config_parser.py 156.0 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
#
# 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
102
    format='[%(levelname)s %(asctime)s %(filename)s:%(lineno)s] %(message)s', )
Z
zhangjinchao01 已提交
103 104 105
logger = logging.getLogger('paddle')
logger.setLevel(logging.INFO)
__real_print__ = print
Q
qijun 已提交
106
print = logger.info
Z
zhangjinchao01 已提交
107 108 109 110

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

Q
qijun 已提交
111

Z
zhangjinchao01 已提交
112 113 114
# 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
115 116 117 118 119 120 121 122 123 124 125 126 127 128
        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={},
X
xuwei06 已提交
129
        g_parameter_initializer_map={},
Q
qijun 已提交
130
        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
324 325 326 327 328 329 330 331
    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
    in_links_count = 0
332
    for linkid, link in enumerate(in_links):
Z
zhangjinchao01 已提交
333 334 335 336
        if isinstance(link, basestring):
            name = link
        else:
            name = link.link_name
337

Z
zhangjinchao01 已提交
338 339 340
        in_links_count += 1
        layer_name = MakeLayerNameInParentSubmodel(name)
        layer = g_layer_map[layer_name]
341 342
        ScatterAgentLayer(
            name=name, size=layer.size, width=layer.width, height=layer.height)
343

Z
zhangjinchao01 已提交
344 345 346 347
        pair = g_current_submodel.in_links.add()
        pair.layer_name = layer_name
        pair.link_name = MakeLayerNameInSubmodel(name)

Q
qijun 已提交
348

Z
zhangjinchao01 已提交
349 350 351 352 353 354 355 356 357 358 359 360 361
@config_func
def RecurrentLayerGroupSetOutLink(link):
    if isinstance(link, basestring):
        name = link
    else:
        name = link.link_name
    layer_name = MakeLayerNameInParentSubmodel(name)
    pair = g_current_submodel.out_links.add()
    pair.layer_name = MakeLayerNameInSubmodel(name)
    pair.link_name = layer_name


def RecurrentLayerGroupSetGenerator(generator=None):
Q
qijun 已提交
362
    generator.eos_layer_name = MakeLayerNameInSubmodel(generator.eos_layer_name)
Z
zhangjinchao01 已提交
363 364 365 366 367 368 369 370
    g_current_submodel.generator.CopyFrom(generator)


@config_func
def RecurrentLayerGroupBegin(name,
                             in_links,
                             out_links,
                             generator=None,
371
                             target_inlinkname="",
Z
zhangjinchao01 已提交
372
                             seq_reversed=False):
373
    RecurrentLayerGroupWithoutOutLinksBegin(name, in_links, seq_reversed)
Z
zhangjinchao01 已提交
374 375 376 377 378
    for link in out_links:
        RecurrentLayerGroupSetOutLink(link)

    if generator is not None:
        RecurrentLayerGroupSetGenerator(generator)
Q
qijun 已提交
379 380 381 382 383
        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
384 385 386 387 388 389 390


@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
391
    for pair in g_current_submodel.memories:  #check exist
Z
zhangjinchao01 已提交
392
        layer = g_layer_map[pair.layer_name]
Y
Yu Yang 已提交
393 394
        config_assert(layer is not None,
                      "memory declare wrong name:%s" % pair.layer_name)
Z
zhangjinchao01 已提交
395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410
        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)
        else:
            GatherAgentLayer(name=agent_name, size=layer.size)

Q
qijun 已提交
411

Z
zhangjinchao01 已提交
412 413 414 415 416 417
# 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
418

Z
zhangjinchao01 已提交
419 420
@config_class
class Bias(Cfg):
X
xuwei06 已提交
421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436
    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,
                 is_shared=None,
                 initializer=None):
Z
zhangjinchao01 已提交
437 438
        self.add_keys(locals())

Q
qijun 已提交
439

Z
zhangjinchao01 已提交
440 441 442 443 444 445 446
# Define one input for a layer
@config_class
class Input(Cfg):
    def __init__(
            self,
            input_layer_name,
            parameter_name=None,
X
xuwei06 已提交
447
            initializer=None,
Z
zhangjinchao01 已提交
448 449 450 451 452 453 454 455 456 457 458 459 460
            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
461
            bilinear_interp=None,
Z
zhangjinchao01 已提交
462 463 464 465
            norm=None,
            pool=None,
            image=None,
            block_expand=None,
466
            maxout=None,
Q
qijun 已提交
467
            spp=None,
D
dangqingqing 已提交
468
            pad=None,
Z
zhangjinchao01 已提交
469 470 471 472 473
            format=None,
            nnz=None,
            is_static=None,
            is_shared=None,
            update_hooks=None,
474
            input_layer_argument=None,
D
dangqingqing 已提交
475 476 477 478 479
            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
480
        self.add_keys(locals())
D
dangqingqing 已提交
481 482 483
        self.input_layer_name = MakeLayerNameInSubmodel(
            input_layer_name
        ) if make_layer_name_in_submodel else input_layer_name
Z
zhangjinchao01 已提交
484

Q
qijun 已提交
485

Z
zhangjinchao01 已提交
486 487 488
# Define a projection for iexed layer
@config_class
class Projection(Input):
Q
qijun 已提交
489 490
    type = None  # subclass should set it correctly

Z
zhangjinchao01 已提交
491 492 493
    def __init__(
            self,
            input_layer_name,
Q
qijun 已提交
494
            size=0,  # projection output size
Z
zhangjinchao01 已提交
495 496 497 498 499 500 501 502 503
            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,
X
xuwei06 已提交
504
            initializer=None,
Z
zhangjinchao01 已提交
505 506 507 508 509 510 511 512 513 514
            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
515
            input_layer_argument=None, ):
Z
zhangjinchao01 已提交
516 517 518 519 520 521 522 523 524 525 526 527 528
        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
529

Z
zhangjinchao01 已提交
530 531
    def calc_parameter_size(self, input_size, output_size):
        raise NotimplementedError
Q
qijun 已提交
532

Z
zhangjinchao01 已提交
533 534 535 536 537 538 539 540 541 542
    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
543

Z
zhangjinchao01 已提交
544 545
    def calc_parameter_size(self, input_size, output_size):
        return 0
Q
qijun 已提交
546

Z
zhangjinchao01 已提交
547 548 549
    def calc_parameter_dims(self, input_size, output_size):
        return []

Q
qijun 已提交
550

Z
zhangjinchao01 已提交
551 552 553 554 555 556
# 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
557 558 559
    def __init__(self, input_layer_name, offset, **xargs):
        super(IdentityOffsetProjection, self).__init__(input_layer_name,
                                                       **xargs)
Z
zhangjinchao01 已提交
560 561
        self.proj_conf.offset = offset

562 563 564
    def calc_output_size(self, input_layer_config):
        return 0  # depends on the outside MixedLayer

Z
zhangjinchao01 已提交
565 566
    def calc_parameter_size(self, input_size, output_size):
        return 0
Q
qijun 已提交
567

Z
zhangjinchao01 已提交
568 569 570
    def calc_parameter_dims(self, input_size, output_size):
        return []

Q
qijun 已提交
571

572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600
@config_class
class SliceProjection(Projection):
    type = 'slice'

    def __init__(self, input_layer_name, slices, **xargs):
        super(SliceProjection, self).__init__(input_layer_name, **xargs)
        input = g_layer_map[input_layer_name]
        if input.type in ["exconv", "cudnn_conv"]:
            # the slice operator is for the channel dimension
            assert input.num_filters is not None
            channels = input.num_filters
            image_size = input.size / channels
            assert slices[len(slices) - 1][1] <= channels
            for i in xrange(len(slices)):
                slice = self.proj_conf.slices.add()
                slice.start = slices[i][0] * image_size
                slice.end = slices[i][1] * image_size
                self.size += slice.end - slice.start
        else:
            config_assert(False,
                          'Currently the input should be convolution layer')

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

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


Z
zhangjinchao01 已提交
601 602 603 604 605 606 607
# 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
608

Z
zhangjinchao01 已提交
609 610
    def calc_parameter_size(self, input_size, output_size):
        return output_size
Q
qijun 已提交
611

Z
zhangjinchao01 已提交
612 613 614
    def calc_parameter_dims(self, input_size, output_size):
        return [1, output_size]

L
Luo Tao 已提交
615

X
xuwei06 已提交
616 617 618 619 620 621 622 623 624 625 626 627 628 629
# 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
630

Z
zhangjinchao01 已提交
631 632 633 634 635 636
@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
637

Z
zhangjinchao01 已提交
638 639 640
    def calc_parameter_dims(self, input_size, output_size):
        return [input_size, output_size]

Q
qijun 已提交
641

Z
zhangjinchao01 已提交
642 643 644 645 646 647
@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
648

Z
zhangjinchao01 已提交
649 650 651
    def calc_parameter_dims(self, input_size, output_size):
        return [input_size, output_size]

Q
qijun 已提交
652

Z
zhangjinchao01 已提交
653 654 655 656 657 658
@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
659

Z
zhangjinchao01 已提交
660 661 662
    def calc_parameter_dims(self, input_size, output_size):
        return [output_size, input_size]

Q
qijun 已提交
663

Z
zhangjinchao01 已提交
664 665 666 667
@config_class
class ContextProjection(Projection):
    type = 'context'

Q
qijun 已提交
668 669
    def __init__(self, input_layer_name, context_start, context_length,
                 trainable_padding, **xargs):
Z
zhangjinchao01 已提交
670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692
        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


693
@config_class
694
class ConvBaseProjection(Projection):
Q
qijun 已提交
695 696 697 698 699
    def __init__(self,
                 input_layer_name,
                 num_filters=None,
                 conv_conf=None,
                 **xargs):
700
        super(ConvBaseProjection, self).__init__(input_layer_name, **xargs)
701 702 703 704 705 706 707 708 709 710 711 712

        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
713 714
        gr = self.proj_conf.conv_conf.groups
        return co * ci * fh * fw / gr
715 716 717 718 719 720 721

    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
722

723 724 725 726 727 728 729 730 731
@config_class
class ConvProjection(ConvBaseProjection):
    type = 'conv'

    def __init__(self,
                 input_layer_name,
                 num_filters=None,
                 conv_conf=None,
                 **xargs):
732 733
        super(ConvProjection, self).__init__(input_layer_name, num_filters,
                                             conv_conf, **xargs)
734

735
        parse_conv(conv_conf, self.input_layer_name, self.proj_conf.conv_conf,
736 737 738 739 740 741 742 743 744 745 746 747 748 749 750
                   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):
751 752
        super(ConvTransProjection, self).__init__(input_layer_name, num_filters,
                                                  conv_conf, **xargs)
753 754 755

        parse_conv(
            conv_conf,
756
            self.input_layer_name,
757 758 759 760 761 762 763 764
            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
765 766 767
# Define a operator for mixed layer
@config_class
class Operator(Cfg):
Q
qijun 已提交
768 769
    type = None  # subclass should set it correctly

Z
zhangjinchao01 已提交
770 771
    def __init__(
            self,
Q
qijun 已提交
772
            input_layer_names, ):
Z
zhangjinchao01 已提交
773 774 775 776 777 778 779 780 781 782
        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
783

Z
zhangjinchao01 已提交
784 785 786
@config_class
class DotMulOperator(Operator):
    type = 'dot_mul'
Q
qijun 已提交
787 788 789

    def __init__(self, input_layer_names, scale=None, **xargs):
        super(DotMulOperator, self).__init__(input_layer_names, **xargs)
Z
zhangjinchao01 已提交
790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807
        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
808 809 810 811 812 813 814

    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
815 816 817
        if num_filters is not None:
            self.operator_conf.num_filters = num_filters

818 819
        parse_conv(conv_conf,
                   MakeLayerNameInSubmodel(input_layer_names[0]),
Q
qijun 已提交
820
                   self.operator_conf.conv_conf, num_filters)
L
Luo Tao 已提交
821 822 823
        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
824 825 826

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

827 828
    def calc_output_size(self, input_sizes):
        return self.operator_conf.output_size
Z
zhangjinchao01 已提交
829 830


831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860
@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
861 862 863
# please refer to the comments in proto/ModelConfig.proto
@config_class
class Conv(Cfg):
Q
qijun 已提交
864 865 866 867 868 869 870 871 872 873 874 875
    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,
W
wanghaoshuang 已提交
876 877 878
                 stride_y=None,
                 dilation=None,
                 dilation_y=None):
Z
zhangjinchao01 已提交
879 880
        self.add_keys(locals())
        if filter_size_y is None:
Q
qijun 已提交
881
            self.filter_size_y = filter_size
Z
zhangjinchao01 已提交
882
        if padding_y is None:
Q
qijun 已提交
883
            self.padding_y = padding
884 885
        if dilation_y is None:
            self.dilation_y = dilation
Z
zhangjinchao01 已提交
886
        if stride_y is None:
Q
qijun 已提交
887
            self.stride_y = stride
Z
zhangjinchao01 已提交
888
        if output_x is not None:
Q
qijun 已提交
889 890
            config_assert(output_x <= 0)

Z
zhangjinchao01 已提交
891

892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911
# please refer to the comments in proto/ModelConfig.proto
@config_class
class Conv3D(Cfg):
    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,
                 filter_size_z=None,
                 padding_z=None,
                 stride_z=None):
        self.add_keys(locals())
C
chengduoZH 已提交
912 913 914 915 916 917
        self.filter_size_y = filter_size_y if filter_size_y else filter_size
        self.filter_size_z = filter_size_z if filter_size_z else filter_size
        self.padding_y = padding_y if padding_y else padding
        self.padding_z = padding_z if padding_z else padding
        self.stride_y = stride_y if stride_y else stride
        self.stride_z = stride_z if stride_z else stride
918 919 920 921
        if output_x is not None:
            config_assert(output_x <= 0)


L
liaogang 已提交
922 923
@config_class
class BilinearInterp(Cfg):
L
Luo Tao 已提交
924
    def __init__(self, out_size_x=None, out_size_y=None, channels=None):
L
liaogang 已提交
925 926
        self.add_keys(locals())

Q
qijun 已提交
927

Z
zhangjinchao01 已提交
928 929
@config_class
class Pool(Cfg):
D
dangqingqing 已提交
930 931 932 933 934 935 936 937 938 939 940
    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
941
        self.add_keys(locals())
Q
qijun 已提交
942 943


C
chengduoZH 已提交
944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968
@config_class
class Pool3d(Cfg):
    def __init__(
            self,
            pool_type,
            channels,
            size_x,
            size_y=None,
            size_z=None,
            start=None,
            stride=None,  # 1 by defalut in protobuf
            stride_y=None,
            stride_z=None,
            padding=None,  # 0 by defalut in protobuf
            padding_y=None,
            padding_z=None):
        self.add_keys(locals())
        self.filter_size_y = size_y if size_y else size_x
        self.filter_size_z = size_z if size_z else size_x
        self.padding_y = padding_y if padding_y else padding
        self.padding_z = padding_z if padding_z else padding
        self.stride_y = stride_y if stride_y else stride
        self.stride_z = stride_z if stride_z else stride


Q
qijun 已提交
969
@config_class
Q
qijun 已提交
970
class SpatialPyramidPool(Cfg):
L
Luo Tao 已提交
971
    def __init__(self, pool_type, pyramid_height, channels):
Q
qijun 已提交
972
        self.add_keys(locals())
Z
zhangjinchao01 已提交
973

Q
qijun 已提交
974

D
dangqingqing 已提交
975 976 977 978 979 980
@config_class
class Pad(Cfg):
    def __init__(self, channels, pad_c, pad_h, pad_w):
        self.add_keys(locals())


Z
zhangjinchao01 已提交
981 982
@config_class
class Norm(Cfg):
Q
qijun 已提交
983 984 985 986 987 988 989 990 991
    def __init__(self,
                 norm_type,
                 channels,
                 size,
                 scale,
                 pow,
                 output_x=None,
                 img_size=None,
                 blocked=None):
Z
zhangjinchao01 已提交
992 993
        self.add_keys(locals())

Q
qijun 已提交
994

Z
zhangjinchao01 已提交
995 996
@config_class
class Image(Cfg):
Q
qijun 已提交
997
    def __init__(self, channels, img_size=None):
Z
zhangjinchao01 已提交
998 999
        self.add_keys(locals())

Q
qijun 已提交
1000

Z
zhangjinchao01 已提交
1001 1002
@config_class
class BlockExpand(Cfg):
Q
qijun 已提交
1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014
    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
1015 1016
        self.add_keys(locals())

Q
qijun 已提交
1017

1018 1019
@config_class
class MaxOut(Cfg):
Q
qijun 已提交
1020
    def __init__(self, channels, groups, img_size_x=0, img_size_y=0):
1021 1022
        self.add_keys(locals())

Q
qijun 已提交
1023

1024
def create_data_config_proto(async_load_data=False,
1025
                             constant_slots=None,
王益 已提交
1026 1027 1028
                             data_ratio=1,
                             is_main_data=True,
                             usage_ratio=None):
Z
zhangjinchao01 已提交
1029 1030 1031 1032 1033 1034 1035 1036
    # 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
1037 1038
    data_config.data_ratio = data_ratio
    data_config.is_main_data = is_main_data
Z
zhangjinchao01 已提交
1039

Q
qijun 已提交
1040
    usage_ratio = default(usage_ratio, settings_deprecated["usage_ratio"])
Z
zhangjinchao01 已提交
1041 1042 1043 1044 1045 1046
    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
1047

Z
zhangjinchao01 已提交
1048
@config_func
Q
qijun 已提交
1049 1050 1051 1052 1053
def SimpleData(files=None,
               feat_dim=None,
               context_len=None,
               buffer_capacity=None,
               **xargs):
1054
    data_config = create_data_config_proto(**xargs)
Z
zhangjinchao01 已提交
1055 1056 1057 1058 1059 1060 1061 1062 1063
    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
1064

Z
zhangjinchao01 已提交
1065
@config_func
Q
qijun 已提交
1066 1067 1068 1069 1070 1071 1072 1073 1074 1075
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):
1076
    data_config = create_data_config_proto(**xargs)
Z
zhangjinchao01 已提交
1077 1078
    data_config.type = 'py'
    if load_data_module in g_py_module_name_list:
Q
qijun 已提交
1079

Z
zhangjinchao01 已提交
1080 1081 1082
        def get_path(module):
            m = __import__(load_data_module)
            return os.path.split(os.path.realpath(m.__file__))[0]
Q
qijun 已提交
1083

Z
zhangjinchao01 已提交
1084 1085 1086
        # 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
1087 1088
        module_new_name = "%s_copy_%d" % (load_data_module,
                                          len(g_py_module_name_list))
Z
zhangjinchao01 已提交
1089
        g_py_module_name_list.append(module_new_name)
Q
qijun 已提交
1090 1091 1092 1093
        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
1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117
        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
1118

Z
zhangjinchao01 已提交
1119
@config_func
Q
qijun 已提交
1120 1121 1122 1123 1124 1125 1126
def ProtoData(files=None,
              type=None,
              file_group_queue_capacity=None,
              load_file_count=None,
              constant_slots=None,
              load_thread_num=None,
              **xargs):
1127
    data_config = create_data_config_proto(**xargs)
Z
zhangjinchao01 已提交
1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146
    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
1147

Z
zhangjinchao01 已提交
1148 1149
#real data for training is actually provided by "sub_data" data providers.
@config_func
Q
qijun 已提交
1150
def MultiData(sub_data=[]):
Z
zhangjinchao01 已提交
1151 1152 1153 1154 1155
    data_config = DataConfig()
    data_config.type = 'multi'
    data_config.sub_data_configs.extend(sub_data)
    return data_config

Q
qijun 已提交
1156

Z
zhangjinchao01 已提交
1157
@config_func
Q
qijun 已提交
1158 1159 1160 1161 1162 1163 1164
def Data(type,
         files=None,
         feat_dim=None,
         slot_dims=None,
         context_len=None,
         buffer_capacity=None,
         **xargs):
Z
zhangjinchao01 已提交
1165

1166
    data_config = create_data_config_proto(**xargs)
Z
zhangjinchao01 已提交
1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199
    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
1200

L
Luo Tao 已提交
1201 1202
#caffe_mode: compute the output size using floor instead of ceil,
#            which is consistent of caffe and CuDNN's convention.
1203 1204 1205 1206 1207 1208 1209
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
1210

1211
#calcualte image_size based on output_size for de-convolution (ConvTransLayer).
L
Luo Tao 已提交
1212
#It is the reverse function of cnn_output_size
1213
def cnn_image_size(output_size, filter_size, padding, stride, caffe_mode):
L
Luo Tao 已提交
1214 1215 1216
    img_size = (output_size - 1) * stride + filter_size - 2 * padding
    if not caffe_mode:
        img_size = img_size + 1
1217 1218
    return img_size

Q
qijun 已提交
1219

L
Luo Tao 已提交
1220
def get_img_size(input_layer_name, channels):
L
Luo Tao 已提交
1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232
    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


C
chengduoZH 已提交
1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246
def get_img3d_size(input_layer_name, channels):
    input = g_layer_map[input_layer_name]
    img_pixels = input.size / channels
    img_size = input.width
    img_size_y = input.height
    img_size_z = input.depth

    config_assert(
        img_size * img_size_y * img_size_z == img_pixels,
        "Input layer %s: Incorrect input image size %d * %d * %d for input image pixels %d"
        % (input_layer_name, img_size, img_size_y, img_size_z, img_pixels))
    return img_size, img_size_y, img_size_z


L
Luo Tao 已提交
1247 1248 1249 1250 1251 1252
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


1253
def parse_pool(pool, input_layer_name, pool_conf, ceil_mode):
Z
zhangjinchao01 已提交
1254
    pool_conf.pool_type = pool.pool_type
Q
qijun 已提交
1255 1256 1257
    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
1258
                  "['max-projection', 'avg-projection', "
Q
qijun 已提交
1259
                  "'cudnn-max-pool', 'cudnn-avg-pool']" % pool.pool_type)
Z
zhangjinchao01 已提交
1260 1261 1262 1263 1264 1265

    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
1266
    pool_conf.stride_y = default(pool.stride_y, pool_conf.stride)
Z
zhangjinchao01 已提交
1267

L
Luo Tao 已提交
1268
    pool_conf.img_size, pool_conf.img_size_y = \
L
Luo Tao 已提交
1269
        get_img_size(input_layer_name, pool.channels)
Z
zhangjinchao01 已提交
1270

1271
    config_assert(not pool.start, "start is deprecated in pooling.")
Z
zhangjinchao01 已提交
1272

1273
    if pool.padding is not None:
Z
zhangjinchao01 已提交
1274
        pool_conf.padding = pool.padding
1275
    pool_conf.padding_y = default(pool.padding_y, pool_conf.padding)
D
dangqingqing 已提交
1276 1277
    pool_conf.output_x = cnn_output_size(pool_conf.img_size, pool_conf.size_x,
                                         pool_conf.padding, pool_conf.stride,
1278
                                         not ceil_mode)
D
dangqingqing 已提交
1279 1280
    pool_conf.output_y = cnn_output_size(pool_conf.img_size_y, pool_conf.size_y,
                                         pool_conf.padding_y,
1281
                                         pool_conf.stride_y, not ceil_mode)
Q
qijun 已提交
1282

Z
zhangjinchao01 已提交
1283

C
chengduoZH 已提交
1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322
def parse_pool3d(pool, input_layer_name, pool_conf, ceil_mode):
    pool_conf.pool_type = pool.pool_type
    config_assert(pool.pool_type in ['max-projection', 'avg-projection'],
                  "pool-type %s is not in "
                  "['max-projection', 'avg-projection']" % pool.pool_type)

    pool_conf.channels = pool.channels

    pool_conf.size_x = pool.size_x
    pool_conf.stride = pool.stride
    pool_conf.padding = pool.padding

    pool_conf.size_y = default(pool.size_y, pool_conf.size_x)
    pool_conf.size_z = default(pool.size_z, pool_conf.size_x)
    pool_conf.stride_y = default(pool.stride_y, pool_conf.stride)
    pool_conf.stride_z = default(pool.stride_z, pool_conf.stride)
    pool_conf.padding_y = default(pool.padding_y, pool_conf.padding)
    pool_conf.padding_z = default(pool.padding_z, pool_conf.padding)

    pool_conf.img_size, pool_conf.img_size_y, pool_conf.img_size_z = \
        get_img3d_size(input_layer_name, pool.channels)

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

    if pool.padding is not None:
        pool_conf.padding = pool.padding
    pool_conf.padding_y = default(pool.padding_y, pool_conf.padding)
    pool_conf.padding_z = default(pool.padding_z, pool_conf.padding)
    pool_conf.output_x = cnn_output_size(pool_conf.img_size, pool_conf.size_x,
                                         pool_conf.padding, pool_conf.stride,
                                         not ceil_mode)
    pool_conf.output_y = cnn_output_size(pool_conf.img_size_y, pool_conf.size_y,
                                         pool_conf.padding_y,
                                         pool_conf.stride_y, not ceil_mode)
    pool_conf.output_z = cnn_output_size(pool_conf.img_size_z, pool_conf.size_z,
                                         pool_conf.padding_z,
                                         pool_conf.stride_z, not ceil_mode)


Q
qijun 已提交
1323
def parse_spp(spp, input_layer_name, spp_conf):
L
Luo Tao 已提交
1324
    parse_image(spp, input_layer_name, spp_conf.image_conf)
Q
qijun 已提交
1325 1326
    spp_conf.pool_type = spp.pool_type
    config_assert(spp.pool_type in ['max-projection', 'avg-projection'],
Q
qijun 已提交
1327 1328
                  "pool-type %s is not in "
                  "['max-projection', 'avg-projection']" % spp.pool_type)
Q
qijun 已提交
1329
    spp_conf.pyramid_height = spp.pyramid_height
Q
qijun 已提交
1330

Q
qijun 已提交
1331

Z
zhangjinchao01 已提交
1332 1333
def parse_image(image, input_layer_name, image_conf):
    image_conf.channels = image.channels
L
Luo Tao 已提交
1334
    image_conf.img_size, image_conf.img_size_y = \
L
Luo Tao 已提交
1335
        get_img_size(input_layer_name, image_conf.channels)
Q
qijun 已提交
1336

Z
zhangjinchao01 已提交
1337

C
chengduoZH 已提交
1338 1339 1340 1341 1342 1343
def parse_image3d(image, input_layer_name, image_conf):
    image_conf.channels = image.channels
    image_conf.img_size, image_conf.img_size_y, image_conf.img_size_z = \
        get_img3d_size(input_layer_name, image_conf.channels)


Z
zhangjinchao01 已提交
1344 1345
def parse_norm(norm, input_layer_name, norm_conf):
    norm_conf.norm_type = norm.norm_type
1346 1347 1348 1349 1350
    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
1351 1352 1353 1354 1355 1356
    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
1357
    norm_conf.img_size, norm_conf.img_size_y = \
L
Luo Tao 已提交
1358
        get_img_size(input_layer_name, norm.channels)
Z
zhangjinchao01 已提交
1359
    norm_conf.output_x = norm_conf.img_size
L
Luo Tao 已提交
1360
    norm_conf.output_y = norm_conf.img_size_y
Z
zhangjinchao01 已提交
1361 1362 1363
    if norm.norm_type in ['cmrnorm-projection']:
        norm_conf.scale /= norm.size
    else:
Q
qijun 已提交
1364 1365
        norm_conf.scale /= norm.size**2

1366

L
Luo Tao 已提交
1367 1368
#caffe_mode: compute the output size using floor instead of ceil,
#            which is consistent of caffe and CuDNN's convention.
1369
def parse_conv(conv, input_layer_name, conv_conf, num_filters, trans=False):
Z
zhangjinchao01 已提交
1370 1371 1372 1373 1374 1375 1376 1377 1378
    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
1379

1380
    if not trans:
1381
        conv_conf.filter_channels = conv.channels / conv.groups
L
Luo Tao 已提交
1382
        conv_conf.img_size, conv_conf.img_size_y = \
L
Luo Tao 已提交
1383
            get_img_size(input_layer_name, conv.channels)
1384
        conv_conf.output_x = cnn_output_size(
Q
qijun 已提交
1385 1386
            conv_conf.img_size, conv_conf.filter_size, conv_conf.padding,
            conv_conf.stride, conv_conf.caffe_mode)
L
Luo Tao 已提交
1387 1388 1389
        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)
1390
    else:
1391
        conv_conf.filter_channels = num_filters / conv.groups
L
Luo Tao 已提交
1392
        conv_conf.output_x, conv_conf.output_y = \
L
Luo Tao 已提交
1393
            get_img_size(input_layer_name, conv.channels)
1394
        conv_conf.img_size = cnn_image_size(
Q
qijun 已提交
1395 1396
            conv_conf.output_x, conv_conf.filter_size, conv_conf.padding,
            conv_conf.stride, conv_conf.caffe_mode)
L
Luo Tao 已提交
1397
        conv_conf.img_size_y = cnn_image_size(
L
Luo Tao 已提交
1398 1399
            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
1400

1401

1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445
#caffe_mode: compute the output size using floor instead of ceil,
#            which is consistent of caffe and CuDNN's convention.
def parse_conv3d(conv, input_layer_name, conv_conf, num_filters, trans=False):
    conv_conf.filter_size = conv.filter_size
    conv_conf.filter_size_y = conv.filter_size_y
    conv_conf.filter_size_z = conv.filter_size_z
    conv_conf.channels = conv.channels
    conv_conf.padding = conv.padding
    conv_conf.padding_y = conv.padding_y
    conv_conf.padding_z = conv.padding_z
    conv_conf.stride = conv.stride
    conv_conf.stride_y = conv.stride_y
    conv_conf.stride_z = conv.stride_z
    conv_conf.groups = conv.groups
    conv_conf.caffe_mode = conv.caffe_mode

    if not trans:
        conv_conf.filter_channels = conv.channels / conv.groups
        conv_conf.img_size, conv_conf.img_size_y, conv_conf.img_size_z = \
            get_img3d_size(input_layer_name, conv.channels)
        conv_conf.output_x = cnn_output_size(
            conv_conf.img_size, conv_conf.filter_size, conv_conf.padding,
            conv_conf.stride, conv_conf.caffe_mode)
        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)
        conv_conf.output_z = cnn_output_size(
            conv_conf.img_size_z, conv_conf.filter_size_z, conv_conf.padding_z,
            conv_conf.stride_z, conv_conf.caffe_mode)
    else:
        conv_conf.filter_channels = num_filters / conv.groups
        conv_conf.output_x, conv_conf.output_y, conv_conf.output_z = \
            get_img3d_size(input_layer_name, conv.channels)
        conv_conf.img_size = cnn_image_size(
            conv_conf.output_x, conv_conf.filter_size, conv_conf.padding,
            conv_conf.stride, conv_conf.caffe_mode)
        conv_conf.img_size_y = cnn_image_size(
            conv_conf.output_y, conv_conf.filter_size_y, conv_conf.padding_y,
            conv_conf.stride_y, conv_conf.caffe_mode)
        conv_conf.img_size_z = cnn_image_size(
            conv_conf.output_z, conv_conf.filter_size_z, conv_conf.padding_z,
            conv_conf.stride_z, conv_conf.caffe_mode)


Z
zhangjinchao01 已提交
1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458
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:
1459
        block_expand_conf.output_x = cnn_output_size(
1460
            block_expand.img_size_x, block_expand.block_x,
1461
            block_expand.padding_x, block_expand.stride_x, False)
Z
zhangjinchao01 已提交
1462 1463

    if block_expand_conf.img_size_y == 0:
1464
        block_expand_conf.output_y = 0
Z
zhangjinchao01 已提交
1465
    else:
1466
        block_expand_conf.output_y = cnn_output_size(
1467
            block_expand.img_size_y, block_expand.block_y,
1468
            block_expand.padding_y, block_expand.stride_y, False)
Z
zhangjinchao01 已提交
1469

Q
qijun 已提交
1470

1471
def parse_maxout(maxout, input_layer_name, maxout_conf):
L
Luo Tao 已提交
1472
    parse_image(maxout, input_layer_name, maxout_conf.image_conf)
1473
    maxout_conf.groups = maxout.groups
1474

Q
qijun 已提交
1475

Z
zhangjinchao01 已提交
1476 1477
# Define an evaluator
@config_func
Y
yangyaming 已提交
1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494
def Evaluator(name,
              type,
              inputs,
              chunk_scheme=None,
              num_chunk_types=None,
              classification_threshold=None,
              positive_label=None,
              dict_file=None,
              result_file=None,
              num_results=None,
              top_k=None,
              delimited=None,
              excluded_chunk_types=None,
              overlap_threshold=None,
              background_id=None,
              evaluate_difficult=None,
              ap_type=None):
Z
zhangjinchao01 已提交
1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508
    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)

1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519
    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
1520 1521
    if top_k is not None:
        evaluator.top_k = top_k
1522 1523
    if delimited is not None:
        evaluator.delimited = delimited
Z
zhangjinchao01 已提交
1524

1525 1526 1527
    if excluded_chunk_types:
        evaluator.excluded_chunk_types.extend(excluded_chunk_types)

Y
yangyaming 已提交
1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539
    if overlap_threshold is not None:
        evaluator.overlap_threshold = overlap_threshold

    if background_id is not None:
        evaluator.background_id = background_id

    if evaluate_difficult is not None:
        evaluator.evaluate_difficult = evaluate_difficult

    if ap_type is not None:
        evaluator.ap_type = ap_type

Q
qijun 已提交
1540

Z
zhangjinchao01 已提交
1541 1542 1543 1544 1545
class LayerBase(object):
    def __init__(
            self,
            name,
            type,
Q
qijun 已提交
1546
            size,  # size can be 0. In this case, subclass should set it.
Z
zhangjinchao01 已提交
1547 1548 1549 1550
            inputs,
            device=None,
            active_type="",
            drop_rate=0.,
C
caoying03 已提交
1551 1552
            coeff=None,
            error_clipping_threshold=None):
Z
zhangjinchao01 已提交
1553
        config_assert('@' not in name,
Q
qijun 已提交
1554
                      "layer name: %s contain special character @" % name)
Z
zhangjinchao01 已提交
1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569
        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()
1570
        assert isinstance(self.config, LayerConfig)
1571
        use_mkldnn = bool(int(g_command_config_args.get("use_mkldnn", 0)))
T
tensor-tang 已提交
1572
        mkldnn_acts = ['relu', 'tanh', 'softmax']
1573 1574
        if use_mkldnn and active_type in mkldnn_acts:
            active_type = "mkldnn_" + active_type
Z
zhangjinchao01 已提交
1575 1576 1577
        self.config.name = name
        self.config.type = type
        self.config.active_type = active_type
1578 1579
        if coeff is not None:
            self.config.coeff = float(coeff)
Z
zhangjinchao01 已提交
1580 1581 1582 1583 1584 1585 1586
        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
1587
        elif g_default_device is not None:
Z
zhangjinchao01 已提交
1588 1589
            self.config.device = g_default_device

C
caoying03 已提交
1590 1591 1592
        if error_clipping_threshold is not None:
            self.config.error_clipping_threshold = error_clipping_threshold

Z
zhangjinchao01 已提交
1593 1594 1595 1596 1597 1598 1599
        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
1600 1601
                    input_layer_name=input,
                    parameter_name=gen_parameter_name(name, input_index))
Z
zhangjinchao01 已提交
1602 1603 1604 1605 1606 1607 1608 1609
                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
1610
                self.operators.append(input)
Z
zhangjinchao01 已提交
1611 1612 1613 1614
                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
1615
                raise ValueError('Wrong type for inputs: %s' % type_of(input))
Z
zhangjinchao01 已提交
1616
            config_assert(input_layer_name in g_layer_map,
Q
qijun 已提交
1617 1618
                          "Unknown input layer '%s' for layer %s" %
                          (input_layer_name, name))
Z
zhangjinchao01 已提交
1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635
            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
1636
            bias,  # True/False or BiasCfg
Z
zhangjinchao01 已提交
1637
            size,
Q
qijun 已提交
1638 1639 1640
            dims=None,
            for_self=True,  # whether create bias for layer self
    ):
Z
zhangjinchao01 已提交
1641 1642 1643 1644 1645 1646

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

Q
qijun 已提交
1647 1648 1649
        config_assert(
            type_of(bias) == bool or type_of(bias) == Bias,
            'Incorrect type for bias: %s' % type_of(bias))
Z
zhangjinchao01 已提交
1650 1651 1652 1653 1654 1655 1656 1657 1658

        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:
1659 1660
                assert isinstance(self.config, LayerConfig)

Z
zhangjinchao01 已提交
1661 1662 1663
                Parameter(
                    bias.parameter_name,
                    size,
Q
qijun 已提交
1664 1665
                    self.config.device
                    if self.config.HasField('device') else None,
Z
zhangjinchao01 已提交
1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676
                    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
1677 1678
                    gradient_clipping_threshold=bias.
                    gradient_clipping_threshold,
Z
zhangjinchao01 已提交
1679
                    is_static=bias.is_static,
X
xuwei06 已提交
1680 1681
                    is_shared=bias.is_shared,
                    initializer=bias.initializer)
Z
zhangjinchao01 已提交
1682 1683 1684 1685 1686
            if for_self:
                self.config.bias_parameter_name = bias.parameter_name
            else:
                return bias.parameter_name

Q
qijun 已提交
1687 1688 1689 1690 1691 1692
    def create_input_parameter(self,
                               input_index,
                               size,
                               dims=None,
                               sparse=None,
                               format=None):
Z
zhangjinchao01 已提交
1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706
        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
1707 1708
            config_assert(size == para.size, (
                'Shared parameter "%s" does not ' + 'have same size: %s vs. %s')
Z
zhangjinchao01 已提交
1709 1710
                          % (input_config.parameter_name, para.size, size))

Q
qijun 已提交
1711 1712
            config_assert(dims == para.dims, (
                'Shared parameter "%s" does not ' + 'have same dims: %s vs. %s')
Z
zhangjinchao01 已提交
1713 1714 1715 1716 1717 1718
                          % (input_config.parameter_name, para.dims, dims))
            return

        Parameter(
            input_config.parameter_name,
            size,
1719
            self.config.device if self.config.HasField("device") else None,
Z
zhangjinchao01 已提交
1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731
            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
1732 1733
            gradient_clipping_threshold=input_config.
            gradient_clipping_threshold,
Z
zhangjinchao01 已提交
1734 1735 1736 1737
            sparse=sparse,
            format=format,
            is_static=input_config.is_static,
            is_shared=input_config.is_shared,
X
xuwei06 已提交
1738 1739
            update_hooks=input_config.update_hooks,
            initializer=input_config.initializer)
Z
zhangjinchao01 已提交
1740 1741 1742 1743 1744 1745 1746 1747 1748

    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
1749 1750 1751 1752
    def set_layer_height_width(self, height, width):
        self.config.height = height
        self.config.width = width

C
chengduoZH 已提交
1753 1754 1755
    def set_layer_depth(self, depth):
        self.config.depth = depth

L
Luo Tao 已提交
1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768
    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
1769

Z
zhangjinchao01 已提交
1770 1771
@config_layer('multi_class_cross_entropy_with_selfnorm')
class MultiClassCrossEntropySelfNormCostLayer(LayerBase):
Q
qijun 已提交
1772 1773 1774
    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
1775 1776
        self.config.softmax_selfnorm_alpha = softmax_selfnorm_alpha

Q
qijun 已提交
1777

C
caoying03 已提交
1778 1779 1780
@config_layer('cross_entropy_over_beam')
class CrossEntropyOverBeamLayer(LayerBase):
    def __init__(self, name, inputs, **xargs):
C
caoying03 已提交
1781
        config_assert(len(inputs) % 3 == 0, "Error input number.")
C
caoying03 已提交
1782 1783 1784 1785
        super(CrossEntropyOverBeamLayer, self).__init__(
            name, 'cross_entropy_over_beam', 0, inputs, **xargs)
        input_num = len(inputs) / 3
        for i in range(input_num):
C
caoying03 已提交
1786 1787 1788 1789 1790
            input_layer = self.get_input_layer(i * 3)
            config_assert(input_layer.size == 1, (
                "Inputs for this layer are made up of "
                "several triples, in which the first one is scores over "
                "all candidate paths, whose size should be equal to 1."))
C
caoying03 已提交
1791 1792


Z
zhangjinchao01 已提交
1793 1794
@config_layer('fc')
class FCLayer(LayerBase):
T
tensor-tang 已提交
1795 1796
    layer_type = 'fc'

L
lianxiaochen 已提交
1797 1798 1799 1800 1801 1802 1803
    def __init__(self,
                 name,
                 size,
                 inputs,
                 bias=True,
                 error_clipping_threshold=None,
                 **xargs):
T
tensor-tang 已提交
1804
        use_mkldnn = bool(int(g_command_config_args.get("use_mkldnn", 0)))
1805 1806
        use_mkldnn_wgt = bool(
            int(g_command_config_args.get("use_mkldnn_wgt", 0)))
T
tensor-tang 已提交
1807 1808 1809 1810 1811 1812 1813
        if use_mkldnn:
            self.layer_type = 'mkldnn_fc'
            config_assert(
                len(inputs) == 1,
                "MkldnnFCLayer support one and only one input!")
        super(FCLayer, self).__init__(
            name, self.layer_type, size, inputs=inputs, **xargs)
Z
zhangjinchao01 已提交
1814 1815 1816 1817 1818 1819
        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"
T
tensor-tang 已提交
1820 1821 1822
            if use_mkldnn:
                config_assert(not sparse,
                              "MkldnnFCLayer do not support sparse format yet")
T
tensor-tang 已提交
1823 1824
                if use_mkldnn_wgt:
                    dims = [self.config.size, input_layer.size]
Z
zhangjinchao01 已提交
1825 1826
            if sparse:
                psize = self.inputs[input_index].nnz
1827 1828
            else:
                sparse = None
Z
zhangjinchao01 已提交
1829

Q
qijun 已提交
1830 1831
            self.create_input_parameter(input_index, psize, dims, sparse,
                                        format)
Z
zhangjinchao01 已提交
1832
        self.create_bias_parameter(bias, self.config.size)
L
lianxiaochen 已提交
1833 1834
        if error_clipping_threshold is not None:
            self.config.error_clipping_threshold = error_clipping_threshold
Z
zhangjinchao01 已提交
1835

Q
qijun 已提交
1836

T
tensor-tang 已提交
1837 1838 1839 1840 1841
@config_layer('mkldnn_fc')
class MkldnnFcLayer(FCLayer):
    layer_type = 'mkldnn_fc'


Z
zhangjinchao01 已提交
1842 1843
@config_layer('selective_fc')
class SelectiveFCLayer(LayerBase):
Q
qijun 已提交
1844 1845 1846 1847 1848 1849 1850 1851 1852 1853
    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
1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873
        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
1874 1875
                          ("if indices of selected columns are not specified, "
                           "selective_fc Layer has at least two inputs"))
Z
zhangjinchao01 已提交
1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887
            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
1888 1889
            self.create_input_parameter(input_index, psize, dims, sparse,
                                        format)
Z
zhangjinchao01 已提交
1890 1891
        self.create_bias_parameter(bias, self.config.size)

Q
qijun 已提交
1892

1893 1894
@config_layer('print')
class PrintLayer(LayerBase):
1895
    def __init__(self, name, inputs, format=None):
1896
        super(PrintLayer, self).__init__(name, 'print', 0, inputs)
1897 1898 1899 1900 1901 1902
        if format is None:
            format = "\n".join([
                "layer=" + input.input_layer_name + " %s"
                for input in self.inputs
            ])
        self.config.user_arg = format
1903

Q
qijun 已提交
1904

Y
yuan 已提交
1905 1906
@config_layer('priorbox')
class PriorBoxLayer(LayerBase):
G
gaoyuan 已提交
1907 1908
    def __init__(self, name, inputs, size, min_size, max_size, aspect_ratio,
                 variance):
Y
yuan 已提交
1909
        super(PriorBoxLayer, self).__init__(name, 'priorbox', 0, inputs)
G
gaoyuan 已提交
1910
        config_assert(len(inputs) == 2, 'PriorBoxLayer must have 2 inputs')
G
gaoyuan 已提交
1911 1912 1913 1914 1915 1916 1917
        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
1918
        config_assert(len(variance) == 4, 'The variance must have 4 inputs')
Y
yuan 已提交
1919 1920 1921 1922 1923 1924
        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
1925

1926 1927 1928
@config_layer('multibox_loss')
class MultiBoxLossLayer(LayerBase):
    def __init__(self, name, inputs, input_num, num_classes, overlap_threshold,
1929
                 neg_pos_ratio, neg_overlap, background_id, **xargs):
1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950
        super(MultiBoxLossLayer, self).__init__(name, 'multibox_loss', 0,
                                                inputs)
        config_assert(
            len(inputs) == (input_num * 2 + 2),
            'MultiBoxLossLayer does not have enough inputs')
        config_assert(num_classes > background_id,
                      'Classes number must greater than background ID')
        self.config.inputs[0].multibox_loss_conf.num_classes = num_classes
        self.config.inputs[
            0].multibox_loss_conf.overlap_threshold = overlap_threshold
        self.config.inputs[0].multibox_loss_conf.neg_pos_ratio = neg_pos_ratio
        self.config.inputs[0].multibox_loss_conf.neg_overlap = neg_overlap
        self.config.inputs[0].multibox_loss_conf.background_id = background_id
        self.config.inputs[0].multibox_loss_conf.input_num = input_num
        self.config.size = 1


@config_layer('detection_output')
class DetectionOutputLayer(LayerBase):
    def __init__(self, name, inputs, size, input_num, num_classes,
                 nms_threshold, nms_top_k, keep_top_k, confidence_threshold,
1951
                 background_id, **xargs):
1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971
        super(DetectionOutputLayer, self).__init__(name, 'detection_output', 0,
                                                   inputs)
        config_assert(
            len(inputs) == (input_num * 2 + 1),
            'DetectionOutputLayer does not have enough inputs')
        config_assert(num_classes > background_id,
                      'Classes number must greater than background ID')
        self.config.inputs[0].detection_output_conf.num_classes = num_classes
        self.config.inputs[
            0].detection_output_conf.nms_threshold = nms_threshold
        self.config.inputs[0].detection_output_conf.nms_top_k = nms_top_k
        self.config.inputs[0].detection_output_conf.keep_top_k = keep_top_k
        self.config.inputs[
            0].detection_output_conf.confidence_threshold = confidence_threshold
        self.config.inputs[
            0].detection_output_conf.background_id = background_id
        self.config.inputs[0].detection_output_conf.input_num = input_num
        self.config.size = size


G
guosheng 已提交
1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982
@config_layer('roi_pool')
class ROIPoolLayer(LayerBase):
    def __init__(self, name, inputs, pooled_width, pooled_height,
                 spatial_scale):
        super(ROIPoolLayer, self).__init__(name, 'roi_pool', 0, inputs)
        config_assert(len(inputs) == 2, 'ROIPoolLayer must have 2 inputs')
        self.config.inputs[0].roi_pool_conf.pooled_width = pooled_width
        self.config.inputs[0].roi_pool_conf.pooled_height = pooled_height
        self.config.inputs[0].roi_pool_conf.spatial_scale = spatial_scale


Z
zhangjinchao01 已提交
1983 1984
@config_layer('data')
class DataLayer(LayerBase):
C
chengduoZH 已提交
1985 1986 1987 1988 1989 1990 1991
    def __init__(self,
                 name,
                 size,
                 depth=None,
                 height=None,
                 width=None,
                 device=None):
Q
qijun 已提交
1992 1993
        super(DataLayer, self).__init__(
            name, 'data', size, inputs=[], device=device)
L
Luo Tao 已提交
1994 1995
        if height and width:
            self.set_layer_height_width(height, width)
C
chengduoZH 已提交
1996 1997
        if depth:
            self.set_layer_depth(depth)
Q
qijun 已提交
1998

Z
zhangjinchao01 已提交
1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025

'''
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
2026 2027


Z
zhangjinchao01 已提交
2028 2029
@config_layer('data_norm')
class DataNormLayer(LayerBase):
Q
qijun 已提交
2030
    def __init__(self, name, inputs, data_norm_strategy="z-score", device=None):
Z
zhangjinchao01 已提交
2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041
        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
2042

Z
zhangjinchao01 已提交
2043 2044 2045
@config_layer('prelu')
class ParameterReluLayer(LayerBase):
    layer_type = 'prelu'
Q
qijun 已提交
2046 2047

    def __init__(self, name, inputs, partial_sum=1, **args):
Z
zhangjinchao01 已提交
2048 2049 2050
        super(ParameterReluLayer, self).__init__(
            name, self.layer_type, 0, inputs=inputs, **args)
        input_layer = self.get_input_layer(0)
2051 2052 2053
        config_assert(len(self.inputs) == 1, "prelu layer has only one input.")
        config_assert(input_layer.size % partial_sum == 0,
                      "a wrong setting for partial_sum")
Z
zhangjinchao01 已提交
2054
        self.set_layer_size(input_layer.size)
C
caoying03 已提交
2055
        self.config.partial_sum = partial_sum
Z
zhangjinchao01 已提交
2056 2057
        self.create_input_parameter(0, input_layer.size / partial_sum)

Q
qijun 已提交
2058

Z
zhangjinchao01 已提交
2059 2060 2061
@config_layer('conv')
class ConvLayerBase(LayerBase):
    layer_type = 'conv'
Q
qijun 已提交
2062 2063 2064 2065 2066 2067 2068 2069

    def __init__(self,
                 name,
                 inputs=[],
                 bias=True,
                 num_filters=None,
                 shared_biases=False,
                 **xargs):
Z
zhangjinchao01 已提交
2070 2071 2072 2073 2074 2075
        super(ConvLayerBase, self).__init__(
            name, self.layer_type, 0, inputs=inputs, **xargs)

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

2076
        use_mkldnn = int(g_command_config_args.get("use_mkldnn", 0))
Z
zhangjinchao01 已提交
2077 2078 2079
        use_gpu = int(g_command_config_args.get("use_gpu", 0))
        parallel_nn = int(g_command_config_args.get("parallel_nn", 0))

2080 2081
        # Automatically select cudnn_type for GPU, exconv for CPU
        # and mkldnn_conv for MKLDNN
Z
zhangjinchao01 已提交
2082
        # if set type=conv, but still reserve the way user specify
2083
        # exconv, mkldnn_conv or cudnn_conv manually.
Z
zhangjinchao01 已提交
2084 2085 2086
        if self.layer_type == "cudnn_conv":
            config_assert(use_gpu, "cudnn_conv only support GPU")

2087 2088 2089
        if self.layer_type == "mkldnn_conv":
            config_assert(use_mkldnn, "mkldnn_conv only support MKLDNN")

Z
zhangjinchao01 已提交
2090
        if (use_gpu == 1 and self.layer_type != "exconv" and
2091
                self.layer_type != "mkldnn_conv" and
Q
qijun 已提交
2092
            (parallel_nn == 0 or self.config.device > -1)):
Z
zhangjinchao01 已提交
2093 2094
            self.layer_type = "cudnn_conv"
        else:
2095
            self.layer_type = "mkldnn_conv" if use_mkldnn else "exconv"
Z
zhangjinchao01 已提交
2096 2097 2098 2099 2100 2101 2102 2103 2104
        # 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
2105 2106
            parse_conv(self.inputs[input_index].conv, input_layer.name,
                       conv_conf, num_filters)
Z
zhangjinchao01 已提交
2107 2108
            psize = self.calc_parameter_size(conv_conf)
            self.create_input_parameter(input_index, psize)
L
Luo Tao 已提交
2109 2110
            self.set_cnn_layer(name, conv_conf.output_y, conv_conf.output_x,
                               self.config.num_filters)
Z
zhangjinchao01 已提交
2111 2112 2113 2114 2115 2116 2117 2118

        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 \
2119
               * (conv_conf.filter_size * conv_conf.filter_size_y)
Z
zhangjinchao01 已提交
2120

Q
qijun 已提交
2121

Z
zhangjinchao01 已提交
2122 2123 2124 2125
@config_layer('exconv')
class ConvLayer(ConvLayerBase):
    layer_type = 'exconv'

Q
qijun 已提交
2126

2127 2128 2129 2130 2131
@config_layer('mkldnn_conv')
class ConvLayer(ConvLayerBase):
    layer_type = 'mkldnn_conv'


Z
zhangjinchao01 已提交
2132 2133 2134 2135
@config_layer('cudnn_conv')
class ConvLayer(ConvLayerBase):
    layer_type = 'cudnn_conv'

2136 2137 2138 2139

@config_layer('convt')
class ConvTransLayerBase(LayerBase):
    layer_type = 'convt'
Q
qijun 已提交
2140 2141 2142 2143 2144 2145 2146 2147

    def __init__(self,
                 name,
                 inputs=[],
                 bias=True,
                 num_filters=None,
                 shared_biases=False,
                 **xargs):
2148
        super(ConvTransLayerBase, self).__init__(
2149 2150 2151 2152 2153 2154 2155 2156
            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))

2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167
        # 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"
2168 2169 2170 2171 2172 2173 2174 2175
        # 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)
2176
            parse_conv(
2177 2178
                self.inputs[input_index].conv,
                input_layer.name,
2179
                self.config.inputs[input_index].conv_conf,
2180
                num_filters,
2181
                trans=True)
2182 2183 2184
            conv_conf = self.config.inputs[input_index].conv_conf
            psize = self.calc_parameter_size(conv_conf)
            self.create_input_parameter(input_index, psize)
2185 2186
            self.set_cnn_layer(name, conv_conf.img_size_y, conv_conf.img_size,
                               self.config.num_filters)
2187 2188 2189 2190 2191 2192 2193

        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):
2194
        return conv_conf.channels * conv_conf.filter_channels \
2195 2196
                    * (conv_conf.filter_size * conv_conf.filter_size_y)

Q
qijun 已提交
2197

2198 2199 2200 2201
@config_layer('exconvt')
class ConvTransLayer(ConvTransLayerBase):
    layer_type = 'exconvt'

Q
qijun 已提交
2202

2203 2204 2205 2206 2207
@config_layer('cudnn_convt')
class ConvTransLayer(ConvTransLayerBase):
    layer_type = 'cudnn_convt'


C
chengduoZH 已提交
2208 2209
@config_layer('conv_3d')
class Conv3DLayerBase(LayerBase):
2210 2211 2212 2213 2214
    def __init__(self,
                 name,
                 inputs=[],
                 bias=True,
                 num_filters=None,
C
chengduoZH 已提交
2215
                 shared_biases=True,
2216
                 **xargs):
C
chengduoZH 已提交
2217
        super(Conv3DLayerBase, self).__init__(
2218 2219 2220 2221 2222 2223 2224 2225
            name, self.layer_type, 0, inputs=inputs, **xargs)

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

        # need to specify layer in config
        self.config.type = self.layer_type

C
chengduoZH 已提交
2226 2227 2228 2229
        trans = False
        if self.config.type == "deconv3d":
            trans = True

2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240
        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
            parse_conv3d(
                self.inputs[input_index].conv,
                input_layer.name,
                conv_conf,
                num_filters,
C
chengduoZH 已提交
2241
                trans=trans
2242 2243 2244
            )  # for z-axis pad:0, strid:1, filter_size:1, img_size:1
            psize = self.calc_parameter_size(conv_conf)
            self.create_input_parameter(input_index, psize)
C
chengduoZH 已提交
2245 2246 2247 2248 2249 2250 2251
            if trans:
                self.set_cnn_layer(name, conv_conf.img_size_z,
                                   conv_conf.img_size_y, conv_conf.img_size,
                                   self.config.num_filters)
            else:
                self.set_cnn_layer(name, conv_conf.output_z, conv_conf.output_y,
                                   conv_conf.output_x, self.config.num_filters)
2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271

        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 \
                  * conv_conf.filter_size_z)

    def set_cnn_layer(self,
                      input_layer_name,
                      depth,
                      height,
                      width,
                      channels,
                      is_print=True):
        size = depth * height * width * channels
        self.set_layer_size(size)
C
chengduoZH 已提交
2272 2273
        self.set_layer_height_width(height, width)
        self.set_layer_depth(depth)
2274 2275 2276 2277 2278
        if is_print:
            print("output for %s: c = %d, d = %d, h = %d, w = %d, size = %d" %
                  (input_layer_name, channels, depth, height, width, size))


C
chengduoZH 已提交
2279 2280 2281
@config_layer('conv3d')
class Conv3DLayer(Conv3DLayerBase):
    layer_type = 'conv3d'
2282

Q
qijun 已提交
2283

C
chengduoZH 已提交
2284 2285 2286
@config_layer('deconv3d')
class Conv3DLayer(Conv3DLayerBase):
    layer_type = 'deconv3d'
2287 2288


Z
zhangjinchao01 已提交
2289 2290
@config_layer('norm')
class NormLayer(LayerBase):
2291 2292
    def __init__(self, name, inputs, **xargs):
        super(NormLayer, self).__init__(name, 'norm', 0, inputs=inputs, **xargs)
Z
zhangjinchao01 已提交
2293 2294 2295
        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
2296 2297 2298 2299
            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)
2300 2301 2302
            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
2303

Z
zhangjinchao01 已提交
2304 2305 2306

@config_layer('pool')
class PoolLayer(LayerBase):
2307 2308
    layer_type = 'pool'

2309
    def __init__(self, name, inputs, ceil_mode=True, **xargs):
2310 2311 2312 2313 2314 2315
        use_mkldnn = int(g_command_config_args.get("use_mkldnn", 0))
        if self.layer_type == "mkldnn_pool":
            config_assert(use_mkldnn, "mkldnn_pool only support MKLDNN")
        self.layer_type = 'mkldnn_pool' if use_mkldnn else 'pool'
        super(PoolLayer, self).__init__(
            name, self.layer_type, 0, inputs=inputs, **xargs)
Z
zhangjinchao01 已提交
2316 2317 2318
        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
2319
            parse_pool(self.inputs[input_index].pool, input_layer.name,
2320
                       pool_conf, ceil_mode)
L
Luo Tao 已提交
2321 2322
            self.set_cnn_layer(name, pool_conf.output_y, pool_conf.output_x,
                               pool_conf.channels)
Q
qijun 已提交
2323

Z
zhangjinchao01 已提交
2324

2325 2326 2327 2328 2329
@config_layer('mkldnn_pool')
class MKLDNNPoolLayer(PoolLayer):
    layer_type = 'mkldnn_pool'


C
chengduoZH 已提交
2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358
@config_layer('pool3d')
class Pool3DLayer(LayerBase):
    def __init__(self, name, inputs, ceil_mode=True, **xargs):
        super(Pool3DLayer, self).__init__(
            name, 'pool3d', 0, inputs=inputs, **xargs)
        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
            parse_pool3d(self.inputs[input_index].pool, input_layer.name,
                         pool_conf, ceil_mode)
            self.set_cnn_layer(name, pool_conf.output_z, pool_conf.output_y,
                               pool_conf.output_x, pool_conf.channels)

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


Q
qijun 已提交
2359 2360
@config_layer('spp')
class SpatialPyramidPoolLayer(LayerBase):
2361
    def __init__(self, name, inputs, **xargs):
Q
qijun 已提交
2362
        super(SpatialPyramidPoolLayer, self).__init__(
2363
            name, 'spp', 0, inputs=inputs, **xargs)
Q
qijun 已提交
2364 2365 2366
        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
2367 2368 2369
            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
2370

Q
qijun 已提交
2371

D
dangqingqing 已提交
2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390
@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


2391 2392
@config_layer('crop')
class CropLayer(LayerBase):
2393
    def __init__(self, name, inputs, axis, offset, shape, **xargs):
2394
        super(CropLayer, self).__init__(name, 'crop', 0, inputs=inputs, **xargs)
2395 2396 2397
        self.config.axis = axis
        self.config.offset.extend(offset)
        self.config.shape.extend(shape)
2398 2399 2400 2401 2402 2403 2404 2405 2406 2407

        # get channel, width and height from input_0 layer
        input_layer = self.get_input_layer(0)
        image_conf = self.config.inputs[0].image_conf
        image_conf.img_size = input_layer.width
        image_conf.img_size_y = input_layer.height
        image_conf.channels = input_layer.size / (input_layer.width *
                                                  input_layer.height)


Z
zhangjinchao01 已提交
2408 2409 2410
@config_layer('batch_norm')
class BatchNormLayer(LayerBase):
    layer_type = 'batch_norm'
Q
qijun 已提交
2411 2412 2413 2414 2415

    def __init__(self,
                 name,
                 inputs,
                 bias=True,
C
chengduoZH 已提交
2416
                 img3D=False,
Q
qijun 已提交
2417 2418 2419
                 use_global_stats=True,
                 moving_average_fraction=0.9,
                 batch_norm_type=None,
C
chengduoZH 已提交
2420
                 mean_var_names=None,
Q
qijun 已提交
2421
                 **xargs):
Z
zhangjinchao01 已提交
2422 2423 2424 2425
        if inputs is None:
            inputs = []
        elif not isinstance(inputs, list):
            inputs = [inputs]
Q
qijun 已提交
2426 2427
        config_assert(
            len(inputs) == 1, "BatchNormLayer must have one and only one input")
Z
zhangjinchao01 已提交
2428 2429 2430 2431 2432 2433
        # 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)))
2434
        use_mkldnn = bool(int(g_command_config_args.get("use_mkldnn", 0)))
Z
zhangjinchao01 已提交
2435 2436
        is_shared = True if not use_gpu else False
        for i in xrange(2):
Q
qijun 已提交
2437 2438 2439 2440 2441 2442
            inputs.append(
                Input(
                    inputs[0].input_layer_name,
                    initial_std=0.0,
                    initial_mean=0.0,
                    is_static=True,
2443
                    is_shared=is_shared,
D
dangqingqing 已提交
2444
                    make_layer_name_in_submodel=False, ))
Z
zhangjinchao01 已提交
2445 2446 2447

        parallel_nn = bool(int(g_command_config_args.get("parallel_nn", 0)))
        cudnn_version = int(g_command_config_args.get("cudnn_version", 0))
2448 2449 2450 2451
        # Automatically select cudnn_batch_norm for GPU, batch_norm for CPU
        # and mkldnn_batch_norm for MKLDNN. Also based on cudnn version.
        if batch_norm_type == "mkldnn_batch_norm":
            config_assert(use_mkldnn, "mkldnn_batch_norm only support MKLDNN")
Z
zhangjinchao01 已提交
2452
        use_cudnn = use_gpu and batch_norm_type != "batch_norm" and \
2453
                not use_mkldnn and batch_norm_type != "mkldnn_batch_norm" and \
2454
                ((not parallel_nn) or self.config.device > -1)
2455 2456 2457 2458
        if use_cudnn:
            self.layer_type = "cudnn_batch_norm"
        else:
            self.layer_type = "mkldnn_batch_norm" if use_mkldnn else "batch_norm"
Q
qijun 已提交
2459
        super(BatchNormLayer, self).__init__(
X
xuwei06 已提交
2460
            name, self.layer_type, 0, inputs=inputs, **xargs)
Z
zhangjinchao01 已提交
2461 2462 2463 2464 2465 2466

        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
2467
        input_layer = self.get_input_layer(0)
Z
zhangjinchao01 已提交
2468
        image_conf = self.config.inputs[0].image_conf
C
chengduoZH 已提交
2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482
        if img3D:
            parse_image3d(self.inputs[0].image, input_layer.name, image_conf)
            # Only pass the width and height of input to batch_norm layer
            # when either of it is non-zero.
            if input_layer.width != 0 or input_layer.height != 0:
                self.set_cnn_layer(
                    input_layer_name=name,
                    depth=image_conf.img_size_z,
                    height=image_conf.img_size_y,
                    width=image_conf.img_size,
                    channels=image_conf.channels,
                    is_print=True)
            else:
                self.set_layer_size(input_layer.size)
2483
        else:
C
chengduoZH 已提交
2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495
            parse_image(self.inputs[0].image, input_layer.name, image_conf)
            # Only pass the width and height of input to batch_norm layer
            # when either of it is non-zero.
            if input_layer.width != 0 or input_layer.height != 0:
                self.set_cnn_layer(
                    input_layer_name=name,
                    height=image_conf.img_size_y,
                    width=image_conf.img_size,
                    channels=image_conf.channels,
                    is_print=True)
            else:
                self.set_layer_size(input_layer.size)
Z
zhangjinchao01 已提交
2496 2497 2498

        psize = self.calc_parameter_size(image_conf)
        dims = [1, psize]
C
chengduoZH 已提交
2499 2500 2501 2502
        if mean_var_names is not None:
            assert len(mean_var_names) == 2
            self.inputs[1].parameter_name = mean_var_names[0]
            self.inputs[2].parameter_name = mean_var_names[1]
C
chengduoZH 已提交
2503

Z
zhangjinchao01 已提交
2504 2505 2506 2507 2508 2509
        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)

C
chengduoZH 已提交
2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531
    def set_cnn_layer(self,
                      input_layer_name,
                      depth=None,
                      height=None,
                      width=None,
                      channels=None,
                      is_print=True):
        depthIsNone = False
        if depth is None:
            depth = 1
            depthIsNone = True
        size = depth * height * width * channels
        self.set_layer_size(size)
        self.set_layer_height_width(height, width)
        self.set_layer_depth(depth)
        if is_print and depthIsNone:
            print("output for %s: c = %d, h = %d, w = %d, size = %d" %
                  (input_layer_name, channels, height, width, size))
        elif is_print:
            print("output for %s: c = %d, d = %d, h = %d, w = %d, size = %d" %
                  (input_layer_name, channels, depth, height, width, size))

Z
zhangjinchao01 已提交
2532 2533 2534
    def calc_parameter_size(self, image_conf):
        return image_conf.channels

Q
qijun 已提交
2535

Z
zhangjinchao01 已提交
2536 2537
@config_layer('trans')
class TransLayer(LayerBase):
2538
    def __init__(self, name, inputs, **xargs):
Q
qijun 已提交
2539
        super(TransLayer, self).__init__(
2540
            name, 'trans', 0, inputs=inputs, **xargs)
Q
qijun 已提交
2541 2542 2543
        config_assert(
            len(self.inputs) == 1,
            'TransLayer must have one and only one input')
Z
zhangjinchao01 已提交
2544 2545
        self.set_layer_size(self.get_input_layer(0).size)

Q
qijun 已提交
2546

Z
zhangjinchao01 已提交
2547 2548
@config_layer('resize')
class ResizeLayer(LayerBase):
2549
    def __init__(self, name, size, inputs, **xargs):
Q
qijun 已提交
2550
        super(ResizeLayer, self).__init__(
2551
            name, 'resize', size=size, inputs=inputs, **xargs)
Q
qijun 已提交
2552 2553 2554 2555
        config_assert(
            len(self.inputs) == 1,
            'ResizeLayer must have one and only one input')

Z
zhangjinchao01 已提交
2556

2557 2558
@config_layer('rotate')
class RotateLayer(LayerBase):
H
Haonan 已提交
2559
    def __init__(self, name, inputs, height, width, device=None):
2560 2561 2562 2563 2564
        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
2565
        self.set_layer_height_width(height, width)
2566 2567 2568
        self.set_layer_size(self.get_input_layer(0).size)


Z
zhangjinchao01 已提交
2569 2570
@config_layer('blockexpand')
class BlockExpandLayer(LayerBase):
2571
    def __init__(self, name, inputs, **xargs):
Q
qijun 已提交
2572
        super(BlockExpandLayer, self).__init__(
2573
            name, 'blockexpand', 0, inputs=inputs, **xargs)
Z
zhangjinchao01 已提交
2574 2575
        for input_index in xrange(len(self.inputs)):
            input_layer = self.get_input_layer(input_index)
Q
qijun 已提交
2576 2577
            parse_block_expand(
                self.inputs[input_index].block_expand, input_layer.name,
Z
zhangjinchao01 已提交
2578
                self.config.inputs[input_index].block_expand_conf)
Q
qijun 已提交
2579 2580 2581 2582 2583 2584
            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
2585

2586 2587
@config_layer('maxout')
class MaxOutLayer(LayerBase):
Q
qijun 已提交
2588 2589 2590
    def __init__(self, name, inputs, **xargs):
        super(MaxOutLayer, self).__init__(
            name, 'maxout', 0, inputs=inputs, **xargs)
2591 2592
        input_layer = self.get_input_layer(0)
        maxout_conf = self.config.inputs[0].maxout_conf
L
Luo Tao 已提交
2593
        parse_maxout(self.inputs[0].maxout, input_layer.name, maxout_conf)
L
Luo Tao 已提交
2594
        out_channels = maxout_conf.image_conf.channels / maxout_conf.groups
2595 2596
        self.set_cnn_layer(name, maxout_conf.image_conf.img_size_y,
                           maxout_conf.image_conf.img_size, out_channels)
Q
qijun 已提交
2597

2598

D
dangqingqing 已提交
2599 2600 2601 2602
@config_layer('row_conv')
class RowConvLayer(LayerBase):
    def __init__(self, name, inputs, context_length, **xargs):
        super(RowConvLayer, self).__init__(
2603
            name, 'row_conv', 0, inputs=inputs, **xargs)
D
dangqingqing 已提交
2604 2605
        config_assert(
            len(self.inputs) == 1,
2606
            'row convolution layer must have one and only one input.')
D
dangqingqing 已提交
2607 2608 2609 2610 2611 2612 2613 2614 2615
        input_layer = self.get_input_layer(0)
        row_conv_conf = self.config.inputs[0].row_conv_conf
        row_conv_conf.context_length = context_length
        self.set_layer_size(input_layer.size)
        psize = context_length * input_layer.size
        dims = [context_length, input_layer.size]
        self.create_input_parameter(0, psize, dims)


G
guosheng 已提交
2616 2617
@config_layer('clip')
class ClipLayer(LayerBase):
2618 2619
    def __init__(self, name, inputs, min, max, **xargs):
        super(ClipLayer, self).__init__(name, 'clip', 0, inputs=inputs, **xargs)
G
guosheng 已提交
2620 2621
        config_assert(
            len(self.inputs) == 1,
2622 2623
            'ClipLayer must have one and only one input.')
        config_assert(min < max, 'min must be less than max.')
G
guosheng 已提交
2624 2625
        input_layer = self.get_input_layer(0)
        self.set_layer_size(input_layer.size)
2626 2627
        self.config.inputs[0].clip_conf.min = min
        self.config.inputs[0].clip_conf.max = max
G
guosheng 已提交
2628 2629


G
guosheng 已提交
2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643
@config_layer('scale_shift')
class ScaleShiftLayer(LayerBase):
    def __init__(self, name, inputs, bias=True, **xargs):
        super(ScaleShiftLayer, self).__init__(
            name, 'scale_shift', 0, inputs=inputs, **xargs)
        config_assert(
            len(self.inputs) == 1,
            'ScaleShiftLayer must have one and only one input.')
        input_layer = self.get_input_layer(0)
        self.set_layer_size(input_layer.size)
        self.create_input_parameter(0, 1, [1, 1])
        self.create_bias_parameter(bias, 1)


Z
zhangjinchao01 已提交
2644 2645 2646 2647
# key: cost type
# value: cost class
g_cost_map = {}

Q
qijun 已提交
2648

Z
zhangjinchao01 已提交
2649 2650 2651
# 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
2652 2653
        super(type(cls), cls).__init__(
            name, cost_type, 1, inputs, device=device, coeff=coeff)
Z
zhangjinchao01 已提交
2654

Q
qijun 已提交
2655
    cls = type(class_name, (LayerBase, ), dict(__init__=init))
Z
zhangjinchao01 已提交
2656 2657 2658
    global g_cost_map
    g_cost_map[cost_type] = cls

Q
qijun 已提交
2659

Z
zhangjinchao01 已提交
2660
define_cost('MultiClassCrossEntropy', 'multi-class-cross-entropy')
C
caoying03 已提交
2661
define_cost('CrossEntropyOverBeamCostLayer', 'cross_entropy_over_beam')
Z
zhangjinchao01 已提交
2662 2663 2664 2665 2666 2667
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')
2668
define_cost('HuberTwoClassification', 'huber_classification')
X
xuwei06 已提交
2669
define_cost('SumCost', 'sum_cost')
D
dangqingqing 已提交
2670
define_cost('SmoothL1Cost', 'smooth_l1')
Z
zhangjinchao01 已提交
2671

Q
qijun 已提交
2672

Z
zhangjinchao01 已提交
2673 2674
@config_layer('hsigmoid')
class HierarchicalSigmoidLayer(LayerBase):
Q
qijun 已提交
2675
    def __init__(self, name, num_classes, inputs, device=None, bias=True):
Z
zhangjinchao01 已提交
2676 2677
        super(HierarchicalSigmoidLayer, self).__init__(
            name, 'hsigmoid', 1, inputs=inputs, device=device)
Q
qijun 已提交
2678 2679 2680
        config_assert(
            len(self.inputs) >= 2,
            'HierarchicalSigmoidLayer must have at least 2 inputs')
Z
zhangjinchao01 已提交
2681 2682 2683 2684 2685 2686 2687 2688
        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
2689

Z
zhangjinchao01 已提交
2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713
'''
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
2714 2715


Z
zhangjinchao01 已提交
2716 2717
@config_layer('lambda_cost')
class LambdaCost(LayerBase):
Q
qijun 已提交
2718
    def __init__(self, name, inputs, NDCG_num=5, max_sort_size=-1, device=None):
Z
zhangjinchao01 已提交
2719 2720
        super(LambdaCost, self).__init__(
            name, 'lambda_cost', 1, inputs=inputs, device=device)
Q
qijun 已提交
2721
        config_assert(len(self.inputs) == 2, 'lambdaCost must have 2 inputs')
Z
zhangjinchao01 已提交
2722 2723
        self.config.NDCG_num = NDCG_num
        if max_sort_size != -1:
Q
qijun 已提交
2724 2725 2726
            config_assert(
                NDCG_num <= max_sort_size,
                'NDCG_num must be less than or equal to max_sort_size')
Z
zhangjinchao01 已提交
2727 2728
        self.config.max_sort_size = max_sort_size

Q
qijun 已提交
2729

L
Luo Tao 已提交
2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740
@config_layer('huber_regression')
class HuberRegressionLoss(LayerBase):
    def __init__(self, name, inputs, delta=1., coeff=1., device=None):
        super(HuberRegressionLoss, self).__init__(
            name, 'huber_regression', 1, inputs=inputs, device=device)
        config_assert(
            len(self.inputs) == 2, 'HuberRegression must have 2 inputs')
        self.config.delta = delta
        self.config.coeff = coeff


Z
zhangjinchao01 已提交
2741 2742
@config_layer('nce')
class NCELayer(LayerBase):
Q
qijun 已提交
2743 2744 2745 2746 2747 2748 2749 2750
    def __init__(self,
                 name,
                 num_classes,
                 inputs,
                 num_neg_samples=10,
                 neg_sampling_dist=None,
                 bias=True,
                 **xargs):
Z
zhangjinchao01 已提交
2751
        super(NCELayer, self).__init__(name, 'nce', 1, inputs=inputs, **xargs)
Q
qijun 已提交
2752 2753
        config_assert(
            len(self.inputs) >= 2, 'NCELayer must have at least 2 inputs')
Z
zhangjinchao01 已提交
2754 2755
        self.config.num_classes = num_classes
        if neg_sampling_dist is not None:
Q
qijun 已提交
2756 2757 2758 2759
            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
2760
            s = sum(neg_sampling_dist)
Q
qijun 已提交
2761 2762 2763
            config_assert(
                abs(s - 1) < 1e-5,
                'The sum of neg_sampling_dist (%s) is not 1' % s)
Z
zhangjinchao01 已提交
2764 2765 2766 2767 2768

            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
2769
        input_layer = self.get_input_layer(num_real_inputs)
Z
zhangjinchao01 已提交
2770 2771 2772 2773
        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
2774 2775
        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
2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788
            # 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):
T
tensor-tang 已提交
2789 2790
    layer_type = 'addto'

Q
qijun 已提交
2791
    def __init__(self, name, inputs, bias=True, **xargs):
T
tensor-tang 已提交
2792 2793 2794 2795
        use_mkldnn = bool(int(g_command_config_args.get("use_mkldnn", 0)))
        if self.layer_type == "mkldnn_addto":
            config_assert(use_mkldnn, "mkldnn_addto only support MKLDNN")
        self.layer_type = 'mkldnn_addto' if use_mkldnn else 'addto'
Z
zhangjinchao01 已提交
2796
        super(AddToLayer, self).__init__(
T
tensor-tang 已提交
2797
            name, self.layer_type, 0, inputs=inputs, **xargs)
Q
qijun 已提交
2798
        config_assert(len(inputs) > 0, 'inputs cannot be empty for AddToLayer')
2799 2800

        if len(self.inputs) > 1:
2801 2802 2803 2804 2805 2806 2807
            for input_index in xrange(len(self.inputs)):
                assert self.get_input_layer(0).height == self.get_input_layer(
                    input_index).height
                assert self.get_input_layer(0).width == self.get_input_layer(
                    input_index).width
                assert self.get_input_layer(0).depth == self.get_input_layer(
                    input_index).depth
2808 2809 2810 2811 2812

        self.set_layer_size(self.get_input_layer(0).size)
        self.set_layer_height_width(self.get_input_layer(0).height, \
                                        self.get_input_layer(0).width)
        self.set_layer_depth(self.get_input_layer(0).depth)
Z
zhangjinchao01 已提交
2813 2814
        self.create_bias_parameter(bias, self.config.size)

Q
qijun 已提交
2815

T
tensor-tang 已提交
2816 2817 2818 2819 2820
@config_layer('mkldnn_addto')
class MKLDNNAddtoLayer(AddToLayer):
    layer_type = 'mkldnn_addto'


Z
zhangjinchao01 已提交
2821 2822
@config_layer('agent')
class AgentLayer(LayerBase):
Q
qijun 已提交
2823 2824 2825 2826
    def __init__(self, name, size, device=None):
        super(AgentLayer, self).__init__(
            name, 'agent', size, inputs=[], device=device)

Z
zhangjinchao01 已提交
2827 2828 2829

@config_layer('gather_agent')
class GatherAgentLayer(LayerBase):
Q
qijun 已提交
2830
    def __init__(self, name, size, device=None):
Z
zhangjinchao01 已提交
2831 2832 2833
        super(GatherAgentLayer, self).__init__(
            name, 'gather_agent', size, inputs=[], device=device)

Q
qijun 已提交
2834

Z
zhangjinchao01 已提交
2835 2836
@config_layer('scatter_agent')
class ScatterAgentLayer(LayerBase):
2837
    def __init__(self, name, size, width=None, height=None, device=None):
Z
zhangjinchao01 已提交
2838 2839
        super(ScatterAgentLayer, self).__init__(
            name, 'scatter_agent', size, inputs=[], device=device)
2840 2841
        if height and width:
            self.set_layer_height_width(height, width)
Z
zhangjinchao01 已提交
2842

Q
qijun 已提交
2843

Z
zhangjinchao01 已提交
2844 2845
@config_layer('multiplex')
class MultiplexLayer(LayerBase):
Q
qijun 已提交
2846 2847 2848 2849 2850
    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
2851
        for i in range(1, len(inputs)):
Q
qijun 已提交
2852 2853 2854 2855 2856
            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
2857 2858

@config_func
2859 2860 2861 2862
def Link(name, has_subseq=False):
    """
    Still keeping has_subseq for backward compatibility
    """
Z
zhangjinchao01 已提交
2863 2864 2865 2866
    link_config = LinkConfig()
    link_config.link_name = name
    return link_config

Q
qijun 已提交
2867

Z
zhangjinchao01 已提交
2868 2869
# memory for recurrent layer group.
# *name* and *size* are actual layer's name and size.
2870 2871 2872 2873
# 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
2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884
# 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
2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896
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
2897
    agent_layer = AgentLayer(agent_name, size)
Z
zhangjinchao01 已提交
2898
    config_assert(g_current_submodel.is_recurrent_layer_group,
Q
qijun 已提交
2899
                  'Memory should be used in recurrent layer group only')
Z
zhangjinchao01 已提交
2900
    memory = g_current_submodel.memories.add()
2901 2902
    if name is not None:
        memory.layer_name = MakeLayerNameInSubmodel(name)
Z
zhangjinchao01 已提交
2903
    memory.link_name = MakeLayerNameInSubmodel(agent_name)
Q
qijun 已提交
2904
    options = sum((boot_layer is not None, bool(boot_bias),
Z
zhangjinchao01 已提交
2905
                   boot_with_const_id is not None))
Q
qijun 已提交
2906 2907 2908 2909
    config_assert(
        options <= 1,
        'take one option at most from boot_layer, boot_bias, or boot_with_const_id'
    )
Z
zhangjinchao01 已提交
2910 2911 2912
    if boot_layer is not None:
        boot_layer = MakeLayerNameInParentSubmodel(boot_layer)
        config_assert(boot_layer in g_layer_map,
Q
qijun 已提交
2913 2914
                      'boot_layer "%s" does not correspond to a layer name' %
                      boot_layer)
Z
zhangjinchao01 已提交
2915 2916 2917
        memory.boot_layer_name = boot_layer
    elif boot_bias:
        memory.boot_bias_parameter_name = agent_layer.create_bias_parameter(
Q
qijun 已提交
2918
            boot_bias, size, for_self=False)
Z
zhangjinchao01 已提交
2919 2920 2921 2922 2923
        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
2924

2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935
@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
2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946
# 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
2947 2948 2949 2950
        eos_layer_name="eos_check",
        num_results_per_sample=1,
        beam_size=1,
        log_prob=None, ):
Z
zhangjinchao01 已提交
2951 2952 2953 2954 2955 2956 2957 2958 2959
    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
2960

Z
zhangjinchao01 已提交
2961 2962
@config_layer('expand')
class ExpandLayer(LayerBase):
2963
    def __init__(self, name, inputs, trans_type='non-seq', bias=False, **xargs):
Q
qijun 已提交
2964
        super(ExpandLayer, self).__init__(
2965
            name, 'expand', 0, inputs=inputs, **xargs)
Q
qijun 已提交
2966 2967 2968 2969 2970 2971 2972 2973
        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
2974 2975 2976

@config_layer('featmap_expand')
class FeatMapExpandLayer(LayerBase):
X
xuwei06 已提交
2977 2978 2979 2980 2981
    def __init__(self,
                 name,
                 inputs,
                 num_filters=None,
                 as_row_vector=True,
X
xuwei06 已提交
2982 2983
                 bias=False,
                 **xargs):
Q
qijun 已提交
2984
        super(FeatMapExpandLayer, self).__init__(
X
xuwei06 已提交
2985
            name, 'featmap_expand', 0, inputs=inputs, **xargs)
Q
qijun 已提交
2986 2987 2988
        config_assert(
            len(self.inputs) == 1, 'ExpandLayer takes 1 and only 1 inputs')
        if num_filters is not None:
Z
zhangjinchao01 已提交
2989
            self.config.num_filters = num_filters
Q
qijun 已提交
2990
        else:
Z
zhangjinchao01 已提交
2991
            logger.fatal("FeatMapExpandLayer must specify num_filters.")
X
xuwei06 已提交
2992 2993
        if not as_row_vector:
            self.config.user_arg = "as_col_vec"
Q
qijun 已提交
2994
        self.set_layer_size(self.get_input_layer(0).size * num_filters)
Z
zhangjinchao01 已提交
2995 2996 2997 2998


@config_layer('max')
class MaxLayer(LayerBase):
Q
qijun 已提交
2999 3000 3001 3002 3003
    def __init__(self,
                 name,
                 inputs,
                 trans_type='non-seq',
                 bias=False,
3004
                 output_max_index=None,
3005
                 stride=-1,
3006
                 **xargs):
3007
        super(MaxLayer, self).__init__(name, 'max', 0, inputs=inputs, **xargs)
Z
zhangjinchao01 已提交
3008
        config_assert(len(self.inputs) == 1, 'MaxLayer must have 1 input')
3009 3010
        if trans_type == 'seq':
            config_assert(stride == -1, 'subseq does not support stride window')
Q
qijun 已提交
3011
        self.config.trans_type = trans_type
3012
        self.config.seq_pool_stride = stride
Z
zhangjinchao01 已提交
3013 3014 3015 3016
        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)
3017 3018
        if output_max_index is not None:
            self.config.output_max_index = output_max_index
Z
zhangjinchao01 已提交
3019 3020 3021 3022


@config_layer('maxid')
class MaxIdLayer(LayerBase):
Q
qijun 已提交
3023
    def __init__(self, name, inputs, beam_size=None, device=None):
Z
zhangjinchao01 已提交
3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040
        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
3041
    def __init__(self, name, inputs, eos_id, device=None):
Z
zhangjinchao01 已提交
3042 3043 3044
        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
3045
        self.set_layer_size(2)  # boolean output
Z
zhangjinchao01 已提交
3046 3047
        self.config.eos_id = eos_id

Q
qijun 已提交
3048

Z
zhangjinchao01 已提交
3049 3050
@config_layer('seqlastins')
class SequenceLastInstanceLayer(LayerBase):
Q
qijun 已提交
3051 3052 3053 3054
    def __init__(self,
                 name,
                 inputs,
                 trans_type='non-seq',
3055
                 bias=False,
3056
                 stride=-1,
3057
                 **xargs):
Q
qijun 已提交
3058
        super(SequenceLastInstanceLayer, self).__init__(
X
xuwei06 已提交
3059
            name, 'seqlastins', 0, inputs=inputs, **xargs)
Q
qijun 已提交
3060 3061
        config_assert(
            len(inputs) == 1, 'SequenceLastInstanceLayer must have 1 input')
3062
        if trans_type == 'seq':
L
Luo Tao 已提交
3063
            config_assert(stride == -1, 'subseq does not support stride window')
Q
qijun 已提交
3064
        self.config.trans_type = trans_type
3065 3066
        self.config.seq_pool_stride = stride
        self.set_layer_size(self.get_input_layer(0).size)
Z
zhangjinchao01 已提交
3067 3068
        self.create_bias_parameter(bias, self.config.size)

Q
qijun 已提交
3069

Z
zhangjinchao01 已提交
3070 3071
@config_layer('seqfirstins')
class SequenceFirstInstanceLayer(SequenceLastInstanceLayer):
3072 3073 3074 3075 3076
    def __init__(self,
                 name,
                 inputs,
                 trans_type='non-seq',
                 bias=False,
3077
                 stride=-1,
3078
                 **xargs):
Q
qijun 已提交
3079
        super(SequenceFirstInstanceLayer, self).__init__(
3080 3081 3082 3083 3084 3085
            name,
            inputs=inputs,
            trans_type=trans_type,
            bias=bias,
            stride=stride,
            **xargs)
Z
zhangjinchao01 已提交
3086 3087
        self.config.select_first = True

Q
qijun 已提交
3088

Z
zhangjinchao01 已提交
3089 3090
@config_layer('seqconcat')
class SequenceConcatLayer(LayerBase):
X
xuwei06 已提交
3091
    def __init__(self, name, inputs, bias=False, **xargs):
Q
qijun 已提交
3092
        super(SequenceConcatLayer, self).__init__(
X
xuwei06 已提交
3093
            name, 'seqconcat', 0, inputs=inputs, **xargs)
Q
qijun 已提交
3094 3095
        config_assert(
            len(inputs) == 2, 'SequenceConcatLayer must have 2 inputs')
Z
zhangjinchao01 已提交
3096 3097 3098 3099 3100
        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
3101

Z
zhangjinchao01 已提交
3102 3103
@config_layer('seqreshape')
class SequenceReshapeLayer(LayerBase):
X
xuwei06 已提交
3104
    def __init__(self, name, size, inputs, bias=False, **xargs):
Q
qijun 已提交
3105
        super(SequenceReshapeLayer, self).__init__(
X
xuwei06 已提交
3106
            name, 'seqreshape', size, inputs=inputs, **xargs)
Q
qijun 已提交
3107 3108
        config_assert(
            len(inputs) == 1, 'SequenceReshapeLayer must have 1 inputs')
Z
zhangjinchao01 已提交
3109 3110 3111
        self.set_layer_size(size)
        self.create_bias_parameter(bias, size)

Q
qijun 已提交
3112

Z
zhangjinchao01 已提交
3113 3114
@config_layer('subseq')
class SubSequenceLayer(LayerBase):
X
xuwei06 已提交
3115
    def __init__(self, name, inputs, bias=False, **xargs):
Q
qijun 已提交
3116
        super(SubSequenceLayer, self).__init__(
X
xuwei06 已提交
3117
            name, 'subseq', 0, inputs=inputs, **xargs)
Z
zhangjinchao01 已提交
3118 3119 3120 3121 3122 3123
        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
3124

3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153
@config_layer('seq_slice')
class SeqSliceLayer(LayerBase):
    def __init__(self, name, inputs, starts, ends, bias=False, **xargs):
        if isinstance(inputs, list):
            assert len(inputs) == 1, ('the first input of sequence slice layer '
                                      'is a single sequence input.')
        else:
            inputs = [inputs]

        if starts is not None:
            if isinstance(starts, list):
                assert len(starts) == 1, (
                    'the start indices for sequence slice layer cannot '
                    'be a list having more than one element.')
                starts = starts[0]
            inputs.append(starts)

        if ends is not None:
            if isinstance(ends, list):
                assert len(ends) == 1, (
                    'the end indices for sequence slice layer cannot '
                    'be a list having more than one element.')
                ends = ends[0]
            inputs.append(ends)
        assert len(inputs) >= 2, (
            'the sequence slice layer has at least two inputs.')

        super(SeqSliceLayer, self).__init__(
            name, 'seq_slice', 0, inputs=inputs, **xargs)
3154

3155 3156 3157 3158 3159 3160 3161 3162 3163 3164
        input_layer0 = self.get_input_layer(0)
        size = input_layer0.size
        self.set_layer_size(size)

        if len(inputs) == 3:
            assert (
                self.get_input_layer(1).size == self.get_input_layer(2).size), (
                    'If start and end indices are both given to'
                    'sequence slice layer, they should have the same width.')
        elif len(inputs) == 2:
C
caoying03 已提交
3165
            self.config.select_first = (starts is not None)
3166 3167


C
caoying03 已提交
3168 3169
@config_layer('sub_nested_seq')
class SubNestedSequenceLayer(LayerBase):
3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181
    def __init__(self, name, inputs, selected_indices, bias=False, **xargs):
        if isinstance(inputs, list):
            assert len(inputs) == 1, ('the first input of sub_nested_seq '
                                      'layer is a single nested sequence.')
            inputs = inputs[0]
        if isinstance(selected_indices, list):
            assert len(selected_indices) == 1, (
                'the second input of '
                'sub_nested_seq layer is a single layer which is a '
                'set of selected indices.')
            selected_indices = selected_indices[0]

C
caoying03 已提交
3182
        super(SubNestedSequenceLayer, self).__init__(
3183 3184 3185 3186 3187
            name,
            'sub_nested_seq',
            0,
            inputs=[inputs, selected_indices],
            **xargs)
C
caoying03 已提交
3188 3189 3190 3191 3192
        input_layer0 = self.get_input_layer(0)
        size = input_layer0.size
        self.set_layer_size(size)


Z
zhangjinchao01 已提交
3193 3194
@config_layer('out_prod')
class OuterProdLayer(LayerBase):
Q
qijun 已提交
3195 3196 3197
    def __init__(self, name, inputs, device=None):
        super(OuterProdLayer, self).__init__(
            name, 'out_prod', 0, inputs=inputs, device=device)
Z
zhangjinchao01 已提交
3198 3199 3200 3201 3202
        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
3203

Z
zhangjinchao01 已提交
3204 3205
@config_layer('power')
class PowerLayer(LayerBase):
Q
qijun 已提交
3206 3207 3208
    def __init__(self, name, inputs, device=None):
        super(PowerLayer, self).__init__(
            name, 'power', 0, inputs=inputs, device=device)
Z
zhangjinchao01 已提交
3209 3210 3211 3212
        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
3213 3214 3215
        config_assert(1 == input_layer0.size,
                      'The left input is the exponent and should be of size 1')

Z
zhangjinchao01 已提交
3216 3217 3218

@config_layer('slope_intercept')
class SlopeInterceptLayer(LayerBase):
Q
qijun 已提交
3219 3220 3221
    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
3222 3223 3224 3225 3226 3227
        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
3228

Z
zhangjinchao01 已提交
3229 3230
@config_layer('scaling')
class ScalingLayer(LayerBase):
Q
qijun 已提交
3231 3232 3233
    def __init__(self, name, inputs, device=None):
        super(ScalingLayer, self).__init__(
            name, 'scaling', 0, inputs=inputs, device=device)
Z
zhangjinchao01 已提交
3234 3235 3236 3237
        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
3238 3239 3240
        config_assert(1 == input_layer0.size,
                      'The left input should be of size 1')

Z
zhangjinchao01 已提交
3241 3242 3243

@config_layer('conv_shift')
class ConvShiftLayer(LayerBase):
Q
qijun 已提交
3244 3245 3246
    def __init__(self, name, inputs, device=None):
        super(ConvShiftLayer, self).__init__(
            name, 'conv_shift', 0, inputs=inputs, device=device)
Z
zhangjinchao01 已提交
3247 3248 3249 3250
        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
3251

Z
zhangjinchao01 已提交
3252 3253
@config_layer('convex_comb')
class ConvexCombinationLayer(LayerBase):
Q
qijun 已提交
3254
    def __init__(self, name, size, inputs, device=None):
Z
zhangjinchao01 已提交
3255
        super(ConvexCombinationLayer, self).__init__(
Q
qijun 已提交
3256 3257 3258
            name, 'convex_comb', size, inputs=inputs, device=device)
        config_assert(
            len(self.inputs) == 2, 'ConvexCombinationLayer must have 2 inputs')
3259 3260 3261
        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
3262 3263
        self.set_layer_size(size)

Q
qijun 已提交
3264

Z
zhangjinchao01 已提交
3265 3266
@config_layer('interpolation')
class InterpolationLayer(LayerBase):
Q
qijun 已提交
3267
    def __init__(self, name, inputs, device=None):
Z
zhangjinchao01 已提交
3268 3269
        super(InterpolationLayer, self).__init__(
            name, 'interpolation', 0, inputs=inputs, device=device)
Q
qijun 已提交
3270 3271
        config_assert(
            len(self.inputs) == 3, 'InterpolationLayer must have 3 inputs')
Z
zhangjinchao01 已提交
3272 3273 3274 3275 3276 3277 3278 3279
        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
3280

L
liaogang 已提交
3281 3282
@config_layer('bilinear_interp')
class BilinearInterpLayer(LayerBase):
Q
qijun 已提交
3283
    def __init__(self, name, inputs, **xargs):
L
liaogang 已提交
3284
        super(BilinearInterpLayer, self).__init__(
L
liaogang 已提交
3285
            name, 'bilinear_interp', 0, inputs=inputs, **xargs)
L
liaogang 已提交
3286
        input_layer = self.get_input_layer(0)
L
Luo Tao 已提交
3287 3288 3289 3290
        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
3291

L
liaogang 已提交
3292

Z
zhangjinchao01 已提交
3293 3294
@config_layer('sum_to_one_norm')
class SumToOneNormLayer(LayerBase):
Q
qijun 已提交
3295
    def __init__(self, name, inputs, device=None):
Z
zhangjinchao01 已提交
3296
        super(SumToOneNormLayer, self).__init__(
Q
qijun 已提交
3297 3298 3299
            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
3300 3301 3302
        input_layer0 = self.get_input_layer(0)
        self.set_layer_size(input_layer0.size)

Q
qijun 已提交
3303

G
guosheng 已提交
3304 3305
@config_layer('row_l2_norm')
class RowL2NormLayer(LayerBase):
3306
    def __init__(self, name, inputs, **xargs):
G
guosheng 已提交
3307
        super(RowL2NormLayer, self).__init__(
3308
            name, 'row_l2_norm', 0, inputs=inputs, **xargs)
G
guosheng 已提交
3309
        config_assert(len(self.inputs) == 1, 'RowL2NormLayer must have 1 input')
3310 3311
        input_layer = self.get_input_layer(0)
        self.set_layer_size(input_layer.size)
G
guosheng 已提交
3312 3313


Z
zhangjinchao01 已提交
3314 3315
@config_layer('cos_vm')
class CosSimVecMatLayer(LayerBase):
Q
qijun 已提交
3316
    def __init__(self, name, size, inputs, cos_scale=1.0, device=None):
Z
zhangjinchao01 已提交
3317
        super(CosSimVecMatLayer, self).__init__(
Q
qijun 已提交
3318
            name, 'cos_vm', size, inputs=inputs, device=device)
Z
zhangjinchao01 已提交
3319
        self.config.cos_scale = cos_scale
Q
qijun 已提交
3320 3321
        config_assert(
            len(self.inputs) == 2, 'CosSimVecMatLayer must have 2 inputs')
3322 3323 3324
        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
3325

Q
qijun 已提交
3326

Z
zhangjinchao01 已提交
3327 3328
@config_layer('sampling_id')
class SamplingIdLayer(LayerBase):
Q
qijun 已提交
3329
    def __init__(self, name, inputs, device=None):
Z
zhangjinchao01 已提交
3330 3331
        super(SamplingIdLayer, self).__init__(
            name, 'sampling_id', 0, inputs=inputs, device=device)
Q
qijun 已提交
3332 3333
        config_assert(
            len(self.inputs) == 1, 'SamplingIdLayer must have 1 input')
Z
zhangjinchao01 已提交
3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345
        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
3346 3347 3348 3349 3350
    def __init__(self,
                 name,
                 inputs,
                 average_strategy='average',
                 trans_type='non-seq',
3351
                 bias=False,
3352
                 stride=-1,
3353
                 **xargs):
Q
qijun 已提交
3354
        super(AverageLayer, self).__init__(
X
xuwei06 已提交
3355
            name, 'average', 0, inputs=inputs, **xargs)
Z
zhangjinchao01 已提交
3356
        self.config.average_strategy = average_strategy
3357 3358
        if trans_type == 'seq':
            config_assert(stride == -1, 'subseq does not support stride window')
Q
qijun 已提交
3359
        self.config.trans_type = trans_type
3360
        self.config.seq_pool_stride = stride
Z
zhangjinchao01 已提交
3361 3362 3363 3364 3365 3366
        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
3367

Z
zhangjinchao01 已提交
3368 3369
@config_layer('cos')
class CosSimLayer(LayerBase):
3370
    def __init__(self, name, inputs, cos_scale=1, device=None):
Z
zhangjinchao01 已提交
3371 3372 3373 3374 3375 3376
        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')
3377
        self.config.cos_scale = cos_scale
Z
zhangjinchao01 已提交
3378 3379 3380 3381


@config_layer('tensor')
class TensorLayer(LayerBase):
3382
    def __init__(self, name, size, inputs, bias=True, **xargs):
Q
qijun 已提交
3383
        super(TensorLayer, self).__init__(
3384
            name, 'tensor', size, inputs=inputs, **xargs)
Z
zhangjinchao01 已提交
3385 3386
        config_assert(len(self.inputs) == 2, 'TensorLayer must have 2 inputs')
        config_assert(size > 0, 'size must be positive')
Q
qijun 已提交
3387 3388
        config_assert(inputs[1].parameter_name == None,
                      'second parameter should be None.')
Z
zhangjinchao01 已提交
3389 3390 3391 3392 3393 3394 3395 3396 3397 3398
        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):
C
caoying03 已提交
3399
    def __init__(self, name, inputs, size=0, bias=True, **xargs):
Z
zhangjinchao01 已提交
3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416
        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)
3417
            if self.config.size == 0:
Z
zhangjinchao01 已提交
3418 3419 3420
                size = operator.calc_output_size(operator_conf.input_sizes)
                if size != 0:
                    self.set_layer_size(size)
3421
            else:
3422 3423
                sz = operator.calc_output_size(operator_conf.input_sizes)
                if sz != 0:
Q
qijun 已提交
3424 3425 3426 3427
                    config_assert(
                        sz == self.config.size,
                        "different inputs have different size: %s vs. %s" %
                        (sz, self.config.size))
Z
zhangjinchao01 已提交
3428 3429 3430 3431
        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
3432 3433 3434
                config_assert(
                    isinstance(input, Projection),
                    "input should be projection or operation")
3435
            if self.config.size == 0 and isinstance(input, Projection):
Z
zhangjinchao01 已提交
3436 3437 3438
                size = input.calc_output_size(input_layer)
                if size != 0:
                    self.set_layer_size(size)
3439
            elif isinstance(input, Projection):
Q
qijun 已提交
3440 3441 3442 3443 3444 3445
                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
3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456
        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
3457 3458
                input_config.proj_conf.name = gen_parameter_name(name,
                                                                 input_index)
Z
zhangjinchao01 已提交
3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469
                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)

3470 3471 3472 3473 3474 3475
        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
3476

3477 3478 3479
        if bias:
            self.config.bias_size = psize
            self.create_bias_parameter(bias, psize)
Z
zhangjinchao01 已提交
3480

Q
qijun 已提交
3481

Z
zhangjinchao01 已提交
3482 3483
# like MixedLayer, but no bias parameter
@config_func
Q
qijun 已提交
3484
def ExpressionLayer(name, inputs, **xargs):
Z
zhangjinchao01 已提交
3485 3486
    MixedLayer(name, inputs, bias=False, **xargs)

Q
qijun 已提交
3487

Z
zhangjinchao01 已提交
3488 3489
@config_layer('concat')
class ConcatenateLayer(LayerBase):
Q
qijun 已提交
3490
    def __init__(self, name, inputs, bias=False, **xargs):
Z
zhangjinchao01 已提交
3491
        config_assert(inputs, 'inputs cannot be empty')
3492
        config_assert(not bias, 'ConcatenateLayer cannot support bias.')
Z
zhangjinchao01 已提交
3493 3494 3495 3496
        super(ConcatenateLayer, self).__init__(
            name, 'concat', 0, inputs=inputs, **xargs)
        size = 0
        for input_index in xrange(len(self.inputs)):
3497 3498 3499 3500 3501 3502
            assert self.get_input_layer(0).height == self.get_input_layer(
                input_index).height
            assert self.get_input_layer(0).width == self.get_input_layer(
                input_index).width
            assert self.get_input_layer(0).depth == self.get_input_layer(
                input_index).depth
Z
zhangjinchao01 已提交
3503 3504
            input_layer = self.get_input_layer(input_index)
            input = self.inputs[input_index]
Q
qijun 已提交
3505
            if self.config.size == 0:
Z
zhangjinchao01 已提交
3506 3507
                size += input_layer.size

3508 3509 3510
        self.set_layer_height_width(self.get_input_layer(0).height, \
                                    self.get_input_layer(0).width)
        self.set_layer_depth(self.get_input_layer(0).depth)
Z
zhangjinchao01 已提交
3511 3512
        self.set_layer_size(size)

Q
qijun 已提交
3513

Z
zhangjinchao01 已提交
3514 3515 3516
# like concat layer, but each input layer was processed by a Projection.
@config_layer('concat2')
class ConcatenateLayer2(LayerBase):
Q
qijun 已提交
3517
    def __init__(self, name, inputs, bias=False, **xargs):
Z
zhangjinchao01 已提交
3518 3519 3520
        config_assert(inputs, 'inputs cannot be empty')
        super(ConcatenateLayer2, self).__init__(
            name, 'concat2', 0, inputs=inputs, **xargs)
3521 3522

        if isinstance(self.inputs[0], ConvProjection):
Q
qijun 已提交
3523 3524 3525 3526 3527 3528
            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.")
3529

Z
zhangjinchao01 已提交
3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549
        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
3550
                                              input.proj_conf.output_size)
Z
zhangjinchao01 已提交
3551
            dims = input.calc_parameter_dims(input.proj_conf.input_size,
Q
qijun 已提交
3552
                                             input.proj_conf.output_size)
Z
zhangjinchao01 已提交
3553 3554
            self.create_input_parameter(input_index, psize, dims)

3555 3556 3557 3558 3559 3560 3561
        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()

3562 3563 3564
        if bias:
            self.config.bias_size = psize
            self.create_bias_parameter(bias, psize)
3565

Q
qijun 已提交
3566

Z
zhangjinchao01 已提交
3567 3568
@config_layer('recurrent')
class RecurrentLayer(LayerBase):
Q
qijun 已提交
3569
    def __init__(self, name, inputs, reversed=False, bias=True, **xargs):
Y
Yu Yang 已提交
3570 3571
        super(RecurrentLayer, self).__init__(name, 'recurrent', 0, inputs,
                                             **xargs)
Z
zhangjinchao01 已提交
3572 3573 3574 3575 3576 3577 3578 3579 3580
        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
3581

Z
zhangjinchao01 已提交
3582 3583
@config_layer('lstmemory')
class LstmLayer(LayerBase):
Q
qijun 已提交
3584 3585 3586 3587 3588 3589 3590 3591
    def __init__(self,
                 name,
                 inputs,
                 reversed=False,
                 active_gate_type="sigmoid",
                 active_state_type="sigmoid",
                 bias=True,
                 **xargs):
Z
zhangjinchao01 已提交
3592 3593 3594 3595 3596 3597 3598 3599
        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
3600
        self.config.active_gate_type = active_gate_type
Z
zhangjinchao01 已提交
3601 3602 3603 3604 3605
        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
3606

Z
zhangjinchao01 已提交
3607 3608
@config_layer('lstm_step')
class LstmStepLayer(LayerBase):
Q
qijun 已提交
3609 3610 3611 3612 3613 3614 3615 3616 3617 3618
    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
3619 3620 3621
        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
3622 3623 3624 3625 3626
        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
3627 3628 3629
        self.config.active_state_type = active_state_type
        self.create_bias_parameter(bias, size * 3)

Q
qijun 已提交
3630

Z
zhangjinchao01 已提交
3631 3632 3633
# get the specific output from the input layer.
@config_layer('get_output')
class GetOutputLayer(LayerBase):
Q
qijun 已提交
3634 3635 3636 3637
    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
3638 3639 3640 3641
        inputs = self.inputs[0]
        config_assert(inputs.input_layer_argument,
                      'input_layer_argument cannot be empty')

Q
qijun 已提交
3642

Z
zhangjinchao01 已提交
3643 3644
@config_layer('mdlstmemory')
class MDLstmLayer(LayerBase):
Q
qijun 已提交
3645 3646 3647 3648 3649 3650 3651 3652
    def __init__(self,
                 name,
                 inputs,
                 directions=True,
                 active_gate_type="sigmoid",
                 active_state_type="sigmoid",
                 bias=True,
                 **xargs):
Y
Yu Yang 已提交
3653 3654
        super(MDLstmLayer, self).__init__(name, 'mdlstmemory', 0, inputs,
                                          **xargs)
Z
zhangjinchao01 已提交
3655 3656 3657 3658
        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
3659 3660
        config_assert(input_layer.size % (3 + dim_num) == 0,
                      "size % (dim_num) should be 0!")
Q
qijun 已提交
3661
        size = input_layer.size / (3 + dim_num)
Z
zhangjinchao01 已提交
3662
        self.set_layer_size(size)
Q
qijun 已提交
3663
        self.config.active_gate_type = active_gate_type
Z
zhangjinchao01 已提交
3664 3665 3666
        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
3667 3668
        self.create_input_parameter(0, size * size * (3 + dim_num),
                                    [size, size, 3 + dim_num])
Z
zhangjinchao01 已提交
3669
        #bias includes 3 kinds of peephole, 3+dim_num+2+dim_num
Q
qijun 已提交
3670 3671
        self.create_bias_parameter(bias, size * (5 + 2 * dim_num))

Z
zhangjinchao01 已提交
3672 3673 3674

@config_layer('gated_recurrent')
class GatedRecurrentLayer(LayerBase):
Q
qijun 已提交
3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685
    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
3686 3687 3688 3689 3690 3691
        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
3692
        self.config.active_gate_type = active_gate_type
Z
zhangjinchao01 已提交
3693 3694 3695
        self.create_input_parameter(0, size * size * 3, [size, size * 3])
        self.create_bias_parameter(bias, size * 3)

Q
qijun 已提交
3696

Z
zhangjinchao01 已提交
3697 3698
@config_layer('gru_step')
class GruStepLayer(LayerBase):
Q
qijun 已提交
3699 3700 3701 3702 3703 3704 3705
    def __init__(self,
                 name,
                 size,
                 inputs,
                 active_gate_type="sigmoid",
                 bias=True,
                 **xargs):
Y
Yu Yang 已提交
3706 3707
        super(GruStepLayer, self).__init__(name, 'gru_step', size, inputs,
                                           **xargs)
Z
zhangjinchao01 已提交
3708 3709 3710
        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
3711 3712 3713 3714 3715
        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
3716
        self.create_input_parameter(0, size * size * 3, [size, size * 3])
Z
zhangjinchao01 已提交
3717 3718
        self.create_bias_parameter(bias, size * 3)

Q
qijun 已提交
3719

Z
zhangjinchao01 已提交
3720 3721 3722 3723 3724 3725 3726
'''
 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
3727 3728


Z
zhangjinchao01 已提交
3729 3730
@config_layer('crf')
class CRFLayer(LayerBase):
Q
qijun 已提交
3731
    def __init__(self, name, size, inputs, coeff=1.0, device=None):
Z
zhangjinchao01 已提交
3732
        super(CRFLayer, self).__init__(name, 'crf', size, inputs, device=device)
Q
qijun 已提交
3733 3734
        config_assert(2 <= len(self.inputs) <= 3,
                      'CRFLayer must have 2 or 3 inputs')
3735
        self.create_input_parameter(0, size * (size + 2), [size + 2, size])
Z
zhangjinchao01 已提交
3736 3737
        self.config.coeff = coeff

Q
qijun 已提交
3738

Z
zhangjinchao01 已提交
3739 3740 3741 3742 3743 3744 3745 3746
'''
 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
3747 3748


Z
zhangjinchao01 已提交
3749 3750
@config_layer('crf_decoding')
class CRFDecodingLayer(LayerBase):
Q
qijun 已提交
3751
    def __init__(self, name, size, inputs, device=None):
Z
zhangjinchao01 已提交
3752 3753 3754 3755 3756
        super(CRFDecodingLayer, self).__init__(
            name, 'crf_decoding', size, inputs, device=device)
        config_assert(
            len(self.inputs) <= 2,
            'CRFDecodingLayer cannot have more than 2 inputs')
3757
        self.create_input_parameter(0, size * (size + 2), [size + 2, size])
Z
zhangjinchao01 已提交
3758

Q
qijun 已提交
3759

Z
zhangjinchao01 已提交
3760 3761
@config_layer('ctc')
class CTCLayer(LayerBase):
Q
qijun 已提交
3762
    def __init__(self, name, size, inputs, norm_by_times=False, device=None):
Z
zhangjinchao01 已提交
3763 3764 3765 3766
        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
3767

3768 3769 3770 3771 3772 3773 3774 3775 3776 3777
@config_layer('kmax_seq_score')
class KmaxSeqScoreLayer(LayerBase):
    def __init__(self, name, inputs, beam_size, **xargs):
        super(KmaxSeqScoreLayer, self).__init__(
            name, 'kmax_seq_score', 0, inputs=inputs, **xargs)
        config_assert(
            len(self.inputs) == 1, 'KmaxSeqScoreLayer has only one input.')
        self.config.beam_size = beam_size


3778 3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798
@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
3799 3800
@config_layer('recurrent_layer_group')
class RecurrentLayerGroup(LayerBase):
Q
qijun 已提交
3801
    def __init__(self, name, device=None):
Z
zhangjinchao01 已提交
3802 3803 3804 3805
        super(RecurrentLayerGroup, self).__init__(
            name, 'recurrent_layer_group', 0, inputs=[], device=device)


3806 3807 3808 3809 3810
@config_layer('switch_order')
class SwitchOrderLayer(LayerBase):
    def __init__(self, name, inputs, reshape, **xargs):
        super(SwitchOrderLayer, self).__init__(
            name, 'switch_order', 0, inputs=inputs, **xargs)
W
wanghaoshuang 已提交
3811 3812
        self.config.reshape_conf.height_axis.extend(reshape['height'])
        self.config.reshape_conf.width_axis.extend(reshape['width'])
3813 3814


Y
yangyaming 已提交
3815 3816
@config_layer('scale_sub_region')
class ScaleSubRegionLayer(LayerBase):
Y
yangyaming 已提交
3817
    def __init__(self, name, inputs, value, **xargs):
Y
yangyaming 已提交
3818 3819 3820 3821
        super(ScaleSubRegionLayer, self).__init__(
            name, 'scale_sub_region', 0, inputs=inputs, **xargs)
        scale_sub_region_conf = self.config.inputs[0].scale_sub_region_conf
        scale_sub_region_conf.value = value
Y
yangyaming 已提交
3822 3823 3824

        # get channel, width and height from input_0 layer
        input_layer = self.get_input_layer(0)
Y
yangyaming 已提交
3825
        image_conf = scale_sub_region_conf.image_conf
Y
yangyaming 已提交
3826 3827 3828 3829
        image_conf.img_size = input_layer.width
        image_conf.img_size_y = input_layer.height
        image_conf.channels = input_layer.size / (input_layer.width *
                                                  input_layer.height)
Y
yangyaming 已提交
3830 3831
        self.set_cnn_layer(name, image_conf.img_size_y, image_conf.img_size,
                           image_conf.channels)
Y
yangyaming 已提交
3832 3833


Z
zhangjinchao01 已提交
3834 3835
# Deprecated, use a new layer specific class instead
@config_func
Q
qijun 已提交
3836
def Layer(name, type, **xargs):
Z
zhangjinchao01 已提交
3837 3838 3839 3840
    layers = {}
    layers.update(g_cost_map)
    layers.update(g_layer_type_map)
    layer_func = layers.get(type)
Q
qijun 已提交
3841
    config_assert(layer_func, "layer type '%s' not supported." % type)
X
xuwei06 已提交
3842
    return layer_func(name, **xargs)
Z
zhangjinchao01 已提交
3843

Q
qijun 已提交
3844

Z
zhangjinchao01 已提交
3845
@config_func
Q
qijun 已提交
3846
def ParameterHook(type, **kwargs):
3847
    if type == 'pruning':
Z
zhangjinchao01 已提交
3848 3849
        hook = ParameterUpdaterHookConfig()
        hook.type = type
X
xzl 已提交
3850
        sparsity_ratio = kwargs.get('sparsity_ratio', None)
X
xzl 已提交
3851 3852
        if sparsity_ratio is not None:
            hook.sparsity_ratio = sparsity_ratio
Z
zhangjinchao01 已提交
3853
        return hook
3854 3855 3856 3857
    elif type == 'dpruning':
        hook = ParameterUpdaterHookConfig()
        hook.type = type
        return hook
Z
zhangjinchao01 已提交
3858 3859 3860 3861 3862
    else:
        return None


@config_func
Q
qijun 已提交
3863 3864 3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883
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,
X
xuwei06 已提交
3884 3885
              update_hooks=None,
              initializer=None):
Z
zhangjinchao01 已提交
3886 3887 3888 3889 3890 3891 3892

    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
3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903
    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
3904 3905
    config_assert(not momentum or not decay_rate_l1,
                  "momentum and decay_rate_l1 cannot both be non-zero")
3906 3907 3908 3909 3910

    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
3911 3912 3913 3914
    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)
3915

Q
qijun 已提交
3916 3917
    num_batches_regularization = default(num_batches_regularization,
                                         g_default_num_batches_regularization)
3918 3919 3920
    if num_batches_regularization is not None:
        para.num_batches_regularization = int(num_batches_regularization)

Z
zhangjinchao01 已提交
3921 3922 3923 3924 3925 3926
    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
3927 3928
    gradient_clipping_threshold = default(gradient_clipping_threshold,
                                          g_default_gradient_clipping_threshold)
3929 3930
    if gradient_clipping_threshold is not None:
        para.gradient_clipping_threshold = gradient_clipping_threshold
Q
qijun 已提交
3931 3932
    para.initial_strategy = default(initial_strategy,
                                    g_default_initial_strategy)
Z
zhangjinchao01 已提交
3933 3934 3935 3936 3937 3938
    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
3939 3940 3941
            print(
                "Use initial_smart, but dims not set. Initial_smart may not be used in this layer"
            )
Z
zhangjinchao01 已提交
3942 3943 3944 3945
            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)
3946 3947 3948 3949 3950 3951 3952

    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
3953 3954 3955 3956
    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")
3957 3958
    if is_shared is not None:
        para.is_shared = is_shared
Z
zhangjinchao01 已提交
3959 3960 3961 3962 3963

    update_hooks = default(update_hooks, g_default_update_hooks)

    if update_hooks is not None:
        if hasattr(update_hooks, '__call__'):
X
xzl 已提交
3964
            update_hooks = update_hooks()
Z
zhangjinchao01 已提交
3965 3966 3967 3968 3969

        if isinstance(update_hooks, list):
            for hook in update_hooks:
                para.update_hooks.extend([hook])
        else:
X
xzl 已提交
3970
            para.update_hooks.extend([update_hooks])
Z
zhangjinchao01 已提交
3971 3972

    g_parameter_map[name] = para
X
xuwei06 已提交
3973 3974 3975 3976 3977
    if initializer is not None:
        config_assert(
            callable(initializer),
            "parameter initializer should be a callable object")
        g_parameter_initializer_map[name] = initializer
Z
zhangjinchao01 已提交
3978 3979 3980 3981 3982 3983 3984


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

Q
qijun 已提交
3985

Z
zhangjinchao01 已提交
3986 3987 3988 3989 3990
@config_func
def default_initial_mean(val):
    global g_default_initial_mean
    g_default_initial_mean = val

Q
qijun 已提交
3991

Z
zhangjinchao01 已提交
3992 3993 3994 3995 3996
@config_func
def default_initial_strategy(val):
    global g_default_initial_strategy
    g_default_initial_strategy = val

Q
qijun 已提交
3997

Z
zhangjinchao01 已提交
3998 3999 4000 4001 4002
@config_func
def default_initial_smart(val):
    global g_default_initial_smart
    g_default_initial_smart = val

Q
qijun 已提交
4003

Z
zhangjinchao01 已提交
4004 4005 4006 4007 4008
@config_func
def default_momentum(val):
    global g_default_momentum
    g_default_momentum = val

Q
qijun 已提交
4009

Z
zhangjinchao01 已提交
4010 4011 4012 4013 4014
@config_func
def default_decay_rate(val):
    global g_default_decay_rate
    g_default_decay_rate = val

Q
qijun 已提交
4015

Z
zhangjinchao01 已提交
4016 4017 4018 4019 4020
@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
4021

Z
zhangjinchao01 已提交
4022 4023 4024 4025 4026
@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
4027

Z
zhangjinchao01 已提交
4028 4029 4030 4031 4032
@config_func
def default_device(val):
    global g_default_device
    g_default_device = val

Q
qijun 已提交
4033

Z
zhangjinchao01 已提交
4034 4035 4036 4037 4038
@config_func
def default_update_hooks(val):
    global g_default_update_hooks
    g_default_update_hooks = val

Q
qijun 已提交
4039

Z
zhangjinchao01 已提交
4040 4041 4042 4043 4044
@config_func
def default_compact_func(val):
    global g_default_compact_func
    g_default_compact_func = val

Q
qijun 已提交
4045

Z
zhangjinchao01 已提交
4046 4047 4048 4049 4050
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
4051 4052 4053
        execfile(config_file,
                 make_config_environment(config_file, config_args), local_args)

Z
zhangjinchao01 已提交
4054 4055
    return Import

Q
qijun 已提交
4056

X
xuwei06 已提交
4057
DEFAULT_SETTING = dict(
Z
zhangjinchao01 已提交
4058 4059 4060 4061 4062
    batch_size=None,
    mini_batch_size=None,
    algorithm='async_sgd',
    async_lagged_grad_discard_ratio=1.5,
    learning_method='momentum',
4063
    gradient_clipping_threshold=None,
Z
zhangjinchao01 已提交
4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077 4078 4079 4080 4081 4082 4083 4084 4085
    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
4086 4087 4088
    adam_beta1=0.9,
    adam_beta2=0.999,
    adam_epsilon=1e-8, )
Z
zhangjinchao01 已提交
4089

X
xuwei06 已提交
4090
settings = copy.deepcopy(DEFAULT_SETTING)
X
xuwei06 已提交
4091

Q
qijun 已提交
4092
settings_deprecated = dict(usage_ratio=1., )
Z
zhangjinchao01 已提交
4093 4094 4095 4096

trainer_settings = dict(
    save_dir="./output/model",
    init_model_path=None,
Q
qijun 已提交
4097 4098
    start_pass=0, )

Z
zhangjinchao01 已提交
4099 4100 4101 4102 4103

@config_func
def Settings(**args):
    for k, v in args.iteritems():
        if k == "usage_ratio":
Q
qijun 已提交
4104 4105
            logger.warning(
                "Deprecated: define usage_ratio in DataConfig instead")
Z
zhangjinchao01 已提交
4106 4107 4108 4109 4110 4111 4112 4113 4114 4115 4116
            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
4117

Z
zhangjinchao01 已提交
4118 4119 4120 4121
@config_func
def cluster_config(**args):
    pass

Q
qijun 已提交
4122

Z
zhangjinchao01 已提交
4123 4124 4125 4126 4127 4128 4129 4130 4131
@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
4132

Z
zhangjinchao01 已提交
4133 4134 4135 4136
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
4137

Z
zhangjinchao01 已提交
4138 4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149 4150 4151 4152
        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
4153
        get_config_arg=make_get_config_arg(config_args), )
Z
zhangjinchao01 已提交
4154 4155 4156 4157 4158

    funcs.update(g_extended_config_funcs)

    return funcs

Q
qijun 已提交
4159

Z
zhangjinchao01 已提交
4160 4161 4162 4163 4164 4165 4166 4167 4168 4169 4170 4171 4172 4173 4174 4175
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
4176

Z
zhangjinchao01 已提交
4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188
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
4189

Z
zhangjinchao01 已提交
4190 4191 4192 4193
def my_fatal(s):
    logger.critical(s)
    raise Exception()

Y
Yu Yang 已提交
4194

4195
_parse_config_hooks = set()
Y
Yu Yang 已提交
4196 4197


4198 4199 4200 4201 4202 4203 4204
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
4205

Y
Yu Yang 已提交
4206

4207
def update_g_config():
Z
zhangjinchao01 已提交
4208
    '''
4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229 4230 4231
    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


4232
def begin_parse():
Z
zhangjinchao01 已提交
4233
    init_config_environment()
4234 4235
    for hook in _parse_config_hooks:
        hook()
Z
zhangjinchao01 已提交
4236 4237 4238 4239 4240

    logger.findCaller = find_caller
    logger.fatal = my_fatal

    g_config.model_config.type = "nn"
X
xuwei06 已提交
4241 4242 4243 4244 4245 4246 4247 4248 4249

    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


def parse_config(trainer_config, config_arg_str):
4250 4251 4252 4253
    '''
    @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
    '''
X
xuwei06 已提交
4254

4255
    begin_parse()
X
xuwei06 已提交
4256 4257
    config_args = {}

Z
zhangjinchao01 已提交
4258 4259 4260 4261 4262 4263 4264 4265 4266 4267 4268 4269
    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)

4270 4271
    if hasattr(trainer_config, '__call__'):
        trainer_config.func_globals.update(
L
Luo Tao 已提交
4272
            make_config_environment("", config_args))
4273
        trainer_config()
H
hanchao 已提交
4274
    else:
4275 4276
        execfile(trainer_config,
                 make_config_environment(trainer_config, config_args))
Z
zhangjinchao01 已提交
4277

4278
    return update_g_config()
Z
zhangjinchao01 已提交
4279 4280


4281
def parse_config_and_serialize(trainer_config, config_arg_str):
Z
zhangjinchao01 已提交
4282
    try:
4283
        config = parse_config(trainer_config, config_arg_str)
Z
zhangjinchao01 已提交
4284 4285 4286 4287 4288 4289
        #logger.info(config)
        return config.SerializeToString()
    except:
        traceback.print_exc()
        raise

Q
qijun 已提交
4290

Z
zhangjinchao01 已提交
4291 4292 4293 4294 4295 4296 4297 4298
if __name__ == '__main__':
    try:
        config = parse_config(sys.argv[1], '')
        config.SerializeToString()
        __real_print__(str(config))
    except:
        traceback.print_exc()
        raise