config_parser.py 120.0 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]
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]

595

596 597 598 599 600 601
@config_class
class TableProjection(Projection):
    type = 'table'

    def calc_parameter_size(self, input_size, output_size):
        return input_size * output_size
602

603 604 605
    def calc_parameter_dims(self, input_size, output_size):
        return [input_size, output_size]

606

607 608 609 610 611 612
@config_class
class FullMatrixProjection(Projection):
    type = 'fc'

    def calc_parameter_size(self, input_size, output_size):
        return input_size * output_size
613

614 615 616
    def calc_parameter_dims(self, input_size, output_size):
        return [input_size, output_size]

617

618 619 620 621 622 623
@config_class
class TransposedFullMatrixProjection(Projection):
    type = 'trans_fc'

    def calc_parameter_size(self, input_size, output_size):
        return input_size * output_size
624

625 626 627
    def calc_parameter_dims(self, input_size, output_size):
        return [output_size, input_size]

628

629 630 631 632
@config_class
class ContextProjection(Projection):
    type = 'context'

633 634
    def __init__(self, input_layer_name, context_start, context_length,
                 trainable_padding, **xargs):
635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657
        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


658 659 660 661
@config_class
class ConvProjection(Projection):
    type = 'conv'

662 663 664 665 666
    def __init__(self,
                 input_layer_name,
                 num_filters=None,
                 conv_conf=None,
                 **xargs):
667 668 669 670 671
        super(ConvProjection, self).__init__(input_layer_name, **xargs)

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

672
        parse_conv(conv_conf, input_layer_name, self.proj_conf.conv_conf,
673
                   num_filters)
674
        # TODO: support rectangle input
675 676
        self.proj_conf.output_size = (self.proj_conf.conv_conf.output_x**
                                      2) * num_filters
677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693

    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
        return co * ci * fh * fw

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

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

694

695 696 697
# Define a operator for mixed layer
@config_class
class Operator(Cfg):
698 699
    type = None  # subclass should set it correctly

700 701
    def __init__(
            self,
702
            input_layer_names, ):
703 704 705 706 707 708 709 710 711 712
        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

713

714 715 716
@config_class
class DotMulOperator(Operator):
    type = 'dot_mul'
717 718 719

    def __init__(self, input_layer_names, scale=None, **xargs):
        super(DotMulOperator, self).__init__(input_layer_names, **xargs)
720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737
        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'
738 739 740 741 742 743 744

    def __init__(self,
                 input_layer_names,
                 num_filters=None,
                 conv_conf=None,
                 **xargs):
        super(ConvOperator, self).__init__(input_layer_names, **xargs)
745 746 747
        if num_filters is not None:
            self.operator_conf.num_filters = num_filters

748 749
        parse_conv(conv_conf,
                   MakeLayerNameInSubmodel(input_layer_names[0]),
750 751 752
                   self.operator_conf.conv_conf, num_filters)
        self.operator_conf.output_size = (self.operator_conf.conv_conf.output_x
                                          **2) * num_filters
753 754 755

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

756 757
    def calc_output_size(self, input_sizes):
        return self.operator_conf.output_size
758 759 760 761 762


# please refer to the comments in proto/ModelConfig.proto
@config_class
class Conv(Cfg):
763 764 765 766 767 768 769 770 771 772 773 774 775
    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):
776 777
        self.add_keys(locals())
        if filter_size_y is None:
778
            self.filter_size_y = filter_size
779
        if padding_y is None:
780
            self.padding_y = padding
781
        if stride_y is None:
782
            self.stride_y = stride
783
        if output_x is not None:
784 785
            config_assert(output_x <= 0)

786

787 788 789
# please refer to the comments in proto/ModelConfig.proto
@config_class
class BilinearInterp(Cfg):
790
    def __init__(self, out_size_x=None, out_size_y=None, num_channels=None):
791 792
        self.add_keys(locals())

793

794 795 796
# please refer to the comments in proto/ModelConfig.proto
@config_class
class Pool(Cfg):
797 798 799 800 801 802 803 804 805 806 807
    def __init__(self,
                 pool_type,
                 channels,
                 size_x,
                 size_y=None,
                 img_width=None,
                 start=None,
                 stride=None,
                 stride_y=None,
                 padding=None,
                 padding_y=None):
808
        self.add_keys(locals())
809 810


Q
qijun 已提交
811 812
# please refer to the comments in proto/ModelConfig.proto
@config_class
Q
qijun 已提交
813
class SpatialPyramidPool(Cfg):
814
    def __init__(self, pool_type, pyramid_height, channels, img_width=None):
Q
qijun 已提交
815
        self.add_keys(locals())
816

817

818 819 820
# please refer to the comments in proto/ModelConfig.proto
@config_class
class Norm(Cfg):
821 822 823 824 825 826 827 828 829
    def __init__(self,
                 norm_type,
                 channels,
                 size,
                 scale,
                 pow,
                 output_x=None,
                 img_size=None,
                 blocked=None):
830 831
        self.add_keys(locals())

832

833 834 835
# please refer to the comments in proto/ModelConfig.proto
@config_class
class Image(Cfg):
836
    def __init__(self, channels, img_size=None):
837 838
        self.add_keys(locals())

839

840 841
@config_class
class BlockExpand(Cfg):
842 843 844 845 846 847 848 849 850 851 852 853
    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):
854 855
        self.add_keys(locals())

856

857 858
@config_class
class MaxOut(Cfg):
859
    def __init__(self, channels, groups, img_size_x=0, img_size_y=0):
860 861
        self.add_keys(locals())

862

863 864 865 866 867 868 869 870 871 872 873 874 875
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)
876 877
    data_config.data_ratio = data_ratio
    data_config.is_main_data = is_main_data
878

879
    usage_ratio = default(usage_ratio, settings_deprecated["usage_ratio"])
880 881 882 883 884 885
    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

886

887
@config_func
888 889 890 891 892
def SimpleData(files=None,
               feat_dim=None,
               context_len=None,
               buffer_capacity=None,
               **xargs):
893 894 895 896 897 898 899 900 901 902
    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

903

904
@config_func
905 906 907 908 909 910 911 912 913 914
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):
915 916 917
    data_config = DataBase(**xargs)
    data_config.type = 'py'
    if load_data_module in g_py_module_name_list:
918

919 920 921
        def get_path(module):
            m = __import__(load_data_module)
            return os.path.split(os.path.realpath(m.__file__))[0]
922

923 924 925
        # 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.
926 927
        module_new_name = "%s_copy_%d" % (load_data_module,
                                          len(g_py_module_name_list))
928
        g_py_module_name_list.append(module_new_name)
929 930 931 932
        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)
933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956
        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

957

958
@config_func
959 960 961 962 963 964 965
def ProtoData(files=None,
              type=None,
              file_group_queue_capacity=None,
              load_file_count=None,
              constant_slots=None,
              load_thread_num=None,
              **xargs):
966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985
    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

986

987 988
#real data for training is actually provided by "sub_data" data providers.
@config_func
989
def MultiData(sub_data=[]):
990 991 992 993 994
    data_config = DataConfig()
    data_config.type = 'multi'
    data_config.sub_data_configs.extend(sub_data)
    return data_config

995

996
@config_func
997 998 999 1000 1001 1002 1003
def Data(type,
         files=None,
         feat_dim=None,
         slot_dims=None,
         context_len=None,
         buffer_capacity=None,
         **xargs):
1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038

    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

1039

1040
def parse_bilinear(bilinear, input_layer_name, bilinear_conf):
1041 1042 1043 1044
    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

1045

1046 1047 1048 1049
'''
caffe_mode: compute the output size using floor instead of ceil,
            which is consistent of caffe and CuDNN's convention.
'''
1050 1051


1052 1053 1054 1055 1056 1057 1058
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))

1059

1060 1061 1062 1063
'''
calcualte image_size based on output_size for convolution. 
It is the reverse function of cnn_output_size
'''
1064 1065


1066 1067 1068 1069
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:
1070
        img_size = (output_size - 2) * stride + filter_size - 2 * padding + 1
1071 1072
    return img_size

1073

1074 1075
def parse_pool(pool, input_layer_name, pool_conf):
    pool_conf.pool_type = pool.pool_type
1076 1077 1078
    config_assert(pool.pool_type in [
        'max-projection', 'avg-projection', 'cudnn-max-pool', 'cudnn-avg-pool'
    ], "pool-type %s is not in "
1079
                  "['max-projection', 'avg-projection', "
1080
                  "'cudnn-max-pool', 'cudnn-avg-pool']" % pool.pool_type)
1081 1082 1083 1084 1085 1086

    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)
1087
    pool_conf.stride_y = default(pool.stride_y, pool_conf.stride)
1088 1089

    img_pixels = g_layer_map[input_layer_name].size / pool.channels
1090 1091
    # the img_width may be removed,
    # and it can be calculated automatically later.
1092
    pool_conf.img_size = default(pool.img_width, int(img_pixels**0.5))
1093 1094
    pool_conf.img_size_y = img_pixels / pool_conf.img_size
    config_assert(pool_conf.img_size * pool_conf.img_size_y == img_pixels,
1095 1096
                  "Incorrect input image size %d for input image pixels %d" %
                  (pool_conf.img_size, img_pixels))
1097

1098
    config_assert(not pool.start, "start is deprecated in pooling.")
1099

1100
    if pool.padding is not None:
1101 1102
        pool_conf.padding = pool.padding
        pool_conf.padding_y = default(pool.padding_y, pool_conf.padding)
1103 1104 1105 1106 1107 1108 1109
        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)

1110

Q
qijun 已提交
1111 1112 1113
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'],
1114 1115
                  "pool-type %s is not in "
                  "['max-projection', 'avg-projection']" % spp.pool_type)
Q
qijun 已提交
1116 1117 1118 1119 1120
    spp_conf.pyramid_height = spp.pyramid_height
    spp_conf.channels = spp.channels

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

1121
    spp_conf.img_size = default(spp.img_width, int(img_pixels**0.5))
Q
qijun 已提交
1122 1123
    spp_conf.img_size_y = img_pixels / spp_conf.img_size
    config_assert(spp_conf.img_size * spp_conf.img_size_y == img_pixels,
1124 1125 1126
                  "Incorrect input image size %d for input image pixels %d" %
                  (spp_conf.img_size, img_pixels))

Q
qijun 已提交
1127

1128 1129 1130
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
1131 1132 1133 1134 1135
    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))

1136 1137 1138 1139

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'],
1140 1141
                  "norm-type %s is not in [rnorm, 'cmrnorm-projection']" %
                  norm.norm_type)
1142 1143 1144 1145 1146 1147 1148
    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
1149 1150 1151 1152
    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))
1153 1154 1155 1156
    norm_conf.output_x = norm_conf.img_size
    if norm.norm_type in ['cmrnorm-projection']:
        norm_conf.scale /= norm.size
    else:
1157 1158
        norm_conf.scale /= norm.size**2

1159

1160 1161 1162 1163
'''
caffe_mode: compute the output size using floor instead of ceil,
            which is consistent of caffe and CuDNN's convention.
'''
1164 1165


1166
def parse_conv(conv, input_layer_name, conv_conf, num_filters, trans=False):
1167 1168 1169 1170 1171 1172 1173 1174 1175
    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
1176

1177
    if not trans:
1178 1179
        conv_conf.filter_channels = conv.channels / conv.groups

1180
        img_pixels = g_layer_map[input_layer_name].size / conv.channels
1181 1182 1183 1184 1185 1186 1187 1188
        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))

1189
        conv_conf.output_x = cnn_output_size(
1190 1191
            conv_conf.img_size, conv_conf.filter_size, conv_conf.padding,
            conv_conf.stride, conv_conf.caffe_mode)
1192
    else:
1193
        conv_conf.filter_channels = num_filters / conv.groups
1194

1195
        outputSize = g_layer_map[input_layer_name].size / conv.channels
1196 1197 1198 1199 1200 1201 1202
        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))
1203
        conv_conf.img_size = cnn_image_size(
1204 1205 1206
            conv_conf.output_x, conv_conf.filter_size, conv_conf.padding,
            conv_conf.stride, conv_conf.caffe_mode)

1207

1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220
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:
1221
        block_expand_conf.output_x = cnn_output_size(
1222
            block_expand.img_size_x, block_expand.block_x,
1223
            block_expand.padding_x, block_expand.stride_x, False)
1224 1225

    if block_expand_conf.img_size_y == 0:
1226
        block_expand_conf.output_y = 0
1227
    else:
1228
        block_expand_conf.output_y = cnn_output_size(
1229
            block_expand.img_size_y, block_expand.block_y,
1230
            block_expand.padding_y, block_expand.stride_y, False)
1231

1232

1233 1234 1235 1236 1237
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
1238

1239

1240 1241 1242 1243 1244 1245
# Define an evaluator
@config_func
def Evaluator(
        name,
        type,
        inputs,
1246 1247 1248 1249 1250 1251 1252 1253
        chunk_scheme=None,
        num_chunk_types=None,
        classification_threshold=None,
        positive_label=None,
        dict_file=None,
        result_file=None,
        num_results=None,
        delimited=None, ):
1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267
    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)

1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280
    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
1281

1282

1283 1284 1285 1286 1287
class LayerBase(object):
    def __init__(
            self,
            name,
            type,
1288
            size,  # size can be 0. In this case, subclass should set it.
1289 1290 1291 1292
            inputs,
            device=None,
            active_type="",
            drop_rate=0.,
1293
            coeff=None):
1294
        config_assert('@' not in name,
1295
                      "layer name: %s contain special character @" % name)
1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310
        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()
1311
        assert isinstance(self.config, LayerConfig)
1312 1313 1314
        self.config.name = name
        self.config.type = type
        self.config.active_type = active_type
1315 1316
        if coeff is not None:
            self.config.coeff = float(coeff)
1317 1318 1319 1320 1321 1322 1323
        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
1324
        elif g_default_device is not None:
1325 1326 1327 1328 1329 1330 1331 1332 1333
            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(
1334 1335
                    input_layer_name=input,
                    parameter_name=gen_parameter_name(name, input_index))
1336 1337 1338 1339 1340 1341 1342 1343
                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):
1344
                self.operators.append(input)
1345 1346 1347 1348
                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:
1349
                raise ValueError('Wrong type for inputs: %s' % type_of(input))
1350
            config_assert(input_layer_name in g_layer_map,
1351 1352
                          "Unknown input layer '%s' for layer %s" %
                          (input_layer_name, name))
1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369
            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,
1370
            bias,  # True/False or BiasCfg
1371
            size,
1372 1373 1374
            dims=None,
            for_self=True,  # whether create bias for layer self
    ):
1375 1376 1377 1378 1379 1380

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

1381 1382 1383
        config_assert(
            type_of(bias) == bool or type_of(bias) == Bias,
            'Incorrect type for bias: %s' % type_of(bias))
1384 1385 1386 1387 1388 1389 1390 1391 1392

        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:
1393 1394
                assert isinstance(self.config, LayerConfig)

1395 1396 1397
                Parameter(
                    bias.parameter_name,
                    size,
1398 1399
                    self.config.device
                    if self.config.HasField('device') else None,
1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410
                    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,
1411 1412
                    gradient_clipping_threshold=bias.
                    gradient_clipping_threshold,
1413
                    is_static=bias.is_static,
1414
                    is_shared=bias.is_shared, )
1415 1416 1417 1418 1419
            if for_self:
                self.config.bias_parameter_name = bias.parameter_name
            else:
                return bias.parameter_name

1420 1421 1422 1423 1424 1425
    def create_input_parameter(self,
                               input_index,
                               size,
                               dims=None,
                               sparse=None,
                               format=None):
1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439
        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]
1440 1441
            config_assert(size == para.size, (
                'Shared parameter "%s" does not ' + 'have same size: %s vs. %s')
1442 1443
                          % (input_config.parameter_name, para.size, size))

1444 1445
            config_assert(dims == para.dims, (
                'Shared parameter "%s" does not ' + 'have same dims: %s vs. %s')
1446 1447 1448 1449 1450 1451
                          % (input_config.parameter_name, para.dims, dims))
            return

        Parameter(
            input_config.parameter_name,
            size,
1452
            self.config.device if self.config.HasField("device") else None,
1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464
            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,
1465 1466
            gradient_clipping_threshold=input_config.
            gradient_clipping_threshold,
1467 1468 1469 1470
            sparse=sparse,
            format=format,
            is_static=input_config.is_static,
            is_shared=input_config.is_shared,
1471
            update_hooks=input_config.update_hooks)
1472 1473 1474 1475 1476 1477 1478 1479 1480

    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)

1481

1482 1483
@config_layer('multi_class_cross_entropy_with_selfnorm')
class MultiClassCrossEntropySelfNormCostLayer(LayerBase):
1484 1485 1486
    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)
1487 1488
        self.config.softmax_selfnorm_alpha = softmax_selfnorm_alpha

1489

1490 1491
@config_layer('fc')
class FCLayer(LayerBase):
1492
    def __init__(self, name, size, inputs, bias=True, **xargs):
1493 1494 1495 1496 1497 1498 1499 1500 1501 1502
        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
1503 1504
            else:
                sparse = None
1505

1506 1507
            self.create_input_parameter(input_index, psize, dims, sparse,
                                        format)
1508 1509
        self.create_bias_parameter(bias, self.config.size)

1510

1511 1512
@config_layer('selective_fc')
class SelectiveFCLayer(LayerBase):
1513 1514 1515 1516 1517 1518 1519 1520 1521 1522
    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):
1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542
        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,
1543 1544
                          ("if indices of selected columns are not specified, "
                           "selective_fc Layer has at least two inputs"))
1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556
            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

1557 1558
            self.create_input_parameter(input_index, psize, dims, sparse,
                                        format)
1559 1560
        self.create_bias_parameter(bias, self.config.size)

1561

1562 1563
@config_layer('print')
class PrintLayer(LayerBase):
1564
    def __init__(self, name, inputs):
1565 1566
        super(PrintLayer, self).__init__(name, 'print', 0, inputs)

1567

1568 1569
@config_layer('data')
class DataLayer(LayerBase):
1570 1571 1572 1573
    def __init__(self, name, size, device=None):
        super(DataLayer, self).__init__(
            name, 'data', size, inputs=[], device=device)

1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600

'''
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
'''
1601 1602


1603 1604
@config_layer('data_norm')
class DataNormLayer(LayerBase):
1605
    def __init__(self, name, inputs, data_norm_strategy="z-score", device=None):
1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616
        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)

1617

1618 1619 1620
@config_layer('prelu')
class ParameterReluLayer(LayerBase):
    layer_type = 'prelu'
1621 1622

    def __init__(self, name, inputs, partial_sum=1, **args):
1623 1624 1625 1626 1627 1628 1629 1630
        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)

1631

1632 1633 1634
@config_layer('conv')
class ConvLayerBase(LayerBase):
    layer_type = 'conv'
1635 1636 1637 1638 1639 1640 1641 1642

    def __init__(self,
                 name,
                 inputs=[],
                 bias=True,
                 num_filters=None,
                 shared_biases=False,
                 **xargs):
1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658
        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
1659
            (parallel_nn == 0 or self.config.device > -1)):
1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670
            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)
1671 1672
            parse_conv(self.inputs[input_index].conv, input_layer.name,
                       self.config.inputs[input_index].conv_conf, num_filters)
1673 1674 1675 1676 1677
            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(
1678
                (conv_conf.output_x**2) * self.config.num_filters)
1679 1680 1681 1682 1683 1684 1685 1686 1687 1688

        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)

1689

1690 1691 1692 1693
@config_layer('exconv')
class ConvLayer(ConvLayerBase):
    layer_type = 'exconv'

1694

1695 1696 1697 1698
@config_layer('cudnn_conv')
class ConvLayer(ConvLayerBase):
    layer_type = 'cudnn_conv'

1699 1700 1701 1702

@config_layer('convt')
class ConvTransLayerBase(LayerBase):
    layer_type = 'convt'
1703 1704 1705 1706 1707 1708 1709 1710

    def __init__(self,
                 name,
                 inputs=[],
                 bias=True,
                 num_filters=None,
                 shared_biases=False,
                 **xargs):
1711
        super(ConvTransLayerBase, self).__init__(
1712 1713 1714 1715 1716 1717 1718 1719
            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))

1720 1721
        # cudnn_convt has not been implemented so use exconvt only
        self.layer_type = "exconvt"
1722 1723 1724 1725 1726 1727 1728 1729
        # 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)
1730
            parse_conv(
1731 1732
                self.inputs[input_index].conv,
                input_layer.name,
1733
                self.config.inputs[input_index].conv_conf,
1734
                num_filters,
1735
                trans=True)
1736 1737 1738 1739 1740
            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(
1741
                (conv_conf.img_size**2) * self.config.num_filters)
1742 1743 1744 1745 1746 1747 1748

        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):
1749
        return conv_conf.channels * conv_conf.filter_channels \
1750 1751
                    * (conv_conf.filter_size * conv_conf.filter_size_y)

1752

1753 1754 1755 1756
@config_layer('exconvt')
class ConvTransLayer(ConvTransLayerBase):
    layer_type = 'exconvt'

1757

1758 1759
@config_layer('norm')
class NormLayer(LayerBase):
1760 1761 1762
    def __init__(self, name, inputs, device=None):
        super(NormLayer, self).__init__(
            name, 'norm', 0, inputs=inputs, device=device)
1763 1764
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
1765 1766
            parse_norm(self.inputs[input_index].norm, input_layer.name,
                       self.config.inputs[input_index].norm_conf)
1767
            norm_conf = self.config.inputs[input_index].norm_conf
1768 1769
            self.set_layer_size((norm_conf.output_x**2) * norm_conf.channels)

1770 1771 1772

@config_layer('pool')
class PoolLayer(LayerBase):
1773 1774 1775
    def __init__(self, name, inputs, device=None):
        super(PoolLayer, self).__init__(
            name, 'pool', 0, inputs=inputs, device=device)
1776 1777
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
1778 1779
            parse_pool(self.inputs[input_index].pool, input_layer.name,
                       self.config.inputs[input_index].pool_conf)
1780
            pool_conf = self.config.inputs[input_index].pool_conf
1781 1782 1783 1784 1785
            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)

1786

Q
qijun 已提交
1787 1788
@config_layer('spp')
class SpatialPyramidPoolLayer(LayerBase):
1789 1790 1791
    def __init__(self, name, inputs, device=None):
        super(SpatialPyramidPoolLayer, self).__init__(
            name, 'spp', 0, inputs=inputs, device=device)
Q
qijun 已提交
1792 1793
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
1794 1795
            parse_spp(self.inputs[input_index].spp, input_layer.name,
                      self.config.inputs[input_index].spp_conf)
Q
qijun 已提交
1796 1797 1798 1799 1800
            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)

1801

1802 1803 1804
@config_layer('batch_norm')
class BatchNormLayer(LayerBase):
    layer_type = 'batch_norm'
1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815

    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):
1816 1817 1818 1819
        if inputs is None:
            inputs = []
        elif not isinstance(inputs, list):
            inputs = [inputs]
1820 1821
        config_assert(
            len(inputs) == 1, "BatchNormLayer must have one and only one input")
1822 1823 1824 1825 1826 1827 1828 1829
        # 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):
1830 1831 1832 1833 1834 1835 1836
            inputs.append(
                Input(
                    inputs[0].input_layer_name,
                    initial_std=0.0,
                    initial_mean=0.0,
                    is_static=True,
                    is_shared=is_shared, ))
1837 1838 1839 1840 1841 1842 1843

        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 \
1844
            cudnn_version >= 4007
1845
        self.layer_type = "cudnn_batch_norm" if use_cudnn else "batch_norm"
1846 1847 1848 1849 1850 1851 1852 1853
        super(BatchNormLayer, self).__init__(
            name,
            self.layer_type,
            0,
            active_type=active_type,
            inputs=inputs,
            device=device,
            **xargs)
1854 1855 1856 1857 1858 1859

        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

1860 1861
        input_layer = self.get_input_layer(0)
        parse_image(self.inputs[0].image, input_layer.name,
1862 1863
                    self.config.inputs[0].image_conf)
        image_conf = self.config.inputs[0].image_conf
1864
        self.set_layer_size((image_conf.img_size**2) * image_conf.channels)
1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876

        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

1877

1878 1879
@config_layer('trans')
class TransLayer(LayerBase):
1880 1881 1882 1883 1884 1885
    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')
1886 1887
        self.set_layer_size(self.get_input_layer(0).size)

1888

1889 1890
@config_layer('resize')
class ResizeLayer(LayerBase):
1891 1892 1893 1894 1895 1896 1897
    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')

1898 1899 1900

@config_layer('blockexpand')
class BlockExpandLayer(LayerBase):
1901 1902 1903
    def __init__(self, name, inputs, device=None):
        super(BlockExpandLayer, self).__init__(
            name, 'blockexpand', 0, inputs=inputs, device=device)
1904 1905
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
1906 1907
            parse_block_expand(
                self.inputs[input_index].block_expand, input_layer.name,
1908
                self.config.inputs[input_index].block_expand_conf)
1909 1910 1911 1912 1913 1914
            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)

1915

1916 1917
@config_layer('maxout')
class MaxOutLayer(LayerBase):
1918 1919 1920
    def __init__(self, name, inputs, **xargs):
        super(MaxOutLayer, self).__init__(
            name, 'maxout', 0, inputs=inputs, **xargs)
1921
        input_layer = self.get_input_layer(0)
1922
        parse_maxout(self.inputs[0].maxout, input_layer.name,
1923 1924
                     self.config.inputs[0].maxout_conf)
        maxout_conf = self.config.inputs[0].maxout_conf
1925 1926 1927
        self.set_layer_size(g_layer_map[input_layer.name].size /
                            maxout_conf.groups)

1928

1929 1930 1931 1932
# key: cost type
# value: cost class
g_cost_map = {}

1933

1934 1935 1936
# define a cost layer without any parameters
def define_cost(class_name, cost_type):
    def init(cls, name, inputs, device=None, coeff=1.):
1937 1938
        super(type(cls), cls).__init__(
            name, cost_type, 1, inputs, device=device, coeff=coeff)
1939

1940
    cls = type(class_name, (LayerBase, ), dict(__init__=init))
1941 1942 1943
    global g_cost_map
    g_cost_map[cost_type] = cls

1944

1945 1946 1947 1948 1949 1950 1951 1952
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
1953
define_cost('SumCost', 'sum_cost')
1954

1955

1956 1957
@config_layer('hsigmoid')
class HierarchicalSigmoidLayer(LayerBase):
1958
    def __init__(self, name, num_classes, inputs, device=None, bias=True):
1959 1960
        super(HierarchicalSigmoidLayer, self).__init__(
            name, 'hsigmoid', 1, inputs=inputs, device=device)
1961 1962 1963
        config_assert(
            len(self.inputs) >= 2,
            'HierarchicalSigmoidLayer must have at least 2 inputs')
1964 1965 1966 1967 1968 1969 1970 1971
        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)

1972

1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996
'''
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.
'''
1997 1998


1999 2000
@config_layer('lambda_cost')
class LambdaCost(LayerBase):
2001
    def __init__(self, name, inputs, NDCG_num=5, max_sort_size=-1, device=None):
2002 2003
        super(LambdaCost, self).__init__(
            name, 'lambda_cost', 1, inputs=inputs, device=device)
2004
        config_assert(len(self.inputs) == 2, 'lambdaCost must have 2 inputs')
2005 2006
        self.config.NDCG_num = NDCG_num
        if max_sort_size != -1:
2007 2008 2009
            config_assert(
                NDCG_num <= max_sort_size,
                'NDCG_num must be less than or equal to max_sort_size')
2010 2011
        self.config.max_sort_size = max_sort_size

2012

2013 2014
@config_layer('nce')
class NCELayer(LayerBase):
2015 2016 2017 2018 2019 2020 2021 2022
    def __init__(self,
                 name,
                 num_classes,
                 inputs,
                 num_neg_samples=10,
                 neg_sampling_dist=None,
                 bias=True,
                 **xargs):
2023
        super(NCELayer, self).__init__(name, 'nce', 1, inputs=inputs, **xargs)
2024 2025
        config_assert(
            len(self.inputs) >= 2, 'NCELayer must have at least 2 inputs')
2026 2027
        self.config.num_classes = num_classes
        if neg_sampling_dist is not None:
2028 2029 2030 2031
            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))
2032
            s = sum(neg_sampling_dist)
2033 2034 2035
            config_assert(
                abs(s - 1) < 1e-5,
                'The sum of neg_sampling_dist (%s) is not 1' % s)
2036 2037 2038 2039 2040

            self.config.neg_sampling_dist.extend(neg_sampling_dist)

        self.config.num_neg_samples = num_neg_samples
        num_real_inputs = len(self.inputs) - 1
2041
        input_layer = self.get_input_layer(num_real_inputs)
2042 2043 2044 2045
        config_assert(input_layer.type == 'data',
                      'Expecting the last input layer of an nce layer to be '
                      'a data layer')

2046 2047
        if (num_real_inputs > 1 and input_layer.size == 1 and
                self.get_input_layer(num_real_inputs - 1).type == 'data'):
2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060
            # 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):
2061
    def __init__(self, name, inputs, bias=True, **xargs):
2062 2063
        super(AddToLayer, self).__init__(
            name, 'addto', 0, inputs=inputs, **xargs)
2064
        config_assert(len(inputs) > 0, 'inputs cannot be empty for AddToLayer')
2065 2066 2067 2068 2069
        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)

2070

2071 2072
@config_layer('agent')
class AgentLayer(LayerBase):
2073 2074 2075 2076
    def __init__(self, name, size, device=None):
        super(AgentLayer, self).__init__(
            name, 'agent', size, inputs=[], device=device)

2077 2078 2079

@config_layer('sequence_agent')
class SequenceAgentLayer(LayerBase):
2080
    def __init__(self, name, size, device=None):
2081 2082 2083
        super(SequenceAgentLayer, self).__init__(
            name, 'sequence_agent', size, inputs=[], device=device)

2084

2085 2086
@config_layer('gather_agent')
class GatherAgentLayer(LayerBase):
2087
    def __init__(self, name, size, device=None):
2088 2089 2090
        super(GatherAgentLayer, self).__init__(
            name, 'gather_agent', size, inputs=[], device=device)

2091

2092 2093
@config_layer('scatter_agent')
class ScatterAgentLayer(LayerBase):
2094
    def __init__(self, name, size, device=None):
2095 2096 2097
        super(ScatterAgentLayer, self).__init__(
            name, 'scatter_agent', size, inputs=[], device=device)

2098

2099 2100
@config_layer('sequence_gather_agent')
class SequenceGatherAgentLayer(LayerBase):
2101
    def __init__(self, name, size, device=None):
2102
        super(SequenceGatherAgentLayer, self).__init__(
2103 2104
            name, 'sequence_gather_agent', size, inputs=[], device=device)

2105 2106 2107

@config_layer('sequence_scatter_agent')
class SequenceScatterAgentLayer(LayerBase):
2108
    def __init__(self, name, size, device=None):
2109
        super(SequenceScatterAgentLayer, self).__init__(
2110 2111
            name, 'sequence_scatter_agent', size, inputs=[], device=device)

2112 2113 2114

@config_layer('multiplex')
class MultiplexLayer(LayerBase):
2115 2116 2117 2118 2119
    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.')
2120
        for i in range(1, len(inputs)):
2121 2122 2123 2124 2125
            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.")

2126 2127

@config_func
2128 2129 2130
def Link(
        name,
        has_subseq=False, ):
2131 2132 2133 2134 2135
    link_config = LinkConfig()
    link_config.link_name = name
    link_config.has_subseq = has_subseq
    return link_config

2136

2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150
# 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
2151 2152 2153 2154 2155 2156 2157 2158
def Memory(
        name,
        size,
        is_sequence=False,
        boot_layer=None,
        boot_bias=False,
        boot_bias_active_type="",
        boot_with_const_id=None, ):
2159 2160 2161 2162 2163 2164
    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,
2165
                  'Memory should be used in recurrent layer group only')
2166 2167 2168 2169
    memory = g_current_submodel.memories.add()
    memory.layer_name = MakeLayerNameInSubmodel(name)
    memory.link_name = MakeLayerNameInSubmodel(agent_name)
    memory.is_sequence = is_sequence
2170
    options = sum((boot_layer is not None, bool(boot_bias),
2171
                   boot_with_const_id is not None))
2172 2173 2174 2175
    config_assert(
        options <= 1,
        'take one option at most from boot_layer, boot_bias, or boot_with_const_id'
    )
2176 2177 2178
    if boot_layer is not None:
        boot_layer = MakeLayerNameInParentSubmodel(boot_layer)
        config_assert(boot_layer in g_layer_map,
2179 2180
                      'boot_layer "%s" does not correspond to a layer name' %
                      boot_layer)
2181 2182 2183
        memory.boot_layer_name = boot_layer
    elif boot_bias:
        memory.boot_bias_parameter_name = agent_layer.create_bias_parameter(
2184
            boot_bias, size, for_self=False)
2185 2186 2187 2188 2189
        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

2190

2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201
# 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,
2202 2203 2204 2205
        eos_layer_name="eos_check",
        num_results_per_sample=1,
        beam_size=1,
        log_prob=None, ):
2206 2207 2208 2209 2210 2211 2212 2213 2214
    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

2215

2216 2217
@config_layer('expand')
class ExpandLayer(LayerBase):
2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233
    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)

2234 2235 2236

@config_layer('featmap_expand')
class FeatMapExpandLayer(LayerBase):
2237 2238 2239 2240 2241 2242
    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:
2243
            self.config.num_filters = num_filters
2244
        else:
2245
            logger.fatal("FeatMapExpandLayer must specify num_filters.")
2246
        self.set_layer_size(self.get_input_layer(0).size * num_filters)
2247 2248 2249 2250


@config_layer('max')
class MaxLayer(LayerBase):
2251 2252 2253 2254 2255 2256 2257 2258 2259 2260
    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)
2261
        config_assert(len(self.inputs) == 1, 'MaxLayer must have 1 input')
2262 2263
        self.config.trans_type = trans_type
        self.config.active_type = active_type
2264 2265 2266 2267
        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)
2268 2269
        if output_max_index is not None:
            self.config.output_max_index = output_max_index
2270 2271 2272 2273


@config_layer('maxid')
class MaxIdLayer(LayerBase):
2274
    def __init__(self, name, inputs, beam_size=None, device=None):
2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291
        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):
2292
    def __init__(self, name, inputs, eos_id, device=None):
2293 2294 2295
        super(EosIdLayer, self).__init__(
            name, 'eos_id', 0, inputs=inputs, device=device)
        config_assert(len(self.inputs) == 1, 'EosIdLayer must have 1 input')
2296
        self.set_layer_size(2)  # boolean output
2297 2298
        self.config.eos_id = eos_id

2299

2300 2301
@config_layer('seqlastins')
class SequenceLastInstanceLayer(LayerBase):
2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318
    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
2319 2320 2321 2322 2323
        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)

2324

2325 2326 2327 2328 2329 2330 2331 2332 2333
@config_layer('seqfirstins')
class SequenceFirstInstanceLayer(SequenceLastInstanceLayer):
    def __init__(
            self,
            name,
            inputs,
            active_type='linear',
            trans_type='non-seq',
            device=None,
2334 2335 2336 2337 2338 2339 2340 2341
            bias=False, ):
        super(SequenceFirstInstanceLayer, self).__init__(
            name,
            inputs=inputs,
            active_type=active_type,
            device=device,
            bias=bias)
        self.config.trans_type = trans_type
2342 2343
        self.config.select_first = True

2344

2345 2346
@config_layer('seqconcat')
class SequenceConcatLayer(LayerBase):
2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361
    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')
2362 2363 2364 2365 2366
        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)

2367

2368 2369
@config_layer('seqreshape')
class SequenceReshapeLayer(LayerBase):
2370 2371 2372 2373 2374 2375 2376 2377 2378 2379
    def __init__(self,
                 name,
                 size,
                 inputs,
                 active_type='linear',
                 device=None,
                 bias=False):
        super(SequenceReshapeLayer, self).__init__(
            name,
            'seqreshape',
2380
            size,
2381 2382 2383 2384 2385
            inputs=inputs,
            device=device,
            active_type=active_type)
        config_assert(
            len(inputs) == 1, 'SequenceReshapeLayer must have 1 inputs')
2386 2387 2388
        self.set_layer_size(size)
        self.create_bias_parameter(bias, size)

2389

2390 2391
@config_layer('subseq')
class SubSequenceLayer(LayerBase):
2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404
    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)
2405 2406 2407 2408 2409 2410
        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)

2411

2412 2413
@config_layer('out_prod')
class OuterProdLayer(LayerBase):
2414 2415 2416
    def __init__(self, name, inputs, device=None):
        super(OuterProdLayer, self).__init__(
            name, 'out_prod', 0, inputs=inputs, device=device)
2417 2418 2419 2420 2421
        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)

2422

2423 2424
@config_layer('power')
class PowerLayer(LayerBase):
2425 2426 2427
    def __init__(self, name, inputs, device=None):
        super(PowerLayer, self).__init__(
            name, 'power', 0, inputs=inputs, device=device)
2428 2429 2430 2431
        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)
2432 2433 2434
        config_assert(1 == input_layer0.size,
                      'The left input is the exponent and should be of size 1')

2435 2436 2437

@config_layer('slope_intercept')
class SlopeInterceptLayer(LayerBase):
2438 2439 2440
    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)
2441 2442 2443 2444 2445 2446
        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)

2447

2448 2449
@config_layer('scaling')
class ScalingLayer(LayerBase):
2450 2451 2452
    def __init__(self, name, inputs, device=None):
        super(ScalingLayer, self).__init__(
            name, 'scaling', 0, inputs=inputs, device=device)
2453 2454 2455 2456
        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)
2457 2458 2459
        config_assert(1 == input_layer0.size,
                      'The left input should be of size 1')

2460 2461 2462

@config_layer('conv_shift')
class ConvShiftLayer(LayerBase):
2463 2464 2465
    def __init__(self, name, inputs, device=None):
        super(ConvShiftLayer, self).__init__(
            name, 'conv_shift', 0, inputs=inputs, device=device)
2466 2467 2468 2469
        config_assert(len(inputs) == 2, 'ConvShiftLayer must have 2 inputs')
        input_layer0 = self.get_input_layer(0)
        self.set_layer_size(input_layer0.size)

2470

2471 2472
@config_layer('convex_comb')
class ConvexCombinationLayer(LayerBase):
2473
    def __init__(self, name, size, inputs, device=None):
2474
        super(ConvexCombinationLayer, self).__init__(
2475 2476 2477
            name, 'convex_comb', size, inputs=inputs, device=device)
        config_assert(
            len(self.inputs) == 2, 'ConvexCombinationLayer must have 2 inputs')
2478 2479 2480
        config_assert(
            size * self.get_input_layer(0).size == self.get_input_layer(1).size,
            'Wrong input size for ConvexCombinationLayer')
2481 2482
        self.set_layer_size(size)

2483

2484 2485
@config_layer('interpolation')
class InterpolationLayer(LayerBase):
2486
    def __init__(self, name, inputs, device=None):
2487 2488
        super(InterpolationLayer, self).__init__(
            name, 'interpolation', 0, inputs=inputs, device=device)
2489 2490
        config_assert(
            len(self.inputs) == 3, 'InterpolationLayer must have 3 inputs')
2491 2492 2493 2494 2495 2496 2497 2498
        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')

2499

2500 2501
@config_layer('bilinear_interp')
class BilinearInterpLayer(LayerBase):
2502
    def __init__(self, name, inputs, **xargs):
2503
        super(BilinearInterpLayer, self).__init__(
L
liaogang 已提交
2504
            name, 'bilinear_interp', 0, inputs=inputs, **xargs)
2505
        input_layer = self.get_input_layer(0)
2506 2507
        parse_bilinear(self.inputs[0].bilinear_interp, input_layer.name,
                       self.config.inputs[0].bilinear_interp_conf)
L
liaogang 已提交
2508
        conf = self.inputs[0].bilinear_interp
2509 2510 2511
        self.set_layer_size(conf.out_size_x * conf.out_size_y *
                            conf.num_channels)

L
liaogang 已提交
2512

2513 2514
@config_layer('sum_to_one_norm')
class SumToOneNormLayer(LayerBase):
2515
    def __init__(self, name, inputs, device=None):
2516
        super(SumToOneNormLayer, self).__init__(
2517 2518 2519
            name, 'sum_to_one_norm', 0, inputs=inputs, device=device)
        config_assert(
            len(self.inputs) == 1, 'SumToOneNormLayer must have 1 input')
2520 2521 2522
        input_layer0 = self.get_input_layer(0)
        self.set_layer_size(input_layer0.size)

2523

2524 2525
@config_layer('cos_vm')
class CosSimVecMatLayer(LayerBase):
2526
    def __init__(self, name, size, inputs, cos_scale=1.0, device=None):
2527
        super(CosSimVecMatLayer, self).__init__(
2528
            name, 'cos_vm', size, inputs=inputs, device=device)
2529
        self.config.cos_scale = cos_scale
2530 2531
        config_assert(
            len(self.inputs) == 2, 'CosSimVecMatLayer must have 2 inputs')
2532 2533 2534
        config_assert(
            size * self.get_input_layer(0).size == self.get_input_layer(1).size,
            'Wrong input size for CosSimVecMatLayer')
2535

2536

2537 2538
@config_layer('sampling_id')
class SamplingIdLayer(LayerBase):
2539
    def __init__(self, name, inputs, device=None):
2540 2541
        super(SamplingIdLayer, self).__init__(
            name, 'sampling_id', 0, inputs=inputs, device=device)
2542 2543
        config_assert(
            len(self.inputs) == 1, 'SamplingIdLayer must have 1 input')
2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555
        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):
2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570
    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)
2571
        self.config.average_strategy = average_strategy
2572
        self.config.trans_type = trans_type
2573 2574 2575 2576 2577 2578
        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)

2579

2580 2581
@config_layer('cos')
class CosSimLayer(LayerBase):
2582
    def __init__(self, name, inputs, cos_scale=5, device=None):
2583 2584 2585 2586 2587 2588
        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')
2589
        self.config.cos_scale = cos_scale
2590 2591 2592 2593


@config_layer('tensor')
class TensorLayer(LayerBase):
2594 2595 2596
    def __init__(self, name, size, inputs, device=None, bias=True, **xargs):
        super(TensorLayer, self).__init__(
            name, 'tensor', size, inputs=inputs, device=device, **xargs)
2597 2598
        config_assert(len(self.inputs) == 2, 'TensorLayer must have 2 inputs')
        config_assert(size > 0, 'size must be positive')
2599 2600
        config_assert(inputs[1].parameter_name == None,
                      'second parameter should be None.')
2601 2602 2603 2604 2605 2606 2607 2608 2609 2610
        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):
2611 2612 2613 2614 2615 2616 2617
    def __init__(self,
                 name,
                 inputs,
                 size=0,
                 bias=True,
                 error_clipping_threshold=None,
                 **xargs):
2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634
        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)
2635
            if self.config.size == 0:
2636 2637 2638
                size = operator.calc_output_size(operator_conf.input_sizes)
                if size != 0:
                    self.set_layer_size(size)
2639
            else:
2640 2641
                sz = operator.calc_output_size(operator_conf.input_sizes)
                if sz != 0:
2642 2643 2644 2645
                    config_assert(
                        sz == self.config.size,
                        "different inputs have different size: %s vs. %s" %
                        (sz, self.config.size))
2646 2647 2648 2649
        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:
2650 2651 2652
                config_assert(
                    isinstance(input, Projection),
                    "input should be projection or operation")
2653
            if self.config.size == 0 and isinstance(input, Projection):
2654 2655 2656
                size = input.calc_output_size(input_layer)
                if size != 0:
                    self.set_layer_size(size)
2657
            elif isinstance(input, Projection):
2658 2659 2660 2661 2662 2663
                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))
2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674
        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)
2675 2676
                input_config.proj_conf.name = gen_parameter_name(name,
                                                                 input_index)
2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687
                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)

2688 2689 2690 2691 2692 2693
        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()
2694

2695 2696 2697
        if bias:
            self.config.bias_size = psize
            self.create_bias_parameter(bias, psize)
2698

2699 2700
        if error_clipping_threshold is not None:
            self.config.error_clipping_threshold = error_clipping_threshold
2701

2702

2703 2704
# like MixedLayer, but no bias parameter
@config_func
2705
def ExpressionLayer(name, inputs, **xargs):
2706 2707
    MixedLayer(name, inputs, bias=False, **xargs)

2708

2709 2710
@config_layer('concat')
class ConcatenateLayer(LayerBase):
2711
    def __init__(self, name, inputs, bias=False, **xargs):
2712
        config_assert(inputs, 'inputs cannot be empty')
2713
        config_assert(not bias, 'ConcatenateLayer cannot support bias.')
2714 2715 2716 2717 2718 2719
        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]
2720
            if self.config.size == 0:
2721 2722 2723 2724
                size += input_layer.size

        self.set_layer_size(size)

2725

2726 2727 2728
# like concat layer, but each input layer was processed by a Projection.
@config_layer('concat2')
class ConcatenateLayer2(LayerBase):
2729
    def __init__(self, name, inputs, bias=False, **xargs):
2730 2731 2732
        config_assert(inputs, 'inputs cannot be empty')
        super(ConcatenateLayer2, self).__init__(
            name, 'concat2', 0, inputs=inputs, **xargs)
2733 2734

        if isinstance(self.inputs[0], ConvProjection):
2735 2736 2737 2738 2739 2740
            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.")
2741

2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761
        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,
2762
                                              input.proj_conf.output_size)
2763
            dims = input.calc_parameter_dims(input.proj_conf.input_size,
2764
                                             input.proj_conf.output_size)
2765 2766
            self.create_input_parameter(input_index, psize, dims)

2767 2768 2769 2770 2771 2772 2773
        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()

2774 2775 2776
        if bias:
            self.config.bias_size = psize
            self.create_bias_parameter(bias, psize)
2777

2778

2779 2780
@config_layer('recurrent')
class RecurrentLayer(LayerBase):
2781 2782 2783
    def __init__(self, name, inputs, reversed=False, bias=True, **xargs):
        super(RecurrentLayer, self).__init__(name, 'recurrent', 0, inputs, **
                                             xargs)
2784 2785 2786 2787 2788 2789 2790 2791 2792
        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)

2793

2794 2795
@config_layer('lstmemory')
class LstmLayer(LayerBase):
2796 2797 2798 2799 2800 2801 2802 2803
    def __init__(self,
                 name,
                 inputs,
                 reversed=False,
                 active_gate_type="sigmoid",
                 active_state_type="sigmoid",
                 bias=True,
                 **xargs):
2804 2805 2806 2807 2808 2809 2810 2811
        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
2812
        self.config.active_gate_type = active_gate_type
2813 2814 2815 2816 2817
        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)

2818

2819 2820
@config_layer('lstm_step')
class LstmStepLayer(LayerBase):
2821 2822 2823 2824 2825 2826 2827 2828 2829 2830
    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)
2831 2832 2833
        config_assert(len(inputs) == 2, 'LstmStepLayer must have 2 inputs')
        input_layer0 = self.get_input_layer(0)
        input_layer1 = self.get_input_layer(1)
2834 2835 2836 2837 2838
        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
2839 2840 2841
        self.config.active_state_type = active_state_type
        self.create_bias_parameter(bias, size * 3)

2842

2843 2844 2845
# get the specific output from the input layer.
@config_layer('get_output')
class GetOutputLayer(LayerBase):
2846 2847 2848 2849
    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')
2850 2851 2852 2853
        inputs = self.inputs[0]
        config_assert(inputs.input_layer_argument,
                      'input_layer_argument cannot be empty')

2854

2855 2856
@config_layer('mdlstmemory')
class MDLstmLayer(LayerBase):
2857 2858 2859 2860 2861 2862 2863 2864 2865 2866
    def __init__(self,
                 name,
                 inputs,
                 directions=True,
                 active_gate_type="sigmoid",
                 active_state_type="sigmoid",
                 bias=True,
                 **xargs):
        super(MDLstmLayer, self).__init__(name, 'mdlstmemory', 0, inputs, **
                                          xargs)
2867 2868 2869 2870
        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)
2871 2872 2873
        config_assert(input_layer.size %
                      (3 + dim_num) == 0, "size % (dim_num) should be 0!")
        size = input_layer.size / (3 + dim_num)
2874
        self.set_layer_size(size)
2875
        self.config.active_gate_type = active_gate_type
2876 2877 2878
        self.config.active_state_type = active_state_type
        for i in xrange(len(directions)):
            self.config.directions.append(int(directions[i]))
2879 2880
        self.create_input_parameter(0, size * size *
                                    (3 + dim_num), [size, size, 3 + dim_num])
2881
        #bias includes 3 kinds of peephole, 3+dim_num+2+dim_num
2882 2883
        self.create_bias_parameter(bias, size * (5 + 2 * dim_num))

2884 2885 2886

@config_layer('gated_recurrent')
class GatedRecurrentLayer(LayerBase):
2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897
    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')
2898 2899 2900 2901 2902 2903
        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
2904
        self.config.active_gate_type = active_gate_type
2905 2906 2907
        self.create_input_parameter(0, size * size * 3, [size, size * 3])
        self.create_bias_parameter(bias, size * 3)

2908

2909 2910
@config_layer('gru_step')
class GruStepLayer(LayerBase):
2911 2912 2913 2914 2915 2916 2917 2918 2919
    def __init__(self,
                 name,
                 size,
                 inputs,
                 active_gate_type="sigmoid",
                 bias=True,
                 **xargs):
        super(GruStepLayer, self).__init__(name, 'gru_step', size, inputs, **
                                           xargs)
2920 2921 2922
        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)
2923 2924 2925 2926 2927
        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
2928 2929 2930
        self.create_input_parameter(0, size * size * 3, [size, size * 3])
        self.create_bias_parameter(bias, size * 3)

2931

2932 2933 2934 2935 2936 2937 2938
'''
 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
'''
2939 2940


2941 2942
@config_layer('crf')
class CRFLayer(LayerBase):
2943
    def __init__(self, name, size, inputs, coeff=1.0, device=None):
2944
        super(CRFLayer, self).__init__(name, 'crf', size, inputs, device=device)
2945 2946
        config_assert(2 <= len(self.inputs) <= 3,
                      'CRFLayer must have 2 or 3 inputs')
2947 2948 2949
        self.create_input_parameter(0, size * (size + 2), [size, size + 2])
        self.config.coeff = coeff

2950

2951 2952 2953 2954 2955 2956 2957 2958
'''
 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
'''
2959 2960


2961 2962
@config_layer('crf_decoding')
class CRFDecodingLayer(LayerBase):
2963
    def __init__(self, name, size, inputs, device=None):
2964 2965 2966 2967 2968 2969 2970
        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])

2971

2972 2973
@config_layer('ctc')
class CTCLayer(LayerBase):
2974
    def __init__(self, name, size, inputs, norm_by_times=False, device=None):
2975 2976 2977 2978
        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')

2979

2980 2981
@config_layer('recurrent_layer_group')
class RecurrentLayerGroup(LayerBase):
2982
    def __init__(self, name, device=None):
2983 2984 2985 2986 2987 2988
        super(RecurrentLayerGroup, self).__init__(
            name, 'recurrent_layer_group', 0, inputs=[], device=device)


# Deprecated, use a new layer specific class instead
@config_func
2989
def Layer(name, type, **xargs):
2990 2991 2992 2993
    layers = {}
    layers.update(g_cost_map)
    layers.update(g_layer_type_map)
    layer_func = layers.get(type)
2994
    config_assert(layer_func, "layer type '%s' not supported." % type)
2995
    return layer_func(name, **xargs)
2996

2997

2998
@config_func
2999
def ParameterHook(type, **kwargs):
3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011
    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
3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033
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):
3034 3035 3036 3037 3038 3039 3040

    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
3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051
    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)

3052 3053
    config_assert(not momentum or not decay_rate_l1,
                  "momentum and decay_rate_l1 cannot both be non-zero")
3054 3055 3056 3057 3058

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

3059 3060 3061 3062
    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)
3063

3064 3065
    num_batches_regularization = default(num_batches_regularization,
                                         g_default_num_batches_regularization)
3066 3067 3068
    if num_batches_regularization is not None:
        para.num_batches_regularization = int(num_batches_regularization)

3069 3070 3071 3072 3073 3074
    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
3075 3076
    gradient_clipping_threshold = default(gradient_clipping_threshold,
                                          g_default_gradient_clipping_threshold)
3077 3078
    if gradient_clipping_threshold is not None:
        para.gradient_clipping_threshold = gradient_clipping_threshold
3079 3080
    para.initial_strategy = default(initial_strategy,
                                    g_default_initial_strategy)
3081 3082 3083 3084 3085 3086
    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:
3087 3088 3089
            print(
                "Use initial_smart, but dims not set. Initial_smart may not be used in this layer"
            )
3090 3091 3092 3093
            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)
3094 3095 3096 3097 3098 3099 3100

    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
3101 3102 3103 3104
    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")
3105 3106
    if is_shared is not None:
        para.is_shared = is_shared
3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127

    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

3128

3129 3130 3131 3132 3133
@config_func
def default_initial_mean(val):
    global g_default_initial_mean
    g_default_initial_mean = val

3134

3135 3136 3137 3138 3139
@config_func
def default_initial_strategy(val):
    global g_default_initial_strategy
    g_default_initial_strategy = val

3140

3141 3142 3143 3144 3145
@config_func
def default_initial_smart(val):
    global g_default_initial_smart
    g_default_initial_smart = val

3146

3147 3148 3149 3150 3151
@config_func
def default_momentum(val):
    global g_default_momentum
    g_default_momentum = val

3152

3153 3154 3155 3156 3157
@config_func
def default_decay_rate(val):
    global g_default_decay_rate
    g_default_decay_rate = val

3158

3159 3160 3161 3162 3163
@config_func
def default_num_batches_regularization(val):
    global g_default_num_batches_regularization
    g_default_num_batches_regularization = val

3164

3165 3166 3167 3168 3169
@config_func
def default_gradient_clipping_threshold(val):
    global g_default_gradient_clipping_threshold
    g_default_gradient_clipping_threshold = val

3170

3171 3172 3173 3174 3175
@config_func
def default_device(val):
    global g_default_device
    g_default_device = val

3176

3177 3178 3179 3180 3181
@config_func
def default_update_hooks(val):
    global g_default_update_hooks
    g_default_update_hooks = val

3182

3183 3184 3185 3186 3187
@config_func
def default_compact_func(val):
    global g_default_compact_func
    g_default_compact_func = val

3188

3189 3190 3191 3192 3193
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)
3194 3195 3196
        execfile(config_file,
                 make_config_environment(config_file, config_args), local_args)

3197 3198
    return Import

3199

3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227
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,
3228 3229 3230
    adam_beta1=0.9,
    adam_beta2=0.999,
    adam_epsilon=1e-8, )
3231

3232
settings_deprecated = dict(usage_ratio=1., )
3233 3234 3235 3236

trainer_settings = dict(
    save_dir="./output/model",
    init_model_path=None,
3237 3238
    start_pass=0, )

3239 3240 3241 3242 3243

@config_func
def Settings(**args):
    for k, v in args.iteritems():
        if k == "usage_ratio":
3244 3245
            logger.warning(
                "Deprecated: define usage_ratio in DataConfig instead")
3246 3247 3248 3249 3250 3251 3252 3253 3254 3255 3256
            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)

3257

3258 3259 3260 3261
@config_func
def cluster_config(**args):
    pass

3262

3263 3264 3265 3266 3267 3268 3269 3270 3271
@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

3272

3273 3274 3275 3276
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))
3277

3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292
        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),
3293
        get_config_arg=make_get_config_arg(config_args), )
3294 3295 3296 3297 3298

    funcs.update(g_extended_config_funcs)

    return funcs

3299

3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315
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

3316

3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328
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)"

3329

3330 3331 3332 3333
def my_fatal(s):
    logger.critical(s)
    raise Exception()

3334

3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371
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
3372
        g_config.opt_config.__setattr__(k, v)
3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398

    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

3399

3400 3401 3402 3403 3404 3405 3406 3407
if __name__ == '__main__':
    try:
        config = parse_config(sys.argv[1], '')
        config.SerializeToString()
        __real_print__(str(config))
    except:
        traceback.print_exc()
        raise