reader.py 43.2 KB
Newer Older
S
sneaxiy 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

15
from . import core, dygraph
16
import sys
S
sneaxiy 已提交
17
import six
18 19
import warnings
import numpy as np
S
sneaxiy 已提交
20
import threading
21
import paddle
Z
Zeng Jinle 已提交
22
from .framework import Program, Variable, program_guard, default_main_program, default_startup_program, in_dygraph_mode, cpu_places
S
sneaxiy 已提交
23
from .executor import global_scope
24
from .data_feeder import DataFeeder, BatchedTensorProvider
S
sneaxiy 已提交
25
from .layers.io import monkey_patch_reader_methods, _copy_reader_var_, double_buffer
S
sneaxiy 已提交
26
from .unique_name import UniqueNameGenerator
27
import logging
Z
Zeng Jinle 已提交
28
from .dataset import DatasetBase, InMemoryDataset
S
sneaxiy 已提交
29

Z
Zeng Jinle 已提交
30 31 32
__all__ = ['PyReader', 'DataLoader']

data_loader_unique_name_generator = UniqueNameGenerator()
S
sneaxiy 已提交
33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49


def _convert_places(places):
    if not isinstance(places, (list, tuple)):
        places = [places]

    ret = []
    for p in places:
        if not isinstance(p, core.Place):
            tmp = core.Place()
            tmp.set_place(p)
            p = tmp

        ret.append(p)
    return ret


Z
Zeng Jinle 已提交
50 51 52
class DataLoaderBase(object):
    def __init__(self):
        self._places = None
S
sneaxiy 已提交
53

Z
Zeng Jinle 已提交
54 55
    def __call__(self):
        return self
S
sneaxiy 已提交
56

Z
Zeng Jinle 已提交
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
    def next(self):
        '''
        Get the next item in the DataLoader object. This method    
        should not be called by users directly. It is used for
        implementing iterator protocol of Python 2.x inside
        PaddlePaddle framework.
        '''
        return self.__next__()

    def __iter__(self):
        raise NotImplementedError()

    def __next__(self):
        raise NotImplementedError()


class DataLoader(object):
    @staticmethod
    def from_generator(feed_list=None,
                       capacity=None,
                       use_double_buffer=True,
                       iterable=True,
                       return_list=False):
        """
        Create a DataLoader object for loading data from Python generator. 
        Data would be prefetched using Python thread and be pushed
        into a queue asynchronously.

        The created DataLoader object provides 3 methods to set the data source
        :code:`set_sample_generator` , :code:`set_sample_list_generator` and 
        :code:`set_batch_generator` . Please see the following example codes
        to know their usages.

        If iterable = True, the created DataLoader object is a Python generator
        object, which is iterable using for-range loop.

        If iterable = False, the created DataLoader object provides 
        :code:`start()` and :code:`reset()` method to control the data reading
        process. This mode is designed to be compatible with the 
        :code:`fluid.layers.py_reader` interface. Users can migrate the codes   
        from :code:`fluid.layers.py_reader` to :code:`fluid.io.DataLoader` 
        easily when using iterable=False. 

        Args:  
            feed_list (list(Variable)|tuple(Variable)): feed variable list.
102
                The variables should be created by :code:`fluid.data()`.
Z
Zeng Jinle 已提交
103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126
            capacity (int): capacity of the queue maintained in DataLoader.
                The unit is batch number. Set larger capacity if your reader 
                is fast. 
            use_double_buffer (bool): whether to use double_buffer_reader. 
                If use_double_buffer=True, the DataLoader would prefetch next 
                batch data asynchronously, so it would speed up data feeding 
                and occupies a little more CPU or GPU memory, i.e., the memory
                of one batch input data. 
            iterable (bool): whether the created DataLoader is iterable. 
            return_list (bool): whether the return value on each device is 
                presented as a list. It is only valid when iterable=True. 
                If return_list=False, the return value on each device would 
                be a dict of str -> LoDTensor, where the key of the dict is 
                the name of each feeded variables. If return_list=True, the 
                return value on each device would be a list(LoDTensor). It is
                recommended to use return_list=False in static graph mode and
                use return_list=True in dygraph mode.   

        Returns:
            loader (DataLoader): the created DataLoader object.

        Examples:
            
            .. code-block:: python
S
sneaxiy 已提交
127

Z
Zeng Jinle 已提交
128 129
                import paddle.fluid as fluid
                import numpy as np
130

Z
Zeng Jinle 已提交
131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185
                BATCH_NUM = 10 
                BATCH_SIZE = 16
                EPOCH_NUM = 4

                CLASS_NUM = 10

                ITERABLE = True # whether the created DataLoader object is iterable
                USE_GPU = False # whether to use GPU

                DATA_FORMAT = 'batch_generator' # data format of data source user provides 

                def simple_net(image, label):
                    fc_tmp = fluid.layers.fc(image, size=CLASS_NUM)
                    cross_entropy = fluid.layers.softmax_with_cross_entropy(image, label)
                    loss = fluid.layers.reduce_mean(cross_entropy)
                    sgd = fluid.optimizer.SGD(learning_rate=1e-3)
                    sgd.minimize(loss)
                    return loss

                def get_random_images_and_labels(image_shape, label_shape):
                    image = np.random.random(size=image_shape).astype('float32')
                    label = np.random.random(size=label_shape).astype('int64')
                    return image, label

                # If the data generator yields one sample each time,
                # use DataLoader.set_sample_generator to set the data source.
                def sample_generator_creator(): 
                    def __reader__():
                        for _ in range(BATCH_NUM * BATCH_SIZE):
                            image, label = get_random_images_and_labels([784], [1])
                            yield image, label

                    return __reader__

                # If the data generator yield list of samples each time,
                # use DataLoader.set_sample_list_generator to set the data source.
                def sample_list_generator_creator():
                    def __reader__():
                        for _ in range(BATCH_NUM): 
                            sample_list = []
                            for _ in range(BATCH_SIZE):
                                image, label = get_random_images_and_labels([784], [1])
                                sample_list.append([image, label])

                            yield sample_list

                    return __reader__ 

                # If the data generator yields a batch each time, 
                # use DataLoader.set_batch_generator to set the data source.
                def batch_generator_creator():
                    def __reader__():
                        for _ in range(BATCH_NUM):
                            batch_image, batch_label = get_random_images_and_labels([BATCH_SIZE, 784], [BATCH_SIZE, 1]) 
                            yield batch_image, batch_label
H
Huihuang Zheng 已提交
186

Z
Zeng Jinle 已提交
187
                    return __reader__
188

Z
Zeng Jinle 已提交
189 190 191 192 193
                # If DataLoader is iterable, use for loop to train the network 
                def train_iterable(exe, prog, loss, loader):
                    for _ in range(EPOCH_NUM):
                        for data in loader():
                            exe.run(prog, feed=data, fetch_list=[loss])
194

Z
Zeng Jinle 已提交
195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213
                # If DataLoader is not iterable, use start() and reset() method to control the process 
                def train_non_iterable(exe, prog, loss, loader):
                    for _ in range(EPOCH_NUM):
                        loader.start() # call DataLoader.start() before each epoch starts
                        try:
                            while True:
                                exe.run(prog, fetch_list=[loss])
                        except fluid.core.EOFException:
                            loader.reset() # call DataLoader.reset() after catching EOFException 

                def set_data_source(loader, places):
                    if DATA_FORMAT == 'sample_generator':
                        loader.set_sample_generator(sample_generator_creator(), batch_size=BATCH_SIZE, drop_last=True, places=places)
                    elif DATA_FORMAT == 'sample_list_generator':
                        loader.set_sample_list_generator(sample_list_generator_creator(), places=places)
                    elif DATA_FORMAT == 'batch_generator':
                        loader.set_batch_generator(batch_generator_creator(), places=places)
                    else:
                        raise ValueError('Unsupported data format')
214

215 216
                image = fluid.data(name='image', shape=[None, 784], dtype='float32')
                label = fluid.data(name='label', shape=[None, 1], dtype='int64')
217

Z
Zeng Jinle 已提交
218 219
                # Define DataLoader 
                loader = fluid.io.DataLoader.from_generator(feed_list=[image, label], capacity=16, iterable=ITERABLE)
220

Z
Zeng Jinle 已提交
221 222
                # Define network
                loss = simple_net(image, label)
S
sneaxiy 已提交
223

Z
Zeng Jinle 已提交
224 225 226 227 228 229 230 231 232
                # Set data source of DataLoader
                #
                # If DataLoader is iterable, places must be given and the number of places must be the same with device number.  
                #  - If you are using GPU, call `fluid.cuda_places()` to get all GPU places. 
                #  - If you are using CPU, call `fluid.cpu_places()` to get all CPU places. 
                # 
                # If DataLoader is not iterable, places can be None.
                places = fluid.cuda_places() if USE_GPU else fluid.cpu_places()
                set_data_source(loader, places)
S
sneaxiy 已提交
233

Z
Zeng Jinle 已提交
234 235
                exe = fluid.Executor(places[0])
                exe.run(fluid.default_startup_program())
H
Huihuang Zheng 已提交
236

Z
Zeng Jinle 已提交
237
                prog = fluid.CompiledProgram(fluid.default_main_program()).with_data_parallel(loss_name=loss.name)
238

Z
Zeng Jinle 已提交
239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264
                if loader.iterable:
                    train_iterable(exe, prog, loss, loader)
                else:
                    train_non_iterable(exe, prog, loss, loader)


                '''
                Users can use return_list = True in dygraph mode. 
                '''
                with fluid.dygraph.guard(places[0]):
                    loader = fluid.io.DataLoader.from_generator(capacity=2, return_list=True)
                    set_data_source(loader, places[0]) 
                    for image, label in loader():
                        relu = fluid.layers.relu(image)
                        assert image.shape == [BATCH_SIZE, 784] 
                        assert label.shape == [BATCH_SIZE, 1]
                        assert relu.shape == [BATCH_SIZE, 784]
        """
        return GeneratorLoader(feed_list, capacity, use_double_buffer, iterable,
                               return_list)

    @staticmethod
    def from_dataset(dataset, places, drop_last=True):
        """
        Create an iterable DataLoader object for loading data from Dataset.    
        Dataset is only supported in Linux system currently.
265

Z
Zeng Jinle 已提交
266 267 268 269 270 271 272
        Args:
            dataset (InMemoryDataset|QueueDataset): the dataset object.
            places (list(CUDAPlace)|list(CPUPlace)): places where the result 
                data should be converted.   
            drop_last (bool): whether to drop the last batch whose sample 
                number is less than batch size. If drop_last = True, they
                would be dropped. If drop_last = False, they would be kept. 
273

Z
Zeng Jinle 已提交
274 275 276
        Returns:
            loader (DataLoader): the created DataLoader object, which can be 
                treated as a Python generator.   
277

Z
Zeng Jinle 已提交
278 279 280
        Examples:

            .. code-block:: python
281

Z
Zeng Jinle 已提交
282
                import paddle.fluid as fluid
283

284 285
                image = fluid.data(name='image', shape=[None, 784], dtype='float32')
                label = fluid.data(name='label', shape=[None, 1], dtype='int64')
286

Z
Zeng Jinle 已提交
287 288 289 290 291
                dataset = fluid.DatasetFactory().create_dataset("QueueDataset")
                dataset.set_batch_size(32)
                dataset.set_filelist(['a.txt', 'b.txt', 'c.txt'])
                dataset.set_use_var([image, label])
                dataset.set_pipe_command('cat') 
292

Z
Zeng Jinle 已提交
293 294 295
                loader = fluid.io.DataLoader.from_dataset(dataset, fluid.cpu_places())
        """
        return DatasetLoader(dataset, places, drop_last)
S
sneaxiy 已提交
296

S
sneaxiy 已提交
297

Z
Zeng Jinle 已提交
298
class GeneratorLoader(DataLoaderBase):
S
sneaxiy 已提交
299
    def __init__(self,
300 301
                 feed_list=None,
                 capacity=None,
S
sneaxiy 已提交
302
                 use_double_buffer=True,
303 304
                 iterable=True,
                 return_list=False):
S
sneaxiy 已提交
305
        self._tensor_reader = None
Z
Zeng Jinle 已提交
306
        self._places = None
S
sneaxiy 已提交
307
        self._thread = None
308 309 310 311 312 313 314
        self._feed_list = feed_list
        if not capacity:
            raise ValueError("Please give value to capacity.")
        # force to use iterable mode under dygraph mode
        if in_dygraph_mode():
            if not iterable:
                warnings.warn(
Z
Zeng Jinle 已提交
315 316
                    "Please NOTE: dygraph can support iterable mode only. Change to iterable mode."
                )
317 318 319
            self._iterable = True
            if not return_list:
                warnings.warn(
Z
Zeng Jinle 已提交
320 321
                    "Please NOTE: dygraph can support return as list only. Change to return as list."
                )
322 323 324 325 326 327
            self._return_list = True
        else:
            self._iterable = iterable
            self._return_list = return_list
            if not self._feed_list:
                raise Exception("Feed list must be given under static mode.")
S
sneaxiy 已提交
328 329 330 331
        self._use_double_buffer = use_double_buffer
        self._capacity = capacity
        if not self._iterable:
            self._init_non_iterable()
S
sneaxiy 已提交
332

Z
Zeng Jinle 已提交
333
    def _wait_thread_ends(self):
334
        # Get self._thread first to prevent data race, because __thread_main__
Z
Zeng Jinle 已提交
335 336 337 338 339 340 341 342
        # would set self._thread be None at the end
        thread = self._thread
        if thread is not None and self._iterable:
            self._queue.close()
            thread.join()

    def _init_iterable(self):
        self._wait_thread_ends()
343 344
        if in_dygraph_mode():
            self._var_names = []
345 346 347
            self._shapes = []
            self._dtypes = []
            self._need_check_feed = []
348 349
        else:
            self._var_names = [v.name for v in self._feed_list]
350 351 352 353 354
            self._shapes = [v.shape for v in self._feed_list]
            self._dtypes = [v.dtype for v in self._feed_list]
            self._need_check_feed = [
                v.desc.need_check_feed() for v in self._feed_list
            ]
S
sneaxiy 已提交
355 356 357
        self._queue = core.init_lod_tensor_blocking_queue(core.Variable(),
                                                          self._capacity)
        self._reader = core.create_py_reader(
358 359
            self.queue, self._var_names, self._shapes, self._dtypes,
            self._need_check_feed, self._places, self._use_double_buffer)
S
sneaxiy 已提交
360 361 362 363 364 365 366

    def _init_non_iterable(self):
        lod_levels = []
        dtypes = []
        shape_concat = []
        ranks = []
        shapes = []
367
        need_check_feed = []
S
sneaxiy 已提交
368 369 370 371 372 373 374

        for feed_data in self._feed_list:
            dtypes.append(feed_data.dtype)
            shape_concat.extend(feed_data.shape)
            ranks.append(len(feed_data.shape))
            shapes.append(feed_data.shape)
            lod_levels.append(feed_data.lod_level)
375
            need_check_feed.append(int(feed_data.desc.need_check_feed()))
S
sneaxiy 已提交
376

Z
Zeng Jinle 已提交
377 378 379 380
        queue_name = data_loader_unique_name_generator(
            'lod_tensor_blocking_queue')
        reader_name = data_loader_unique_name_generator('create_py_reader')
        double_buffer_name = data_loader_unique_name_generator('double_buffer')
S
sneaxiy 已提交
381

S
sneaxiy 已提交
382
        var = global_scope().var(queue_name)
S
sneaxiy 已提交
383 384 385 386 387
        self._queue = core.init_lod_tensor_blocking_queue(var, self._capacity)

        startup_blk = default_startup_program().current_block()
        startup_var = startup_blk.create_var(name=reader_name)

388
        dtype_int = [int(t) for t in dtypes]
S
sneaxiy 已提交
389 390 391 392 393 394 395
        startup_blk.append_op(
            type='create_py_reader',
            inputs={'blocking_queue': [queue_name]},
            outputs={'Out': [startup_var]},
            attrs={
                'shape_concat': shape_concat,
                'lod_levels': lod_levels,
396 397
                'dtypes': dtype_int,
                'need_check_feed': need_check_feed,
S
sneaxiy 已提交
398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432
                'ranks': ranks
            })

        startup_var.desc.set_dtypes(dtypes)
        startup_var.persistable = True

        main_prog_var = _copy_reader_var_(
            default_main_program().current_block(), startup_var)

        main_prog_var.stop_gradient = True
        main_prog_var.persistable = True

        reader = monkey_patch_reader_methods(main_prog_var)
        if self._use_double_buffer:
            double_buffer_reader = double_buffer(
                reader, name=double_buffer_name)
            # we return a double buffer reader. However, the reset method comes from
            # py_reader.
            double_buffer_reader.reset = reader.reset
            reader = double_buffer_reader

        self._reader = reader

        default_main_program().current_block().append_op(
            type='read',
            inputs={'Reader': [self._reader]},
            outputs={'Out': self._feed_list})

    @property
    def queue(self):
        return self._queue

    @property
    def iterable(self):
        return self._iterable
S
sneaxiy 已提交
433

Z
Zeng Jinle 已提交
434 435
    def __iter__(self):
        assert self.iterable, "DataLoader is not iterable"
S
sneaxiy 已提交
436
        assert self._tensor_reader is not None, \
Z
Zeng Jinle 已提交
437
            "Data source of DataLoader has not set yet"
S
sneaxiy 已提交
438

Z
Zeng Jinle 已提交
439
        self._init_iterable()
S
sneaxiy 已提交
440
        self._start()
Z
Zeng Jinle 已提交
441 442 443 444
        return self

    def __next__(self):
        try:
445 446 447
            if in_dygraph_mode():
                return self._reader.read_next_var_list()
            else:
Z
Zeng Jinle 已提交
448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466
                if self._return_list:
                    return self._reader.read_next_list()
                else:
                    return self._reader.read_next()
        except StopIteration:
            self._queue.close()
            self._reset()
            six.reraise(*sys.exc_info())

    def start(self):
        if not in_dygraph_mode():
            assert not self._iterable, "start() cannot be called when DataLoader is iterable"
            self._start()

    def reset(self):
        if not in_dygraph_mode():
            assert not self._iterable, "reset() cannot be called when DataLoader is iterable"
            self._reset()

467 468 469 470 471 472 473 474 475 476 477
    @classmethod
    def _check_input_array(cls, item):
        arr = np.array(item)
        if arr.dtype == np.object:
            raise TypeError((
                "\n\tFaild to convert input data to a regular ndarray :\n\t* Usually "
                "this means the input data contains nested lists with different lengths. "
                "\n\t* Check the reader function passed to 'decorate_batch_generator'"
                " to locate the data causes this issue.\n\t* Please consider using "
                "'fluid.create_lod_tensor' to convert it to a LoD-Tensor."))

Z
Zeng Jinle 已提交
478 479 480 481 482 483 484
    def _start(self):
        def __thread_main__():
            try:
                for tensors in self._tensor_reader():
                    array = core.LoDTensorArray()
                    for item in tensors:
                        if not isinstance(item, core.LoDTensor):
485
                            self._check_input_array(item)
Z
Zeng Jinle 已提交
486 487 488 489 490 491 492 493 494 495 496 497
                            tmp = core.LoDTensor()
                            tmp.set(item, core.CPUPlace())
                            item = tmp

                        array.append(item)

                    if not self._queue.push(array):
                        break

                self._queue.close()
                self._thread = None
            except Exception as ex:
Z
Zeng Jinle 已提交
498
                self._queue.kill()
Z
Zeng Jinle 已提交
499 500 501 502 503 504 505
                self._thread = None
                logging.warn('Your reader has raised an exception!')
                six.reraise(*sys.exc_info())

        self._thread = threading.Thread(target=__thread_main__)
        self._thread.daemon = True
        self._thread.start()
S
sneaxiy 已提交
506

S
sneaxiy 已提交
507
    def _reset(self):
S
sneaxiy 已提交
508
        self._reader.reset()
Z
Zeng Jinle 已提交
509 510 511 512 513 514 515 516 517 518
        thread = self._thread
        if thread is not None:
            thread.join()

    def set_sample_generator(self,
                             reader,
                             batch_size,
                             drop_last=True,
                             places=None):
        assert batch_size > 0, "batch_size must be larger than 0"
519 520 521 522 523 524
        if in_dygraph_mode():
            self.set_sample_list_generator(
                paddle.batch(
                    reader, batch_size=batch_size, drop_last=drop_last),
                places=places)
        else:
Z
Zeng Jinle 已提交
525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546
            has_lod = False
            for f in self._feed_list:
                if f.lod_level != 0:
                    has_lod = True
                    break

            if has_lod:
                self.set_sample_list_generator(
                    paddle.batch(
                        reader, batch_size=batch_size, drop_last=drop_last),
                    places=places)
            else:
                reader = BatchedTensorProvider(
                    feed_list=self._feed_list,
                    place=core.CPUPlace(),
                    batch_size=batch_size,
                    generator=reader,
                    drop_last=drop_last)
                self.set_batch_generator(reader, places=places)
        return self

    def set_sample_list_generator(self, reader, places=None):
547 548 549
        if in_dygraph_mode():

            def __tensor_reader_impl__():
550 551 552 553 554 555 556 557 558
                for batch in reader():
                    slots = []
                    for items in batch:
                        for i, item in enumerate(items):
                            if len(slots) < len(items):
                                slots.append([item])
                            else:
                                slots[i].append(item)
                    yield slots
559
        else:
Z
Zeng Jinle 已提交
560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578
            with program_guard(Program(), Program()):
                feeder = DataFeeder(
                    feed_list=self._feed_list, place=core.CPUPlace())
                paddle_reader = feeder.decorate_reader(
                    reader, multi_devices=False)

            def __tensor_reader_impl__():
                for slots in paddle_reader():
                    yield [slots[var.name] for var in self._feed_list]

        self.set_batch_generator(__tensor_reader_impl__, places)
        return self

    def set_batch_generator(self, reader, places=None):
        self._tensor_reader = reader
        if self._iterable:
            assert places is not None, "Places cannot be None when DataLoader is iterable"
            self._places = _convert_places(places)
            if in_dygraph_mode():
579 580
                assert len(self._places) == 1, \
                    "Number of places must be 1 in dygraph mode"
Z
Zeng Jinle 已提交
581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616
        else:
            if places is not None:
                logging.info(
                    'places would be ommited when DataLoader is not iterable')
        return self


class PyReader(DataLoaderBase):
    """
    Create a reader object for data feeding in Python. 
    Data would be prefetched using Python thread and be pushed
    into a queue asynchronously. Data in the queue would be extracted 
    automatically when `Executor.run(...)` is called.

    Args:  
        feed_list (list(Variable)|tuple(Variable)): feed variable list.
            The variables should be created by :code:`fluid.layers.data()`.
        capacity (int): capacity of the queue maintained in PyReader.
            The unit is batch number. Set larger capacity if your reader 
            is fast. 
        use_double_buffer (bool): whether to use double_buffer_reader. 
            If use_double_buffer=True, PyReader would prefetch next 
            batch data asynchronously, so it would speed up data feeding 
            and occupies a little more CPU or GPU memory, i.e., the memory
            of one batch input data. 
        iterable (bool): whether the created PyReader is iterable. 
        return_list (bool): whether the return value on each device is 
            presented as a list. It is only valid when iterable=True. 
            If return_list=False, the return value on each device would 
            be a dict of str -> LoDTensor, where the key of the dict is 
            the name of each feeded variables. If return_list=True, the 
            return value on each device would be a list(LoDTensor). It is
            recommended to use return_list=False in static graph mode and
            use return_list=True in dygraph mode. 

    Returns:
G
guofei 已提交
617 618 619 620
        the created reader object.

    Return type:
        reader(Reader)
Z
Zeng Jinle 已提交
621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639

    Examples:
        1. If iterable = False, the created PyReader object is almost the
           same as :code:`fluid.layers.py_reader()`. Operators would be 
           inserted into the program. User should call :code:`start()` 
           before each epoch and catch :code:`fluid.core.EOFException`
           thrown by :code:`Executor.run()` when epoch ends. Once the 
           exception is caught, user should call :code:`reset()` to reset 
           the reader manually.

        .. code-block:: python

           import paddle
           import paddle.fluid as fluid
           import numpy as np

           EPOCH_NUM = 3
           ITER_NUM = 5
           BATCH_SIZE = 3
G
guofei 已提交
640 641 642 643 644
           
           def network(image, label):
               # User-defined network, here is an example of softmax regression.
               predict = fluid.layers.fc(input=image, size=10, act='softmax')           
               return fluid.layers.cross_entropy(input=predict, label=label)
Z
Zeng Jinle 已提交
645 646 647 648 649 650 651 652 653 654 655

           def reader_creator_random_image_and_label(height, width):
               def reader():
                   for i in range(ITER_NUM):
                       fake_image = np.random.uniform(low=0,
                                                      high=255,
                                                      size=[height, width])
                       fake_label = np.ones([1])
                       yield fake_image, fake_label
               return reader

G
guofei 已提交
656 657
           image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
           label = fluid.data(name='label', shape=[None, 1], dtype='int64')
Z
Zeng Jinle 已提交
658 659 660 661 662 663 664 665

           reader = fluid.io.PyReader(feed_list=[image, label],
                                      capacity=4,
                                      iterable=False)

           user_defined_reader = reader_creator_random_image_and_label(784, 784)
           reader.decorate_sample_list_generator(
               paddle.batch(user_defined_reader, batch_size=BATCH_SIZE))
G
guofei 已提交
666 667
           loss = network(image, label)
           executor = fluid.Executor(fluid.CPUPlace())
Z
Zeng Jinle 已提交
668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694
           executor.run(fluid.default_startup_program())
           for i in range(EPOCH_NUM):
               reader.start()
               while True:
                   try:
                       executor.run(feed=None)
                   except fluid.core.EOFException:
                       reader.reset()
                       break

 
        2. If iterable=True, the created PyReader object is decoupled with
           the program. No operator would be inserted into the program. 
           In this case, the created reader is a Python generator, which 
           is iterable. User should feed the data yielded from PyReader 
           object into :code:`Executor.run(feed=...)`.  

        .. code-block:: python

           import paddle
           import paddle.fluid as fluid
           import numpy as np

           EPOCH_NUM = 3
           ITER_NUM = 5
           BATCH_SIZE = 10

G
guofei 已提交
695 696 697 698 699
           def network(image, label):
               # User-defined network, here is an example of softmax regression.
               predict = fluid.layers.fc(input=image, size=10, act='softmax')           
               return fluid.layers.cross_entropy(input=predict, label=label)

Z
Zeng Jinle 已提交
700 701 702
           def reader_creator_random_image(height, width):
               def reader():
                   for i in range(ITER_NUM):
G
guofei 已提交
703 704 705
                       fake_image = np.random.uniform(low=0, high=255, size=[height, width])
                       fake_label = np.ones([1])
                       yield fake_image, fake_label 
Z
Zeng Jinle 已提交
706 707
               return reader

G
guofei 已提交
708 709 710
           image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
           label = fluid.data(name='label', shape=[None, 1], dtype='int64')
           reader = fluid.io.PyReader(feed_list=[image, label], capacity=4, iterable=True, return_list=False)
Z
Zeng Jinle 已提交
711 712 713 714

           user_defined_reader = reader_creator_random_image(784, 784)
           reader.decorate_sample_list_generator(
               paddle.batch(user_defined_reader, batch_size=BATCH_SIZE),
G
guofei 已提交
715 716 717 718 719 720
                   fluid.core.CPUPlace())
           
           loss = network(image, label)
           executor = fluid.Executor(fluid.CPUPlace())
           executor.run(fluid.default_startup_program())
           
Z
Zeng Jinle 已提交
721 722
           for _ in range(EPOCH_NUM):
               for data in reader():
G
guofei 已提交
723
                   executor.run(feed=data, fetch_list=[loss])
Z
Zeng Jinle 已提交
724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777


        3. If return_list=True, the return values would be presented as list instead of dict. 
           This is usually used in dygraph mode.

        .. code-block:: python

           import paddle
           import paddle.fluid as fluid
           import numpy as np

           ITER_NUM = 5
           BATCH_SIZE = 10

           def reader_creator_random_image(height, width):
               def reader():
                   for i in range(ITER_NUM):
                       yield np.random.uniform(low=0, high=255, size=[height, width]), \
                           np.random.random_integers(low=0, high=9, size=[1])
               return reader

           place = fluid.CPUPlace()
           with fluid.dygraph.guard(place):
               py_reader = fluid.io.PyReader(capacity=2, return_list=True)
               user_defined_reader = reader_creator_random_image(784, 784)
               py_reader.decorate_sample_list_generator(
                   paddle.batch(user_defined_reader, batch_size=BATCH_SIZE),
                   place)
               for image, label in py_reader():
                   relu = fluid.layers.relu(image)
    """

    def __init__(self,
                 feed_list=None,
                 capacity=None,
                 use_double_buffer=True,
                 iterable=True,
                 return_list=False):
        self._loader = DataLoader.from_generator(
            feed_list, capacity, use_double_buffer, iterable, return_list)

    @property
    def queue(self):
        return self._loader.queue

    @property
    def iterable(self):
        return self._loader.iterable

    def __iter__(self):
        return self._loader.__iter__()

    def __next__(self):
        return self._loader.__next__()
S
sneaxiy 已提交
778 779

    def start(self):
S
add doc  
sneaxiy 已提交
780 781 782
        '''
        Start the data feeding thread. 
        Can only call when the reader object is not iterable.  
783
        
G
guofei 已提交
784 785
	Example:
	    .. code-block:: python
Z
Zeng Jinle 已提交
786
    
H
Huihuang Zheng 已提交
787 788 789 790
                import paddle
                import paddle.fluid as fluid
                import numpy as np

791 792 793 794 795 796
                BATCH_SIZE = 10

                def generator():
                    for i in range(5):
                        yield np.random.uniform(low=0, high=255, size=[784, 784]),

G
guofei 已提交
797
                image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
798 799 800 801
                reader = fluid.io.PyReader(feed_list=[image], capacity=4, iterable=False)
                reader.decorate_sample_list_generator(
                    paddle.batch(generator, batch_size=BATCH_SIZE))

G
guofei 已提交
802
                executor = fluid.Executor(fluid.CPUPlace())
803 804 805 806 807 808 809 810 811 812
                executor.run(fluid.default_startup_program())
                for i in range(3):
                    reader.start()
                    while True:
                        try:
                            executor.run(feed=None)
                        except fluid.core.EOFException:
                            reader.reset()
                            break

Z
Zeng Jinle 已提交
813 814
	    '''
        self._loader.start()
S
sneaxiy 已提交
815

S
sneaxiy 已提交
816
    def reset(self):
S
add doc  
sneaxiy 已提交
817 818 819
        '''
        Reset the reader object when :code:`fluid.core.EOFException` raises. 
        Can only call when the reader object is not iterable.
820 821 822 823
        
        Example:
            .. code-block:: python

H
Huihuang Zheng 已提交
824 825 826 827
                import paddle
                import paddle.fluid as fluid
                import numpy as np

828 829 830 831 832 833
                BATCH_SIZE = 10

                def generator():
                    for i in range(5):
                        yield np.random.uniform(low=0, high=255, size=[784, 784]),

G
guofei 已提交
834
                image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
835 836 837 838
                reader = fluid.io.PyReader(feed_list=[image], capacity=4, iterable=False)
                reader.decorate_sample_list_generator(
                    paddle.batch(generator, batch_size=BATCH_SIZE))

G
guofei 已提交
839
                executor = fluid.Executor(fluid.CPUPlace())
840 841 842 843 844 845 846 847 848 849
                executor.run(fluid.default_startup_program())
                for i in range(3):
                    reader.start()
                    while True:
                        try:
                            executor.run(feed=None)
                        except fluid.core.EOFException:
                            reader.reset()
                            break        

S
add doc  
sneaxiy 已提交
850
        '''
Z
Zeng Jinle 已提交
851
        self._loader.reset()
S
sneaxiy 已提交
852

S
sneaxiy 已提交
853 854 855 856 857 858 859 860 861
    def decorate_sample_generator(self,
                                  sample_generator,
                                  batch_size,
                                  drop_last=True,
                                  places=None):
        '''
        Set the data source of the PyReader object.
        
        The provided :code:`sample_generator` should be a Python generator,
862
        which yields list(numpy.ndarray)-typed data of each sample.
S
sneaxiy 已提交
863 864 865 866

        :code:`places` must be set when the PyReader object is iterable.

        If all inputs have no lods, this method is faster than 
S
sneaxiy 已提交
867
        :code:`decorate_sample_list_generator(paddle.batch(sample_generator, ...))` .
S
sneaxiy 已提交
868 869 870

        Args:
            sample_generator (generator): Python generator that yields
871
                list(numpy.ndarray)-typed sample data.
S
sneaxiy 已提交
872 873 874 875 876
            batch_size (int): batch size. Must be larger than 0.
            drop_last (bool): Whether to drop the last batch when sample number
                is less than batch_size. 
            places (None|list(CUDAPlace)|list(CPUPlace)): place list. Must
                be provided when PyReader is iterable.
877 878 879 880

        Example:
            .. code-block:: python

H
Huihuang Zheng 已提交
881 882 883
                import paddle.fluid as fluid
                import numpy as np

884 885 886
                EPOCH_NUM = 3
                ITER_NUM = 15
                BATCH_SIZE = 3
G
guofei 已提交
887 888 889 890 891
        
                def network(image, label):
                    # User-defined network, here is an example of softmax regression.
                    predict = fluid.layers.fc(input=image, size=10, act='softmax')           
                    return fluid.layers.cross_entropy(input=predict, label=label)
892 893 894 895 896 897 898 899 900 901 902

                def random_image_and_label_generator(height, width):
                    def generator():
                        for i in range(ITER_NUM):
                            fake_image = np.random.uniform(low=0,
                                                           high=255,
                                                           size=[height, width])
                            fake_label = np.array([1])
                            yield fake_image, fake_label
                    return generator

G
guofei 已提交
903 904
                image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
                label = fluid.data(name='label', shape=[None, 1], dtype='int64')
905 906 907 908 909
                reader = fluid.io.PyReader(feed_list=[image, label], capacity=4, iterable=True)

                user_defined_generator = random_image_and_label_generator(784, 784)
                reader.decorate_sample_generator(user_defined_generator,
                                                 batch_size=BATCH_SIZE,
G
guofei 已提交
910 911 912 913
                                                 places=[fluid.CPUPlace()])
                loss = network(image, label)
                executor = fluid.Executor(fluid.CPUPlace())
                executor.run(fluid.default_startup_program())
914 915 916

                for _ in range(EPOCH_NUM):
                    for data in reader():
G
guofei 已提交
917
                        executor.run(feed=data, fetch_list=[loss])
918
    
S
sneaxiy 已提交
919
        '''
Z
Zeng Jinle 已提交
920 921
        self._loader.set_sample_generator(sample_generator, batch_size,
                                          drop_last, places)
S
sneaxiy 已提交
922

S
sneaxiy 已提交
923
    def decorate_sample_list_generator(self, reader, places=None):
S
add doc  
sneaxiy 已提交
924 925 926 927
        '''
        Set the data source of the PyReader object. 

        The provided :code:`reader` should be a Python generator,
S
sneaxiy 已提交
928
        which yields list(numpy.ndarray) typed batched data. 
S
add doc  
sneaxiy 已提交
929 930 931 932
        
        :code:`places` must be set when the PyReader object is iterable.

        Args:
S
sneaxiy 已提交
933 934 935 936
            reader (generator): Python generator that yields 
                list(numpy.ndarray)-typed batched data. 
            places (None|list(CUDAPlace)|list(CPUPlace)): place list. Must
                be provided when PyReader is iterable.
937 938 939 940
        
        Example:
            .. code-block:: python

H
Huihuang Zheng 已提交
941 942 943 944
                import paddle
                import paddle.fluid as fluid
                import numpy as np

945 946 947 948
                EPOCH_NUM = 3
                ITER_NUM = 15
                BATCH_SIZE = 3

G
guofei 已提交
949 950 951 952 953
                def network(image, label):
                    # User-defined network, here is an example of softmax regression.
                    predict = fluid.layers.fc(input=image, size=10, act='softmax')           
                    return fluid.layers.cross_entropy(input=predict, label=label)

954 955 956 957 958 959 960 961 962 963
                def random_image_and_label_generator(height, width):
                    def generator():
                        for i in range(ITER_NUM):
                            fake_image = np.random.uniform(low=0,
                                                           high=255,
                                                           size=[height, width])
                            fake_label = np.ones([1])
                            yield fake_image, fake_label
                    return generator

G
guofei 已提交
964 965
                image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
                label = fluid.data(name='label', shape=[None, 1], dtype='int64')
966 967 968 969 970
                reader = fluid.io.PyReader(feed_list=[image, label], capacity=4, iterable=True)

                user_defined_generator = random_image_and_label_generator(784, 784)
                reader.decorate_sample_list_generator(
                    paddle.batch(user_defined_generator, batch_size=BATCH_SIZE),
G
guofei 已提交
971 972 973 974 975
                    fluid.core.CPUPlace())
                
                loss = network(image, label)
                executor = fluid.Executor(fluid.core.CPUPlace())
                executor.run(fluid.default_startup_program())
976 977 978

                for _ in range(EPOCH_NUM):
                    for data in reader():
G
guofei 已提交
979
                        executor.run(feed=data, fetch_list=[loss])
980
                 
S
add doc  
sneaxiy 已提交
981
        '''
Z
Zeng Jinle 已提交
982
        self._loader.set_sample_list_generator(reader, places)
S
sneaxiy 已提交
983

S
sneaxiy 已提交
984
    def decorate_batch_generator(self, reader, places=None):
S
add doc  
sneaxiy 已提交
985 986 987 988
        '''
        Set the data source of the PyReader object.

        The provided :code:`reader` should be a Python generator,
S
sneaxiy 已提交
989
        which yields numpy.ndarray-typed or LoDTensor-typed batched data.
S
add doc  
sneaxiy 已提交
990 991 992 993 994 995

        :code:`places` must be set when the PyReader object is iterable.

        Args:
            reader (generator): Python generator that yields LoDTensor-typed
                batched data.
S
sneaxiy 已提交
996
            places (None|list(CUDAPlace)|list(CPUPlace)): place list. Must
S
sneaxiy 已提交
997
                be provided when PyReader is iterable.
998 999 1000 1001

        Example:
            .. code-block:: python

H
Huihuang Zheng 已提交
1002 1003 1004
                import paddle.fluid as fluid
                import numpy as np

1005 1006 1007
                EPOCH_NUM = 3
                ITER_NUM = 15
                BATCH_SIZE = 3
G
guofei 已提交
1008 1009 1010 1011 1012
               
                def network(image, label):
                    # User-defined network, here is an example of softmax regression.
                    predict = fluid.layers.fc(input=image, size=10, act='softmax')           
                    return fluid.layers.cross_entropy(input=predict, label=label)
1013 1014 1015 1016 1017 1018 1019 1020

                def random_image_and_label_generator(height, width):
                    def generator():
                        for i in range(ITER_NUM):
                            batch_image = np.random.uniform(low=0,
                                                            high=255,
                                                            size=[BATCH_SIZE, height, width])
                            batch_label = np.ones([BATCH_SIZE, 1])
G
guofei 已提交
1021 1022
                            batch_image = batch_image.astype('float32')
                            batch_label = batch_label.astype('int64')
1023 1024 1025
                            yield batch_image, batch_label
                    return generator

G
guofei 已提交
1026 1027
                image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
                label = fluid.data(name='label', shape=[None, 1], dtype='int64')
1028 1029 1030
                reader = fluid.io.PyReader(feed_list=[image, label], capacity=4, iterable=True)

                user_defined_generator = random_image_and_label_generator(784, 784)
G
guofei 已提交
1031 1032 1033 1034 1035
                reader.decorate_batch_generator(user_defined_generator, fluid.CPUPlace())
                
                loss = network(image, label)
                executor = fluid.Executor(fluid.CPUPlace())
                executor.run(fluid.default_startup_program())
1036 1037 1038

                for _ in range(EPOCH_NUM):
                    for data in reader():
G
guofei 已提交
1039
                        executor.run(feed=data, fetch_list=[loss])
1040

S
add doc  
sneaxiy 已提交
1041
        '''
Z
Zeng Jinle 已提交
1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086
        self._loader.set_batch_generator(reader, places)


class DatasetLoader(DataLoaderBase):
    def __init__(self, dataset, places, drop_last):
        assert isinstance(dataset,
                          DatasetBase), "dataset must be type of DatasetBase"
        assert not in_dygraph_mode(
        ), "DatasetLoader is not supported in dygraph mode yet"

        thread_num = len(places)

        assert len(dataset.filelist) >= thread_num, \
            "Filelist number of dataset {} must be not less than place number {}".format(len(dataset.filelist), thread_num)

        if dataset.thread_num != 0 and dataset.thread_num != thread_num:
            logging.warn('thread_num {} which is set in Dataset is ignored'.
                         format(dataset.thread_num))

        dataset.set_thread(thread_num)

        if isinstance(dataset,
                      InMemoryDataset) and dataset.queue_num > thread_num:
            logging.warn("queue_num {} which is set in Dataset is ignored".
                         format(dataset.queue_num))
            dataset.set_queue_num(thread_num)

        self._dataset = dataset
        use_slots = [
            slot.name for slot in dataset.proto_desc.multi_slot_desc.slots
            if slot.is_used
        ]

        self._iterable_dataset = core.IterableDatasetWrapper(
            dataset.dataset, use_slots,
            _convert_places(places), dataset.proto_desc.batch_size, drop_last)

    def __iter__(self):
        self._dataset._finish_to_run()
        self._dataset._prepare_to_run()
        self._iterable_dataset._start()
        return self

    def __next__(self):
        return self._iterable_dataset._next()