config_parser.py 127.7 KB
Newer Older
1
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
Z
zhangjinchao01 已提交
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
#
# 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(
Q
qijun 已提交
103
    format='[%(levelname)s %(asctime)s %(filename)s:%(lineno)s] %(message)s', )
Z
zhangjinchao01 已提交
104 105 106
logger = logging.getLogger('paddle')
logger.setLevel(logging.INFO)
__real_print__ = print
Q
qijun 已提交
107
print = logger.info
Z
zhangjinchao01 已提交
108 109 110 111

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

Q
qijun 已提交
112

Z
zhangjinchao01 已提交
113 114 115
# Initialize global variables. We use this function so that we can
# call parse_config() multiple times
def init_config_environment(
Q
qijun 已提交
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={},
Z
zhangjinchao01 已提交
131 132

        # store command args of paddle_trainer
Q
qijun 已提交
133
        g_command_config_args={},
Z
zhangjinchao01 已提交
134 135

        # Used for PyDataProvider to avoid duplicate module name
Q
qijun 已提交
136 137 138 139 140
        g_py_module_name_list=[],
        g_current_submodel=None,
        g_root_submodel=None,
        g_submodel_map={},
        g_submodel_stack=[],
141
        g_add_submodel_suffix=False, ):
Z
zhangjinchao01 已提交
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)

Q
qijun 已提交
158

Z
zhangjinchao01 已提交
159 160
g_config_funcs = {}

Q
qijun 已提交
161

Z
zhangjinchao01 已提交
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

Q
qijun 已提交
167

Z
zhangjinchao01 已提交
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

Q
qijun 已提交
173

Z
zhangjinchao01 已提交
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
Q
qijun 已提交
180

Z
zhangjinchao01 已提交
181 182
    return wrap

Q
qijun 已提交
183

Z
zhangjinchao01 已提交
184 185 186
def gen_parameter_name(layer_name, input_index):
    return '_%s.w%d' % (layer_name, input_index)

Q
qijun 已提交
187

Z
zhangjinchao01 已提交
188 189 190
def gen_bias_parameter_name(layer_name):
    return '_%s.wbias' % layer_name

Q
qijun 已提交
191

Z
zhangjinchao01 已提交
192 193 194
def default(x, default_value):
    return default_value if x is None else x

Q
qijun 已提交
195

Z
zhangjinchao01 已提交
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)

Q
qijun 已提交
202

Z
zhangjinchao01 已提交
203 204
# functions available in config file

Q
qijun 已提交
205

Z
zhangjinchao01 已提交
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)

Q
qijun 已提交
224

225 226
@config_func
def HasInputsSet():
227
    return len(g_current_submodel.input_layer_names) != 0
228

Z
zhangjinchao01 已提交
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)

Q
qijun 已提交
253
    name = MakeLayerNameInParentSubmodel(name)  #rename in nested submodel
Z
zhangjinchao01 已提交
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

Q
qijun 已提交
263

Z
zhangjinchao01 已提交
264
@config_func
Q
qijun 已提交
265
def SubModelEnd(name=None):
Z
zhangjinchao01 已提交
266
    global g_current_submodel, g_root_submodel, g_submodel_stack
Q
qijun 已提交
267 268
    config_assert(g_current_submodel is not g_root_submodel,
                  "submodel not begin")
Z
zhangjinchao01 已提交
269
    if name is not None:
Q
qijun 已提交
270 271 272
        config_assert(
            g_current_submodel.name == MakeLayerNameInParentSubmodel(name),
            "submodel name error")
Z
zhangjinchao01 已提交
273 274 275

    g_current_submodel = g_submodel_stack.pop()

Q
qijun 已提交
276

Z
zhangjinchao01 已提交
277 278
def MakeLayerNameInParentSubmodel(name):
    suffix = ""
279 280
    if len(g_submodel_stack) > 1:
        suffix = "@" + g_submodel_stack[-1].name
Z
zhangjinchao01 已提交
281 282
    return name + suffix

Q
qijun 已提交
283

Z
zhangjinchao01 已提交
284 285 286
def GetLayerBaseName(name):
    return name.split('@')[0]

Q
qijun 已提交
287 288

def MakeLayerNameInSubmodel(name, submodel_name=None):
Z
zhangjinchao01 已提交
289 290
    global g_current_submodel
    global g_add_submodel_suffix
Q
qijun 已提交
291 292
    if (submodel_name is None and not g_add_submodel_suffix and
            not g_current_submodel.is_recurrent_layer_group):
Z
zhangjinchao01 已提交
293 294 295 296 297
        return name
    if submodel_name is None:
        submodel_name = g_current_submodel.name
    return name + "@" + submodel_name

Q
qijun 已提交
298

Z
zhangjinchao01 已提交
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=""):
Z
zhangjinchao01 已提交
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
Z
zhangjinchao01 已提交
332
    in_links_count = 0
333
    for linkid, link in enumerate(in_links):
Z
zhangjinchao01 已提交
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

Z
zhangjinchao01 已提交
344 345 346
        if in_links_count == 0:
            in_links_has_subseq = has_subseq
        else:
Q
qijun 已提交
347 348 349 350
            config_assert(
                in_links_has_subseq == has_subseq,
                "The sequence type of in_links should be the same in RecurrentLayerGroup"
            )
Z
zhangjinchao01 已提交
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

Z
zhangjinchao01 已提交
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

Q
qijun 已提交
364

Z
zhangjinchao01 已提交
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):
Q
qijun 已提交
381
    generator.eos_layer_name = MakeLayerNameInSubmodel(generator.eos_layer_name)
Z
zhangjinchao01 已提交
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="",
Z
zhangjinchao01 已提交
391
                             seq_reversed=False):
Q
qijun 已提交
392
    RecurrentLayerGroupWithoutOutLinksBegin(name, in_links, seq_reversed,
393
                                            target_inlinkname)
Z
zhangjinchao01 已提交
394 395 396 397 398
    for link in out_links:
        RecurrentLayerGroupSetOutLink(link)

    if generator is not None:
        RecurrentLayerGroupSetGenerator(generator)
Q
qijun 已提交
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")
Z
zhangjinchao01 已提交
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")
Q
qijun 已提交
411
    for pair in g_current_submodel.memories:  #check exist
Z
zhangjinchao01 已提交
412
        layer = g_layer_map[pair.layer_name]
Y
Yu Yang 已提交
413 414
        config_assert(layer is not None,
                      "memory declare wrong name:%s" % pair.layer_name)
Z
zhangjinchao01 已提交
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)

Q
qijun 已提交
433

Z
zhangjinchao01 已提交
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

Q
qijun 已提交
440

Z
zhangjinchao01 已提交
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,
Q
qijun 已提交
458
            is_shared=None, ):
Z
zhangjinchao01 已提交
459 460
        self.add_keys(locals())

Q
qijun 已提交
461

Z
zhangjinchao01 已提交
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,
L
liaogang 已提交
482
            bilinear_interp=None,
Z
zhangjinchao01 已提交
483 484 485 486
            norm=None,
            pool=None,
            image=None,
            block_expand=None,
487
            maxout=None,
Q
qijun 已提交
488
            spp=None,
D
dangqingqing 已提交
489
            pad=None,
Z
zhangjinchao01 已提交
490 491 492 493 494
            format=None,
            nnz=None,
            is_static=None,
            is_shared=None,
            update_hooks=None,
495
            input_layer_argument=None,
D
dangqingqing 已提交
496 497 498 499 500
            make_layer_name_in_submodel=True, ):
        """
        @param make_layer_name_in_submodel True by defalut, you might need to
        set it carefully when adding Input in config_parser.py.
        """
Z
zhangjinchao01 已提交
501
        self.add_keys(locals())
D
dangqingqing 已提交
502 503 504
        self.input_layer_name = MakeLayerNameInSubmodel(
            input_layer_name
        ) if make_layer_name_in_submodel else input_layer_name
Z
zhangjinchao01 已提交
505

Q
qijun 已提交
506

Z
zhangjinchao01 已提交
507 508 509
# Define a projection for iexed layer
@config_class
class Projection(Input):
Q
qijun 已提交
510 511
    type = None  # subclass should set it correctly

Z
zhangjinchao01 已提交
512 513 514
    def __init__(
            self,
            input_layer_name,
Q
qijun 已提交
515
            size=0,  # projection output size
Z
zhangjinchao01 已提交
516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534
            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,
Q
qijun 已提交
535
            input_layer_argument=None, ):
Z
zhangjinchao01 已提交
536 537 538 539 540 541 542 543 544 545 546 547 548
        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
Q
qijun 已提交
549

Z
zhangjinchao01 已提交
550 551
    def calc_parameter_size(self, input_size, output_size):
        raise NotimplementedError
Q
qijun 已提交
552

Z
zhangjinchao01 已提交
553 554 555 556 557 558 559 560 561 562
    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
Q
qijun 已提交
563

Z
zhangjinchao01 已提交
564 565
    def calc_parameter_size(self, input_size, output_size):
        return 0
Q
qijun 已提交
566

Z
zhangjinchao01 已提交
567 568 569
    def calc_parameter_dims(self, input_size, output_size):
        return []

Q
qijun 已提交
570

Z
zhangjinchao01 已提交
571 572 573 574 575 576
# 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'

Q
qijun 已提交
577 578 579
    def __init__(self, input_layer_name, offset, **xargs):
        super(IdentityOffsetProjection, self).__init__(input_layer_name,
                                                       **xargs)
Z
zhangjinchao01 已提交
580 581 582 583
        self.proj_conf.offset = offset

    def calc_parameter_size(self, input_size, output_size):
        return 0
Q
qijun 已提交
584

Z
zhangjinchao01 已提交
585 586 587
    def calc_parameter_dims(self, input_size, output_size):
        return []

Q
qijun 已提交
588

Z
zhangjinchao01 已提交
589 590 591 592 593 594 595
# 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
Q
qijun 已提交
596

Z
zhangjinchao01 已提交
597 598
    def calc_parameter_size(self, input_size, output_size):
        return output_size
Q
qijun 已提交
599

Z
zhangjinchao01 已提交
600 601 602
    def calc_parameter_dims(self, input_size, output_size):
        return [1, output_size]

L
Luo Tao 已提交
603

X
xuwei06 已提交
604 605 606 607 608 609 610 611 612 613 614 615 616 617
# ScalingProjection
@config_class
class ScalingProjection(Projection):
    type = 'scaling'

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

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

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

Q
qijun 已提交
618

Z
zhangjinchao01 已提交
619 620 621 622 623 624
@config_class
class TableProjection(Projection):
    type = 'table'

    def calc_parameter_size(self, input_size, output_size):
        return input_size * output_size
Q
qijun 已提交
625

Z
zhangjinchao01 已提交
626 627 628
    def calc_parameter_dims(self, input_size, output_size):
        return [input_size, output_size]

Q
qijun 已提交
629

Z
zhangjinchao01 已提交
630 631 632 633 634 635
@config_class
class FullMatrixProjection(Projection):
    type = 'fc'

    def calc_parameter_size(self, input_size, output_size):
        return input_size * output_size
Q
qijun 已提交
636

Z
zhangjinchao01 已提交
637 638 639
    def calc_parameter_dims(self, input_size, output_size):
        return [input_size, output_size]

Q
qijun 已提交
640

Z
zhangjinchao01 已提交
641 642 643 644 645 646
@config_class
class TransposedFullMatrixProjection(Projection):
    type = 'trans_fc'

    def calc_parameter_size(self, input_size, output_size):
        return input_size * output_size
Q
qijun 已提交
647

Z
zhangjinchao01 已提交
648 649 650
    def calc_parameter_dims(self, input_size, output_size):
        return [output_size, input_size]

Q
qijun 已提交
651

Z
zhangjinchao01 已提交
652 653 654 655
@config_class
class ContextProjection(Projection):
    type = 'context'

Q
qijun 已提交
656 657
    def __init__(self, input_layer_name, context_start, context_length,
                 trainable_padding, **xargs):
Z
zhangjinchao01 已提交
658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680
        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


681
@config_class
682
class ConvBaseProjection(Projection):
Q
qijun 已提交
683 684 685 686 687
    def __init__(self,
                 input_layer_name,
                 num_filters=None,
                 conv_conf=None,
                 **xargs):
688
        super(ConvBaseProjection, self).__init__(input_layer_name, **xargs)
689 690 691 692 693 694 695 696 697 698 699 700

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

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

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

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

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

Q
qijun 已提交
710

711 712 713 714 715 716 717 718 719
@config_class
class ConvProjection(ConvBaseProjection):
    type = 'conv'

    def __init__(self,
                 input_layer_name,
                 num_filters=None,
                 conv_conf=None,
                 **xargs):
720 721
        super(ConvProjection, self).__init__(input_layer_name, num_filters,
                                             conv_conf, **xargs)
722

723
        parse_conv(conv_conf, self.input_layer_name, self.proj_conf.conv_conf,
724 725 726 727 728 729 730 731 732 733 734 735 736 737 738
                   num_filters)
        self.proj_conf.output_size = self.proj_conf.conv_conf.output_x * \
                                     self.proj_conf.conv_conf.output_y * \
                                     num_filters


@config_class
class ConvTransProjection(ConvBaseProjection):
    type = 'convt'

    def __init__(self,
                 input_layer_name,
                 num_filters=None,
                 conv_conf=None,
                 **xargs):
739 740
        super(ConvTransProjection, self).__init__(input_layer_name, num_filters,
                                                  conv_conf, **xargs)
741 742 743

        parse_conv(
            conv_conf,
744
            self.input_layer_name,
745 746 747 748 749 750 751 752
            self.proj_conf.conv_conf,
            num_filters,
            trans=True)
        self.proj_conf.output_size = self.proj_conf.conv_conf.img_size_y * \
                                     self.proj_conf.conv_conf.img_size * \
                                     num_filters


Z
zhangjinchao01 已提交
753 754 755
# Define a operator for mixed layer
@config_class
class Operator(Cfg):
Q
qijun 已提交
756 757
    type = None  # subclass should set it correctly

Z
zhangjinchao01 已提交
758 759
    def __init__(
            self,
Q
qijun 已提交
760
            input_layer_names, ):
Z
zhangjinchao01 已提交
761 762 763 764 765 766 767 768 769 770
        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

Q
qijun 已提交
771

Z
zhangjinchao01 已提交
772 773 774
@config_class
class DotMulOperator(Operator):
    type = 'dot_mul'
Q
qijun 已提交
775 776 777

    def __init__(self, input_layer_names, scale=None, **xargs):
        super(DotMulOperator, self).__init__(input_layer_names, **xargs)
Z
zhangjinchao01 已提交
778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795
        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'
Q
qijun 已提交
796 797 798 799 800 801 802

    def __init__(self,
                 input_layer_names,
                 num_filters=None,
                 conv_conf=None,
                 **xargs):
        super(ConvOperator, self).__init__(input_layer_names, **xargs)
Z
zhangjinchao01 已提交
803 804 805
        if num_filters is not None:
            self.operator_conf.num_filters = num_filters

806 807
        parse_conv(conv_conf,
                   MakeLayerNameInSubmodel(input_layer_names[0]),
Q
qijun 已提交
808
                   self.operator_conf.conv_conf, num_filters)
L
Luo Tao 已提交
809 810 811
        self.operator_conf.output_size = self.operator_conf.conv_conf.output_x * \
                                         self.operator_conf.conv_conf.output_y * \
                                         num_filters
Z
zhangjinchao01 已提交
812 813 814

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

815 816
    def calc_output_size(self, input_sizes):
        return self.operator_conf.output_size
Z
zhangjinchao01 已提交
817 818


819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848
@config_class
class ConvTransOperator(Operator):
    type = 'convt'

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

        parse_conv(
            conv_conf,
            MakeLayerNameInSubmodel(input_layer_names[0]),
            self.operator_conf.conv_conf,
            num_filters,
            trans=True)
        self.operator_conf.output_size = \
            self.operator_conf.conv_conf.img_size * \
            self.operator_conf.conv_conf.img_size_y * \
            num_filters

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

    def calc_output_size(self, input_sizes):
        return self.operator_conf.output_size


Z
zhangjinchao01 已提交
849 850 851
# please refer to the comments in proto/ModelConfig.proto
@config_class
class Conv(Cfg):
Q
qijun 已提交
852 853 854 855 856 857 858 859 860 861 862 863 864
    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):
Z
zhangjinchao01 已提交
865 866
        self.add_keys(locals())
        if filter_size_y is None:
Q
qijun 已提交
867
            self.filter_size_y = filter_size
Z
zhangjinchao01 已提交
868
        if padding_y is None:
Q
qijun 已提交
869
            self.padding_y = padding
Z
zhangjinchao01 已提交
870
        if stride_y is None:
Q
qijun 已提交
871
            self.stride_y = stride
Z
zhangjinchao01 已提交
872
        if output_x is not None:
Q
qijun 已提交
873 874
            config_assert(output_x <= 0)

Z
zhangjinchao01 已提交
875

L
liaogang 已提交
876 877
@config_class
class BilinearInterp(Cfg):
L
Luo Tao 已提交
878
    def __init__(self, out_size_x=None, out_size_y=None, channels=None):
L
liaogang 已提交
879 880
        self.add_keys(locals())

Q
qijun 已提交
881

Z
zhangjinchao01 已提交
882 883
@config_class
class Pool(Cfg):
D
dangqingqing 已提交
884 885 886 887 888 889 890 891 892 893 894
    def __init__(
            self,
            pool_type,
            channels,
            size_x,
            size_y=None,
            start=None,
            stride=None,  # 1 by defalut in protobuf
            stride_y=None,
            padding=None,  # 0 by defalut in protobuf
            padding_y=None):
Z
zhangjinchao01 已提交
895
        self.add_keys(locals())
Q
qijun 已提交
896 897


Q
qijun 已提交
898
@config_class
Q
qijun 已提交
899
class SpatialPyramidPool(Cfg):
L
Luo Tao 已提交
900
    def __init__(self, pool_type, pyramid_height, channels):
Q
qijun 已提交
901
        self.add_keys(locals())
Z
zhangjinchao01 已提交
902

Q
qijun 已提交
903

D
dangqingqing 已提交
904 905 906 907 908 909
@config_class
class Pad(Cfg):
    def __init__(self, channels, pad_c, pad_h, pad_w):
        self.add_keys(locals())


Z
zhangjinchao01 已提交
910 911
@config_class
class Norm(Cfg):
Q
qijun 已提交
912 913 914 915 916 917 918 919 920
    def __init__(self,
                 norm_type,
                 channels,
                 size,
                 scale,
                 pow,
                 output_x=None,
                 img_size=None,
                 blocked=None):
Z
zhangjinchao01 已提交
921 922
        self.add_keys(locals())

Q
qijun 已提交
923

Z
zhangjinchao01 已提交
924 925
@config_class
class Image(Cfg):
Q
qijun 已提交
926
    def __init__(self, channels, img_size=None):
Z
zhangjinchao01 已提交
927 928
        self.add_keys(locals())

Q
qijun 已提交
929

Z
zhangjinchao01 已提交
930 931
@config_class
class BlockExpand(Cfg):
Q
qijun 已提交
932 933 934 935 936 937 938 939 940 941 942 943
    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):
Z
zhangjinchao01 已提交
944 945
        self.add_keys(locals())

Q
qijun 已提交
946

947 948
@config_class
class MaxOut(Cfg):
Q
qijun 已提交
949
    def __init__(self, channels, groups, img_size_x=0, img_size_y=0):
950 951
        self.add_keys(locals())

Q
qijun 已提交
952

953
def create_data_config_proto(async_load_data=False,
954
                             constant_slots=None,
王益 已提交
955 956 957
                             data_ratio=1,
                             is_main_data=True,
                             usage_ratio=None):
Z
zhangjinchao01 已提交
958 959 960 961 962 963 964 965
    # 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)
Q
qijun 已提交
966 967
    data_config.data_ratio = data_ratio
    data_config.is_main_data = is_main_data
Z
zhangjinchao01 已提交
968

Q
qijun 已提交
969
    usage_ratio = default(usage_ratio, settings_deprecated["usage_ratio"])
Z
zhangjinchao01 已提交
970 971 972 973 974 975
    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

Q
qijun 已提交
976

Z
zhangjinchao01 已提交
977
@config_func
Q
qijun 已提交
978 979 980 981 982
def SimpleData(files=None,
               feat_dim=None,
               context_len=None,
               buffer_capacity=None,
               **xargs):
983
    data_config = create_data_config_proto(**xargs)
Z
zhangjinchao01 已提交
984 985 986 987 988 989 990 991 992
    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

Q
qijun 已提交
993

Z
zhangjinchao01 已提交
994
@config_func
Q
qijun 已提交
995 996 997 998 999 1000 1001 1002 1003 1004
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):
1005
    data_config = create_data_config_proto(**xargs)
Z
zhangjinchao01 已提交
1006 1007
    data_config.type = 'py'
    if load_data_module in g_py_module_name_list:
Q
qijun 已提交
1008

Z
zhangjinchao01 已提交
1009 1010 1011
        def get_path(module):
            m = __import__(load_data_module)
            return os.path.split(os.path.realpath(m.__file__))[0]
Q
qijun 已提交
1012

Z
zhangjinchao01 已提交
1013 1014 1015
        # 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.
Q
qijun 已提交
1016 1017
        module_new_name = "%s_copy_%d" % (load_data_module,
                                          len(g_py_module_name_list))
Z
zhangjinchao01 已提交
1018
        g_py_module_name_list.append(module_new_name)
Q
qijun 已提交
1019 1020 1021 1022
        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)
Z
zhangjinchao01 已提交
1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046
        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

Q
qijun 已提交
1047

Z
zhangjinchao01 已提交
1048
@config_func
Q
qijun 已提交
1049 1050 1051 1052 1053 1054 1055
def ProtoData(files=None,
              type=None,
              file_group_queue_capacity=None,
              load_file_count=None,
              constant_slots=None,
              load_thread_num=None,
              **xargs):
1056
    data_config = create_data_config_proto(**xargs)
Z
zhangjinchao01 已提交
1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075
    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

Q
qijun 已提交
1076

Z
zhangjinchao01 已提交
1077 1078
#real data for training is actually provided by "sub_data" data providers.
@config_func
Q
qijun 已提交
1079
def MultiData(sub_data=[]):
Z
zhangjinchao01 已提交
1080 1081 1082 1083 1084
    data_config = DataConfig()
    data_config.type = 'multi'
    data_config.sub_data_configs.extend(sub_data)
    return data_config

Q
qijun 已提交
1085

Z
zhangjinchao01 已提交
1086
@config_func
Q
qijun 已提交
1087 1088 1089 1090 1091 1092 1093
def Data(type,
         files=None,
         feat_dim=None,
         slot_dims=None,
         context_len=None,
         buffer_capacity=None,
         **xargs):
Z
zhangjinchao01 已提交
1094

1095
    data_config = create_data_config_proto(**xargs)
Z
zhangjinchao01 已提交
1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128
    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

Q
qijun 已提交
1129

L
Luo Tao 已提交
1130 1131
#caffe_mode: compute the output size using floor instead of ceil,
#            which is consistent of caffe and CuDNN's convention.
1132 1133 1134 1135 1136 1137 1138
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))

Q
qijun 已提交
1139

1140
#calcualte image_size based on output_size for de-convolution (ConvTransLayer).
L
Luo Tao 已提交
1141
#It is the reverse function of cnn_output_size
1142
def cnn_image_size(output_size, filter_size, padding, stride, caffe_mode):
L
Luo Tao 已提交
1143 1144 1145
    img_size = (output_size - 1) * stride + filter_size - 2 * padding
    if not caffe_mode:
        img_size = img_size + 1
1146 1147
    return img_size

Q
qijun 已提交
1148

L
Luo Tao 已提交
1149
def get_img_size(input_layer_name, channels):
L
Luo Tao 已提交
1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167
    input = g_layer_map[input_layer_name]
    img_pixels = input.size / channels
    img_size = input.width if input.width > 0 else int(img_pixels**0.5)
    img_size_y = input.height if input.height > 0 else int(img_pixels /
                                                           img_size)
    config_assert(
        img_size * img_size_y == img_pixels,
        "Input layer %s: Incorrect input image size %d * %d for input image pixels %d"
        % (input_layer_name, img_size, img_size_y, img_pixels))
    return img_size, img_size_y


def parse_bilinear(bilinear, input_layer_name, bilinear_conf):
    parse_image(bilinear, input_layer_name, bilinear_conf.image_conf)
    bilinear_conf.out_size_x = bilinear.out_size_x
    bilinear_conf.out_size_y = bilinear.out_size_y


1168
def parse_pool(pool, input_layer_name, pool_conf, ceil_mode):
Z
zhangjinchao01 已提交
1169
    pool_conf.pool_type = pool.pool_type
Q
qijun 已提交
1170 1171 1172
    config_assert(pool.pool_type in [
        'max-projection', 'avg-projection', 'cudnn-max-pool', 'cudnn-avg-pool'
    ], "pool-type %s is not in "
Z
zhangjinchao01 已提交
1173
                  "['max-projection', 'avg-projection', "
Q
qijun 已提交
1174
                  "'cudnn-max-pool', 'cudnn-avg-pool']" % pool.pool_type)
Z
zhangjinchao01 已提交
1175 1176 1177 1178 1179 1180

    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)
Q
qijun 已提交
1181
    pool_conf.stride_y = default(pool.stride_y, pool_conf.stride)
Z
zhangjinchao01 已提交
1182

L
Luo Tao 已提交
1183
    pool_conf.img_size, pool_conf.img_size_y = \
L
Luo Tao 已提交
1184
        get_img_size(input_layer_name, pool.channels)
Z
zhangjinchao01 已提交
1185

1186
    config_assert(not pool.start, "start is deprecated in pooling.")
Z
zhangjinchao01 已提交
1187

1188
    if pool.padding is not None:
Z
zhangjinchao01 已提交
1189
        pool_conf.padding = pool.padding
1190
    pool_conf.padding_y = default(pool.padding_y, pool_conf.padding)
D
dangqingqing 已提交
1191 1192
    pool_conf.output_x = cnn_output_size(pool_conf.img_size, pool_conf.size_x,
                                         pool_conf.padding, pool_conf.stride,
1193
                                         not ceil_mode)
D
dangqingqing 已提交
1194 1195
    pool_conf.output_y = cnn_output_size(pool_conf.img_size_y, pool_conf.size_y,
                                         pool_conf.padding_y,
1196
                                         pool_conf.stride_y, not ceil_mode)
Q
qijun 已提交
1197

Z
zhangjinchao01 已提交
1198

Q
qijun 已提交
1199
def parse_spp(spp, input_layer_name, spp_conf):
L
Luo Tao 已提交
1200
    parse_image(spp, input_layer_name, spp_conf.image_conf)
Q
qijun 已提交
1201 1202
    spp_conf.pool_type = spp.pool_type
    config_assert(spp.pool_type in ['max-projection', 'avg-projection'],
Q
qijun 已提交
1203 1204
                  "pool-type %s is not in "
                  "['max-projection', 'avg-projection']" % spp.pool_type)
Q
qijun 已提交
1205
    spp_conf.pyramid_height = spp.pyramid_height
Q
qijun 已提交
1206

Q
qijun 已提交
1207

Z
zhangjinchao01 已提交
1208 1209
def parse_image(image, input_layer_name, image_conf):
    image_conf.channels = image.channels
L
Luo Tao 已提交
1210
    image_conf.img_size, image_conf.img_size_y = \
L
Luo Tao 已提交
1211
        get_img_size(input_layer_name, image_conf.channels)
Q
qijun 已提交
1212

Z
zhangjinchao01 已提交
1213 1214 1215

def parse_norm(norm, input_layer_name, norm_conf):
    norm_conf.norm_type = norm.norm_type
1216 1217 1218 1219 1220
    config_assert(
        norm.norm_type in
        ['rnorm', 'cmrnorm-projection', 'cross-channel-norm'],
        "norm-type %s is not in [rnorm, cmrnorm-projection, cross-channel-norm]"
        % norm.norm_type)
Z
zhangjinchao01 已提交
1221 1222 1223 1224 1225 1226
    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

L
Luo Tao 已提交
1227
    norm_conf.img_size, norm_conf.img_size_y = \
L
Luo Tao 已提交
1228
        get_img_size(input_layer_name, norm.channels)
Z
zhangjinchao01 已提交
1229
    norm_conf.output_x = norm_conf.img_size
L
Luo Tao 已提交
1230
    norm_conf.output_y = norm_conf.img_size_y
Z
zhangjinchao01 已提交
1231 1232 1233
    if norm.norm_type in ['cmrnorm-projection']:
        norm_conf.scale /= norm.size
    else:
Q
qijun 已提交
1234 1235
        norm_conf.scale /= norm.size**2

1236

L
Luo Tao 已提交
1237 1238
#caffe_mode: compute the output size using floor instead of ceil,
#            which is consistent of caffe and CuDNN's convention.
1239
def parse_conv(conv, input_layer_name, conv_conf, num_filters, trans=False):
Z
zhangjinchao01 已提交
1240 1241 1242 1243 1244 1245 1246 1247 1248
    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
Q
qijun 已提交
1249

1250
    if not trans:
1251
        conv_conf.filter_channels = conv.channels / conv.groups
L
Luo Tao 已提交
1252
        conv_conf.img_size, conv_conf.img_size_y = \
L
Luo Tao 已提交
1253
            get_img_size(input_layer_name, conv.channels)
1254
        conv_conf.output_x = cnn_output_size(
Q
qijun 已提交
1255 1256
            conv_conf.img_size, conv_conf.filter_size, conv_conf.padding,
            conv_conf.stride, conv_conf.caffe_mode)
L
Luo Tao 已提交
1257 1258 1259
        conv_conf.output_y = cnn_output_size(
            conv_conf.img_size_y, conv_conf.filter_size_y, conv_conf.padding_y,
            conv_conf.stride_y, conv_conf.caffe_mode)
1260
    else:
1261
        conv_conf.filter_channels = num_filters / conv.groups
L
Luo Tao 已提交
1262
        conv_conf.output_x, conv_conf.output_y = \
L
Luo Tao 已提交
1263
            get_img_size(input_layer_name, conv.channels)
1264
        conv_conf.img_size = cnn_image_size(
Q
qijun 已提交
1265 1266
            conv_conf.output_x, conv_conf.filter_size, conv_conf.padding,
            conv_conf.stride, conv_conf.caffe_mode)
L
Luo Tao 已提交
1267
        conv_conf.img_size_y = cnn_image_size(
L
Luo Tao 已提交
1268 1269
            conv_conf.output_y, conv_conf.filter_size_y, conv_conf.padding_y,
            conv_conf.stride_y, conv_conf.caffe_mode)
Q
qijun 已提交
1270

1271

Z
zhangjinchao01 已提交
1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284
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:
1285
        block_expand_conf.output_x = cnn_output_size(
1286
            block_expand.img_size_x, block_expand.block_x,
1287
            block_expand.padding_x, block_expand.stride_x, False)
Z
zhangjinchao01 已提交
1288 1289

    if block_expand_conf.img_size_y == 0:
1290
        block_expand_conf.output_y = 0
Z
zhangjinchao01 已提交
1291
    else:
1292
        block_expand_conf.output_y = cnn_output_size(
1293
            block_expand.img_size_y, block_expand.block_y,
1294
            block_expand.padding_y, block_expand.stride_y, False)
Z
zhangjinchao01 已提交
1295

Q
qijun 已提交
1296

1297
def parse_maxout(maxout, input_layer_name, maxout_conf):
L
Luo Tao 已提交
1298
    parse_image(maxout, input_layer_name, maxout_conf.image_conf)
1299
    maxout_conf.groups = maxout.groups
1300

Q
qijun 已提交
1301

Z
zhangjinchao01 已提交
1302 1303 1304 1305 1306 1307
# Define an evaluator
@config_func
def Evaluator(
        name,
        type,
        inputs,
Q
qijun 已提交
1308 1309 1310 1311 1312 1313 1314
        chunk_scheme=None,
        num_chunk_types=None,
        classification_threshold=None,
        positive_label=None,
        dict_file=None,
        result_file=None,
        num_results=None,
L
Liang Zhao 已提交
1315
        top_k=None,
1316 1317
        delimited=None,
        excluded_chunk_types=None, ):
Z
zhangjinchao01 已提交
1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331
    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)

1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342
    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
L
Liang Zhao 已提交
1343 1344
    if top_k is not None:
        evaluator.top_k = top_k
1345 1346
    if delimited is not None:
        evaluator.delimited = delimited
Z
zhangjinchao01 已提交
1347

1348 1349 1350
    if excluded_chunk_types:
        evaluator.excluded_chunk_types.extend(excluded_chunk_types)

Q
qijun 已提交
1351

Z
zhangjinchao01 已提交
1352 1353 1354 1355 1356
class LayerBase(object):
    def __init__(
            self,
            name,
            type,
Q
qijun 已提交
1357
            size,  # size can be 0. In this case, subclass should set it.
Z
zhangjinchao01 已提交
1358 1359 1360 1361
            inputs,
            device=None,
            active_type="",
            drop_rate=0.,
1362
            coeff=None):
Z
zhangjinchao01 已提交
1363
        config_assert('@' not in name,
Q
qijun 已提交
1364
                      "layer name: %s contain special character @" % name)
Z
zhangjinchao01 已提交
1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379
        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()
1380
        assert isinstance(self.config, LayerConfig)
Z
zhangjinchao01 已提交
1381 1382 1383
        self.config.name = name
        self.config.type = type
        self.config.active_type = active_type
1384 1385
        if coeff is not None:
            self.config.coeff = float(coeff)
Z
zhangjinchao01 已提交
1386 1387 1388 1389 1390 1391 1392
        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
1393
        elif g_default_device is not None:
Z
zhangjinchao01 已提交
1394 1395 1396 1397 1398 1399 1400 1401 1402
            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(
Q
qijun 已提交
1403 1404
                    input_layer_name=input,
                    parameter_name=gen_parameter_name(name, input_index))
Z
zhangjinchao01 已提交
1405 1406 1407 1408 1409 1410 1411 1412
                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):
Q
qijun 已提交
1413
                self.operators.append(input)
Z
zhangjinchao01 已提交
1414 1415 1416 1417
                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:
Q
qijun 已提交
1418
                raise ValueError('Wrong type for inputs: %s' % type_of(input))
Z
zhangjinchao01 已提交
1419
            config_assert(input_layer_name in g_layer_map,
Q
qijun 已提交
1420 1421
                          "Unknown input layer '%s' for layer %s" %
                          (input_layer_name, name))
Z
zhangjinchao01 已提交
1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438
            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,
Q
qijun 已提交
1439
            bias,  # True/False or BiasCfg
Z
zhangjinchao01 已提交
1440
            size,
Q
qijun 已提交
1441 1442 1443
            dims=None,
            for_self=True,  # whether create bias for layer self
    ):
Z
zhangjinchao01 已提交
1444 1445 1446 1447 1448 1449

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

Q
qijun 已提交
1450 1451 1452
        config_assert(
            type_of(bias) == bool or type_of(bias) == Bias,
            'Incorrect type for bias: %s' % type_of(bias))
Z
zhangjinchao01 已提交
1453 1454 1455 1456 1457 1458 1459 1460 1461

        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:
1462 1463
                assert isinstance(self.config, LayerConfig)

Z
zhangjinchao01 已提交
1464 1465 1466
                Parameter(
                    bias.parameter_name,
                    size,
Q
qijun 已提交
1467 1468
                    self.config.device
                    if self.config.HasField('device') else None,
Z
zhangjinchao01 已提交
1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479
                    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,
Q
qijun 已提交
1480 1481
                    gradient_clipping_threshold=bias.
                    gradient_clipping_threshold,
Z
zhangjinchao01 已提交
1482
                    is_static=bias.is_static,
Q
qijun 已提交
1483
                    is_shared=bias.is_shared, )
Z
zhangjinchao01 已提交
1484 1485 1486 1487 1488
            if for_self:
                self.config.bias_parameter_name = bias.parameter_name
            else:
                return bias.parameter_name

Q
qijun 已提交
1489 1490 1491 1492 1493 1494
    def create_input_parameter(self,
                               input_index,
                               size,
                               dims=None,
                               sparse=None,
                               format=None):
Z
zhangjinchao01 已提交
1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508
        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]
Q
qijun 已提交
1509 1510
            config_assert(size == para.size, (
                'Shared parameter "%s" does not ' + 'have same size: %s vs. %s')
Z
zhangjinchao01 已提交
1511 1512
                          % (input_config.parameter_name, para.size, size))

Q
qijun 已提交
1513 1514
            config_assert(dims == para.dims, (
                'Shared parameter "%s" does not ' + 'have same dims: %s vs. %s')
Z
zhangjinchao01 已提交
1515 1516 1517 1518 1519 1520
                          % (input_config.parameter_name, para.dims, dims))
            return

        Parameter(
            input_config.parameter_name,
            size,
1521
            self.config.device if self.config.HasField("device") else None,
Z
zhangjinchao01 已提交
1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533
            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,
Q
qijun 已提交
1534 1535
            gradient_clipping_threshold=input_config.
            gradient_clipping_threshold,
Z
zhangjinchao01 已提交
1536 1537 1538 1539
            sparse=sparse,
            format=format,
            is_static=input_config.is_static,
            is_shared=input_config.is_shared,
Q
qijun 已提交
1540
            update_hooks=input_config.update_hooks)
Z
zhangjinchao01 已提交
1541 1542 1543 1544 1545 1546 1547 1548 1549

    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)

L
Luo Tao 已提交
1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566
    def set_layer_height_width(self, height, width):
        self.config.height = height
        self.config.width = width

    def set_cnn_layer(self,
                      input_layer_name,
                      height,
                      width,
                      channels,
                      is_print=True):
        size = height * width * channels
        self.set_layer_size(size)
        self.set_layer_height_width(height, width)
        if is_print:
            print("output for %s: c = %d, h = %d, w = %d, size = %d" %
                  (input_layer_name, channels, height, width, size))

Q
qijun 已提交
1567

Z
zhangjinchao01 已提交
1568 1569
@config_layer('multi_class_cross_entropy_with_selfnorm')
class MultiClassCrossEntropySelfNormCostLayer(LayerBase):
Q
qijun 已提交
1570 1571 1572
    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)
Z
zhangjinchao01 已提交
1573 1574
        self.config.softmax_selfnorm_alpha = softmax_selfnorm_alpha

Q
qijun 已提交
1575

Z
zhangjinchao01 已提交
1576 1577
@config_layer('fc')
class FCLayer(LayerBase):
Q
qijun 已提交
1578
    def __init__(self, name, size, inputs, bias=True, **xargs):
Z
zhangjinchao01 已提交
1579 1580 1581 1582 1583 1584 1585 1586 1587 1588
        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
1589 1590
            else:
                sparse = None
Z
zhangjinchao01 已提交
1591

Q
qijun 已提交
1592 1593
            self.create_input_parameter(input_index, psize, dims, sparse,
                                        format)
Z
zhangjinchao01 已提交
1594 1595
        self.create_bias_parameter(bias, self.config.size)

Q
qijun 已提交
1596

Z
zhangjinchao01 已提交
1597 1598
@config_layer('selective_fc')
class SelectiveFCLayer(LayerBase):
Q
qijun 已提交
1599 1600 1601 1602 1603 1604 1605 1606 1607 1608
    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):
Z
zhangjinchao01 已提交
1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628
        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,
Q
qijun 已提交
1629 1630
                          ("if indices of selected columns are not specified, "
                           "selective_fc Layer has at least two inputs"))
Z
zhangjinchao01 已提交
1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642
            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

Q
qijun 已提交
1643 1644
            self.create_input_parameter(input_index, psize, dims, sparse,
                                        format)
Z
zhangjinchao01 已提交
1645 1646
        self.create_bias_parameter(bias, self.config.size)

Q
qijun 已提交
1647

1648 1649
@config_layer('print')
class PrintLayer(LayerBase):
Q
qijun 已提交
1650
    def __init__(self, name, inputs):
1651 1652
        super(PrintLayer, self).__init__(name, 'print', 0, inputs)

Q
qijun 已提交
1653

Y
yuan 已提交
1654 1655
@config_layer('priorbox')
class PriorBoxLayer(LayerBase):
G
gaoyuan 已提交
1656 1657
    def __init__(self, name, inputs, size, min_size, max_size, aspect_ratio,
                 variance):
Y
yuan 已提交
1658
        super(PriorBoxLayer, self).__init__(name, 'priorbox', 0, inputs)
G
gaoyuan 已提交
1659
        config_assert(len(inputs) == 2, 'PriorBoxLayer must have 2 inputs')
G
gaoyuan 已提交
1660 1661 1662 1663 1664 1665 1666
        input_layer = self.get_input_layer(1)
        config_assert(
            input_layer.type == 'data',
            'Expecting the second input layer of an priorbox layer to be '
            'a data layer')
        config_assert(input_layer.width > 0, 'The data layer must set width')
        config_assert(input_layer.height > 0, 'The data layer must set height')
G
gaoyuan 已提交
1667
        config_assert(len(variance) == 4, 'The variance must have 4 inputs')
Y
yuan 已提交
1668 1669 1670 1671 1672 1673
        self.config.inputs[0].priorbox_conf.min_size.extend(min_size)
        self.config.inputs[0].priorbox_conf.max_size.extend(max_size)
        self.config.inputs[0].priorbox_conf.aspect_ratio.extend(aspect_ratio)
        self.config.inputs[0].priorbox_conf.variance.extend(variance)
        self.config.size = size

Q
qijun 已提交
1674

Z
zhangjinchao01 已提交
1675 1676
@config_layer('data')
class DataLayer(LayerBase):
L
Luo Tao 已提交
1677
    def __init__(self, name, size, height=None, width=None, device=None):
Q
qijun 已提交
1678 1679
        super(DataLayer, self).__init__(
            name, 'data', size, inputs=[], device=device)
L
Luo Tao 已提交
1680 1681
        if height and width:
            self.set_layer_height_width(height, width)
Q
qijun 已提交
1682

Z
zhangjinchao01 已提交
1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709

'''
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
'''
Q
qijun 已提交
1710 1711


Z
zhangjinchao01 已提交
1712 1713
@config_layer('data_norm')
class DataNormLayer(LayerBase):
Q
qijun 已提交
1714
    def __init__(self, name, inputs, data_norm_strategy="z-score", device=None):
Z
zhangjinchao01 已提交
1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725
        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)

Q
qijun 已提交
1726

Z
zhangjinchao01 已提交
1727 1728 1729
@config_layer('prelu')
class ParameterReluLayer(LayerBase):
    layer_type = 'prelu'
Q
qijun 已提交
1730 1731

    def __init__(self, name, inputs, partial_sum=1, **args):
Z
zhangjinchao01 已提交
1732 1733 1734 1735 1736 1737 1738 1739
        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)

Q
qijun 已提交
1740

Z
zhangjinchao01 已提交
1741 1742 1743
@config_layer('conv')
class ConvLayerBase(LayerBase):
    layer_type = 'conv'
Q
qijun 已提交
1744 1745 1746 1747 1748 1749 1750 1751

    def __init__(self,
                 name,
                 inputs=[],
                 bias=True,
                 num_filters=None,
                 shared_biases=False,
                 **xargs):
Z
zhangjinchao01 已提交
1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767
        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
Q
qijun 已提交
1768
            (parallel_nn == 0 or self.config.device > -1)):
Z
zhangjinchao01 已提交
1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780
            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)
            conv_conf = self.config.inputs[input_index].conv_conf
L
Luo Tao 已提交
1781 1782
            parse_conv(self.inputs[input_index].conv, input_layer.name,
                       conv_conf, num_filters)
Z
zhangjinchao01 已提交
1783 1784
            psize = self.calc_parameter_size(conv_conf)
            self.create_input_parameter(input_index, psize)
L
Luo Tao 已提交
1785 1786
            self.set_cnn_layer(name, conv_conf.output_y, conv_conf.output_x,
                               self.config.num_filters)
Z
zhangjinchao01 已提交
1787 1788 1789 1790 1791 1792 1793 1794 1795 1796

        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)

Q
qijun 已提交
1797

Z
zhangjinchao01 已提交
1798 1799 1800 1801
@config_layer('exconv')
class ConvLayer(ConvLayerBase):
    layer_type = 'exconv'

Q
qijun 已提交
1802

Z
zhangjinchao01 已提交
1803 1804 1805 1806
@config_layer('cudnn_conv')
class ConvLayer(ConvLayerBase):
    layer_type = 'cudnn_conv'

1807 1808 1809 1810

@config_layer('convt')
class ConvTransLayerBase(LayerBase):
    layer_type = 'convt'
Q
qijun 已提交
1811 1812 1813 1814 1815 1816 1817 1818

    def __init__(self,
                 name,
                 inputs=[],
                 bias=True,
                 num_filters=None,
                 shared_biases=False,
                 **xargs):
1819
        super(ConvTransLayerBase, self).__init__(
1820 1821 1822 1823 1824 1825 1826 1827
            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))

1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838
        # Automatically select cudnn_type for GPU and exconvt for CPU
        # if set type=exconvt, but still reserve the way user specify
        # exconvt or cudnn_convt manually.
        if self.layer_type == "cudnn_convt":
            config_assert(use_gpu, "cudnn_convt only support GPU")

        if (use_gpu == 1 and self.layer_type != "exconvt" and
            (parallel_nn == 0 or self.config.device > -1)):
            self.layer_type = "cudnn_convt"
        else:
            self.layer_type = "exconvt"
1839 1840 1841 1842 1843 1844 1845 1846
        # 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)
1847
            parse_conv(
1848 1849
                self.inputs[input_index].conv,
                input_layer.name,
1850
                self.config.inputs[input_index].conv_conf,
1851
                num_filters,
1852
                trans=True)
1853 1854 1855
            conv_conf = self.config.inputs[input_index].conv_conf
            psize = self.calc_parameter_size(conv_conf)
            self.create_input_parameter(input_index, psize)
1856 1857
            self.set_cnn_layer(name, conv_conf.img_size_y, conv_conf.img_size,
                               self.config.num_filters)
1858 1859 1860 1861 1862 1863 1864

        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):
1865
        return conv_conf.channels * conv_conf.filter_channels \
1866 1867
                    * (conv_conf.filter_size * conv_conf.filter_size_y)

Q
qijun 已提交
1868

1869 1870 1871 1872
@config_layer('exconvt')
class ConvTransLayer(ConvTransLayerBase):
    layer_type = 'exconvt'

Q
qijun 已提交
1873

1874 1875 1876 1877 1878
@config_layer('cudnn_convt')
class ConvTransLayer(ConvTransLayerBase):
    layer_type = 'cudnn_convt'


Z
zhangjinchao01 已提交
1879 1880
@config_layer('norm')
class NormLayer(LayerBase):
1881 1882
    def __init__(self, name, inputs, **xargs):
        super(NormLayer, self).__init__(name, 'norm', 0, inputs=inputs, **xargs)
Z
zhangjinchao01 已提交
1883 1884 1885
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
            norm_conf = self.config.inputs[input_index].norm_conf
L
Luo Tao 已提交
1886 1887 1888 1889
            parse_norm(self.inputs[input_index].norm, input_layer.name,
                       norm_conf)
            self.set_cnn_layer(name, norm_conf.output_y, norm_conf.output_x,
                               norm_conf.channels, False)
1890 1891 1892
            if norm_conf.norm_type == "cross-channel-norm":
                self.create_input_parameter(0, norm_conf.channels,
                                            [norm_conf.channels, 1])
Q
qijun 已提交
1893

Z
zhangjinchao01 已提交
1894 1895 1896

@config_layer('pool')
class PoolLayer(LayerBase):
1897 1898
    def __init__(self, name, inputs, ceil_mode=True, **xargs):
        super(PoolLayer, self).__init__(name, 'pool', 0, inputs=inputs, **xargs)
Z
zhangjinchao01 已提交
1899 1900 1901
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
            pool_conf = self.config.inputs[input_index].pool_conf
L
Luo Tao 已提交
1902
            parse_pool(self.inputs[input_index].pool, input_layer.name,
1903
                       pool_conf, ceil_mode)
L
Luo Tao 已提交
1904 1905
            self.set_cnn_layer(name, pool_conf.output_y, pool_conf.output_x,
                               pool_conf.channels)
Q
qijun 已提交
1906

Z
zhangjinchao01 已提交
1907

Q
qijun 已提交
1908 1909
@config_layer('spp')
class SpatialPyramidPoolLayer(LayerBase):
1910
    def __init__(self, name, inputs, **xargs):
Q
qijun 已提交
1911
        super(SpatialPyramidPoolLayer, self).__init__(
1912
            name, 'spp', 0, inputs=inputs, **xargs)
Q
qijun 已提交
1913 1914 1915
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
            spp_conf = self.config.inputs[input_index].spp_conf
L
Luo Tao 已提交
1916 1917 1918
            parse_spp(self.inputs[input_index].spp, input_layer.name, spp_conf)
            output_x = (pow(4, spp_conf.pyramid_height) - 1) / (4 - 1)
            self.set_cnn_layer(name, 1, output_x, spp_conf.image_conf.channels)
Q
qijun 已提交
1919

Q
qijun 已提交
1920

D
dangqingqing 已提交
1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939
@config_layer('pad')
class PadLayer(LayerBase):
    def __init__(self, name, inputs, **xargs):
        super(PadLayer, self).__init__(name, 'pad', 0, inputs=inputs, **xargs)
        pad = self.inputs[0].pad
        self.config.inputs[0].pad_conf.pad_c.extend(pad.pad_c)
        self.config.inputs[0].pad_conf.pad_h.extend(pad.pad_h)
        self.config.inputs[0].pad_conf.pad_w.extend(pad.pad_w)

        input_layer = self.get_input_layer(0)
        image_conf = self.config.inputs[0].pad_conf.image_conf
        parse_image(pad, input_layer.name, image_conf)
        out_ch = pad.channels + pad.pad_c[0] + pad.pad_c[1]
        out_h = image_conf.img_size_y + pad.pad_h[0] + pad.pad_h[1]
        out_w = image_conf.img_size + pad.pad_w[0] + pad.pad_w[1]
        self.set_cnn_layer(name, out_h, out_w, out_ch)
        self.config.size = out_ch * out_h * out_w


Z
zhangjinchao01 已提交
1940 1941 1942
@config_layer('batch_norm')
class BatchNormLayer(LayerBase):
    layer_type = 'batch_norm'
Q
qijun 已提交
1943 1944 1945 1946 1947 1948 1949 1950 1951 1952

    def __init__(self,
                 name,
                 inputs,
                 active_type="linear",
                 bias=True,
                 use_global_stats=True,
                 moving_average_fraction=0.9,
                 batch_norm_type=None,
                 **xargs):
Z
zhangjinchao01 已提交
1953 1954 1955 1956
        if inputs is None:
            inputs = []
        elif not isinstance(inputs, list):
            inputs = [inputs]
Q
qijun 已提交
1957 1958
        config_assert(
            len(inputs) == 1, "BatchNormLayer must have one and only one input")
Z
zhangjinchao01 已提交
1959 1960 1961 1962 1963 1964 1965 1966
        # 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):
Q
qijun 已提交
1967 1968 1969 1970 1971 1972
            inputs.append(
                Input(
                    inputs[0].input_layer_name,
                    initial_std=0.0,
                    initial_mean=0.0,
                    is_static=True,
1973
                    is_shared=is_shared,
D
dangqingqing 已提交
1974
                    make_layer_name_in_submodel=False, ))
Z
zhangjinchao01 已提交
1975 1976 1977 1978 1979 1980 1981

        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 \
1982
            cudnn_version >= 4007
Z
zhangjinchao01 已提交
1983
        self.layer_type = "cudnn_batch_norm" if use_cudnn else "batch_norm"
Q
qijun 已提交
1984 1985 1986 1987 1988 1989 1990
        super(BatchNormLayer, self).__init__(
            name,
            self.layer_type,
            0,
            active_type=active_type,
            inputs=inputs,
            **xargs)
Z
zhangjinchao01 已提交
1991 1992 1993 1994 1995 1996

        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

Q
qijun 已提交
1997
        input_layer = self.get_input_layer(0)
Z
zhangjinchao01 已提交
1998
        image_conf = self.config.inputs[0].image_conf
L
Luo Tao 已提交
1999
        parse_image(self.inputs[0].image, input_layer.name, image_conf)
2000

2001 2002
        # Only pass the width and height of input to batch_norm layer
        # when either of it is non-zero.
2003 2004
        if input_layer.width != 0 or input_layer.height != 0:
            self.set_cnn_layer(name, image_conf.img_size_y, image_conf.img_size,
D
dangqingqing 已提交
2005
                               image_conf.channels, False)
2006 2007
        else:
            self.set_layer_size(input_layer.size)
Z
zhangjinchao01 已提交
2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019

        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

Q
qijun 已提交
2020

Z
zhangjinchao01 已提交
2021 2022
@config_layer('trans')
class TransLayer(LayerBase):
2023
    def __init__(self, name, inputs, **xargs):
Q
qijun 已提交
2024
        super(TransLayer, self).__init__(
2025
            name, 'trans', 0, inputs=inputs, **xargs)
Q
qijun 已提交
2026 2027 2028
        config_assert(
            len(self.inputs) == 1,
            'TransLayer must have one and only one input')
Z
zhangjinchao01 已提交
2029 2030
        self.set_layer_size(self.get_input_layer(0).size)

Q
qijun 已提交
2031

Z
zhangjinchao01 已提交
2032 2033
@config_layer('resize')
class ResizeLayer(LayerBase):
2034
    def __init__(self, name, size, inputs, **xargs):
Q
qijun 已提交
2035
        super(ResizeLayer, self).__init__(
2036
            name, 'resize', size=size, inputs=inputs, **xargs)
Q
qijun 已提交
2037 2038 2039 2040
        config_assert(
            len(self.inputs) == 1,
            'ResizeLayer must have one and only one input')

Z
zhangjinchao01 已提交
2041

2042 2043
@config_layer('rotate')
class RotateLayer(LayerBase):
H
Haonan 已提交
2044
    def __init__(self, name, inputs, height, width, device=None):
2045 2046 2047 2048 2049
        super(RotateLayer, self).__init__(
            name, 'rotate', 0, inputs=inputs, device=device)
        config_assert(
            len(self.inputs) == 1,
            'RotateLayer must have one and only one input')
H
Haonan 已提交
2050
        self.set_layer_height_width(height, width)
2051 2052 2053
        self.set_layer_size(self.get_input_layer(0).size)


Z
zhangjinchao01 已提交
2054 2055
@config_layer('blockexpand')
class BlockExpandLayer(LayerBase):
2056
    def __init__(self, name, inputs, **xargs):
Q
qijun 已提交
2057
        super(BlockExpandLayer, self).__init__(
2058
            name, 'blockexpand', 0, inputs=inputs, **xargs)
Z
zhangjinchao01 已提交
2059 2060
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
Q
qijun 已提交
2061 2062
            parse_block_expand(
                self.inputs[input_index].block_expand, input_layer.name,
Z
zhangjinchao01 已提交
2063
                self.config.inputs[input_index].block_expand_conf)
Q
qijun 已提交
2064 2065 2066 2067 2068 2069
            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)

Z
zhangjinchao01 已提交
2070

2071 2072
@config_layer('maxout')
class MaxOutLayer(LayerBase):
Q
qijun 已提交
2073 2074 2075
    def __init__(self, name, inputs, **xargs):
        super(MaxOutLayer, self).__init__(
            name, 'maxout', 0, inputs=inputs, **xargs)
2076 2077
        input_layer = self.get_input_layer(0)
        maxout_conf = self.config.inputs[0].maxout_conf
L
Luo Tao 已提交
2078
        parse_maxout(self.inputs[0].maxout, input_layer.name, maxout_conf)
L
Luo Tao 已提交
2079 2080 2081
        out_channels = maxout_conf.image_conf.channels / maxout_conf.groups
        self.set_cnn_layer(name, g_layer_map[input_layer.name].height,
                           g_layer_map[input_layer.name].width, out_channels)
Q
qijun 已提交
2082

2083

Z
zhangjinchao01 已提交
2084 2085 2086 2087
# key: cost type
# value: cost class
g_cost_map = {}

Q
qijun 已提交
2088

Z
zhangjinchao01 已提交
2089 2090 2091
# define a cost layer without any parameters
def define_cost(class_name, cost_type):
    def init(cls, name, inputs, device=None, coeff=1.):
Q
qijun 已提交
2092 2093
        super(type(cls), cls).__init__(
            name, cost_type, 1, inputs, device=device, coeff=coeff)
Z
zhangjinchao01 已提交
2094

Q
qijun 已提交
2095
    cls = type(class_name, (LayerBase, ), dict(__init__=init))
Z
zhangjinchao01 已提交
2096 2097 2098
    global g_cost_map
    g_cost_map[cost_type] = cls

Q
qijun 已提交
2099

Z
zhangjinchao01 已提交
2100 2101 2102 2103 2104 2105 2106 2107
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
2108
define_cost('SumCost', 'sum_cost')
D
dangqingqing 已提交
2109
define_cost('SmoothL1Cost', 'smooth_l1')
Z
zhangjinchao01 已提交
2110

Q
qijun 已提交
2111

Z
zhangjinchao01 已提交
2112 2113
@config_layer('hsigmoid')
class HierarchicalSigmoidLayer(LayerBase):
Q
qijun 已提交
2114
    def __init__(self, name, num_classes, inputs, device=None, bias=True):
Z
zhangjinchao01 已提交
2115 2116
        super(HierarchicalSigmoidLayer, self).__init__(
            name, 'hsigmoid', 1, inputs=inputs, device=device)
Q
qijun 已提交
2117 2118 2119
        config_assert(
            len(self.inputs) >= 2,
            'HierarchicalSigmoidLayer must have at least 2 inputs')
Z
zhangjinchao01 已提交
2120 2121 2122 2123 2124 2125 2126 2127
        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)

Q
qijun 已提交
2128

Z
zhangjinchao01 已提交
2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152
'''
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.
'''
Q
qijun 已提交
2153 2154


Z
zhangjinchao01 已提交
2155 2156
@config_layer('lambda_cost')
class LambdaCost(LayerBase):
Q
qijun 已提交
2157
    def __init__(self, name, inputs, NDCG_num=5, max_sort_size=-1, device=None):
Z
zhangjinchao01 已提交
2158 2159
        super(LambdaCost, self).__init__(
            name, 'lambda_cost', 1, inputs=inputs, device=device)
Q
qijun 已提交
2160
        config_assert(len(self.inputs) == 2, 'lambdaCost must have 2 inputs')
Z
zhangjinchao01 已提交
2161 2162
        self.config.NDCG_num = NDCG_num
        if max_sort_size != -1:
Q
qijun 已提交
2163 2164 2165
            config_assert(
                NDCG_num <= max_sort_size,
                'NDCG_num must be less than or equal to max_sort_size')
Z
zhangjinchao01 已提交
2166 2167
        self.config.max_sort_size = max_sort_size

Q
qijun 已提交
2168

Z
zhangjinchao01 已提交
2169 2170
@config_layer('nce')
class NCELayer(LayerBase):
Q
qijun 已提交
2171 2172 2173 2174 2175 2176 2177 2178
    def __init__(self,
                 name,
                 num_classes,
                 inputs,
                 num_neg_samples=10,
                 neg_sampling_dist=None,
                 bias=True,
                 **xargs):
Z
zhangjinchao01 已提交
2179
        super(NCELayer, self).__init__(name, 'nce', 1, inputs=inputs, **xargs)
Q
qijun 已提交
2180 2181
        config_assert(
            len(self.inputs) >= 2, 'NCELayer must have at least 2 inputs')
Z
zhangjinchao01 已提交
2182 2183
        self.config.num_classes = num_classes
        if neg_sampling_dist is not None:
Q
qijun 已提交
2184 2185 2186 2187
            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))
Z
zhangjinchao01 已提交
2188
            s = sum(neg_sampling_dist)
Q
qijun 已提交
2189 2190 2191
            config_assert(
                abs(s - 1) < 1e-5,
                'The sum of neg_sampling_dist (%s) is not 1' % s)
Z
zhangjinchao01 已提交
2192 2193 2194 2195 2196

            self.config.neg_sampling_dist.extend(neg_sampling_dist)

        self.config.num_neg_samples = num_neg_samples
        num_real_inputs = len(self.inputs) - 1
Q
qijun 已提交
2197
        input_layer = self.get_input_layer(num_real_inputs)
Z
zhangjinchao01 已提交
2198 2199 2200 2201
        config_assert(input_layer.type == 'data',
                      'Expecting the last input layer of an nce layer to be '
                      'a data layer')

Q
qijun 已提交
2202 2203
        if (num_real_inputs > 1 and input_layer.size == 1 and
                self.get_input_layer(num_real_inputs - 1).type == 'data'):
Z
zhangjinchao01 已提交
2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216
            # 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):
Q
qijun 已提交
2217
    def __init__(self, name, inputs, bias=True, **xargs):
Z
zhangjinchao01 已提交
2218 2219
        super(AddToLayer, self).__init__(
            name, 'addto', 0, inputs=inputs, **xargs)
Q
qijun 已提交
2220
        config_assert(len(inputs) > 0, 'inputs cannot be empty for AddToLayer')
Z
zhangjinchao01 已提交
2221 2222 2223 2224 2225
        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)

Q
qijun 已提交
2226

Z
zhangjinchao01 已提交
2227 2228
@config_layer('agent')
class AgentLayer(LayerBase):
Q
qijun 已提交
2229 2230 2231 2232
    def __init__(self, name, size, device=None):
        super(AgentLayer, self).__init__(
            name, 'agent', size, inputs=[], device=device)

Z
zhangjinchao01 已提交
2233 2234 2235

@config_layer('sequence_agent')
class SequenceAgentLayer(LayerBase):
Q
qijun 已提交
2236
    def __init__(self, name, size, device=None):
Z
zhangjinchao01 已提交
2237 2238 2239
        super(SequenceAgentLayer, self).__init__(
            name, 'sequence_agent', size, inputs=[], device=device)

Q
qijun 已提交
2240

Z
zhangjinchao01 已提交
2241 2242
@config_layer('gather_agent')
class GatherAgentLayer(LayerBase):
Q
qijun 已提交
2243
    def __init__(self, name, size, device=None):
Z
zhangjinchao01 已提交
2244 2245 2246
        super(GatherAgentLayer, self).__init__(
            name, 'gather_agent', size, inputs=[], device=device)

Q
qijun 已提交
2247

Z
zhangjinchao01 已提交
2248 2249
@config_layer('scatter_agent')
class ScatterAgentLayer(LayerBase):
Q
qijun 已提交
2250
    def __init__(self, name, size, device=None):
Z
zhangjinchao01 已提交
2251 2252 2253
        super(ScatterAgentLayer, self).__init__(
            name, 'scatter_agent', size, inputs=[], device=device)

Q
qijun 已提交
2254

Z
zhangjinchao01 已提交
2255 2256
@config_layer('sequence_gather_agent')
class SequenceGatherAgentLayer(LayerBase):
Q
qijun 已提交
2257
    def __init__(self, name, size, device=None):
Z
zhangjinchao01 已提交
2258
        super(SequenceGatherAgentLayer, self).__init__(
Q
qijun 已提交
2259 2260
            name, 'sequence_gather_agent', size, inputs=[], device=device)

Z
zhangjinchao01 已提交
2261 2262 2263

@config_layer('sequence_scatter_agent')
class SequenceScatterAgentLayer(LayerBase):
Q
qijun 已提交
2264
    def __init__(self, name, size, device=None):
Z
zhangjinchao01 已提交
2265
        super(SequenceScatterAgentLayer, self).__init__(
Q
qijun 已提交
2266 2267
            name, 'sequence_scatter_agent', size, inputs=[], device=device)

Z
zhangjinchao01 已提交
2268 2269 2270

@config_layer('multiplex')
class MultiplexLayer(LayerBase):
Q
qijun 已提交
2271 2272 2273 2274 2275
    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.')
Z
zhangjinchao01 已提交
2276
        for i in range(1, len(inputs)):
Q
qijun 已提交
2277 2278 2279 2280 2281
            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.")

Z
zhangjinchao01 已提交
2282 2283

@config_func
Q
qijun 已提交
2284 2285 2286
def Link(
        name,
        has_subseq=False, ):
Z
zhangjinchao01 已提交
2287 2288 2289 2290 2291
    link_config = LinkConfig()
    link_config.link_name = name
    link_config.has_subseq = has_subseq
    return link_config

Q
qijun 已提交
2292

Z
zhangjinchao01 已提交
2293 2294
# memory for recurrent layer group.
# *name* and *size* are actual layer's name and size.
2295 2296 2297 2298
# If *name* is None, need to provide *memory_name* and need to use
# SetMemoryInput() later to specify the layer which this memory remembers.
#
# return the name of the memory,
Z
zhangjinchao01 已提交
2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309
# 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
2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321
def Memory(name,
           size,
           is_sequence=False,
           boot_layer=None,
           boot_bias=False,
           boot_bias_active_type="",
           boot_with_const_id=None,
           memory_name=None):
    if not memory_name:
        config_assert(name is not None, "name needs cannot be None")
        memory_name = name + "+delay1"
    agent_name = memory_name
Z
zhangjinchao01 已提交
2322
    if is_sequence:
L
Luo Tao 已提交
2323 2324 2325
        config_assert(
            boot_layer is not None,
            "there must be boot_layer in network when is_sequence = True")
Z
zhangjinchao01 已提交
2326 2327 2328 2329
        agent_layer = SequenceAgentLayer(agent_name, size)
    else:
        agent_layer = AgentLayer(agent_name, size)
    config_assert(g_current_submodel.is_recurrent_layer_group,
Q
qijun 已提交
2330
                  'Memory should be used in recurrent layer group only')
Z
zhangjinchao01 已提交
2331
    memory = g_current_submodel.memories.add()
2332 2333
    if name is not None:
        memory.layer_name = MakeLayerNameInSubmodel(name)
Z
zhangjinchao01 已提交
2334 2335
    memory.link_name = MakeLayerNameInSubmodel(agent_name)
    memory.is_sequence = is_sequence
Q
qijun 已提交
2336
    options = sum((boot_layer is not None, bool(boot_bias),
Z
zhangjinchao01 已提交
2337
                   boot_with_const_id is not None))
Q
qijun 已提交
2338 2339 2340 2341
    config_assert(
        options <= 1,
        'take one option at most from boot_layer, boot_bias, or boot_with_const_id'
    )
Z
zhangjinchao01 已提交
2342 2343 2344
    if boot_layer is not None:
        boot_layer = MakeLayerNameInParentSubmodel(boot_layer)
        config_assert(boot_layer in g_layer_map,
Q
qijun 已提交
2345 2346
                      'boot_layer "%s" does not correspond to a layer name' %
                      boot_layer)
Z
zhangjinchao01 已提交
2347 2348 2349
        memory.boot_layer_name = boot_layer
    elif boot_bias:
        memory.boot_bias_parameter_name = agent_layer.create_bias_parameter(
Q
qijun 已提交
2350
            boot_bias, size, for_self=False)
Z
zhangjinchao01 已提交
2351 2352 2353 2354 2355
        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

Q
qijun 已提交
2356

2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367
@config_func
def SetMemoryInput(memory_name, layer_name):
    memory_name = MakeLayerNameInSubmodel(memory_name)
    layer_name = MakeLayerNameInSubmodel(layer_name)
    for mem in g_current_submodel.memories:
        if mem.link_name == memory_name:
            mem.layer_name = layer_name
            return
    logger.fatal("Nonexistent memory name: " + memory_name)


Z
zhangjinchao01 已提交
2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378
# 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,
Q
qijun 已提交
2379 2380 2381 2382
        eos_layer_name="eos_check",
        num_results_per_sample=1,
        beam_size=1,
        log_prob=None, ):
Z
zhangjinchao01 已提交
2383 2384 2385 2386 2387 2388 2389 2390 2391
    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

Q
qijun 已提交
2392

Z
zhangjinchao01 已提交
2393 2394
@config_layer('expand')
class ExpandLayer(LayerBase):
2395
    def __init__(self, name, inputs, trans_type='non-seq', bias=False, **xargs):
Q
qijun 已提交
2396
        super(ExpandLayer, self).__init__(
2397
            name, 'expand', 0, inputs=inputs, **xargs)
Q
qijun 已提交
2398 2399 2400 2401 2402 2403 2404 2405
        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)

Z
zhangjinchao01 已提交
2406 2407 2408

@config_layer('featmap_expand')
class FeatMapExpandLayer(LayerBase):
Q
qijun 已提交
2409 2410 2411 2412 2413 2414
    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:
Z
zhangjinchao01 已提交
2415
            self.config.num_filters = num_filters
Q
qijun 已提交
2416
        else:
Z
zhangjinchao01 已提交
2417
            logger.fatal("FeatMapExpandLayer must specify num_filters.")
Q
qijun 已提交
2418
        self.set_layer_size(self.get_input_layer(0).size * num_filters)
Z
zhangjinchao01 已提交
2419 2420 2421 2422


@config_layer('max')
class MaxLayer(LayerBase):
Q
qijun 已提交
2423 2424 2425 2426 2427 2428
    def __init__(self,
                 name,
                 inputs,
                 trans_type='non-seq',
                 active_type='linear',
                 bias=False,
2429 2430
                 output_max_index=None,
                 **xargs):
2431
        super(MaxLayer, self).__init__(name, 'max', 0, inputs=inputs, **xargs)
Z
zhangjinchao01 已提交
2432
        config_assert(len(self.inputs) == 1, 'MaxLayer must have 1 input')
Q
qijun 已提交
2433 2434
        self.config.trans_type = trans_type
        self.config.active_type = active_type
Z
zhangjinchao01 已提交
2435 2436 2437 2438
        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)
2439 2440
        if output_max_index is not None:
            self.config.output_max_index = output_max_index
Z
zhangjinchao01 已提交
2441 2442 2443 2444


@config_layer('maxid')
class MaxIdLayer(LayerBase):
Q
qijun 已提交
2445
    def __init__(self, name, inputs, beam_size=None, device=None):
Z
zhangjinchao01 已提交
2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462
        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):
Q
qijun 已提交
2463
    def __init__(self, name, inputs, eos_id, device=None):
Z
zhangjinchao01 已提交
2464 2465 2466
        super(EosIdLayer, self).__init__(
            name, 'eos_id', 0, inputs=inputs, device=device)
        config_assert(len(self.inputs) == 1, 'EosIdLayer must have 1 input')
Q
qijun 已提交
2467
        self.set_layer_size(2)  # boolean output
Z
zhangjinchao01 已提交
2468 2469
        self.config.eos_id = eos_id

Q
qijun 已提交
2470

Z
zhangjinchao01 已提交
2471 2472
@config_layer('seqlastins')
class SequenceLastInstanceLayer(LayerBase):
Q
qijun 已提交
2473 2474 2475 2476 2477
    def __init__(self,
                 name,
                 inputs,
                 active_type='linear',
                 trans_type='non-seq',
2478
                 bias=False,
2479
                 stride=-1,
2480
                 **xargs):
Q
qijun 已提交
2481 2482 2483 2484 2485
        super(SequenceLastInstanceLayer, self).__init__(
            name,
            'seqlastins',
            0,
            inputs=inputs,
2486 2487
            active_type=active_type,
            **xargs)
Q
qijun 已提交
2488 2489
        config_assert(
            len(inputs) == 1, 'SequenceLastInstanceLayer must have 1 input')
2490
        if trans_type == 'seq':
L
Luo Tao 已提交
2491
            config_assert(stride == -1, 'subseq does not support stride window')
Q
qijun 已提交
2492
        self.config.trans_type = trans_type
2493 2494
        self.config.seq_pool_stride = stride
        self.set_layer_size(self.get_input_layer(0).size)
Z
zhangjinchao01 已提交
2495 2496
        self.create_bias_parameter(bias, self.config.size)

Q
qijun 已提交
2497

Z
zhangjinchao01 已提交
2498 2499
@config_layer('seqfirstins')
class SequenceFirstInstanceLayer(SequenceLastInstanceLayer):
2500 2501 2502 2503 2504 2505
    def __init__(self,
                 name,
                 inputs,
                 active_type='linear',
                 trans_type='non-seq',
                 bias=False,
2506
                 stride=-1,
2507
                 **xargs):
Q
qijun 已提交
2508
        super(SequenceFirstInstanceLayer, self).__init__(
2509 2510 2511 2512 2513 2514 2515
            name,
            inputs=inputs,
            active_type=active_type,
            trans_type=trans_type,
            bias=bias,
            stride=stride,
            **xargs)
Z
zhangjinchao01 已提交
2516 2517
        self.config.select_first = True

Q
qijun 已提交
2518

Z
zhangjinchao01 已提交
2519 2520
@config_layer('seqconcat')
class SequenceConcatLayer(LayerBase):
2521
    def __init__(self, name, inputs, active_type='linear', bias=False, **xargs):
Q
qijun 已提交
2522 2523 2524 2525 2526
        super(SequenceConcatLayer, self).__init__(
            name,
            'seqconcat',
            0,
            inputs=inputs,
2527 2528
            active_type=active_type,
            **xargs)
Q
qijun 已提交
2529 2530
        config_assert(
            len(inputs) == 2, 'SequenceConcatLayer must have 2 inputs')
Z
zhangjinchao01 已提交
2531 2532 2533 2534 2535
        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)

Q
qijun 已提交
2536

Z
zhangjinchao01 已提交
2537 2538
@config_layer('seqreshape')
class SequenceReshapeLayer(LayerBase):
Q
qijun 已提交
2539 2540 2541 2542 2543
    def __init__(self,
                 name,
                 size,
                 inputs,
                 active_type='linear',
2544 2545
                 bias=False,
                 **xargs):
Q
qijun 已提交
2546 2547 2548
        super(SequenceReshapeLayer, self).__init__(
            name,
            'seqreshape',
Z
zhangjinchao01 已提交
2549
            size,
Q
qijun 已提交
2550
            inputs=inputs,
2551 2552
            active_type=active_type,
            **xargs)
Q
qijun 已提交
2553 2554
        config_assert(
            len(inputs) == 1, 'SequenceReshapeLayer must have 1 inputs')
Z
zhangjinchao01 已提交
2555 2556 2557
        self.set_layer_size(size)
        self.create_bias_parameter(bias, size)

Q
qijun 已提交
2558

Z
zhangjinchao01 已提交
2559 2560
@config_layer('subseq')
class SubSequenceLayer(LayerBase):
2561
    def __init__(self, name, inputs, active_type='linear', bias=False, **xargs):
Q
qijun 已提交
2562
        super(SubSequenceLayer, self).__init__(
2563
            name, 'subseq', 0, inputs=inputs, active_type=active_type, **xargs)
Z
zhangjinchao01 已提交
2564 2565 2566 2567 2568 2569
        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)

Q
qijun 已提交
2570

Z
zhangjinchao01 已提交
2571 2572
@config_layer('out_prod')
class OuterProdLayer(LayerBase):
Q
qijun 已提交
2573 2574 2575
    def __init__(self, name, inputs, device=None):
        super(OuterProdLayer, self).__init__(
            name, 'out_prod', 0, inputs=inputs, device=device)
Z
zhangjinchao01 已提交
2576 2577 2578 2579 2580
        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)

Q
qijun 已提交
2581

Z
zhangjinchao01 已提交
2582 2583
@config_layer('power')
class PowerLayer(LayerBase):
Q
qijun 已提交
2584 2585 2586
    def __init__(self, name, inputs, device=None):
        super(PowerLayer, self).__init__(
            name, 'power', 0, inputs=inputs, device=device)
Z
zhangjinchao01 已提交
2587 2588 2589 2590
        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)
Q
qijun 已提交
2591 2592 2593
        config_assert(1 == input_layer0.size,
                      'The left input is the exponent and should be of size 1')

Z
zhangjinchao01 已提交
2594 2595 2596

@config_layer('slope_intercept')
class SlopeInterceptLayer(LayerBase):
Q
qijun 已提交
2597 2598 2599
    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)
Z
zhangjinchao01 已提交
2600 2601 2602 2603 2604 2605
        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)

Q
qijun 已提交
2606

Z
zhangjinchao01 已提交
2607 2608
@config_layer('scaling')
class ScalingLayer(LayerBase):
Q
qijun 已提交
2609 2610 2611
    def __init__(self, name, inputs, device=None):
        super(ScalingLayer, self).__init__(
            name, 'scaling', 0, inputs=inputs, device=device)
Z
zhangjinchao01 已提交
2612 2613 2614 2615
        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)
Q
qijun 已提交
2616 2617 2618
        config_assert(1 == input_layer0.size,
                      'The left input should be of size 1')

Z
zhangjinchao01 已提交
2619 2620 2621

@config_layer('conv_shift')
class ConvShiftLayer(LayerBase):
Q
qijun 已提交
2622 2623 2624
    def __init__(self, name, inputs, device=None):
        super(ConvShiftLayer, self).__init__(
            name, 'conv_shift', 0, inputs=inputs, device=device)
Z
zhangjinchao01 已提交
2625 2626 2627 2628
        config_assert(len(inputs) == 2, 'ConvShiftLayer must have 2 inputs')
        input_layer0 = self.get_input_layer(0)
        self.set_layer_size(input_layer0.size)

Q
qijun 已提交
2629

Z
zhangjinchao01 已提交
2630 2631
@config_layer('convex_comb')
class ConvexCombinationLayer(LayerBase):
Q
qijun 已提交
2632
    def __init__(self, name, size, inputs, device=None):
Z
zhangjinchao01 已提交
2633
        super(ConvexCombinationLayer, self).__init__(
Q
qijun 已提交
2634 2635 2636
            name, 'convex_comb', size, inputs=inputs, device=device)
        config_assert(
            len(self.inputs) == 2, 'ConvexCombinationLayer must have 2 inputs')
2637 2638 2639
        config_assert(
            size * self.get_input_layer(0).size == self.get_input_layer(1).size,
            'Wrong input size for ConvexCombinationLayer')
Z
zhangjinchao01 已提交
2640 2641
        self.set_layer_size(size)

Q
qijun 已提交
2642

Z
zhangjinchao01 已提交
2643 2644
@config_layer('interpolation')
class InterpolationLayer(LayerBase):
Q
qijun 已提交
2645
    def __init__(self, name, inputs, device=None):
Z
zhangjinchao01 已提交
2646 2647
        super(InterpolationLayer, self).__init__(
            name, 'interpolation', 0, inputs=inputs, device=device)
Q
qijun 已提交
2648 2649
        config_assert(
            len(self.inputs) == 3, 'InterpolationLayer must have 3 inputs')
Z
zhangjinchao01 已提交
2650 2651 2652 2653 2654 2655 2656 2657
        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')

Q
qijun 已提交
2658

L
liaogang 已提交
2659 2660
@config_layer('bilinear_interp')
class BilinearInterpLayer(LayerBase):
Q
qijun 已提交
2661
    def __init__(self, name, inputs, **xargs):
L
liaogang 已提交
2662
        super(BilinearInterpLayer, self).__init__(
L
liaogang 已提交
2663
            name, 'bilinear_interp', 0, inputs=inputs, **xargs)
L
liaogang 已提交
2664
        input_layer = self.get_input_layer(0)
L
Luo Tao 已提交
2665 2666 2667 2668
        conf = self.config.inputs[0].bilinear_interp_conf
        parse_bilinear(self.inputs[0].bilinear_interp, input_layer.name, conf)
        self.set_cnn_layer(name, conf.out_size_y, conf.out_size_x,
                           conf.image_conf.channels)
Q
qijun 已提交
2669

L
liaogang 已提交
2670

Z
zhangjinchao01 已提交
2671 2672
@config_layer('sum_to_one_norm')
class SumToOneNormLayer(LayerBase):
Q
qijun 已提交
2673
    def __init__(self, name, inputs, device=None):
Z
zhangjinchao01 已提交
2674
        super(SumToOneNormLayer, self).__init__(
Q
qijun 已提交
2675 2676 2677
            name, 'sum_to_one_norm', 0, inputs=inputs, device=device)
        config_assert(
            len(self.inputs) == 1, 'SumToOneNormLayer must have 1 input')
Z
zhangjinchao01 已提交
2678 2679 2680
        input_layer0 = self.get_input_layer(0)
        self.set_layer_size(input_layer0.size)

Q
qijun 已提交
2681

Z
zhangjinchao01 已提交
2682 2683
@config_layer('cos_vm')
class CosSimVecMatLayer(LayerBase):
Q
qijun 已提交
2684
    def __init__(self, name, size, inputs, cos_scale=1.0, device=None):
Z
zhangjinchao01 已提交
2685
        super(CosSimVecMatLayer, self).__init__(
Q
qijun 已提交
2686
            name, 'cos_vm', size, inputs=inputs, device=device)
Z
zhangjinchao01 已提交
2687
        self.config.cos_scale = cos_scale
Q
qijun 已提交
2688 2689
        config_assert(
            len(self.inputs) == 2, 'CosSimVecMatLayer must have 2 inputs')
2690 2691 2692
        config_assert(
            size * self.get_input_layer(0).size == self.get_input_layer(1).size,
            'Wrong input size for CosSimVecMatLayer')
Z
zhangjinchao01 已提交
2693

Q
qijun 已提交
2694

Z
zhangjinchao01 已提交
2695 2696
@config_layer('sampling_id')
class SamplingIdLayer(LayerBase):
Q
qijun 已提交
2697
    def __init__(self, name, inputs, device=None):
Z
zhangjinchao01 已提交
2698 2699
        super(SamplingIdLayer, self).__init__(
            name, 'sampling_id', 0, inputs=inputs, device=device)
Q
qijun 已提交
2700 2701
        config_assert(
            len(self.inputs) == 1, 'SamplingIdLayer must have 1 input')
Z
zhangjinchao01 已提交
2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713
        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):
Q
qijun 已提交
2714 2715 2716 2717 2718 2719
    def __init__(self,
                 name,
                 inputs,
                 average_strategy='average',
                 trans_type='non-seq',
                 active_type='linear',
2720 2721
                 bias=False,
                 **xargs):
Q
qijun 已提交
2722
        super(AverageLayer, self).__init__(
2723
            name, 'average', 0, inputs=inputs, active_type=active_type, **xargs)
Z
zhangjinchao01 已提交
2724
        self.config.average_strategy = average_strategy
Q
qijun 已提交
2725
        self.config.trans_type = trans_type
Z
zhangjinchao01 已提交
2726 2727 2728 2729 2730 2731
        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)

Q
qijun 已提交
2732

Z
zhangjinchao01 已提交
2733 2734
@config_layer('cos')
class CosSimLayer(LayerBase):
2735
    def __init__(self, name, inputs, cos_scale=1, device=None):
Z
zhangjinchao01 已提交
2736 2737 2738 2739 2740 2741
        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')
2742
        self.config.cos_scale = cos_scale
Z
zhangjinchao01 已提交
2743 2744 2745 2746


@config_layer('tensor')
class TensorLayer(LayerBase):
2747
    def __init__(self, name, size, inputs, bias=True, **xargs):
Q
qijun 已提交
2748
        super(TensorLayer, self).__init__(
2749
            name, 'tensor', size, inputs=inputs, **xargs)
Z
zhangjinchao01 已提交
2750 2751
        config_assert(len(self.inputs) == 2, 'TensorLayer must have 2 inputs')
        config_assert(size > 0, 'size must be positive')
Q
qijun 已提交
2752 2753
        config_assert(inputs[1].parameter_name == None,
                      'second parameter should be None.')
Z
zhangjinchao01 已提交
2754 2755 2756 2757 2758 2759 2760 2761 2762 2763
        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):
Q
qijun 已提交
2764 2765 2766 2767 2768 2769 2770
    def __init__(self,
                 name,
                 inputs,
                 size=0,
                 bias=True,
                 error_clipping_threshold=None,
                 **xargs):
Z
zhangjinchao01 已提交
2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787
        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)
2788
            if self.config.size == 0:
Z
zhangjinchao01 已提交
2789 2790 2791
                size = operator.calc_output_size(operator_conf.input_sizes)
                if size != 0:
                    self.set_layer_size(size)
2792
            else:
2793 2794
                sz = operator.calc_output_size(operator_conf.input_sizes)
                if sz != 0:
Q
qijun 已提交
2795 2796 2797 2798
                    config_assert(
                        sz == self.config.size,
                        "different inputs have different size: %s vs. %s" %
                        (sz, self.config.size))
Z
zhangjinchao01 已提交
2799 2800 2801 2802
        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:
Q
qijun 已提交
2803 2804 2805
                config_assert(
                    isinstance(input, Projection),
                    "input should be projection or operation")
2806
            if self.config.size == 0 and isinstance(input, Projection):
Z
zhangjinchao01 已提交
2807 2808 2809
                size = input.calc_output_size(input_layer)
                if size != 0:
                    self.set_layer_size(size)
2810
            elif isinstance(input, Projection):
Q
qijun 已提交
2811 2812 2813 2814 2815 2816
                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))
Z
zhangjinchao01 已提交
2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827
        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)
Q
qijun 已提交
2828 2829
                input_config.proj_conf.name = gen_parameter_name(name,
                                                                 input_index)
Z
zhangjinchao01 已提交
2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840
                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)

2841 2842 2843 2844 2845 2846
        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()
Z
zhangjinchao01 已提交
2847

2848 2849 2850
        if bias:
            self.config.bias_size = psize
            self.create_bias_parameter(bias, psize)
Z
zhangjinchao01 已提交
2851

2852 2853
        if error_clipping_threshold is not None:
            self.config.error_clipping_threshold = error_clipping_threshold
Z
zhangjinchao01 已提交
2854

Q
qijun 已提交
2855

Z
zhangjinchao01 已提交
2856 2857
# like MixedLayer, but no bias parameter
@config_func
Q
qijun 已提交
2858
def ExpressionLayer(name, inputs, **xargs):
Z
zhangjinchao01 已提交
2859 2860
    MixedLayer(name, inputs, bias=False, **xargs)

Q
qijun 已提交
2861

Z
zhangjinchao01 已提交
2862 2863
@config_layer('concat')
class ConcatenateLayer(LayerBase):
Q
qijun 已提交
2864
    def __init__(self, name, inputs, bias=False, **xargs):
Z
zhangjinchao01 已提交
2865
        config_assert(inputs, 'inputs cannot be empty')
2866
        config_assert(not bias, 'ConcatenateLayer cannot support bias.')
Z
zhangjinchao01 已提交
2867 2868 2869 2870 2871 2872
        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]
Q
qijun 已提交
2873
            if self.config.size == 0:
Z
zhangjinchao01 已提交
2874 2875 2876 2877
                size += input_layer.size

        self.set_layer_size(size)

Q
qijun 已提交
2878

Z
zhangjinchao01 已提交
2879 2880 2881
# like concat layer, but each input layer was processed by a Projection.
@config_layer('concat2')
class ConcatenateLayer2(LayerBase):
Q
qijun 已提交
2882
    def __init__(self, name, inputs, bias=False, **xargs):
Z
zhangjinchao01 已提交
2883 2884 2885
        config_assert(inputs, 'inputs cannot be empty')
        super(ConcatenateLayer2, self).__init__(
            name, 'concat2', 0, inputs=inputs, **xargs)
2886 2887

        if isinstance(self.inputs[0], ConvProjection):
Q
qijun 已提交
2888 2889 2890 2891 2892 2893
            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.")
2894

Z
zhangjinchao01 已提交
2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914
        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,
Q
qijun 已提交
2915
                                              input.proj_conf.output_size)
Z
zhangjinchao01 已提交
2916
            dims = input.calc_parameter_dims(input.proj_conf.input_size,
Q
qijun 已提交
2917
                                             input.proj_conf.output_size)
Z
zhangjinchao01 已提交
2918 2919
            self.create_input_parameter(input_index, psize, dims)

2920 2921 2922 2923 2924 2925 2926
        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()

2927 2928 2929
        if bias:
            self.config.bias_size = psize
            self.create_bias_parameter(bias, psize)
2930

Q
qijun 已提交
2931

Z
zhangjinchao01 已提交
2932 2933
@config_layer('recurrent')
class RecurrentLayer(LayerBase):
Q
qijun 已提交
2934
    def __init__(self, name, inputs, reversed=False, bias=True, **xargs):
Y
Yu Yang 已提交
2935 2936
        super(RecurrentLayer, self).__init__(name, 'recurrent', 0, inputs,
                                             **xargs)
Z
zhangjinchao01 已提交
2937 2938 2939 2940 2941 2942 2943 2944 2945
        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)

Q
qijun 已提交
2946

Z
zhangjinchao01 已提交
2947 2948
@config_layer('lstmemory')
class LstmLayer(LayerBase):
Q
qijun 已提交
2949 2950 2951 2952 2953 2954 2955 2956
    def __init__(self,
                 name,
                 inputs,
                 reversed=False,
                 active_gate_type="sigmoid",
                 active_state_type="sigmoid",
                 bias=True,
                 **xargs):
Z
zhangjinchao01 已提交
2957 2958 2959 2960 2961 2962 2963 2964
        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
Q
qijun 已提交
2965
        self.config.active_gate_type = active_gate_type
Z
zhangjinchao01 已提交
2966 2967 2968 2969 2970
        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)

Q
qijun 已提交
2971

Z
zhangjinchao01 已提交
2972 2973
@config_layer('lstm_step')
class LstmStepLayer(LayerBase):
Q
qijun 已提交
2974 2975 2976 2977 2978 2979 2980 2981 2982 2983
    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)
Z
zhangjinchao01 已提交
2984 2985 2986
        config_assert(len(inputs) == 2, 'LstmStepLayer must have 2 inputs')
        input_layer0 = self.get_input_layer(0)
        input_layer1 = self.get_input_layer(1)
Q
qijun 已提交
2987 2988 2989 2990 2991
        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
Z
zhangjinchao01 已提交
2992 2993 2994
        self.config.active_state_type = active_state_type
        self.create_bias_parameter(bias, size * 3)

Q
qijun 已提交
2995

Z
zhangjinchao01 已提交
2996 2997 2998
# get the specific output from the input layer.
@config_layer('get_output')
class GetOutputLayer(LayerBase):
Q
qijun 已提交
2999 3000 3001 3002
    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')
Z
zhangjinchao01 已提交
3003 3004 3005 3006
        inputs = self.inputs[0]
        config_assert(inputs.input_layer_argument,
                      'input_layer_argument cannot be empty')

Q
qijun 已提交
3007

Z
zhangjinchao01 已提交
3008 3009
@config_layer('mdlstmemory')
class MDLstmLayer(LayerBase):
Q
qijun 已提交
3010 3011 3012 3013 3014 3015 3016 3017
    def __init__(self,
                 name,
                 inputs,
                 directions=True,
                 active_gate_type="sigmoid",
                 active_state_type="sigmoid",
                 bias=True,
                 **xargs):
Y
Yu Yang 已提交
3018 3019
        super(MDLstmLayer, self).__init__(name, 'mdlstmemory', 0, inputs,
                                          **xargs)
Z
zhangjinchao01 已提交
3020 3021 3022 3023
        config_assert(len(self.inputs) == 1, 'MDLstmLayer must have 1 input')
        input_layer = self.get_input_layer(0)
        dim_num = len(directions)
        #check input_layer.size is divided by (3+dim_num)
Y
Yu Yang 已提交
3024 3025
        config_assert(input_layer.size % (3 + dim_num) == 0,
                      "size % (dim_num) should be 0!")
Q
qijun 已提交
3026
        size = input_layer.size / (3 + dim_num)
Z
zhangjinchao01 已提交
3027
        self.set_layer_size(size)
Q
qijun 已提交
3028
        self.config.active_gate_type = active_gate_type
Z
zhangjinchao01 已提交
3029 3030 3031
        self.config.active_state_type = active_state_type
        for i in xrange(len(directions)):
            self.config.directions.append(int(directions[i]))
Y
Yu Yang 已提交
3032 3033
        self.create_input_parameter(0, size * size * (3 + dim_num),
                                    [size, size, 3 + dim_num])
Z
zhangjinchao01 已提交
3034
        #bias includes 3 kinds of peephole, 3+dim_num+2+dim_num
Q
qijun 已提交
3035 3036
        self.create_bias_parameter(bias, size * (5 + 2 * dim_num))

Z
zhangjinchao01 已提交
3037 3038 3039

@config_layer('gated_recurrent')
class GatedRecurrentLayer(LayerBase):
Q
qijun 已提交
3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050
    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')
Z
zhangjinchao01 已提交
3051 3052 3053 3054 3055 3056
        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
Q
qijun 已提交
3057
        self.config.active_gate_type = active_gate_type
Z
zhangjinchao01 已提交
3058 3059 3060
        self.create_input_parameter(0, size * size * 3, [size, size * 3])
        self.create_bias_parameter(bias, size * 3)

Q
qijun 已提交
3061

Z
zhangjinchao01 已提交
3062 3063
@config_layer('gru_step')
class GruStepLayer(LayerBase):
Q
qijun 已提交
3064 3065 3066 3067 3068 3069 3070
    def __init__(self,
                 name,
                 size,
                 inputs,
                 active_gate_type="sigmoid",
                 bias=True,
                 **xargs):
Y
Yu Yang 已提交
3071 3072
        super(GruStepLayer, self).__init__(name, 'gru_step', size, inputs,
                                           **xargs)
Z
zhangjinchao01 已提交
3073 3074 3075
        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)
Q
qijun 已提交
3076 3077 3078 3079 3080
        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
H
Haonan 已提交
3081
        self.create_input_parameter(0, size * size * 3, [size, size * 3])
Z
zhangjinchao01 已提交
3082 3083
        self.create_bias_parameter(bias, size * 3)

Q
qijun 已提交
3084

Z
zhangjinchao01 已提交
3085 3086 3087 3088 3089 3090 3091
'''
 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
'''
Q
qijun 已提交
3092 3093


Z
zhangjinchao01 已提交
3094 3095
@config_layer('crf')
class CRFLayer(LayerBase):
Q
qijun 已提交
3096
    def __init__(self, name, size, inputs, coeff=1.0, device=None):
Z
zhangjinchao01 已提交
3097
        super(CRFLayer, self).__init__(name, 'crf', size, inputs, device=device)
Q
qijun 已提交
3098 3099
        config_assert(2 <= len(self.inputs) <= 3,
                      'CRFLayer must have 2 or 3 inputs')
3100
        self.create_input_parameter(0, size * (size + 2), [size + 2, size])
Z
zhangjinchao01 已提交
3101 3102
        self.config.coeff = coeff

Q
qijun 已提交
3103

Z
zhangjinchao01 已提交
3104 3105 3106 3107 3108 3109 3110 3111
'''
 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
'''
Q
qijun 已提交
3112 3113


Z
zhangjinchao01 已提交
3114 3115
@config_layer('crf_decoding')
class CRFDecodingLayer(LayerBase):
Q
qijun 已提交
3116
    def __init__(self, name, size, inputs, device=None):
Z
zhangjinchao01 已提交
3117 3118 3119 3120 3121
        super(CRFDecodingLayer, self).__init__(
            name, 'crf_decoding', size, inputs, device=device)
        config_assert(
            len(self.inputs) <= 2,
            'CRFDecodingLayer cannot have more than 2 inputs')
3122
        self.create_input_parameter(0, size * (size + 2), [size + 2, size])
Z
zhangjinchao01 已提交
3123

Q
qijun 已提交
3124

Z
zhangjinchao01 已提交
3125 3126
@config_layer('ctc')
class CTCLayer(LayerBase):
Q
qijun 已提交
3127
    def __init__(self, name, size, inputs, norm_by_times=False, device=None):
Z
zhangjinchao01 已提交
3128 3129 3130 3131
        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')

Q
qijun 已提交
3132

3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153
@config_layer('warp_ctc')
class WarpCTCLayer(LayerBase):
    def __init__(self,
                 name,
                 size,
                 inputs,
                 blank=0,
                 norm_by_times=False,
                 device=None):
        super(WarpCTCLayer, self).__init__(
            name, 'warp_ctc', size=size, inputs=inputs, device=device)
        self.config.blank = blank
        self.config.norm_by_times = norm_by_times
        config_assert(len(self.inputs) == 2, 'WarpCTCLayer must have 2 inputs')
        input_layer = self.get_input_layer(0)
        config_assert(
            (input_layer.active_type == '' or
             input_layer.active_type == 'linear'),
            "Expecting the active_type of input layer to be linear or null")


Z
zhangjinchao01 已提交
3154 3155
@config_layer('recurrent_layer_group')
class RecurrentLayerGroup(LayerBase):
Q
qijun 已提交
3156
    def __init__(self, name, device=None):
Z
zhangjinchao01 已提交
3157 3158 3159 3160 3161 3162
        super(RecurrentLayerGroup, self).__init__(
            name, 'recurrent_layer_group', 0, inputs=[], device=device)


# Deprecated, use a new layer specific class instead
@config_func
Q
qijun 已提交
3163
def Layer(name, type, **xargs):
Z
zhangjinchao01 已提交
3164 3165 3166 3167
    layers = {}
    layers.update(g_cost_map)
    layers.update(g_layer_type_map)
    layer_func = layers.get(type)
Q
qijun 已提交
3168
    config_assert(layer_func, "layer type '%s' not supported." % type)
X
xuwei06 已提交
3169
    return layer_func(name, **xargs)
Z
zhangjinchao01 已提交
3170

Q
qijun 已提交
3171

Z
zhangjinchao01 已提交
3172
@config_func
Q
qijun 已提交
3173
def ParameterHook(type, **kwargs):
Z
zhangjinchao01 已提交
3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185
    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
Q
qijun 已提交
3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207
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):
Z
zhangjinchao01 已提交
3208 3209 3210 3211 3212 3213 3214

    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
3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225
    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)

Z
zhangjinchao01 已提交
3226 3227
    config_assert(not momentum or not decay_rate_l1,
                  "momentum and decay_rate_l1 cannot both be non-zero")
3228 3229 3230 3231 3232

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

Z
zhangjinchao01 已提交
3233 3234 3235 3236
    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)
3237

Q
qijun 已提交
3238 3239
    num_batches_regularization = default(num_batches_regularization,
                                         g_default_num_batches_regularization)
3240 3241 3242
    if num_batches_regularization is not None:
        para.num_batches_regularization = int(num_batches_regularization)

Z
zhangjinchao01 已提交
3243 3244 3245 3246 3247 3248
    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
Q
qijun 已提交
3249 3250
    gradient_clipping_threshold = default(gradient_clipping_threshold,
                                          g_default_gradient_clipping_threshold)
3251 3252
    if gradient_clipping_threshold is not None:
        para.gradient_clipping_threshold = gradient_clipping_threshold
Q
qijun 已提交
3253 3254
    para.initial_strategy = default(initial_strategy,
                                    g_default_initial_strategy)
Z
zhangjinchao01 已提交
3255 3256 3257 3258 3259 3260
    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:
Q
qijun 已提交
3261 3262 3263
            print(
                "Use initial_smart, but dims not set. Initial_smart may not be used in this layer"
            )
Z
zhangjinchao01 已提交
3264 3265 3266 3267
            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)
3268 3269 3270 3271 3272 3273 3274

    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
Z
zhangjinchao01 已提交
3275 3276 3277 3278
    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")
3279 3280
    if is_shared is not None:
        para.is_shared = is_shared
Z
zhangjinchao01 已提交
3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301

    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

Q
qijun 已提交
3302

Z
zhangjinchao01 已提交
3303 3304 3305 3306 3307
@config_func
def default_initial_mean(val):
    global g_default_initial_mean
    g_default_initial_mean = val

Q
qijun 已提交
3308

Z
zhangjinchao01 已提交
3309 3310 3311 3312 3313
@config_func
def default_initial_strategy(val):
    global g_default_initial_strategy
    g_default_initial_strategy = val

Q
qijun 已提交
3314

Z
zhangjinchao01 已提交
3315 3316 3317 3318 3319
@config_func
def default_initial_smart(val):
    global g_default_initial_smart
    g_default_initial_smart = val

Q
qijun 已提交
3320

Z
zhangjinchao01 已提交
3321 3322 3323 3324 3325
@config_func
def default_momentum(val):
    global g_default_momentum
    g_default_momentum = val

Q
qijun 已提交
3326

Z
zhangjinchao01 已提交
3327 3328 3329 3330 3331
@config_func
def default_decay_rate(val):
    global g_default_decay_rate
    g_default_decay_rate = val

Q
qijun 已提交
3332

Z
zhangjinchao01 已提交
3333 3334 3335 3336 3337
@config_func
def default_num_batches_regularization(val):
    global g_default_num_batches_regularization
    g_default_num_batches_regularization = val

Q
qijun 已提交
3338

Z
zhangjinchao01 已提交
3339 3340 3341 3342 3343
@config_func
def default_gradient_clipping_threshold(val):
    global g_default_gradient_clipping_threshold
    g_default_gradient_clipping_threshold = val

Q
qijun 已提交
3344

Z
zhangjinchao01 已提交
3345 3346 3347 3348 3349
@config_func
def default_device(val):
    global g_default_device
    g_default_device = val

Q
qijun 已提交
3350

Z
zhangjinchao01 已提交
3351 3352 3353 3354 3355
@config_func
def default_update_hooks(val):
    global g_default_update_hooks
    g_default_update_hooks = val

Q
qijun 已提交
3356

Z
zhangjinchao01 已提交
3357 3358 3359 3360 3361
@config_func
def default_compact_func(val):
    global g_default_compact_func
    g_default_compact_func = val

Q
qijun 已提交
3362

Z
zhangjinchao01 已提交
3363 3364 3365 3366 3367
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)
Q
qijun 已提交
3368 3369 3370
        execfile(config_file,
                 make_config_environment(config_file, config_args), local_args)

Z
zhangjinchao01 已提交
3371 3372
    return Import

Q
qijun 已提交
3373

Z
zhangjinchao01 已提交
3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401
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,
Q
qijun 已提交
3402 3403 3404
    adam_beta1=0.9,
    adam_beta2=0.999,
    adam_epsilon=1e-8, )
Z
zhangjinchao01 已提交
3405

Q
qijun 已提交
3406
settings_deprecated = dict(usage_ratio=1., )
Z
zhangjinchao01 已提交
3407 3408 3409 3410

trainer_settings = dict(
    save_dir="./output/model",
    init_model_path=None,
Q
qijun 已提交
3411 3412
    start_pass=0, )

Z
zhangjinchao01 已提交
3413 3414 3415 3416 3417

@config_func
def Settings(**args):
    for k, v in args.iteritems():
        if k == "usage_ratio":
Q
qijun 已提交
3418 3419
            logger.warning(
                "Deprecated: define usage_ratio in DataConfig instead")
Z
zhangjinchao01 已提交
3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430
            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)

Q
qijun 已提交
3431

Z
zhangjinchao01 已提交
3432 3433 3434 3435
@config_func
def cluster_config(**args):
    pass

Q
qijun 已提交
3436

Z
zhangjinchao01 已提交
3437 3438 3439 3440 3441 3442 3443 3444 3445
@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

Q
qijun 已提交
3446

Z
zhangjinchao01 已提交
3447 3448 3449 3450
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))
Q
qijun 已提交
3451

Z
zhangjinchao01 已提交
3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466
        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),
Q
qijun 已提交
3467
        get_config_arg=make_get_config_arg(config_args), )
Z
zhangjinchao01 已提交
3468 3469 3470 3471 3472

    funcs.update(g_extended_config_funcs)

    return funcs

Q
qijun 已提交
3473

Z
zhangjinchao01 已提交
3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489
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

Q
qijun 已提交
3490

Z
zhangjinchao01 已提交
3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502
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)"

Q
qijun 已提交
3503

Z
zhangjinchao01 已提交
3504 3505 3506 3507
def my_fatal(s):
    logger.critical(s)
    raise Exception()

Y
Yu Yang 已提交
3508

3509
_parse_config_hooks = set()
Y
Yu Yang 已提交
3510 3511


3512 3513 3514 3515 3516 3517 3518
def register_parse_config_hook(f):
    """
    Register a hook function for parse_config. parse_config will invoke the hook
    at the beginning of parse. This make it possible to reset global state for
    for constructing the model.
    """
    _parse_config_hooks.add(f)
Q
qijun 已提交
3519

Y
Yu Yang 已提交
3520

3521
def update_g_config():
Z
zhangjinchao01 已提交
3522
    '''
3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549
    Update g_config after execute config_file or config_functions.
    '''
    for k, v in settings.iteritems():
        if v is None:
            continue
        g_config.opt_config.__setattr__(k, v)

    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(trainer_config, config_arg_str):
    '''
    @param trainer_config: can be a string of config file name or a function name
    with config logic
Z
zhangjinchao01 已提交
3550 3551 3552 3553
    @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()
3554 3555
    for hook in _parse_config_hooks:
        hook()
Z
zhangjinchao01 已提交
3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582

    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

3583 3584
    if hasattr(trainer_config, '__call__'):
        trainer_config.func_globals.update(
L
Luo Tao 已提交
3585
            make_config_environment("", config_args))
3586
        trainer_config()
H
hanchao 已提交
3587
    else:
3588 3589
        execfile(trainer_config,
                 make_config_environment(trainer_config, config_args))
Z
zhangjinchao01 已提交
3590

3591
    return update_g_config()
Z
zhangjinchao01 已提交
3592 3593


3594
def parse_config_and_serialize(trainer_config, config_arg_str):
Z
zhangjinchao01 已提交
3595
    try:
3596
        config = parse_config(trainer_config, config_arg_str)
Z
zhangjinchao01 已提交
3597 3598 3599 3600 3601 3602
        #logger.info(config)
        return config.SerializeToString()
    except:
        traceback.print_exc()
        raise

Q
qijun 已提交
3603

Z
zhangjinchao01 已提交
3604 3605 3606 3607 3608 3609 3610 3611
if __name__ == '__main__':
    try:
        config = parse_config(sys.argv[1], '')
        config.SerializeToString()
        __real_print__(str(config))
    except:
        traceback.print_exc()
        raise