Evaluator(
        name,
        type,
        inputs,
1263 1264 1265 1266 1267 1268 1269 1270
        chunk_scheme=None,
        num_chunk_types=None,
        classification_threshold=None,
        positive_label=None,
        dict_file=None,
        result_file=None,
        num_results=None,
        delimited=None, ):
1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284
    evaluator = g_config.model_config.evaluators.add()
    evaluator.type = type
    evaluator.name = MakeLayerNameInSubmodel(name)
    if type_of(inputs) == str:
        inputs = [inputs]

    evaluator.input_layers.extend(
        [MakeLayerNameInSubmodel(name) for name in inputs])

    if chunk_scheme is not None:
        evaluator.chunk_scheme = chunk_scheme
        evaluator.num_chunk_types = num_chunk_types
    g_current_submodel.evaluator_names.append(evaluator.name)

1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297
    if classification_threshold is not None:
        evaluator.classification_threshold = classification_threshold
    if positive_label is not None:
        evaluator.positive_label = positive_label
    if dict_file is not None:
        evaluator.dict_file = dict_file

    if result_file is not None:
        evaluator.result_file = result_file
    if num_results is not None:
        evaluator.num_results = num_results
    if delimited is not None:
        evaluator.delimited = delimited
1298

1299

1300 1301 1302 1303 1304
class LayerBase(object):
    def __init__(
            self,
            name,
            type,
1305
            size,  # size can be 0. In this case, subclass should set it.
1306 1307 1308 1309
            inputs,
            device=None,
            active_type="",
            drop_rate=0.,
1310
            coeff=None):
1311
        config_assert('@' not in name,
1312
                      "layer name: %s contain special character @" % name)
1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327
        global g_current_submodel
        name = MakeLayerNameInSubmodel(name)

        config_assert(name not in g_layer_map,
                      'Duplicated layer name: %s' % name)

        self.inputs = copy.deepcopy(inputs)
        self.operators = []

        if self.inputs is None:
            self.inputs = []
        elif type_of(self.inputs) != list:
            self.inputs = [self.inputs]

        self.config = g_config.model_config.layers.add()
1328
        assert isinstance(self.config, LayerConfig)
1329 1330 1331
        self.config.name = name
        self.config.type = type
        self.config.active_type = active_type
1332 1333
        if coeff is not None:
            self.config.coeff = float(coeff)
1334 1335 1336 1337 1338 1339 1340
        if size != 0:
            self.config.size = size
        if drop_rate != 0:
            self.config.drop_rate = drop_rate

        if device is not None:
            self.config.device = device
1341
        elif g_default_device is not None:
1342 1343 1344 1345 1346 1347 1348 1349 1350
            self.config.device = g_default_device

        for input_index in xrange(len(self.inputs)):
            input = self.inputs[input_index]
            input_config = None
            input_layer_name = ''
            if type_of(input) == str:
                input_layer_name = input
                input_config = Input(
1351 1352
                    input_layer_name=input,
                    parameter_name=gen_parameter_name(name, input_index))
1353 1354 1355 1356 1357 1358 1359 1360
                input_layer_name = input_config.input_layer_name
            elif isinstance(input, Input):
                input_layer_name = input.input_layer_name
                input_config = input
                if input_config.parameter_name is None:
                    input_config.parameter_name = \
                        gen_parameter_name(name, input_index)
            elif isinstance(input, Operator):
1361
                self.operators.append(input)
1362 1363 1364 1365
                input.operator_conf.input_indices.append(input_index)
                input_config = Input(input.input_layer_names[0])
                input_layer_name = input_config.input_layer_name
            else:
1366
                raise ValueError('Wrong type for inputs: %s' % type_of(input))
1367
            config_assert(input_layer_name in g_layer_map,
1368 1369
                          "Unknown input layer '%s' for layer %s" %
                          (input_layer_name, name))
1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386
            self.inputs[input_index] = input_config
            layer_input = self.config.inputs.add()
            layer_input.input_layer_name = input_config.input_layer_name
            if input_config.input_layer_argument is not None:
                layer_input.input_layer_argument = \
                    input_config.input_layer_argument

        g_layer_map[name] = self.config

        g_current_submodel.layer_names.append(self.config.name)

    def get_input_layer(self, input_index):
        return g_layer_map[self.config.inputs[input_index].input_layer_name]

    # will return the bias created if not *for_self*
    def create_bias_parameter(
            self,
1387
            bias,  # True/False or BiasCfg
1388
            size,
1389 1390 1391
            dims=None,
            for_self=True,  # whether create bias for layer self
    ):
1392 1393 1394 1395 1396 1397

        if size == 0:
            return
        if dims is None:
            dims = [1, size]

1398 1399 1400
        config_assert(
            type_of(bias) == bool or type_of(bias) == Bias,
            'Incorrect type for bias: %s' % type_of(bias))
1401 1402 1403 1404 1405 1406 1407 1408 1409

        if type_of(bias) == bool:
            if bias:
                bias = Bias()

        if type_of(bias) == Bias:
            if bias.parameter_name is None:
                bias.parameter_name = gen_bias_parameter_name(self.config.name)
            if bias.parameter_name not in g_parameter_map:
1410 1411
                assert isinstance(self.config, LayerConfig)

1412 1413 1414
                Parameter(
                    bias.parameter_name,
                    size,
1415 1416
                    self.config.device
                    if self.config.HasField('device') else None,
1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427
                    dims,
                    bias.learning_rate,
                    bias.momentum,
                    decay_rate=bias.decay_rate,
                    decay_rate_l1=bias.decay_rate_l1,
                    initial_mean=bias.initial_mean,
                    initial_std=bias.initial_std,
                    initial_strategy=bias.initial_strategy,
                    initial_smart=bias.initial_smart,
                    num_batches_regularization=bias.num_batches_regularization,
                    sparse_remote_update=bias.sparse_remote_update,
1428 1429
                    gradient_clipping_threshold=bias.
                    gradient_clipping_threshold,
1430
                    is_static=bias.is_static,
1431
                    is_shared=bias.is_shared, )
1432 1433 1434 1435 1436
            if for_self:
                self.config.bias_parameter_name = bias.parameter_name
            else:
                return bias.parameter_name

1437 1438 1439 1440 1441 1442
    def create_input_parameter(self,
                               input_index,
                               size,
                               dims=None,
                               sparse=None,
                               format=None):
1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456
        if dims is None:
            # TODO(yuyang18): print warning and callstack here!
            dims = list()

        if size == 0:
            return

        input_config = self.inputs[input_index]

        self.config.inputs[input_index].input_parameter_name = \
            input_config.parameter_name

        if input_config.parameter_name in g_parameter_map:
            para = g_parameter_map[input_config.parameter_name]
1457 1458
            config_assert(size == para.size, (
                'Shared parameter "%s" does not ' + 'have same size: %s vs. %s')
1459 1460
                          % (input_config.parameter_name, para.size, size))

1461 1462
            config_assert(dims == para.dims, (
                'Shared parameter "%s" does not ' + 'have same dims: %s vs. %s')
1463 1464 1465 1466 1467 1468
                          % (input_config.parameter_name, para.dims, dims))
            return

        Parameter(
            input_config.parameter_name,
            size,
1469
            self.config.device if self.config.HasField("device") else None,
1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481
            dims,
            input_config.learning_rate,
            input_config.momentum,
            decay_rate=input_config.decay_rate,
            decay_rate_l1=input_config.decay_rate_l1,
            initial_mean=input_config.initial_mean,
            initial_std=input_config.initial_std,
            initial_strategy=input_config.initial_strategy,
            initial_smart=input_config.initial_smart,
            num_batches_regularization=input_config.num_batches_regularization,
            sparse_remote_update=input_config.sparse_remote_update,
            sparse_update=input_config.sparse_update,
1482 1483
            gradient_clipping_threshold=input_config.
            gradient_clipping_threshold,
1484 1485 1486 1487
            sparse=sparse,
            format=format,
            is_static=input_config.is_static,
            is_shared=input_config.is_shared,
1488
            update_hooks=input_config.update_hooks)
1489 1490 1491 1492 1493 1494 1495 1496 1497

    def set_layer_size(self, size):
        if self.config.size == 0:
            self.config.size = size
        else:
            config_assert(self.config.size == size,
                          'Different inputs result in' +
                          'different layer size at layer %s' % self.config.name)

1498

1499 1500
@config_layer('multi_class_cross_entropy_with_selfnorm')
class MultiClassCrossEntropySelfNormCostLayer(LayerBase):
1501 1502 1503
    def __init__(self, name, inputs, softmax_selfnorm_alpha=0.1, **xargs):
        super(MultiClassCrossEntropySelfNormCostLayer, self).__init__(
            name, 'multi_class_cross_entropy_with_selfnorm', 0, inputs, **xargs)
1504 1505
        self.config.softmax_selfnorm_alpha = softmax_selfnorm_alpha

1506

1507 1508
@config_layer('fc')
class FCLayer(LayerBase):
1509
    def __init__(self, name, size, inputs, bias=True, **xargs):
1510 1511 1512 1513 1514 1515 1516 1517 1518 1519
        super(FCLayer, self).__init__(name, 'fc', size, inputs=inputs, **xargs)
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
            psize = self.config.size * input_layer.size
            dims = [input_layer.size, self.config.size]
            format = self.inputs[input_index].format
            sparse = format == "csr" or format == "csc"

            if sparse:
                psize = self.inputs[input_index].nnz
1520 1521
            else:
                sparse = None
1522

1523 1524
            self.create_input_parameter(input_index, psize, dims, sparse,
                                        format)
1525 1526
        self.create_bias_parameter(bias, self.config.size)

1527

1528 1529
@config_layer('selective_fc')
class SelectiveFCLayer(LayerBase):
1530 1531 1532 1533 1534 1535 1536 1537 1538 1539
    def __init__(self,
                 name,
                 size,
                 inputs,
                 bias=True,
                 selective_fc_pass_generation=False,
                 has_selected_colums=True,
                 selective_fc_full_mul_ratio=0.02,
                 selective_fc_parallel_plain_mul_thread_num=None,
                 **xargs):
1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559
        super(SelectiveFCLayer, self).__init__(
            name, 'selective_fc', size, inputs=inputs, **xargs)
        # user MUST know if selctive fc is used in training,
        # parameter matrices saved by this layer are automatically transposed,
        # BUT bias is not.

        # if selective_fc is used only in testing mode, and parameters for
        # this layer are trained by fully connected layers,
        # then TranposedFullMatrixProjectin MUST be used in training
        # to avoid manual transpose in testing.

        self.config.selective_fc_pass_generation = selective_fc_pass_generation
        self.config.has_selected_colums = has_selected_colums
        self.config.selective_fc_full_mul_ratio = selective_fc_full_mul_ratio
        if selective_fc_parallel_plain_mul_thread_num is not None:
            self.config.selective_fc_parallel_plain_mul_thread_num = selective_fc_parallel_plain_mul_thread_num

        input_num = len(self.inputs)
        if has_selected_colums:
            config_assert(input_num >= 2,
1560 1561
                          ("if indices of selected columns are not specified, "
                           "selective_fc Layer has at least two inputs"))
1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573
            input_num -= 1

        for input_index in xrange(input_num):
            input_layer = self.get_input_layer(input_index)
            psize = self.config.size * input_layer.size
            dims = [input_layer.size, self.config.size]
            dims = dims[::-1]  # transpose the parameter
            format = self.inputs[input_index].format
            sparse = format == "csr" or format == "csc"
            if sparse:
                psize = self.inputs[input_index].nnz

1574 1575
            self.create_input_parameter(input_index, psize, dims, sparse,
                                        format)
1576 1577
        self.create_bias_parameter(bias, self.config.size)

1578

1579 1580
@config_layer('print')
class PrintLayer(LayerBase):
1581
    def __init__(self, name, inputs):
1582 1583
        super(PrintLayer, self).__init__(name, 'print', 0, inputs)

1584

1585 1586
@config_layer('data')
class DataLayer(LayerBase):
1587 1588 1589 1590
    def __init__(self, name, size, device=None):
        super(DataLayer, self).__init__(
            name, 'data', size, inputs=[], device=device)

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

'''
DataNormLayer: A layer for data normalization
Input: One and only one input layer is accepted. The input layer must
       be DataLayer with dense data type
Output: The normalization of the input data

Reference:
    LA Shalabi, Z Shaaban, B Kasasbeh. Data mining: A preprocessing engine

Example:
    Layer(
        name = "norm_input_layer",
        type = "data_norm",
        inputs = [Input("input_layer",
                        parameter_name = "_slot0.stats")],
        data_norm_strategy = "z-score",
    )

Note:
  (1) The parameter has been calculated in the preprocessing stage,
      and should be initialized by --init_model_path when training.
  (2) Three data normalization methoeds are considered
          z-score: y = (x-mean)/std
          min-max: y = (x-min)/(max-min)
          decimal-scaling: y = x/10^j, where j is the smallest integer such that max(|y|)<1
'''
1618 1619


1620 1621
@config_layer('data_norm')
class DataNormLayer(LayerBase):
1622
    def __init__(self, name, inputs, data_norm_strategy="z-score", device=None):
1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633
        super(DataNormLayer, self).__init__(
            name, 'data_norm', 0, inputs=inputs, device=device)
        self.config.data_norm_strategy = data_norm_strategy
        config_assert(len(inputs) == 1, 'DataNormLayer must have 1 input')
        input_layer = self.get_input_layer(0)
        self.set_layer_size(input_layer.size)
        para_size = 5 * input_layer.size
        para_dims = [5, input_layer.size]
        self.inputs[0].is_static = True
        self.create_input_parameter(0, para_size, para_dims)

1634

1635 1636 1637
@config_layer('prelu')
class ParameterReluLayer(LayerBase):
    layer_type = 'prelu'
1638 1639

    def __init__(self, name, inputs, partial_sum=1, **args):
1640 1641 1642 1643 1644 1645 1646 1647
        super(ParameterReluLayer, self).__init__(
            name, self.layer_type, 0, inputs=inputs, **args)
        config_assert(len(self.inputs) == 1)
        config_assert(self.input_layer.size % partial_sum == 0)
        input_layer = self.get_input_layer(0)
        self.set_layer_size(input_layer.size)
        self.create_input_parameter(0, input_layer.size / partial_sum)

1648

1649 1650 1651
@config_layer('conv')
class ConvLayerBase(LayerBase):
    layer_type = 'conv'
1652 1653 1654 1655 1656 1657 1658 1659

    def __init__(self,
                 name,
                 inputs=[],
                 bias=True,
                 num_filters=None,
                 shared_biases=False,
                 **xargs):
1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675
        super(ConvLayerBase, self).__init__(
            name, self.layer_type, 0, inputs=inputs, **xargs)

        if num_filters is not None:
            self.config.num_filters = num_filters

        use_gpu = int(g_command_config_args.get("use_gpu", 0))
        parallel_nn = int(g_command_config_args.get("parallel_nn", 0))

        # Automatically select cudnn_type for GPU and exconv for CPU
        # if set type=conv, but still reserve the way user specify
        # exconv or cudnn_conv manually.
        if self.layer_type == "cudnn_conv":
            config_assert(use_gpu, "cudnn_conv only support GPU")

        if (use_gpu == 1 and self.layer_type != "exconv" and
1676
            (parallel_nn == 0 or self.config.device > -1)):
1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687
            self.layer_type = "cudnn_conv"
        else:
            self.layer_type = "exconv"
        # need to specify layer in config
        self.config.type = self.layer_type

        if shared_biases is not None:
            self.config.shared_biases = shared_biases

        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
1688 1689
            parse_conv(self.inputs[input_index].conv, input_layer.name,
                       self.config.inputs[input_index].conv_conf, num_filters)
1690 1691 1692 1693 1694
            conv_conf = self.config.inputs[input_index].conv_conf
            psize = self.calc_parameter_size(conv_conf)
            print("output size for %s is %d " % (name, conv_conf.output_x))
            self.create_input_parameter(input_index, psize)
            self.set_layer_size(
1695
                (conv_conf.output_x**2) * self.config.num_filters)
1696 1697 1698 1699 1700 1701 1702 1703 1704 1705

        psize = self.config.size
        if shared_biases:
            psize = self.config.num_filters
        self.create_bias_parameter(bias, psize, [psize, 1])

    def calc_parameter_size(self, conv_conf):
        return self.config.num_filters * conv_conf.filter_channels \
                    * (conv_conf.filter_size * conv_conf.filter_size_y)

1706

1707 1708 1709 1710
@config_layer('exconv')
class ConvLayer(ConvLayerBase):
    layer_type = 'exconv'

1711

1712 1713 1714 1715
@config_layer('cudnn_conv')
class ConvLayer(ConvLayerBase):
    layer_type = 'cudnn_conv'

1716 1717 1718 1719

@config_layer('convt')
class ConvTransLayerBase(LayerBase):
    layer_type = 'convt'
1720 1721 1722 1723 1724 1725 1726 1727

    def __init__(self,
                 name,
                 inputs=[],
                 bias=True,
                 num_filters=None,
                 shared_biases=False,
                 **xargs):
1728
        super(ConvTransLayerBase, self).__init__(
1729 1730 1731 1732 1733 1734 1735 1736
            name, self.layer_type, 0, inputs=inputs, **xargs)

        if num_filters is not None:
            self.config.num_filters = num_filters

        use_gpu = int(g_command_config_args.get("use_gpu", 0))
        parallel_nn = int(g_command_config_args.get("parallel_nn", 0))

1737 1738
        # cudnn_convt has not been implemented so use exconvt only
        self.layer_type = "exconvt"
1739 1740 1741 1742 1743 1744 1745 1746
        # need to specify layer in config
        self.config.type = self.layer_type

        if shared_biases is not None:
            self.config.shared_biases = shared_biases

        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
1747
            parse_conv(
1748 1749
                self.inputs[input_index].conv,
                input_layer.name,
1750
                self.config.inputs[input_index].conv_conf,
1751
                num_filters,
1752
                trans=True)
1753 1754 1755 1756 1757
            conv_conf = self.config.inputs[input_index].conv_conf
            psize = self.calc_parameter_size(conv_conf)
            print("output size for %s is %d " % (name, conv_conf.output_x))
            self.create_input_parameter(input_index, psize)
            self.set_layer_size(
1758
                (conv_conf.img_size**2) * self.config.num_filters)
1759 1760 1761 1762 1763 1764 1765

        psize = self.config.size
        if shared_biases:
            psize = self.config.num_filters
        self.create_bias_parameter(bias, psize, [psize, 1])

    def calc_parameter_size(self, conv_conf):
1766
        return conv_conf.channels * conv_conf.filter_channels \
1767 1768
                    * (conv_conf.filter_size * conv_conf.filter_size_y)

1769

1770 1771 1772 1773
@config_layer('exconvt')
class ConvTransLayer(ConvTransLayerBase):
    layer_type = 'exconvt'

1774

1775 1776
@config_layer('norm')
class NormLayer(LayerBase):
1777 1778 1779
    def __init__(self, name, inputs, device=None):
        super(NormLayer, self).__init__(
            name, 'norm', 0, inputs=inputs, device=device)
1780 1781
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
1782 1783
            parse_norm(self.inputs[input_index].norm, input_layer.name,
                       self.config.inputs[input_index].norm_conf)
1784
            norm_conf = self.config.inputs[input_index].norm_conf
1785 1786
            self.set_layer_size((norm_conf.output_x**2) * norm_conf.channels)

1787 1788 1789

@config_layer('pool')
class PoolLayer(LayerBase):
1790 1791 1792
    def __init__(self, name, inputs, device=None):
        super(PoolLayer, self).__init__(
            name, 'pool', 0, inputs=inputs, device=device)
1793 1794
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
1795 1796
            parse_pool(self.inputs[input_index].pool, input_layer.name,
                       self.config.inputs[input_index].pool_conf)
1797
            pool_conf = self.config.inputs[input_index].pool_conf
1798 1799 1800 1801 1802
            print("output size for %s is %d*%d " % (name, pool_conf.output_y,
                                                    pool_conf.output_x))
            self.set_layer_size(
                (pool_conf.output_x * pool_conf.output_y) * pool_conf.channels)

1803

Q
qijun 已提交
1804 1805
@config_layer('spp')
class SpatialPyramidPoolLayer(LayerBase):
1806 1807 1808
    def __init__(self, name, inputs, device=None):
        super(SpatialPyramidPoolLayer, self).__init__(
            name, 'spp', 0, inputs=inputs, device=device)
Q
qijun 已提交
1809 1810
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
1811 1812
            parse_spp(self.inputs[input_index].spp, input_layer.name,
                      self.config.inputs[input_index].spp_conf)
Q
qijun 已提交
1813 1814 1815 1816 1817
            spp_conf = self.config.inputs[input_index].spp_conf
            output_size = (pow(4, spp_conf.pyramid_height) - 1) / (4 - 1)
            print("output size for %s is %d " % (name, output_size))
            self.set_layer_size(output_size * spp_conf.channels)

1818

1819 1820 1821
@config_layer('batch_norm')
class BatchNormLayer(LayerBase):
    layer_type = 'batch_norm'
1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832

    def __init__(self,
                 name,
                 inputs,
                 active_type="linear",
                 bias=True,
                 device=None,
                 use_global_stats=True,
                 moving_average_fraction=0.9,
                 batch_norm_type=None,
                 **xargs):
1833 1834 1835 1836
        if inputs is None:
            inputs = []
        elif not isinstance(inputs, list):
            inputs = [inputs]
1837 1838
        config_assert(
            len(inputs) == 1, "BatchNormLayer must have one and only one input")
1839 1840 1841 1842 1843 1844 1845 1846
        # Create Input for moving mean and std,
        # in batch normalization layer.
        # These paras no need to update, so set is_static is true.
        # If not use is_static, even set learning_rate = 0, decay_rate = 0,
        # these paras will change if set average_window in configure.
        use_gpu = bool(int(g_command_config_args.get("use_gpu", 0)))
        is_shared = True if not use_gpu else False
        for i in xrange(2):
1847 1848 1849 1850 1851 1852 1853
            inputs.append(
                Input(
                    inputs[0].input_layer_name,
                    initial_std=0.0,
                    initial_mean=0.0,
                    is_static=True,
                    is_shared=is_shared, ))
1854 1855 1856 1857 1858 1859 1860

        parallel_nn = bool(int(g_command_config_args.get("parallel_nn", 0)))
        cudnn_version = int(g_command_config_args.get("cudnn_version", 0))
        # Automatically select cudnn_batch_norm for GPU and batch_norm for CPU.
        # Also based on cudnn version.
        use_cudnn = use_gpu and batch_norm_type != "batch_norm" and \
            ((not parallel_nn) or self.config.device > -1) and \
1861
            cudnn_version >= 4007
1862
        self.layer_type = "cudnn_batch_norm" if use_cudnn else "batch_norm"
1863 1864 1865 1866 1867 1868 1869 1870
        super(BatchNormLayer, self).__init__(
            name,
            self.layer_type,
            0,
            active_type=active_type,
            inputs=inputs,
            device=device,
            **xargs)
1871 1872 1873 1874 1875 1876

        if use_global_stats is not None:
            self.config.use_global_stats = use_global_stats
        if moving_average_fraction is not None:
            self.config.moving_average_fraction = moving_average_fraction

1877 1878
        input_layer = self.get_input_layer(0)
        parse_image(self.inputs[0].image, input_layer.name,
1879 1880
                    self.config.inputs[0].image_conf)
        image_conf = self.config.inputs[0].image_conf
1881
        self.set_layer_size((image_conf.img_size**2) * image_conf.channels)
1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893

        psize = self.calc_parameter_size(image_conf)
        dims = [1, psize]
        self.create_input_parameter(0, psize)
        self.create_input_parameter(1, psize, dims)
        self.create_input_parameter(2, psize, dims)

        self.create_bias_parameter(bias, psize)

    def calc_parameter_size(self, image_conf):
        return image_conf.channels

1894

1895 1896
@config_layer('trans')
class TransLayer(LayerBase):
1897 1898 1899 1900 1901 1902
    def __init__(self, name, inputs, device=None):
        super(TransLayer, self).__init__(
            name, 'trans', 0, inputs=inputs, device=device)
        config_assert(
            len(self.inputs) == 1,
            'TransLayer must have one and only one input')
1903 1904
        self.set_layer_size(self.get_input_layer(0).size)

1905

1906 1907
@config_layer('resize')
class ResizeLayer(LayerBase):
1908 1909 1910 1911 1912 1913 1914
    def __init__(self, name, size, inputs, device=None):
        super(ResizeLayer, self).__init__(
            name, 'resize', size=size, inputs=inputs, device=device)
        config_assert(
            len(self.inputs) == 1,
            'ResizeLayer must have one and only one input')

1915 1916 1917

@config_layer('blockexpand')
class BlockExpandLayer(LayerBase):
1918 1919 1920
    def __init__(self, name, inputs, device=None):
        super(BlockExpandLayer, self).__init__(
            name, 'blockexpand', 0, inputs=inputs, device=device)
1921 1922
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
1923 1924
            parse_block_expand(
                self.inputs[input_index].block_expand, input_layer.name,
1925
                self.config.inputs[input_index].block_expand_conf)
1926 1927 1928 1929 1930 1931
            block_expand_conf = self.config.inputs[
                input_index].block_expand_conf
            self.set_layer_size(block_expand_conf.block_x *
                                block_expand_conf.block_y *
                                block_expand_conf.channels)

1932

1933 1934
@config_layer('maxout')
class MaxOutLayer(LayerBase):
1935 1936 1937
    def __init__(self, name, inputs, **xargs):
        super(MaxOutLayer, self).__init__(
            name, 'maxout', 0, inputs=inputs, **xargs)
1938
        input_layer = self.get_input_layer(0)
1939
        parse_maxout(self.inputs[0].maxout, input_layer.name,
1940 1941
                     self.config.inputs[0].maxout_conf)
        maxout_conf = self.config.inputs[0].maxout_conf
1942 1943 1944
        self.set_layer_size(g_layer_map[input_layer.name].size /
                            maxout_conf.groups)

1945

1946 1947 1948 1949
# key: cost type
# value: cost class
g_cost_map = {}

1950

1951 1952 1953
# define a cost layer without any parameters
def define_cost(class_name, cost_type):
    def init(cls, name, inputs, device=None, coeff=1.):
1954 1955
        super(type(cls), cls).__init__(
            name, cost_type, 1, inputs, device=device, coeff=coeff)
1956

1957
    cls = type(class_name, (LayerBase, ), dict(__init__=init))
1958 1959 1960
    global g_cost_map
    g_cost_map[cost_type] = cls

1961

1962 1963 1964 1965 1966 1967 1968 1969
define_cost('MultiClassCrossEntropy', 'multi-class-cross-entropy')
define_cost('RankingCost', 'rank-cost')
define_cost('AucValidation', 'auc-validation')
define_cost('PnpairValidation', 'pnpair-validation')
define_cost('SumOfSquaresCostLayer', 'square_error')
define_cost('MultiBinaryLabelCrossEntropy', 'multi_binary_label_cross_entropy')
define_cost('SoftBinaryClassCrossEntropy', 'soft_binary_class_cross_entropy')
define_cost('HuberTwoClass', 'huber')
X
xuwei06 已提交
1970
define_cost('SumCost', 'sum_cost')
1971

1972

1973 1974
@config_layer('hsigmoid')
class HierarchicalSigmoidLayer(LayerBase):
1975
    def __init__(self, name, num_classes, inputs, device=None, bias=True):
1976 1977
        super(HierarchicalSigmoidLayer, self).__init__(
            name, 'hsigmoid', 1, inputs=inputs, device=device)
1978 1979 1980
        config_assert(
            len(self.inputs) >= 2,
            'HierarchicalSigmoidLayer must have at least 2 inputs')
1981 1982 1983 1984 1985 1986 1987 1988
        self.config.num_classes = num_classes
        for input_index in xrange(len(self.inputs) - 1):
            input_layer = self.get_input_layer(input_index)
            psize = (num_classes - 1) * input_layer.size
            dims = [num_classes - 1, input_layer.size]
            self.create_input_parameter(input_index, psize, dims)
        self.create_bias_parameter(bias, num_classes - 1)

1989

1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013
'''
lambdaCost for lambdaRank LTR approach

Usage:
  Example: Layer(name = "cost", type = "lambda_cost", NDCG_num = 8,
             max_sort_size = -1, inputs = ["output", "score"])

  Input data: Samples of the same query should be loaded as a sequence,
          by ProtoDataProvider or PyDataProvider etc.. User should provide
          scores for each sample. The score slot should be the 2nd
          input of lambdaRank layer.

  NDCG_num = the size of NDCG, e.g., 5 for NDCG@5.
    Note: NDCG_num must be less than or equal to the minimum
          size of lists.

  max_sort_size = the size of partial sorting in calculating gradient.
    Note: 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 equal
          to NDCG_num.
          max_sort_size can be greater than the size of a list, in which
          case the algorithm will sort the entire list to get gradient.
'''
2014 2015


2016 2017
@config_layer('lambda_cost')
class LambdaCost(LayerBase):
2018
    def __init__(self, name, inputs, NDCG_num=5, max_sort_size=-1, device=None):
2019 2020
        super(LambdaCost, self).__init__(
            name, 'lambda_cost', 1, inputs=inputs, device=device)
2021
        config_assert(len(self.inputs) == 2, 'lambdaCost must have 2 inputs')
2022 2023
        self.config.NDCG_num = NDCG_num
        if max_sort_size != -1:
2024 2025 2026
            config_assert(
                NDCG_num <= max_sort_size,
                'NDCG_num must be less than or equal to max_sort_size')
2027 2028
        self.config.max_sort_size = max_sort_size

2029

2030 2031
@config_layer('nce')
class NCELayer(LayerBase):
2032 2033 2034 2035 2036 2037 2038 2039
    def __init__(self,
                 name,
                 num_classes,
                 inputs,
                 num_neg_samples=10,
                 neg_sampling_dist=None,
                 bias=True,
                 **xargs):
2040
        super(NCELayer, self).__init__(name, 'nce', 1, inputs=inputs, **xargs)
2041 2042
        config_assert(
            len(self.inputs) >= 2, 'NCELayer must have at least 2 inputs')
2043 2044
        self.config.num_classes = num_classes
        if neg_sampling_dist is not None:
2045 2046 2047 2048
            config_assert(
                len(neg_sampling_dist) == num_classes,
                'len(neg_sampling_dist)(%s) is not same as num_classes (%s)' %
                (len(neg_sampling_dist), num_classes))
2049
            s = sum(neg_sampling_dist)
2050 2051 2052
            config_assert(
                abs(s - 1) < 1e-5,
                'The sum of neg_sampling_dist (%s) is not 1' % s)
2053 2054 2055 2056 2057

            self.config.neg_sampling_dist.extend(neg_sampling_dist)

        self.config.num_neg_samples = num_neg_samples
        num_real_inputs = len(self.inputs) - 1
2058
        input_layer = self.get_input_layer(num_real_inputs)
2059 2060 2061 2062
        config_assert(input_layer.type == 'data',
                      'Expecting the last input layer of an nce layer to be '
                      'a data layer')

2063 2064
        if (num_real_inputs > 1 and input_layer.size == 1 and
                self.get_input_layer(num_real_inputs - 1).type == 'data'):
2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077
            # This input layer is assumed to be a sample weight layer
            num_real_inputs -= 1

        for input_index in xrange(num_real_inputs):
            input_layer = self.get_input_layer(input_index)
            psize = num_classes * input_layer.size
            dims = [num_classes, input_layer.size]
            self.create_input_parameter(input_index, psize, dims)
        self.create_bias_parameter(bias, num_classes)


@config_layer('addto')
class AddToLayer(LayerBase):
2078
    def __init__(self, name, inputs, bias=True, **xargs):
2079 2080
        super(AddToLayer, self).__init__(
            name, 'addto', 0, inputs=inputs, **xargs)
2081
        config_assert(len(inputs) > 0, 'inputs cannot be empty for AddToLayer')
2082 2083 2084 2085 2086
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
            self.set_layer_size(input_layer.size)
        self.create_bias_parameter(bias, self.config.size)

2087

2088 2089
@config_layer('agent')
class AgentLayer(LayerBase):
2090 2091 2092 2093
    def __init__(self, name, size, device=None):
        super(AgentLayer, self).__init__(
            name, 'agent', size, inputs=[], device=device)

2094 2095 2096

@config_layer('sequence_agent')
class SequenceAgentLayer(LayerBase):
2097
    def __init__(self, name, size, device=None):
2098 2099 2100
        super(SequenceAgentLayer, self).__init__(
            name, 'sequence_agent', size, inputs=[], device=device)

2101

2102 2103
@config_layer('gather_agent')
class GatherAgentLayer(LayerBase):
2104
    def __init__(self, name, size, device=None):
2105 2106 2107
        super(GatherAgentLayer, self).__init__(
            name, 'gather_agent', size, inputs=[], device=device)

2108

2109 2110
@config_layer('scatter_agent')
class ScatterAgentLayer(LayerBase):
2111
    def __init__(self, name, size, device=None):
2112 2113 2114
        super(ScatterAgentLayer, self).__init__(
            name, 'scatter_agent', size, inputs=[], device=device)

2115

2116 2117
@config_layer('sequence_gather_agent')
class SequenceGatherAgentLayer(LayerBase):
2118
    def __init__(self, name, size, device=None):
2119
        super(SequenceGatherAgentLayer, self).__init__(
2120 2121
            name, 'sequence_gather_agent', size, inputs=[], device=device)

2122 2123 2124

@config_layer('sequence_scatter_agent')
class SequenceScatterAgentLayer(LayerBase):
2125
    def __init__(self, name, size, device=None):
2126
        super(SequenceScatterAgentLayer, self).__init__(
2127 2128
            name, 'sequence_scatter_agent', size, inputs=[], device=device)

2129 2130 2131

@config_layer('multiplex')
class MultiplexLayer(LayerBase):
2132 2133 2134 2135 2136
    def __init__(self, name, inputs, size, device=None):
        super(MultiplexLayer, self).__init__(
            name, 'multiplex', size, inputs=inputs, device=device)
        config_assert(
            len(inputs) > 2, 'MultiplexLayer should have more than 2 inputs.')
2137
        for i in range(1, len(inputs)):
2138 2139 2140 2141 2142
            config_assert(
                self.get_input_layer(i).size == size,
                "All the input layers except the first one should"
                "have the same size as the MultiplexLayer.")

2143 2144

@config_func
2145 2146 2147
def Link(
        name,
        has_subseq=False, ):
2148 2149 2150 2151 2152
    link_config = LinkConfig()
    link_config.link_name = name
    link_config.has_subseq = has_subseq
    return link_config

2153

2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167
# memory for recurrent layer group.
# *name* and *size* are actual layer's name and size.
# will return name of the memory,
# use this name if you assign the memory as other layer's input
#
# boot frame of memory is zeroed by default,
# or initialize by boot layer output if *boot_layer* set,
# or initialize by trainable bias if *boot_bias* set,
# or initialize by a constant id if *boot_with_const_id* set
#
# Memory can be a sequence if *is_sequence* set, this type of memory
# can only be initailized by a *boot_layer* which is a sequence.
#
@config_func
2168 2169 2170 2171 2172 2173 2174 2175
def Memory(
        name,
        size,
        is_sequence=False,
        boot_layer=None,
        boot_bias=False,
        boot_bias_active_type="",
        boot_with_const_id=None, ):
2176 2177 2178 2179 2180 2181
    agent_name = name + "+delay1"
    if is_sequence:
        agent_layer = SequenceAgentLayer(agent_name, size)
    else:
        agent_layer = AgentLayer(agent_name, size)
    config_assert(g_current_submodel.is_recurrent_layer_group,
2182
                  'Memory should be used in recurrent layer group only')
2183 2184 2185 2186
    memory = g_current_submodel.memories.add()
    memory.layer_name = MakeLayerNameInSubmodel(name)
    memory.link_name = MakeLayerNameInSubmodel(agent_name)
    memory.is_sequence = is_sequence
2187
    options = sum((boot_layer is not None, bool(boot_bias),
2188
                   boot_with_const_id is not None))
2189 2190 2191 2192
    config_assert(
        options <= 1,
        'take one option at most from boot_layer, boot_bias, or boot_with_const_id'
    )
2193 2194 2195
    if boot_layer is not None:
        boot_layer = MakeLayerNameInParentSubmodel(boot_layer)
        config_assert(boot_layer in g_layer_map,
2196 2197
                      'boot_layer "%s" does not correspond to a layer name' %
                      boot_layer)
2198 2199 2200
        memory.boot_layer_name = boot_layer
    elif boot_bias:
        memory.boot_bias_parameter_name = agent_layer.create_bias_parameter(
2201
            boot_bias, size, for_self=False)
2202 2203 2204 2205 2206
        memory.boot_bias_active_type = boot_bias_active_type
    elif boot_with_const_id is not None:
        memory.boot_with_const_id = boot_with_const_id
    return agent_name

2207

2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218
# Generator for recurrent layer group, to use it:
#  1. define a id layer as output of layer group
#  2. define a memory of this id layer, and assign a boot id(begin of sequence)
#  3. define a eos check layer and fill its name in generator's *eos_layer_name*
# Sequence generation will stop when eos check return 1 or *max_num_frames* reached.
# If *beam_size* is greater than one, generator will use beam search.
#   in beam search, if *num_results_per_sample* set, one sample sequence can output
#   multiple results each with a probility.
@config_func
def Generator(
        max_num_frames,
2219 2220 2221 2222
        eos_layer_name="eos_check",
        num_results_per_sample=1,
        beam_size=1,
        log_prob=None, ):
2223 2224 2225 2226 2227 2228 2229 2230 2231
    generator_config = GeneratorConfig()
    generator_config.max_num_frames = max_num_frames
    generator_config.eos_layer_name = eos_layer_name
    generator_config.num_results_per_sample = num_results_per_sample
    generator_config.beam_size = beam_size
    if log_prob is not None:
        generator_config.log_prob = log_prob
    return generator_config

2232

2233 2234
@config_layer('expand')
class ExpandLayer(LayerBase):
2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250
    def __init__(self,
                 name,
                 inputs,
                 trans_type='non-seq',
                 device=None,
                 bias=False):
        super(ExpandLayer, self).__init__(
            name, 'expand', 0, inputs=inputs, device=device)
        config_assert(
            len(self.inputs) == 2, 'ExpandLayer takes 2 and only 2 inputs')
        self.config.trans_type = trans_type
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
        self.set_layer_size(self.get_input_layer(0).size)
        self.create_bias_parameter(bias, self.config.size)

2251 2252 2253

@config_layer('featmap_expand')
class FeatMapExpandLayer(LayerBase):
2254 2255 2256 2257 2258 2259
    def __init__(self, name, inputs, device=None, num_filters=None, bias=False):
        super(FeatMapExpandLayer, self).__init__(
            name, 'featmap_expand', 0, inputs=inputs, device=device)
        config_assert(
            len(self.inputs) == 1, 'ExpandLayer takes 1 and only 1 inputs')
        if num_filters is not None:
2260
            self.config.num_filters = num_filters
2261
        else:
2262
            logger.fatal("FeatMapExpandLayer must specify num_filters.")
2263
        self.set_layer_size(self.get_input_layer(0).size * num_filters)
2264 2265 2266 2267


@config_layer('max')
class MaxLayer(LayerBase):
2268 2269 2270 2271 2272 2273 2274 2275 2276 2277
    def __init__(self,
                 name,
                 inputs,
                 trans_type='non-seq',
                 active_type='linear',
                 device=None,
                 bias=False,
                 output_max_index=None):
        super(MaxLayer, self).__init__(
            name, 'max', 0, inputs=inputs, device=device)
2278
        config_assert(len(self.inputs) == 1, 'MaxLayer must have 1 input')
2279 2280
        self.config.trans_type = trans_type
        self.config.active_type = active_type
2281 2282 2283 2284
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
            self.set_layer_size(input_layer.size)
        self.create_bias_parameter(bias, self.config.size)
2285 2286
        if output_max_index is not None:
            self.config.output_max_index = output_max_index
2287 2288 2289 2290


@config_layer('maxid')
class MaxIdLayer(LayerBase):
2291
    def __init__(self, name, inputs, beam_size=None, device=None):
2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308
        super(MaxIdLayer, self).__init__(
            name, 'maxid', 0, inputs=inputs, device=device)
        config_assert(len(self.inputs) == 1, 'MaxIdLayer must have 1 input')
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
            self.set_layer_size(input_layer.size)

        if beam_size is None:
            global g_current_submodel
            if g_current_submodel.HasField("generator"):
                self.config.beam_size = g_current_submodel.generator.beam_size
        else:
            self.config.beam_size = beam_size


@config_layer('eos_id')
class EosIdLayer(LayerBase):
2309
    def __init__(self, name, inputs, eos_id, device=None):
2310 2311 2312
        super(EosIdLayer, self).__init__(
            name, 'eos_id', 0, inputs=inputs, device=device)
        config_assert(len(self.inputs) == 1, 'EosIdLayer must have 1 input')
2313
        self.set_layer_size(2)  # boolean output
2314 2315
        self.config.eos_id = eos_id

2316

2317 2318
@config_layer('seqlastins')
class SequenceLastInstanceLayer(LayerBase):
2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335
    def __init__(self,
                 name,
                 inputs,
                 active_type='linear',
                 trans_type='non-seq',
                 device=None,
                 bias=False):
        super(SequenceLastInstanceLayer, self).__init__(
            name,
            'seqlastins',
            0,
            inputs=inputs,
            device=device,
            active_type=active_type)
        config_assert(
            len(inputs) == 1, 'SequenceLastInstanceLayer must have 1 input')
        self.config.trans_type = trans_type
2336 2337 2338 2339 2340
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
            self.set_layer_size(input_layer.size)
        self.create_bias_parameter(bias, self.config.size)

2341

2342 2343 2344 2345 2346 2347 2348 2349 2350
@config_layer('seqfirstins')
class SequenceFirstInstanceLayer(SequenceLastInstanceLayer):
    def __init__(
            self,
            name,
            inputs,
            active_type='linear',
            trans_type='non-seq',
            device=None,
2351 2352 2353 2354 2355 2356 2357 2358
            bias=False, ):
        super(SequenceFirstInstanceLayer, self).__init__(
            name,
            inputs=inputs,
            active_type=active_type,
            device=device,
            bias=bias)
        self.config.trans_type = trans_type
2359 2360
        self.config.select_first = True

2361

2362 2363
@config_layer('seqconcat')
class SequenceConcatLayer(LayerBase):
2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378
    def __init__(self,
                 name,
                 inputs,
                 active_type='linear',
                 device=None,
                 bias=False):
        super(SequenceConcatLayer, self).__init__(
            name,
            'seqconcat',
            0,
            inputs=inputs,
            device=device,
            active_type=active_type)
        config_assert(
            len(inputs) == 2, 'SequenceConcatLayer must have 2 inputs')
2379 2380 2381 2382 2383
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
            self.set_layer_size(input_layer.size)
        self.create_bias_parameter(bias, self.config.size)

2384

2385 2386
@config_layer('seqreshape')
class SequenceReshapeLayer(LayerBase):
2387 2388 2389 2390 2391 2392 2393 2394 2395 2396
    def __init__(self,
                 name,
                 size,
                 inputs,
                 active_type='linear',
                 device=None,
                 bias=False):
        super(SequenceReshapeLayer, self).__init__(
            name,
            'seqreshape',
2397
            size,
2398 2399 2400 2401 2402
            inputs=inputs,
            device=device,
            active_type=active_type)
        config_assert(
            len(inputs) == 1, 'SequenceReshapeLayer must have 1 inputs')
2403 2404 2405
        self.set_layer_size(size)
        self.create_bias_parameter(bias, size)

2406

2407 2408
@config_layer('subseq')
class SubSequenceLayer(LayerBase):
2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421
    def __init__(self,
                 name,
                 inputs,
                 active_type='linear',
                 device=None,
                 bias=False):
        super(SubSequenceLayer, self).__init__(
            name,
            'subseq',
            0,
            inputs=inputs,
            device=device,
            active_type=active_type)
2422 2423 2424 2425 2426 2427
        config_assert(len(inputs) == 3, 'SubSequenceLayer must have 3 inputs')
        input_layer0 = self.get_input_layer(0)
        size = input_layer0.size
        self.set_layer_size(size)
        self.create_bias_parameter(bias, size)

2428

2429 2430
@config_layer('out_prod')
class OuterProdLayer(LayerBase):
2431 2432 2433
    def __init__(self, name, inputs, device=None):
        super(OuterProdLayer, self).__init__(
            name, 'out_prod', 0, inputs=inputs, device=device)
2434 2435 2436 2437 2438
        config_assert(len(inputs) == 2, 'OuterProdLayer must have 2 inputs')
        input_layer0 = self.get_input_layer(0)
        input_layer1 = self.get_input_layer(1)
        self.set_layer_size(input_layer0.size * input_layer1.size)

2439

2440 2441
@config_layer('power')
class PowerLayer(LayerBase):
2442 2443 2444
    def __init__(self, name, inputs, device=None):
        super(PowerLayer, self).__init__(
            name, 'power', 0, inputs=inputs, device=device)
2445 2446 2447 2448
        config_assert(len(inputs) == 2, 'PowerLayer must have 2 inputs')
        input_layer1 = self.get_input_layer(1)
        self.set_layer_size(input_layer1.size)
        input_layer0 = self.get_input_layer(0)
2449 2450 2451
        config_assert(1 == input_layer0.size,
                      'The left input is the exponent and should be of size 1')

2452 2453 2454

@config_layer('slope_intercept')
class SlopeInterceptLayer(LayerBase):
2455 2456 2457
    def __init__(self, name, inputs, slope=1.0, intercept=0.0, device=None):
        super(SlopeInterceptLayer, self).__init__(
            name, 'slope_intercept', 0, inputs=inputs, device=device)
2458 2459 2460 2461 2462 2463
        self.config.slope = slope
        self.config.intercept = intercept
        config_assert(len(inputs) == 1, 'SlopeInterceptLayer must have 1 input')
        input_layer0 = self.get_input_layer(0)
        self.set_layer_size(input_layer0.size)

2464

2465 2466
@config_layer('scaling')
class ScalingLayer(LayerBase):
2467 2468 2469
    def __init__(self, name, inputs, device=None):
        super(ScalingLayer, self).__init__(
            name, 'scaling', 0, inputs=inputs, device=device)
2470 2471 2472 2473
        config_assert(len(inputs) == 2, 'ScalingLayer must have 2 inputs')
        input_layer1 = self.get_input_layer(1)
        self.set_layer_size(input_layer1.size)
        input_layer0 = self.get_input_layer(0)
2474 2475 2476
        config_assert(1 == input_layer0.size,
                      'The left input should be of size 1')

2477 2478 2479

@config_layer('conv_shift')
class ConvShiftLayer(LayerBase):
2480 2481 2482
    def __init__(self, name, inputs, device=None):
        super(ConvShiftLayer, self).__init__(
            name, 'conv_shift', 0, inputs=inputs, device=device)
2483 2484 2485 2486
        config_assert(len(inputs) == 2, 'ConvShiftLayer must have 2 inputs')
        input_layer0 = self.get_input_layer(0)
        self.set_layer_size(input_layer0.size)

2487

2488 2489
@config_layer('convex_comb')
class ConvexCombinationLayer(LayerBase):
2490
    def __init__(self, name, size, inputs, device=None):
2491
        super(ConvexCombinationLayer, self).__init__(
2492 2493 2494
            name, 'convex_comb', size, inputs=inputs, device=device)
        config_assert(
            len(self.inputs) == 2, 'ConvexCombinationLayer must have 2 inputs')
2495 2496 2497
        config_assert(
            size * self.get_input_layer(0).size == self.get_input_layer(1).size,
            'Wrong input size for ConvexCombinationLayer')
2498 2499
        self.set_layer_size(size)

2500

2501 2502
@config_layer('interpolation')
class InterpolationLayer(LayerBase):
2503
    def __init__(self, name, inputs, device=None):
2504 2505
        super(InterpolationLayer, self).__init__(
            name, 'interpolation', 0, inputs=inputs, device=device)
2506 2507
        config_assert(
            len(self.inputs) == 3, 'InterpolationLayer must have 3 inputs')
2508 2509 2510 2511 2512 2513 2514 2515
        input_layer0 = self.get_input_layer(0)
        input_layer1 = self.get_input_layer(1)
        input_layer2 = self.get_input_layer(2)
        self.set_layer_size(input_layer1.size)
        config_assert(input_layer0.size == 1, 'weight should be of size 1')
        config_assert(input_layer1.size == input_layer2.size,
                      'the two vector inputs should be of the same size')

2516

2517 2518
@config_layer('bilinear_interp')
class BilinearInterpLayer(LayerBase):
2519
    def __init__(self, name, inputs, **xargs):
2520
        super(BilinearInterpLayer, self).__init__(
L
liaogang 已提交
2521
            name, 'bilinear_interp', 0, inputs=inputs, **xargs)
2522
        input_layer = self.get_input_layer(0)
2523 2524
        parse_bilinear(self.inputs[0].bilinear_interp, input_layer.name,
                       self.config.inputs[0].bilinear_interp_conf)
L
liaogang 已提交
2525
        conf = self.inputs[0].bilinear_interp
2526 2527 2528
        self.set_layer_size(conf.out_size_x * conf.out_size_y *
                            conf.num_channels)

L
liaogang 已提交
2529

2530 2531
@config_layer('sum_to_one_norm')
class SumToOneNormLayer(LayerBase):
2532
    def __init__(self, name, inputs, device=None):
2533
        super(SumToOneNormLayer, self).__init__(
2534 2535 2536
            name, 'sum_to_one_norm', 0, inputs=inputs, device=device)
        config_assert(
            len(self.inputs) == 1, 'SumToOneNormLayer must have 1 input')
2537 2538 2539
        input_layer0 = self.get_input_layer(0)
        self.set_layer_size(input_layer0.size)

2540

2541 2542
@config_layer('cos_vm')
class CosSimVecMatLayer(LayerBase):
2543
    def __init__(self, name, size, inputs, cos_scale=1.0, device=None):
2544
        super(CosSimVecMatLayer, self).__init__(
2545
            name, 'cos_vm', size, inputs=inputs, device=device)
2546
        self.config.cos_scale = cos_scale
2547 2548
        config_assert(
            len(self.inputs) == 2, 'CosSimVecMatLayer must have 2 inputs')
2549 2550 2551
        config_assert(
            size * self.get_input_layer(0).size == self.get_input_layer(1).size,
            'Wrong input size for CosSimVecMatLayer')
2552

2553

2554 2555
@config_layer('sampling_id')
class SamplingIdLayer(LayerBase):
2556
    def __init__(self, name, inputs, device=None):
2557 2558
        super(SamplingIdLayer, self).__init__(
            name, 'sampling_id', 0, inputs=inputs, device=device)
2559 2560
        config_assert(
            len(self.inputs) == 1, 'SamplingIdLayer must have 1 input')
2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
            self.set_layer_size(input_layer.size)


# AverageLayer: "average" for each sample within a sequence.
# average_stratrgy: set to one of the following:
# 'average': plain average.
# 'sum': sum each sample instead of average (which is divide by sample_num).
# 'squarerootn': sum each sample, but divide by sqrt(sample_num).
@config_layer('average')
class AverageLayer(LayerBase):
2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587
    def __init__(self,
                 name,
                 inputs,
                 average_strategy='average',
                 trans_type='non-seq',
                 active_type='linear',
                 device=None,
                 bias=False):
        super(AverageLayer, self).__init__(
            name,
            'average',
            0,
            inputs=inputs,
            device=device,
            active_type=active_type)
2588
        self.config.average_strategy = average_strategy
2589
        self.config.trans_type = trans_type
2590 2591 2592 2593 2594 2595
        config_assert(len(inputs) == 1, 'AverageLayer must have 1 input')
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
            self.set_layer_size(input_layer.size)
        self.create_bias_parameter(bias, self.config.size)

2596

2597 2598
@config_layer('cos')
class CosSimLayer(LayerBase):
2599
    def __init__(self, name, inputs, cos_scale=5, device=None):
2600 2601 2602 2603 2604 2605
        super(CosSimLayer, self).__init__(
            name, 'cos', 1, inputs=inputs, device=device)
        config_assert(len(self.inputs) == 2, 'CosSimLayer must have 2 inputs')
        config_assert(
            self.get_input_layer(0).size == self.get_input_layer(1).size,
            'inputs of CosSimLayer must have same dim')
2606
        self.config.cos_scale = cos_scale
2607 2608 2609 2610


@config_layer('tensor')
class TensorLayer(LayerBase):
2611 2612 2613
    def __init__(self, name, size, inputs, device=None, bias=True, **xargs):
        super(TensorLayer, self).__init__(
            name, 'tensor', size, inputs=inputs, device=device, **xargs)
2614 2615
        config_assert(len(self.inputs) == 2, 'TensorLayer must have 2 inputs')
        config_assert(size > 0, 'size must be positive')
2616 2617
        config_assert(inputs[1].parameter_name == None,
                      'second parameter should be None.')
2618 2619 2620 2621 2622 2623 2624 2625 2626 2627
        input_layer0 = self.get_input_layer(0)
        input_layer1 = self.get_input_layer(1)
        psize = size * input_layer0.size * input_layer1.size
        dims = [input_layer0.size, input_layer1.size, size]
        self.create_input_parameter(0, psize, dims)
        self.create_bias_parameter(bias, size)


@config_layer('mixed')
class MixedLayer(LayerBase):
2628 2629 2630 2631 2632 2633 2634
    def __init__(self,
                 name,
                 inputs,
                 size=0,
                 bias=True,
                 error_clipping_threshold=None,
                 **xargs):
2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651
        config_assert(inputs, 'inputs cannot be empty')
        super(MixedLayer, self).__init__(
            name, 'mixed', size, inputs=inputs, **xargs)
        operator_input_index = []
        for operator in self.operators:
            operator_conf = operator.operator_conf
            for i in xrange(1, len(operator.input_layer_names)):
                input_index = len(self.config.inputs)
                operator_conf.input_indices.append(input_index)
                input_config = Input(operator.input_layer_names[i])
                self.inputs.append(input_config)
                layer_input = self.config.inputs.add()
                layer_input.input_layer_name = input_config.input_layer_name
            for input_index in operator_conf.input_indices:
                input_layer = self.get_input_layer(input_index)
                operator_conf.input_sizes.append(input_layer.size)
                operator_input_index.append(input_index)
2652
            if self.config.size == 0:
2653 2654 2655
                size = operator.calc_output_size(operator_conf.input_sizes)
                if size != 0:
                    self.set_layer_size(size)
2656
            else:
2657 2658
                sz = operator.calc_output_size(operator_conf.input_sizes)
                if sz != 0:
2659 2660 2661 2662
                    config_assert(
                        sz == self.config.size,
                        "different inputs have different size: %s vs. %s" %
                        (sz, self.config.size))
2663 2664 2665 2666
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
            input = self.inputs[input_index]
            if input_index not in operator_input_index:
2667 2668 2669
                config_assert(
                    isinstance(input, Projection),
                    "input should be projection or operation")
2670
            if self.config.size == 0 and isinstance(input, Projection):
2671 2672 2673
                size = input.calc_output_size(input_layer)
                if size != 0:
                    self.set_layer_size(size)
2674
            elif isinstance(input, Projection):
2675 2676 2677 2678 2679 2680
                sz = input.calc_output_size(input_layer)
                if sz != 0:
                    config_assert(
                        sz == self.config.size,
                        "different inputs have different size: %s vs. %s" %
                        (sz, self.config.size))
2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691
        config_assert(size != 0, "size is not set")

        for input_index in xrange(len(self.inputs)):
            input = self.inputs[input_index]
            if isinstance(input, Projection):
                input_layer = self.get_input_layer(input_index)
                input.proj_conf.input_size = input_layer.size
                input.proj_conf.output_size = size

                input_config = self.config.inputs[input_index]
                input_config.proj_conf.CopyFrom(input.proj_conf)
2692 2693
                input_config.proj_conf.name = gen_parameter_name(name,
                                                                 input_index)
2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704
                psize = input.calc_parameter_size(input_layer.size, size)
                dims = input.calc_parameter_dims(input_layer.size, size)
                self.create_input_parameter(input_index, psize, dims)

        for operator in self.operators:
            operator_conf = operator.operator_conf
            operator_conf.output_size = self.config.size
            operator.check_dims()
            record_operator_conf = self.config.operator_confs.add()
            record_operator_conf.CopyFrom(operator_conf)

2705 2706 2707 2708 2709 2710
        psize = self.config.size
        if isinstance(self.inputs[0], ConvProjection):
            self.config.shared_biases = True
            psize = 0
            for input in self.inputs:
                psize += input.calc_bias_size()
2711

2712 2713 2714
        if bias:
            self.config.bias_size = psize
            self.create_bias_parameter(bias, psize)
2715

2716 2717
        if error_clipping_threshold is not None:
            self.config.error_clipping_threshold = error_clipping_threshold
2718

2719

2720 2721
# like MixedLayer, but no bias parameter
@config_func
2722
def ExpressionLayer(name, inputs, **xargs):
2723 2724
    MixedLayer(name, inputs, bias=False, **xargs)

2725

2726 2727
@config_layer('concat')
class ConcatenateLayer(LayerBase):
2728
    def __init__(self, name, inputs, bias=False, **xargs):
2729
        config_assert(inputs, 'inputs cannot be empty')
2730
        config_assert(not bias, 'ConcatenateLayer cannot support bias.')
2731 2732 2733 2734 2735 2736
        super(ConcatenateLayer, self).__init__(
            name, 'concat', 0, inputs=inputs, **xargs)
        size = 0
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
            input = self.inputs[input_index]
2737
            if self.config.size == 0:
2738 2739 2740 2741
                size += input_layer.size

        self.set_layer_size(size)

2742

2743 2744 2745
# like concat layer, but each input layer was processed by a Projection.
@config_layer('concat2')
class ConcatenateLayer2(LayerBase):
2746
    def __init__(self, name, inputs, bias=False, **xargs):
2747 2748 2749
        config_assert(inputs, 'inputs cannot be empty')
        super(ConcatenateLayer2, self).__init__(
            name, 'concat2', 0, inputs=inputs, **xargs)
2750 2751

        if isinstance(self.inputs[0], ConvProjection):
2752 2753 2754 2755 2756 2757
            for input_index in xrange(len(self.inputs) - 1):
                input = self.inputs[input_index + 1]
                config_assert(
                    isinstance(input, ConvProjection),
                    "The first input of ConcatenateLayer2 is ConvProjection, "
                    "the other inputs should also be ConvProjection.")
2758

2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778
        size = 0
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
            input = self.inputs[input_index]
            output_size = input.calc_output_size(input_layer)
            config_assert(output_size != 0, "proj output size is not set")
            size += output_size

        self.set_layer_size(size)

        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
            input = self.inputs[input_index]
            input.proj_conf.input_size = input_layer.size
            input.proj_conf.output_size = input.calc_output_size(input_layer)

            input_config = self.config.inputs[input_index]
            input_config.proj_conf.CopyFrom(input.proj_conf)
            input_config.proj_conf.name = gen_parameter_name(name, input_index)
            psize = input.calc_parameter_size(input.proj_conf.input_size,
2779
                                              input.proj_conf.output_size)
2780
            dims = input.calc_parameter_dims(input.proj_conf.input_size,
2781
                                             input.proj_conf.output_size)
2782 2783
            self.create_input_parameter(input_index, psize, dims)

2784 2785 2786 2787 2788 2789 2790
        psize = self.config.size
        if isinstance(self.inputs[0], ConvProjection):
            self.config.shared_biases = True
            psize = 0
            for input in self.inputs:
                psize += input.calc_bias_size()

2791 2792 2793
        if bias:
            self.config.bias_size = psize
            self.create_bias_parameter(bias, psize)
2794

2795

2796 2797
@config_layer('recurrent')
class RecurrentLayer(LayerBase):
2798
    def __init__(self, name, inputs, reversed=False, bias=True, **xargs):
Y
Yu Yang 已提交
2799 2800
        super(RecurrentLayer, self).__init__(name, 'recurrent', 0, inputs,
                                             **xargs)
2801 2802 2803 2804 2805 2806 2807 2808 2809
        config_assert(len(self.inputs) == 1, 'RecurrentLayer must have 1 input')
        input_layer = self.get_input_layer(0)
        size = input_layer.size
        self.set_layer_size(size)
        self.config.reversed = reversed
        dims = [size, size]
        self.create_input_parameter(0, size * size, dims)
        self.create_bias_parameter(bias, self.config.size)

2810

2811 2812
@config_layer('lstmemory')
class LstmLayer(LayerBase):
2813 2814 2815 2816 2817 2818 2819 2820
    def __init__(self,
                 name,
                 inputs,
                 reversed=False,
                 active_gate_type="sigmoid",
                 active_state_type="sigmoid",
                 bias=True,
                 **xargs):
2821 2822 2823 2824 2825 2826 2827 2828
        super(LstmLayer, self).__init__(name, 'lstmemory', 0, inputs, **xargs)
        config_assert(len(self.inputs) == 1, 'LstmLayer must have 1 input')
        input_layer = self.get_input_layer(0)
        #check input_layer.size is divided by 4
        config_assert(input_layer.size % 4 == 0, "size % 4 should be 0!")
        size = input_layer.size / 4
        self.set_layer_size(size)
        self.config.reversed = reversed
2829
        self.config.active_gate_type = active_gate_type
2830 2831 2832 2833 2834
        self.config.active_state_type = active_state_type
        self.create_input_parameter(0, size * size * 4, [size, size, 4])
        #bias includes 3 kinds of peephole, 4 + 3 = 7
        self.create_bias_parameter(bias, size * 7)

2835

2836 2837
@config_layer('lstm_step')
class LstmStepLayer(LayerBase):
2838 2839 2840 2841 2842 2843 2844 2845 2846 2847
    def __init__(self,
                 name,
                 size,
                 inputs,
                 active_gate_type="sigmoid",
                 active_state_type="sigmoid",
                 bias=True,
                 **xargs):
        super(LstmStepLayer, self).__init__(name, 'lstm_step', size, inputs,
                                            **xargs)
2848 2849 2850
        config_assert(len(inputs) == 2, 'LstmStepLayer must have 2 inputs')
        input_layer0 = self.get_input_layer(0)
        input_layer1 = self.get_input_layer(1)
2851 2852 2853 2854 2855
        config_assert(input_layer0.size == 4 * size,
                      'input_layer0.size != 4 * layer.size')
        config_assert(input_layer1.size == size,
                      'input_layer1.size != layer.size')
        self.config.active_gate_type = active_gate_type
2856 2857 2858
        self.config.active_state_type = active_state_type
        self.create_bias_parameter(bias, size * 3)

2859

2860 2861 2862
# get the specific output from the input layer.
@config_layer('get_output')
class GetOutputLayer(LayerBase):
2863 2864 2865 2866
    def __init__(self, name, size, inputs):
        super(GetOutputLayer, self).__init__(name, 'get_output', size, inputs)
        config_assert(
            len(self.inputs) == 1, 'GetOutputLayer must have 1 inputs')
2867 2868 2869 2870
        inputs = self.inputs[0]
        config_assert(inputs.input_layer_argument,
                      'input_layer_argument cannot be empty')

2871

2872 2873
@config_layer('mdlstmemory')
class MDLstmLayer(LayerBase):
2874 2875 2876 2877 2878 2879 2880 2881
    def __init__(self,
                 name,
                 inputs,
                 directions=True,
                 active_gate_type="sigmoid",
                 active_state_type="sigmoid",
                 bias=True,
                 **xargs):
Y
Yu Yang 已提交
2882 2883
        super(MDLstmLayer, self).__init__(name, 'mdlstmemory', 0, inputs,
                                          **xargs)
2884 2885 2886 2887
        config_assert(len(self.inputs) == 1, 'MDLstmLayer must have 1 input')
        input_layer = self.get_input_layer(0)
        dim_num = len(directions)
        #check input_layer.size is divided by (3+dim_num)
Y
Yu Yang 已提交
2888 2889
        config_assert(input_layer.size % (3 + dim_num) == 0,
                      "size % (dim_num) should be 0!")
2890
        size = input_layer.size / (3 + dim_num)
2891
        self.set_layer_size(size)
2892
        self.config.active_gate_type = active_gate_type
2893 2894 2895
        self.config.active_state_type = active_state_type
        for i in xrange(len(directions)):
            self.config.directions.append(int(directions[i]))
Y
Yu Yang 已提交
2896 2897
        self.create_input_parameter(0, size * size * (3 + dim_num),
                                    [size, size, 3 + dim_num])
2898
        #bias includes 3 kinds of peephole, 3+dim_num+2+dim_num
2899 2900
        self.create_bias_parameter(bias, size * (5 + 2 * dim_num))

2901 2902 2903

@config_layer('gated_recurrent')
class GatedRecurrentLayer(LayerBase):
2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914
    def __init__(self,
                 name,
                 inputs,
                 reversed=False,
                 active_gate_type="sigmoid",
                 bias=True,
                 **xargs):
        super(GatedRecurrentLayer, self).__init__(name, 'gated_recurrent', 0,
                                                  inputs, **xargs)
        config_assert(
            len(self.inputs) == 1, 'GatedRecurrentLayer must have 1 input')
2915 2916 2917 2918 2919 2920
        input_layer = self.get_input_layer(0)
        #check input_layer.size is divided by 3
        config_assert(input_layer.size % 3 == 0, "size % 3 should be 0!")
        size = input_layer.size / 3
        self.set_layer_size(size)
        self.config.reversed = reversed
2921
        self.config.active_gate_type = active_gate_type
2922 2923 2924
        self.create_input_parameter(0, size * size * 3, [size, size * 3])
        self.create_bias_parameter(bias, size * 3)

2925

2926 2927
@config_layer('gru_step')
class GruStepLayer(LayerBase):
2928 2929 2930 2931 2932 2933 2934
    def __init__(self,
                 name,
                 size,
                 inputs,
                 active_gate_type="sigmoid",
                 bias=True,
                 **xargs):
Y
Yu Yang 已提交
2935 2936
        super(GruStepLayer, self).__init__(name, 'gru_step', size, inputs,
                                           **xargs)
2937 2938 2939
        config_assert(len(self.inputs) == 2, 'GruStepLayer must have 2 input')
        input_layer0 = self.get_input_layer(0)
        input_layer1 = self.get_input_layer(1)
2940 2941 2942 2943 2944
        config_assert(input_layer0.size == 3 * size,
                      'input_layer0.size != 3 * layer.size')
        config_assert(input_layer1.size == size,
                      'input_layer1.size != layer.size')
        self.config.active_gate_type = active_gate_type
2945 2946 2947
        self.create_input_parameter(0, size * size * 3, [size, size * 3])
        self.create_bias_parameter(bias, size * 3)

2948

2949 2950 2951 2952 2953 2954 2955
'''
 A layer for calculating the cost of sequential conditional random field model.
 Example: CRFLayer(name="crf_cost", size=label_num,
                   inputs=["output", "label", "weight"])
          where "weight" is optional, one weight for each sequence
 @param coeff: weight of the layer
'''
2956 2957


2958 2959
@config_layer('crf')
class CRFLayer(LayerBase):
2960
    def __init__(self, name, size, inputs, coeff=1.0, device=None):
2961
        super(CRFLayer, self).__init__(name, 'crf', size, inputs, device=device)
2962 2963
        config_assert(2 <= len(self.inputs) <= 3,
                      'CRFLayer must have 2 or 3 inputs')
2964 2965 2966
        self.create_input_parameter(0, size * (size + 2), [size, size + 2])
        self.config.coeff = coeff

2967

2968 2969 2970 2971 2972 2973 2974 2975
'''
 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
'''
2976 2977


2978 2979
@config_layer('crf_decoding')
class CRFDecodingLayer(LayerBase):
2980
    def __init__(self, name, size, inputs, device=None):
2981 2982 2983 2984 2985 2986 2987
        super(CRFDecodingLayer, self).__init__(
            name, 'crf_decoding', size, inputs, device=device)
        config_assert(
            len(self.inputs) <= 2,
            'CRFDecodingLayer cannot have more than 2 inputs')
        self.create_input_parameter(0, size * (size + 2), [size, size + 2])

2988

2989 2990
@config_layer('ctc')
class CTCLayer(LayerBase):
2991
    def __init__(self, name, size, inputs, norm_by_times=False, device=None):
2992 2993 2994 2995
        super(CTCLayer, self).__init__(name, 'ctc', size, inputs, device=device)
        self.config.norm_by_times = norm_by_times
        config_assert(len(self.inputs) == 2, 'CTCLayer must have 2 inputs')

2996

2997 2998
@config_layer('recurrent_layer_group')
class RecurrentLayerGroup(LayerBase):
2999
    def __init__(self, name, device=None):
3000 3001 3002 3003 3004 3005
        super(RecurrentLayerGroup, self).__init__(
            name, 'recurrent_layer_group', 0, inputs=[], device=device)


# Deprecated, use a new layer specific class instead
@config_func
3006
def Layer(name, type, **xargs):
3007 3008 3009 3010
    layers = {}
    layers.update(g_cost_map)
    layers.update(g_layer_type_map)
    layer_func = layers.get(type)
3011
    config_assert(layer_func, "layer type '%s' not supported." % type)
3012
    return layer_func(name, **xargs)
3013

3014

3015
@config_func
3016
def ParameterHook(type, **kwargs):
3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028
    if type == 'pruning':
        mask_filename = kwargs.get('mask_filename', None)
        assert mask_filename is not None
        hook = ParameterUpdaterHookConfig()
        hook.type = type
        hook.purning_mask_filename = mask_filename
        return hook
    else:
        return None


@config_func
3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050
def Parameter(name,
              size,
              device,
              dims,
              learning_rate=None,
              momentum=None,
              decay_rate=None,
              decay_rate_l1=None,
              initial_mean=None,
              initial_std=None,
              initial_strategy=None,
              initial_smart=None,
              num_batches_regularization=None,
              sparse_remote_update=None,
              sparse_update=None,
              gradient_clipping_threshold=None,
              sparse=None,
              format=None,
              need_compact=None,
              is_static=None,
              is_shared=None,
              update_hooks=None):
3051 3052 3053 3054 3055 3056 3057

    config_assert(name not in g_parameter_map,
                  'Duplicated parameter name: ' + name)

    para = g_config.model_config.parameters.add()
    para.name = name
    para.size = size
3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068
    if device is not None:
        para.device = int(device)
    para.dims.extend(dims)

    if learning_rate is not None:
        para.learning_rate = float(learning_rate)

    momentum = default(momentum, g_default_momentum)
    if momentum is not None:
        para.momentum = float(momentum)

3069 3070
    config_assert(not momentum or not decay_rate_l1,
                  "momentum and decay_rate_l1 cannot both be non-zero")
3071 3072 3073 3074 3075

    decay_rate = default(decay_rate, g_default_decay_rate)
    if decay_rate is not None:
        para.decay_rate = decay_rate

3076 3077 3078 3079
    if decay_rate_l1 is not None:
        para.decay_rate_l1 = decay_rate_l1
    para.initial_std = default(initial_std, g_default_initial_std)
    para.initial_mean = default(initial_mean, g_default_initial_mean)
3080

3081 3082
    num_batches_regularization = default(num_batches_regularization,
                                         g_default_num_batches_regularization)
3083 3084 3085
    if num_batches_regularization is not None:
        para.num_batches_regularization = int(num_batches_regularization)

3086 3087 3088 3089 3090 3091
    if sparse_remote_update is not None:
        para.sparse_remote_update = sparse_remote_update
        if sparse_remote_update:
            g_config.opt_config.use_sparse_remote_updater = True
    if sparse_update is not None:
        para.sparse_update = sparse_update
3092 3093
    gradient_clipping_threshold = default(gradient_clipping_threshold,
                                          g_default_gradient_clipping_threshold)
3094 3095
    if gradient_clipping_threshold is not None:
        para.gradient_clipping_threshold = gradient_clipping_threshold
3096 3097
    para.initial_strategy = default(initial_strategy,
                                    g_default_initial_strategy)
3098 3099 3100 3101 3102 3103
    para.initial_smart = default(initial_smart, g_default_initial_smart)
    if para.initial_smart:
        para.initial_mean = 0.
        if len(para.dims) != 0:
            para.initial_std = 1. / math.sqrt(para.dims[0])
        else:
3104 3105 3106
            print(
                "Use initial_smart, but dims not set. Initial_smart may not be used in this layer"
            )
3107 3108 3109 3110
            traceback.print_exc()
            para.initial_std = 1. / math.sqrt(para.size)
    if g_default_compact_func is not None:
        sparse, format, need_compact = g_default_compact_func(para.name)
3111 3112 3113 3114 3115 3116 3117

    if sparse is not None:
        para.is_sparse = sparse
    if format is not None:
        para.format = format
    if need_compact is not None:
        para.need_compact = need_compact
3118 3119 3120 3121
    if is_static is not None:
        para.is_static = is_static
    config_assert(not para.sparse_remote_update or not para.is_static,
                  "sparse_remote_update and is_static cannot both be true")
3122 3123
    if is_shared is not None:
        para.is_shared = is_shared
3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144

    update_hooks = default(update_hooks, g_default_update_hooks)

    if update_hooks is not None:
        if hasattr(update_hooks, '__call__'):
            update_hooks = update_hooks(para.name)

        if isinstance(update_hooks, list):
            for hook in update_hooks:
                para.update_hooks.extend([hook])
        else:
            para.update_hooks.extend(update_hooks)

    g_parameter_map[name] = para


@config_func
def default_initial_std(val):
    global g_default_initial_std
    g_default_initial_std = val

3145

3146 3147 3148 3149 3150
@config_func
def default_initial_mean(val):
    global g_default_initial_mean
    g_default_initial_mean = val

3151

3152 3153 3154 3155 3156
@config_func
def default_initial_strategy(val):
    global g_default_initial_strategy
    g_default_initial_strategy = val

3157

3158 3159 3160 3161 3162
@config_func
def default_initial_smart(val):
    global g_default_initial_smart
    g_default_initial_smart = val

3163

3164 3165 3166 3167 3168
@config_func
def default_momentum(val):
    global g_default_momentum
    g_default_momentum = val

3169

3170 3171 3172 3173 3174
@config_func
def default_decay_rate(val):
    global g_default_decay_rate
    g_default_decay_rate = val

3175

3176 3177 3178 3179 3180
@config_func
def default_num_batches_regularization(val):
    global g_default_num_batches_regularization
    g_default_num_batches_regularization = val

3181

3182 3183 3184 3185 3186
@config_func
def default_gradient_clipping_threshold(val):
    global g_default_gradient_clipping_threshold
    g_default_gradient_clipping_threshold = val

3187

3188 3189 3190 3191 3192
@config_func
def default_device(val):
    global g_default_device
    g_default_device = val

3193

3194 3195 3196 3197 3198
@config_func
def default_update_hooks(val):
    global g_default_update_hooks
    g_default_update_hooks = val

3199

3200 3201 3202 3203 3204
@config_func
def default_compact_func(val):
    global g_default_compact_func
    g_default_compact_func = val

3205

3206 3207 3208 3209 3210
def make_importer(config_dir, config_args):
    def Import(config_file, local_args={}):
        if not config_file.startswith('/'):
            config_file = config_dir + '/' + config_file
            g_config.config_files.append(config_file)
3211 3212 3213
        execfile(config_file,
                 make_config_environment(config_file, config_args), local_args)

3214 3215
    return Import

3216

3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244
settings = dict(
    batch_size=None,
    mini_batch_size=None,
    algorithm='async_sgd',
    async_lagged_grad_discard_ratio=1.5,
    learning_method='momentum',
    num_batches_per_send_parameter=None,
    num_batches_per_get_parameter=None,
    center_parameter_update_method=None,
    learning_rate=1.,
    learning_rate_decay_a=0.,
    learning_rate_decay_b=0.,
    learning_rate_schedule='poly',
    learning_rate_args='',
    l1weight=0.1,
    l2weight=0.,
    l2weight_zero_iter=0,
    c1=0.0001,
    backoff=0.5,
    owlqn_steps=10,
    max_backoff=5,
    average_window=0,
    do_average_in_cpu=False,
    max_average_window=None,
    ada_epsilon=1e-6,
    ada_rou=0.95,
    delta_add_rate=1.0,
    shrink_parameter_value=0,
3245 3246 3247
    adam_beta1=0.9,
    adam_beta2=0.999,
    adam_epsilon=1e-8, )
3248

3249
settings_deprecated = dict(usage_ratio=1., )
3250 3251 3252 3253

trainer_settings = dict(
    save_dir="./output/model",
    init_model_path=None,
3254 3255
    start_pass=0, )

3256 3257 3258 3259 3260

@config_func
def Settings(**args):
    for k, v in args.iteritems():
        if k == "usage_ratio":
3261 3262
            logger.warning(
                "Deprecated: define usage_ratio in DataConfig instead")
3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273
            if g_config.HasField("data_config"):
                g_config.data_config.__setattr__(k, v)
            settings_deprecated[k] = v
            continue
        elif k in settings:
            settings[k] = v
        elif k in trainer_settings:
            trainer_settings[k] = v
        else:
            logger.fatal('Unkown setting: %s' % k)

3274

3275 3276 3277 3278
@config_func
def cluster_config(**args):
    pass

3279

3280 3281 3282 3283 3284 3285 3286 3287 3288
@config_func
def EnableSubmodelSuffix(flag=True):
    """
    If enabled, the layer and evaluator names in submodel will be automatically
    appended with @submodel_name
    """
    global g_add_submodel_suffix
    g_add_submodel_suffix = flag

3289

3290 3291 3292 3293
def make_config_environment(config_file, config_args):
    def make_setter(k):
        def setter(v):
            logger.fatal("Obsolete: use Settings(%s=%s, ...) instead" % (k, v))
3294

3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309
        return setter

    funcs = {}
    funcs.update(g_config_funcs)

    for k in settings.iterkeys():
        funcs[k] = make_setter(k)
    for k in settings_deprecated.iterkeys():
        funcs[k] = make_setter(k)
    config_dir = os.path.dirname(config_file)
    if not config_dir:
        config_dir = '.'

    funcs.update(
        Import=make_importer(config_dir, config_args),
3310
        get_config_arg=make_get_config_arg(config_args), )
3311 3312 3313 3314 3315

    funcs.update(g_extended_config_funcs)

    return funcs

3316

3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332
def make_get_config_arg(config_args):
    def get_config_arg(name, type, default=None):
        if type == bool:
            s = config_args.get(name)
            if not s:
                return default
            if s == 'True' or s == '1' or s == 'true':
                return True
            if s == 'False' or s == '0' or s == 'false':
                return False
            raise ValueError('Value of config_arg %s is not boolean' % name)
        else:
            return type(config_args.get(name, default))

    return get_config_arg

3333

3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345
def importlib(name):
    __import__(name)
    return sys.modules[name]


def find_caller():
    stack = traceback.extract_stack()
    for s in stack[-4::-1]:
        if not s[0].endswith('config_parser.py'):
            return s[0], s[1], s[2]
    return "(unknown file)", 0, "(unknown function)"

3346

3347 3348 3349 3350
def my_fatal(s):
    logger.critical(s)
    raise Exception()

3351

3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388
def parse_config(config_file, config_arg_str):
    '''
    @param config_arg_str: a string of the form var1=val1,var2=val2. It will be
    passed to config script as a dictionary CONFIG_ARGS
    '''
    init_config_environment()

    config_args = {}

    logger.findCaller = find_caller
    logger.fatal = my_fatal

    g_config.model_config.type = "nn"
    if config_arg_str:
        config_args = dict([f.split('=') for f in config_arg_str.split(',')])

    global g_command_config_args
    g_command_config_args.update(config_args)

    extension_module_name = config_args.get('extension_module_name')
    if extension_module_name:
        global g_extended_config_funcs
        extension_module = importlib(extension_module_name)
        g_extended_config_funcs = extension_module.get_config_funcs(g_config)

    g_config.model_config.type = 'nn'

    global g_current_submodel, g_root_submodel
    g_root_submodel = g_config.model_config.sub_models.add()
    g_root_submodel.name = 'root'
    g_root_submodel.is_recurrent_layer_group = False
    g_current_submodel = g_root_submodel

    execfile(config_file, make_config_environment(config_file, config_args))
    for k, v in settings.iteritems():
        if v is None:
            continue
3389
        g_config.opt_config.__setattr__(k, v)
3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415

    for k, v in trainer_settings.iteritems():
        if v is None:
            continue
        g_config.__setattr__(k, v)

    for name in g_config.model_config.input_layer_names:
        assert name in g_layer_map, \
            'input name "%s" does not correspond to a layer name' % name
        assert (g_layer_map[name].type == "data" or g_layer_map[name].type == "data_trim"), \
            'The type of input layer "%s" is not "data"' % name
    for name in g_config.model_config.output_layer_names:
        assert name in g_layer_map, \
            'input name "%s" does not correspond to a layer name' % name
    return g_config


def parse_config_and_serialize(config_file, config_arg_str):
    try:
        config = parse_config(config_file, config_arg_str)
        #logger.info(config)
        return config.SerializeToString()
    except:
        traceback.print_exc()
        raise

3416

3417 3418 3419 3420 3421 3422 3423 3424
if __name__ == '__main__':
    try:
        config = parse_config(sys.argv[1], '')
        config.SerializeToString()
        __real_print__(str(config))
    except:
        traceback.print_exc()
        raise