PyDataProvider2.py 14.2 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
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import cPickle
import logging
17 18 19 20
import collections
import functools
import itertools

Q
qijun 已提交
21 22
logging.basicConfig(format="[%(levelname)s %(asctime)s %(filename)s:%(lineno)s]"
                    " %(message)s")
Z
zhangjinchao01 已提交
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


class SequenceType(object):
    NO_SEQUENCE = 0
    SEQUENCE = 1
    SUB_SEQUENCE = 2


# TODO(yuyang18): Add string data type here.
class DataType(object):
    Dense = 0
    SparseNonValue = 1
    SparseValue = 2
    Index = 3


class CacheType(object):
    NO_CACHE = 0  # No cache at all

    # First pass, read data from python.  And store them in memory. Read from
    # memory during rest passes.
    CACHE_PASS_IN_MEM = 1


class InputType(object):
    __slots__ = ['dim', 'seq_type', 'type']

    def __init__(self, dim, seq_type, tp):
        self.dim = dim
        self.seq_type = seq_type
        self.type = tp


def dense_slot(dim, seq_type=SequenceType.NO_SEQUENCE):
    return InputType(dim, seq_type, DataType.Dense)


def sparse_non_value_slot(dim, seq_type=SequenceType.NO_SEQUENCE):
    return InputType(dim, seq_type, DataType.SparseNonValue)


def sparse_value_slot(dim, seq_type=SequenceType.NO_SEQUENCE):
    return InputType(dim, seq_type, DataType.SparseValue)


def index_slot(dim, seq_type=SequenceType.NO_SEQUENCE):
    return InputType(dim, seq_type, DataType.Index)


dense_vector = dense_slot
sparse_binary_vector = sparse_non_value_slot
sparse_vector = sparse_value_slot
integer_value = index_slot

77

Z
zhangjinchao01 已提交
78 79 80
def dense_vector_sequence(dim):
    return dense_vector(dim, seq_type=SequenceType.SEQUENCE)

81

Z
zhangjinchao01 已提交
82 83 84
def dense_vector_sub_sequence(dim):
    return dense_vector(dim, seq_type=SequenceType.SUB_SEQUENCE)

85

Z
zhangjinchao01 已提交
86 87 88
def sparse_binary_vector_sequence(dim):
    return sparse_binary_vector(dim, seq_type=SequenceType.SEQUENCE)

89

Z
zhangjinchao01 已提交
90 91 92
def sparse_binary_vector_sub_sequence(dim):
    return sparse_binary_vector(dim, seq_type=SequenceType.SUB_SEQUENCE)

93

Z
zhangjinchao01 已提交
94 95 96
def sparse_vector_sequence(dim):
    return sparse_vector(dim, seq_type=SequenceType.SEQUENCE)

97

Z
zhangjinchao01 已提交
98 99 100
def sparse_vector_sub_sequence(dim):
    return sparse_vector(dim, seq_type=SequenceType.SUB_SEQUENCE)

101

Z
zhangjinchao01 已提交
102 103 104
def integer_value_sequence(dim):
    return integer_value(dim, seq_type=SequenceType.SEQUENCE)

105

Z
zhangjinchao01 已提交
106 107 108
def integer_value_sub_sequence(dim):
    return integer_value(dim, seq_type=SequenceType.SUB_SEQUENCE)

109

Z
zhangjinchao01 已提交
110 111 112 113 114 115 116 117 118 119
def integer_sequence(dim):
    return index_slot(dim, seq_type=SequenceType.SEQUENCE)


class SingleSlotWrapper(object):
    def __init__(self, generator):
        self.generator = generator

    def __call__(self, obj, filename):
        for item in self.generator(obj, filename):
120 121 122 123
            if isinstance(item, dict):
                yield item
            else:
                yield [item]
Z
zhangjinchao01 已提交
124 125


126 127 128 129 130 131 132 133
class InputOrderWrapper(object):
    def __init__(self, generator, input_order):
        self.generator = generator
        self.input_order = input_order

    def __call__(self, obj, filename):
        for item in self.generator(obj, filename):
            if isinstance(item, dict):
Q
qijun 已提交
134 135 136 137
                yield [
                    item.get(input_name, None)
                    for input_name in self.input_order
                ]
138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165
            else:
                yield item


class CheckWrapper(object):
    def __init__(self, generator, input_types, check_fail_continue, logger):
        self.generator = generator
        self.input_types = input_types
        self.check_fail_continue = check_fail_continue
        self.logger = logger

    def __call__(self, obj, filename):
        for items in self.generator(obj, filename):
            try:
                assert len(items) == len(self.input_types)
                assert len(filter(lambda x: x is None, items)) == 0
                for item, input_type in itertools.izip(items, self.input_types):
                    callback = functools.partial(CheckWrapper.loop_callback,
                                                 input_type)

                    for _ in xrange(input_type.seq_type):
                        callback = functools.partial(CheckWrapper.loop_check,
                                                     callback)
                    callback(item)

                yield items
            except AssertionError as e:
                self.logger.warning(
Q
qijun 已提交
166 167
                    "Item (%s) is not fit the input type with error %s" %
                    (repr(item), repr(e)))
168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204

                if self.check_fail_continue:
                    continue
                else:
                    raise

    @staticmethod
    def loop_callback(input_type, each):
        assert isinstance(input_type, InputType)
        if input_type.type == DataType.Dense:
            assert isinstance(each, collections.Sequence)
            for d in each:
                assert isinstance(d, float)
            assert len(each, input_type.dim)
        elif input_type.type == DataType.Index:
            assert isinstance(each, int)
            assert each < input_type.dim
        elif input_type.type == DataType.SparseNonValue \
                or input_type.type == DataType.SparseValue:
            assert isinstance(each, collections.Sequence)
            sparse_id = set()
            for k in each:
                if input_type.type == DataType.SparseValue:
                    k, v = k
                    assert isinstance(v, float)
                assert isinstance(k, int)
                assert k < input_type.dim
                sparse_id.add(k)
            assert len(sparse_id) == len(each)
        else:
            raise RuntimeError("Not support input type")

    @staticmethod
    def loop_check(callback, item):
        for each in item:
            callback(each)

Y
Yu Yang 已提交
205

206 207 208 209 210 211
class CheckInputTypeWrapper(object):
    def __init__(self, generator, input_types, logger):
        self.generator = generator
        self.input_types = input_types
        self.logger = logger

Y
Yu Yang 已提交
212 213 214 215 216 217 218 219 220
    def __call__(self, obj, filename):
        for items in self.generator(obj, filename):
            try:
                # dict type is required for input_types when item is dict type 
                assert (isinstance(items, dict) and \
                        not isinstance(self.input_types, dict))==False
                yield items
            except AssertionError as e:
                self.logger.error(
221 222
                    "%s type is required for input type but got %s" %
                    (repr(type(items)), repr(type(self.input_types))))
Y
Yu Yang 已提交
223 224
                raise

225

Q
qijun 已提交
226 227 228
def provider(input_types=None,
             should_shuffle=None,
             pool_size=-1,
229
             min_pool_size=-1,
Z
zhangjinchao01 已提交
230 231 232
             can_over_batch_size=True,
             calc_batch_size=None,
             cache=CacheType.NO_CACHE,
Q
qijun 已提交
233 234 235 236
             check=False,
             check_fail_continue=False,
             init_hook=None,
             **kwargs):
Z
zhangjinchao01 已提交
237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254
    """
    Provider decorator. Use it to make a function into PyDataProvider2 object.
    In this function, user only need to get each sample for some train/test
    file.

    The basic usage is:

    ..  code-block:: python

        @provider(some data provider config here...)
        def process(settings, file_name):
            while not at end of file_name:
                sample = readOneSampleFromFile(file_name)
                yield sample.

    The configuration of data provider should be setup by\:

    :param input_types: Specify the input types, can also be set in init_hook.
255 256 257 258 259 260 261 262 263
                        It could be a list of InputType object. For example,
                        input_types=[dense_vector(9), integer_value(2)]. Or user
                        can set a dict of InputType object, which key is
                        data_layer's name. For example, input_types=\
                        {'img': img_features, 'label': label}. when using dict of
                        InputType, user could yield a dict of feature values, which
                        key is also data_layer's name.

    :type input_types: list|tuple|dict
264 265 266

    :param should_shuffle: True if data should shuffle. Pass None means shuffle
                           when is training and not to shuffle when is testing.
Z
zhangjinchao01 已提交
267
    :type should_shuffle: bool
268

Z
zhangjinchao01 已提交
269 270
    :param pool_size: Max number of sample in data pool.
    :type pool_size: int
271 272 273 274 275 276

    :param min_pool_size: Set minimal sample in data pool. The PaddlePaddle will
                          random pick sample in pool. So the min_pool_size
                          effect the randomize of data.
    :type min_pool_size: int

Z
zhangjinchao01 已提交
277 278 279 280 281 282
    :param can_over_batch_size: True if paddle can return a mini-batch larger
                                than batch size in settings. It is useful when
                                custom calculate one sample's batch_size.

                                It is very danger to set it to false and use
                                calc_batch_size together. Default is false.
283 284
    :type can_over_batch_size: bool

Z
zhangjinchao01 已提交
285 286 287
    :param calc_batch_size: a method to calculate each sample's batch size.
                            Default each sample's batch size is 1. But to you
                            can customize each sample's batch size.
288 289
    :type calc_batch_size: callable

Z
zhangjinchao01 已提交
290
    :param cache: Cache strategy of Data Provider. Default is CacheType.NO_CACHE
291
    :type cache: int
Z
zhangjinchao01 已提交
292 293 294 295 296

    :param init_hook: Initialize hook. Useful when data provider need load some
                      external data like dictionary. The parameter is
                      (settings, file_list, \*\*kwargs).

297 298 299 300 301
                      - settings. It is the global settings object. User can set
                        settings.input_types here.
                      - file_list. All file names for passed to data provider.
                      - is_train. Is this data provider used for training or not.
                      - kwargs. Other keyword arguments passed from
Z
zhangjinchao01 已提交
302
                        trainer_config's args parameter.
303 304 305 306 307 308 309 310 311 312 313
    :type init_hook: callable

    :param check: Check the yield data format is as same as input_types. Enable
                  this will make data provide process slow but it is very useful
                  for debug. Default is disabled.
    :type check: bool

    :param check_fail_continue: Continue train or not when check failed. Just
                                drop the wrong format data when it is True. Has
                                no effect when check set to False.
    :type check_fail_continue: bool
Z
zhangjinchao01 已提交
314 315 316 317 318 319 320 321 322 323 324 325 326 327
    """

    def __wrapper__(generator):
        class DataProvider(object):
            def __init__(self, file_list, **kwargs):
                self.logger = logging.getLogger("")
                self.logger.setLevel(logging.INFO)
                self.input_types = None
                if 'slots' in kwargs:
                    self.logger.warning('setting slots value is deprecated, '
                                        'please use input_types instead.')
                    self.slots = kwargs['slots']
                self.slots = input_types
                self.should_shuffle = should_shuffle
328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345

                true_table = [1, 't', 'true', 'on']
                false_table = [0, 'f', 'false', 'off']
                if not isinstance(self.should_shuffle, bool) and \
                                self.should_shuffle is not None:

                    if isinstance(self.should_shuffle, basestring):
                        self.should_shuffle = self.should_shuffle.lower()

                    if self.should_shuffle in true_table:
                        self.should_shuffle = True
                    elif self.should_shuffle in false_table:
                        self.should_shuffle = False
                    else:
                        self.logger.warning(
                            "Could not recognize should_shuffle (%s), "
                            "just use default value of should_shuffle."
                            " Please set should_shuffle to bool value or "
Q
qijun 已提交
346 347 348
                            "something in %s" %
                            (repr(self.should_shuffle),
                             repr(true_table + false_table)))
349 350
                        self.should_shuffle = None

Z
zhangjinchao01 已提交
351 352 353 354 355 356
                self.pool_size = pool_size
                self.can_over_batch_size = can_over_batch_size
                self.calc_batch_size = calc_batch_size
                self.file_list = file_list
                self.generator = generator
                self.cache = cache
357 358 359
                self.min_pool_size = min_pool_size
                self.input_order = kwargs['input_order']
                self.check = check
Z
zhangjinchao01 已提交
360 361 362 363 364 365 366
                if init_hook is not None:
                    init_hook(self, file_list=file_list, **kwargs)
                if self.input_types is not None:
                    self.slots = self.input_types
                assert self.slots is not None
                assert self.generator is not None

367 368 369 370 371
                use_dynamic_order = False
                if isinstance(self.slots, dict):  # reorder input_types
                    self.slots = [self.slots[ipt] for ipt in self.input_order]
                    use_dynamic_order = True

Z
zhangjinchao01 已提交
372 373 374
                if len(self.slots) == 1:
                    self.generator = SingleSlotWrapper(self.generator)

375 376 377
                if use_dynamic_order:
                    self.generator = InputOrderWrapper(self.generator,
                                                       self.input_order)
378
                else:
Y
Yu Yang 已提交
379 380
                    self.generator = CheckInputTypeWrapper(
                        self.generator, self.slots, self.logger)
381
                if self.check:
Q
qijun 已提交
382
                    self.generator = CheckWrapper(self.generator, self.slots,
383 384 385
                                                  check_fail_continue,
                                                  self.logger)

Z
zhangjinchao01 已提交
386 387 388 389 390 391 392 393 394 395 396 397
        return DataProvider

    return __wrapper__


def deserialize_args(args):
    """
    Internal use only.
    :param args:
    :return:
    """
    return cPickle.loads(args)