reader.py 69.4 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
16
import sys
S
sneaxiy 已提交
17
import six
18
import numpy as np
S
sneaxiy 已提交
19
import threading
20
import paddle
21
from .framework import Program, Variable, program_guard, default_main_program, default_startup_program, in_dygraph_mode, cpu_places, _current_expected_place
S
sneaxiy 已提交
22
from .executor import global_scope
23
from .data_feeder import DataFeeder, BatchedTensorProvider
24 25 26
from .multiprocess_utils import multiprocess_queue_set, CleanupFuncRegistrar, _cleanup_mmap, _cleanup, _set_SIGCHLD_handler
from .dataloader import BatchSampler, Dataset
from .dataloader.dataloader_iter import _DataLoaderIterSingleProcess, _DataLoaderIterMultiProcess
S
sneaxiy 已提交
27
from .layers.io import monkey_patch_reader_methods, _copy_reader_var_, double_buffer
S
sneaxiy 已提交
28
from .unique_name import UniqueNameGenerator
29
import logging
Z
Zeng Jinle 已提交
30
from .dataset import DatasetBase, InMemoryDataset
S
sneaxiy 已提交
31

32
### Dygraph DataLoader configs ###
33
import os
34 35
import multiprocessing
import signal
36

37
# NOTE: queue has a different name in python2 and python3
38
if six.PY2:
39 40 41
    import Queue as queue
else:
    import queue
42

43 44 45
# NOTE: [ avoid hanging & failed quickly ] These value is used in getting data from another process
QUEUE_GET_TIMEOUT = 60

Z
Zeng Jinle 已提交
46 47 48
__all__ = ['PyReader', 'DataLoader']

data_loader_unique_name_generator = UniqueNameGenerator()
S
sneaxiy 已提交
49

50 51 52 53 54 55 56 57 58 59 60
KEEP_DATA_LOADER_ORDER = True


def keep_data_loader_order(*args):
    global KEEP_DATA_LOADER_ORDER
    if len(args) == 0:
        return KEEP_DATA_LOADER_ORDER
    else:
        assert len(args) == 1 and isinstance(args[0], bool)
        KEEP_DATA_LOADER_ORDER = args[0]

S
sneaxiy 已提交
61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76

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
77 78 79
class DataLoaderBase(object):
    def __init__(self):
        self._places = None
S
sneaxiy 已提交
80

Z
Zeng Jinle 已提交
81 82
    def __call__(self):
        return self
S
sneaxiy 已提交
83

Z
Zeng Jinle 已提交
84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100
    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):
101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358
    """
    DataLoader prodives an iterator which iterates given dataset
    once by the batch_sampler.

    DataLoader supports single-process and multi-prcess data loading,
    multi-process workers will be used to load data asynchronously if
    :attr:`num_workers` is set as a positive number.

    DataLoader only supports map-style dataset(can get a sample from
    dataset with a given index) currently, for a map-style dataset,
    please see :code:`paddle.io.Dataset`.

    batch_sampler please see :code:`paddle.io.BatchSampler`

    Args:  
        dataset(Dataset): the dataset to load data from, should be an
            instance of subclass of :code:`paddle.io.Dataset`.
        feed_list (list(Variable)|tuple(Variable)): feed variable list.
            The variables should be created by :code:`fluid.data()`.
            :attr:`feed_list` must be set if :attr:`return_list` is
            False. Default None.
        places(list(Place)|tuple(Place)): a list of Place, to put data
            onto, :attr:`places` must be set in both static graph and 
            dynamic graph mode, in dynamic graph mode, place number must
            be 1. Default None.
        return_list (bool): whether the return value on each device is 
            presented as a list. If :attr:`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 fed variables. If 
            :attr:`return_list=True`, the return value on each device would
            be a list(LoDTensor). :attr:`return_list` can only be True
            in dynamic graph mode. Default False.
        batch_sampler(BatchSampler): an instance of `paddle.io.BatchSampler`
            to generate batch indices to draw samples from :attr:`dataset`
            and combine a batch. Default None.
        batch_size(int): sample number in a mini-batch, a substitution
            parameter for :attr:`batch_sampler`, if :attr:`batch_sampler`
            is not set, a default `paddle.io.BatchSampler` will be used
            and initialize by :attr:`batch_size`, :attr:`shuffle` and
            :attr:`drop_last`. Default 1.
        shuffle(bool): whther to shuffle indices order before genrate
            batch indices, a substitution parameter for :attr:`batch_sampler`
            see :attr:`batch_size`. Default False.
        drop_last(bool): whether drop the last incomplete batch dataset size
            is not divisible by the batch size, a substitution parameter
            for :attr:`batch_sampler`, see :attr:`batch_size`. Default False
        collate_fn(callable): function to generate mini-batch data by merging
            the sample list, None for only stack each fields of sample in axis
            0(same as :attr::`np.stack(..., axis=0)`). Default None
        num_workers(int): the number of subprocess to load data, 0 for no
            subprocess used and loading data in main process. Default 0
        use_buffer_reader (bool): whether to use bufferred reader. 
            If use_buffer_reader=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. Default True.
        use_shared_memory (bool): whether to use shared memory to speed up
            putting data into inter-process queue, set :attr:`use_shared_memory`
            as True only when the shared memory space on your machine(e.g.
            space of '/dev/shm' on Linux operating sysytem) is large enough.
            Shared memory will only be enabled in multi-process mode(num_workers
            > 0). Default True.
        timeout(int): the timeout value for getting data form output queue
            of subprocesses. Default 0.
        worker_init_fn(callable): init function which will be called with
            worker id on each subproces starting if not set as None. Default
            None.

    Returns:
        DataLoader: an iterable object for data iterating

    Examples:
        
        .. code-block:: python

            import numpy as np
            import paddle.fluid as fluid
            from paddle.io import Dataset, BatchSampler, DataLoader

            BATCH_NUM = 20
            BATCH_SIZE = 16
            EPOCH_NUM = 4

            IMAGE_SIZE = 784
            CLASS_NUM = 10

            USE_GPU = False # whether use GPU to run model

            # define a random dataset
            class RandomDataset(Dataset):
                def __init__(self, num_samples):
                    self.num_samples = num_samples

                def __getitem__(self, idx):
                    image = np.random.random([IMAGE_SIZE]).astype('float32')
                    label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
                    return image, label

                def __len__(self):
                    return self.num_samples

            # get places
            places = fluid.cuda_places() if USE_GPU else fluid.cpu_places()

            # -------------------- static graph ---------------------

            def simple_net(image, label):
                fc_tmp = fluid.layers.fc(image, size=CLASS_NUM, act='softmax')
                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

            image = fluid.data(name='image', shape=[None, IMAGE_SIZE], dtype='float32')
            label = fluid.data(name='label', shape=[None, 1], dtype='int64')

            loss = simple_net(image, label)

            exe = fluid.Executor(places[0])
            exe.run(fluid.default_startup_program())

            prog = fluid.CompiledProgram(fluid.default_main_program()).with_data_parallel(loss_name=loss.name)

            dataset = RandomDataset(BATCH_NUM * BATCH_SIZE)

            loader = DataLoader(dataset,
                                feed_list=[image, label],
                                places=places,
                                batch_size=BATCH_SIZE, 
                                shuffle=True,
                                drop_last=True,
                                num_workers=2)

            for e in range(EPOCH_NUM):
                for i, data in enumerate(loader()):
                    l = exe.run(prog, feed=data, fetch_list=[loss], return_numpy=True)
                    print("Epoch {} batch {}: loss = {}".format(e, i, l[0][0]))

            # -------------------------------------------------------
                
            # --------------------- dygraph mode --------------------

            class SimpleNet(fluid.dygraph.Layer):
                def __init__(self):
                    super(SimpleNet, self).__init__()
                    self.fc = fluid.dygraph.nn.Linear(IMAGE_SIZE, CLASS_NUM, act='softmax')

                def forward(self, image, label=None):
                    return self.fc(image)

            with fluid.dygraph.guard(places[0]):
                simple_net = SimpleNet()
                opt = fluid.optimizer.SGD(learning_rate=1e-3,
                                          parameter_list=simple_net.parameters())

                loader = DataLoader(dataset,
                                    places=places[0],
                                    batch_size=BATCH_SIZE,
                                    shuffle=True,
                                    drop_last=True,
                                    num_workers=2)

                for e in range(EPOCH_NUM):
                    for i, (image, label) in enumerate(loader()):
                        out = simple_net(image)
                        loss = fluid.layers.cross_entropy(out, label)
                        avg_loss = fluid.layers.reduce_mean(loss)
                        avg_loss.backward()
                        opt.minimize(avg_loss)
                        simple_net.clear_gradients()
                        print("Epoch {} batch {}: loss = {}".format(e, i, np.mean(loss.numpy())))

            # -------------------------------------------------------

    """

    def __init__(self,
                 dataset,
                 feed_list=None,
                 places=None,
                 return_list=False,
                 batch_sampler=None,
                 batch_size=1,
                 shuffle=False,
                 drop_last=False,
                 collate_fn=None,
                 num_workers=0,
                 use_buffer_reader=True,
                 use_shared_memory=True,
                 timeout=0,
                 worker_init_fn=None):
        self.return_list = return_list
        self.collate_fn = collate_fn
        self.use_buffer_reader = use_buffer_reader
        self.worker_init_fn = worker_init_fn

        assert isinstance(dataset, Dataset), \
            "dataset should be subclass instance of paddle.io.Dataset"
        self.dataset = dataset

        if not return_list and not in_dygraph_mode():
            assert feed_list is not None, \
                    "feed_list should be set when return_list=False"
        self.feed_list = feed_list

        assert places is not None, "places cannot be None"
        self.places = _convert_places(places)
        if in_dygraph_mode():
            assert len(self.places) == 1, \
                    "Number of places must be 1 in dygraph mode"

        assert num_workers >= 0, "num_workers should be a non-negative value"
        if num_workers > 0 and (sys.platform == 'darwin' or
                                sys.platform == 'win32'):
            logging.warning(
                "multi-process mode not support MacOs and Windows currently." \
                " use signle-process with num_workers = 0 instead")
            num_workers = 0
        self.num_workers = num_workers

        self.use_shared_memory = use_shared_memory
        if use_shared_memory and num_workers == 0:
            self.use_shared_memory = False

        assert timeout >= 0, "timeout should be a non-negative value"
        self.timeout = timeout

        if batch_sampler is not None:
            assert isinstance(batch_sampler, BatchSampler), \
                "batch_sampler should be None or subclass instance " \
                "of paddle.io.BatchSampler"
            assert batch_size == 1 and not shuffle and not drop_last, \
                "batch_size/shuffle/drop_last should not be set when " \
                "batch_sampler is given"
            self.batch_sampler = batch_sampler
        else:
            assert batch_size is not None and batch_size > 0, \
                "batch_size should be a positive value when " \
                "batch_sampler is not given"
            self.batch_sampler = BatchSampler(
                dataset=dataset,
                batch_size=batch_size,
                shuffle=shuffle,
                drop_last=drop_last)

    def __len__(self):
        return len(self.batch_sampler)

    def __iter__(self):
        if self.num_workers == 0:
            return _DataLoaderIterSingleProcess(self)
        else:
            return _DataLoaderIterMultiProcess(self)

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

Z
Zeng Jinle 已提交
359 360 361 362 363
    @staticmethod
    def from_generator(feed_list=None,
                       capacity=None,
                       use_double_buffer=True,
                       iterable=True,
364
                       return_list=False,
365 366
                       use_multiprocess=False,
                       drop_last=True):
Z
Zeng Jinle 已提交
367
        """
368 369 370
        .. note::
          **The framework ensures that the data loading order of DataLoader is exactly the same as the user-defined data source.**

Z
Zeng Jinle 已提交
371 372 373 374 375 376 377 378
        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.
379
        
Z
Zeng Jinle 已提交
380 381 382 383 384 385 386 387 388 389 390 391
        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.
392
                The variables should be created by :code:`fluid.data()`.
Z
Zeng Jinle 已提交
393 394 395 396 397 398 399 400 401 402 403 404 405
            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 
T
tianshuo78520a 已提交
406
                the name of each fed variables. If return_list=True, the 
Z
Zeng Jinle 已提交
407 408
                return value on each device would be a list(LoDTensor). It is
                recommended to use return_list=False in static graph mode and
409 410 411 412 413 414
                use return_list=True in dygraph mode.  
            use_multiprocess (bool): whether to use multi-process to speed up
                the data loading process in dygraph. Note: this parameter only
                can be used in the dygraph mode. In the static graph mode,
                whether this parameter is set or not has no effect.
                The Default value is False.
415 416 417 418 419 420 421
            drop_last (bool): whether to drop the last batches whose number is
                less than the CPU core/GPU card number. The default value is 
                True. In training phase, users should not set drop_last=False,
                because all CPU cores/GPU cards must read data from DataLoader. 
                In inference phase, users can set drop_last=False, so that the
                last batches whose number is less than the CPU core/GPU card
                number can be tested. 
Z
Zeng Jinle 已提交
422 423 424 425

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

426
        Examples 1:
Z
Zeng Jinle 已提交
427 428
            
            .. code-block:: python
S
sneaxiy 已提交
429

Z
Zeng Jinle 已提交
430 431
                import paddle.fluid as fluid
                import numpy as np
432

Z
Zeng Jinle 已提交
433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487
                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
488

Z
Zeng Jinle 已提交
489
                    return __reader__
490

Z
Zeng Jinle 已提交
491 492 493 494 495
                # 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])
496

Z
Zeng Jinle 已提交
497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515
                # 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')
516

517 518
                image = fluid.data(name='image', shape=[None, 784], dtype='float32')
                label = fluid.data(name='label', shape=[None, 1], dtype='int64')
519

Z
Zeng Jinle 已提交
520 521
                # Define DataLoader 
                loader = fluid.io.DataLoader.from_generator(feed_list=[image, label], capacity=16, iterable=ITERABLE)
522

Z
Zeng Jinle 已提交
523 524
                # Define network
                loss = simple_net(image, label)
S
sneaxiy 已提交
525

Z
Zeng Jinle 已提交
526 527 528 529 530 531 532 533 534
                # 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
535

Z
Zeng Jinle 已提交
536 537
                exe = fluid.Executor(places[0])
                exe.run(fluid.default_startup_program())
H
Huihuang Zheng 已提交
538

Z
Zeng Jinle 已提交
539
                prog = fluid.CompiledProgram(fluid.default_main_program()).with_data_parallel(loss_name=loss.name)
540

Z
Zeng Jinle 已提交
541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557
                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]
558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600

        Examples 2:

            .. code-block:: python

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

                # We use 2 CPU cores to run inference network 
                os.environ['CPU_NUM'] = '2'

                # The data source has only 3 batches, which can not be
                # divided evenly to each CPU core
                def batch_generator():  
                    for i in range(3):
                        yield np.array([i+1]).astype('float32'), 

                x = fluid.data(name='x', shape=[None], dtype='float32')  
                y = x * x

                def run_inference(drop_last): 
                    loader = fluid.io.DataLoader.from_generator(feed_list=[x],
                            capacity=8, drop_last=drop_last)
                    loader.set_batch_generator(batch_generator, fluid.cpu_places())

                    exe = fluid.Executor(fluid.CPUPlace())
                    prog = fluid.CompiledProgram(fluid.default_main_program())
                    prog = prog.with_data_parallel()

                    result = []
                    for data in loader():
                        each_ret, = exe.run(prog, feed=data, fetch_list=[y])
                        result.extend(each_ret)
                    return result

                # Set drop_last to True, so that the last batch whose
                # number is less than CPU core number would be discarded.
                print(run_inference(drop_last=True)) # [1.0, 4.0]

                # Set drop_last to False, so that the last batch whose
                # number is less than CPU core number can be tested.
                print(run_inference(drop_last=False)) # [1.0, 4.0, 9.0]
Z
Zeng Jinle 已提交
601
        """
602 603 604 605 606 607
        if in_dygraph_mode():
            return DygraphGeneratorLoader(feed_list, capacity,
                                          use_double_buffer, iterable,
                                          return_list, use_multiprocess)
        else:
            return GeneratorLoader(feed_list, capacity, use_double_buffer,
608
                                   iterable, return_list, drop_last)
Z
Zeng Jinle 已提交
609 610 611 612 613 614

    @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.
615

Z
Zeng Jinle 已提交
616 617 618 619 620 621 622
        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. 
623

Z
Zeng Jinle 已提交
624 625 626
        Returns:
            loader (DataLoader): the created DataLoader object, which can be 
                treated as a Python generator.   
627

Z
Zeng Jinle 已提交
628 629 630
        Examples:

            .. code-block:: python
631

Z
Zeng Jinle 已提交
632
                import paddle.fluid as fluid
633

634 635
                image = fluid.data(name='image', shape=[None, 784], dtype='float32')
                label = fluid.data(name='label', shape=[None, 1], dtype='int64')
636

Z
Zeng Jinle 已提交
637 638 639 640 641
                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') 
642

Z
Zeng Jinle 已提交
643 644 645
                loader = fluid.io.DataLoader.from_dataset(dataset, fluid.cpu_places())
        """
        return DatasetLoader(dataset, places, drop_last)
S
sneaxiy 已提交
646

S
sneaxiy 已提交
647

648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673
class DygraphGeneratorLoader(DataLoaderBase):
    """
    The GeneratorLoader of dygraph

    The multiprocess dygraph GeneratorLoader's most functions are different from 
    static graph GeneratorLoader, Separate implementation to keep code readable.
    """

    def __init__(self,
                 feed_list=None,
                 capacity=None,
                 use_double_buffer=True,
                 iterable=True,
                 return_list=True,
                 use_multiprocess=False):
        self._batch_reader = None
        self._places = None
        self._feed_list = feed_list

        if not capacity:
            raise ValueError("Please give value to capacity.")
        self._capacity = capacity
        self._use_double_buffer = use_double_buffer

        if not iterable:
            logging.warning(
674
                "Please NOTE: imperative mode can support iterable mode only. Change to iterable mode."
675 676 677 678
            )
        self._iterable = True
        if not return_list:
            logging.warning(
679
                "Please NOTE: imperative mode can support return as list only. Change to return as list."
680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713
            )
        self._return_list = True

        # NOTE: the multiprocessing in different platform is incompatible, we will solve it later
        self._use_multiprocess = use_multiprocess
        if self._use_multiprocess and (sys.platform == 'darwin' or
                                       sys.platform == 'win32'):
            logging.warning(
                "NOTE: The multiprocess mode does not currently support MacOs and Windows."
            )
            self._use_multiprocess = False

        if self._use_multiprocess:
            # NOTE: the multiprocessing.Queue used to save loading data in self._process
            self._data_queue = None
            # NOTE: this process is used to load data asynchronously from self._batch_reader
            self._process = None

        # NOTE: the C++ LoDTensorBlockingQueue instance
        self._blocking_queue = None
        # NOTE: 1. In multiprocess mode, this thread is used to get next batch data from
        # self._data_queue, then push it into self._blocking_queue; 2. In singleprocess
        # mode, this thread is used to get next batch data from self._batch_reader, then 
        # push it into self._blocking_queue
        self._thread = None

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

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

714 715 716 717 718 719 720 721 722 723
    def _clear_and_remove_data_queue(self):
        if self._data_queue is not None:
            while True:
                try:
                    self._data_queue.get_nowait()
                except queue.Empty:
                    break
            global multiprocess_queue_set
            multiprocess_queue_set.remove(self._data_queue)

724 725 726 727 728 729 730 731 732 733 734
    def _wait_thread_ends(self):
        thread = self._thread
        if thread is not None:
            self._blocking_queue.close()
            thread.join()

    def _wait_process_ends(self):
        process = self._process
        if process is not None:
            process.join()
            # erase process id
735
            core._erase_process_pids(id(self))
736

737 738 739 740 741 742 743 744 745
    def _init_iterable(self):
        self._wait_thread_ends()
        if self._use_multiprocess:
            self._wait_process_ends()
        self._var_names = []
        self._shapes = []
        self._dtypes = []
        self._need_check_feed = []
        self._blocking_queue = core.init_lod_tensor_blocking_queue(
746
            core.Variable(), self._capacity, False)
747
        self._reader = None
748 749
        self._reader = core.create_py_reader(
            self.queue, self._var_names, self._shapes, self._dtypes,
750
            self._need_check_feed, self._places, self._use_double_buffer, True)
751 752 753

    def _start(self):
        if self._use_multiprocess:
754 755 756
            # clear old _data_queue and remove it from multiprocess_queue_set
            self._clear_and_remove_data_queue()
            # set data_queue and process
757
            self._data_queue = multiprocessing.Queue(self._capacity)
758 759 760
            # add _data_queue into global queue set
            global multiprocess_queue_set
            multiprocess_queue_set.add(self._data_queue)
761 762 763 764 765 766 767 768 769 770 771
            self._process = multiprocessing.Process(
                target=self._reader_process_loop)
            self._process.daemon = True
            self._process.start()

            # Set child process signal handler
            # NOTE: [ avoiding hang ] 1. if the child process dies due to bus error/segfault
            # or just hang, the main process will hang waiting for data, so here need to deal 
            # with SIGSEGV and SIGBUS of child process; 2. if the main process end before child
            # process, it shuts the all its daemonic children down with a SIGTERM (instead of 
            # joining them without a timeout), so here nedd to deal with SIGTERM.
772 773
            core._set_process_pids(id(self), [self._process.pid])
            _set_SIGCHLD_handler()
774 775 776 777

            # Set reader_thread
            self._thread_done_event = threading.Event()
            self._thread = threading.Thread(
778
                target=self._reader_thread_loop_for_multiprocess)
779 780 781
            self._thread.daemon = True
            self._thread.start()
        else:
782 783
            self._thread = threading.Thread(
                target=self._reader_thread_loop_for_singleprocess)
784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819
            self._thread.daemon = True
            self._thread.start()

    def _reset(self):
        self._reader.reset()
        self._wait_thread_ends()
        if self._use_multiprocess:
            self._wait_process_ends()

    def __iter__(self):
        assert self.iterable, "DataLoader is not iterable"
        assert self._batch_reader is not None, \
            "Data source of DataLoader has not set yet"

        self._init_iterable()
        self._start()
        return self

    def __next__(self):
        try:
            return self._reader.read_next_var_list()
        except StopIteration:
            self._reset()
            six.reraise(*sys.exc_info())

    @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.")

820 821 822 823 824 825 826 827 828
    def _exit_thread_expectedly(self):
        self._thread_done_event.set()
        self._blocking_queue.close()

    def _exit_thread_unexpectedly(self):
        self._thread_done_event.set()
        self._blocking_queue.kill()
        logging.error("DataLoader reader thread raised an exception!")

829 830 831 832 833
    def _reader_process_loop(self):
        try:
            # set signal handler
            core._set_process_signal_handler()

834 835 836 837
            # NOTE: [ mmap files clear ] When the child process exits unexpectedly,
            # some shared memory objects may have been applied for but have not yet
            # been put into the inter-process Queue. This part of the object needs
            # to be cleaned up when the process ends.
838
            CleanupFuncRegistrar.register(_cleanup_mmap)
839 840 841 842 843

            for batch in self._batch_reader():
                tensor_list = core._convert_to_tensor_list(batch)
                self._data_queue.put(tensor_list)
                core._remove_tensor_list_mmap_fds(tensor_list)
844 845 846 847 848 849 850
            self._data_queue.put(None)
        except KeyboardInterrupt:
            # NOTE: Main process will raise KeyboardInterrupt anyways, ignore it in child process
            pass
        except:
            six.reraise(*sys.exc_info())

851
    def _reader_thread_loop_for_multiprocess(self):
852 853 854 855 856 857 858
        while not self._thread_done_event.is_set():
            try:
                # NOTE: [ avoid hanging ] Even with carefully designed data dependencies 
                # (i.e., a put() always corresponding to a get()), hanging on get() can 
                # still happen when data in queue is corrupted (e.g., due to 
                # Queue.cancel_join_thread or unexpected exit). So we set a timeout whenever 
                # we try to get data from `data_queue`
859 860 861 862 863 864 865
                # NOTE: [ avoid failed quickly ] Here, the time setting of QUEUE_GET_TIMEOUT
                # is relatively long, currently it is 60 seconds, because in some models,
                # if the reader child process starts with a heavy burden, the child process
                # has no enough time to put the data in the queue when the main process
                # start trying to get data from queue. At this time, the child thread needs
                # to wait slightly longer
                tensor_list = self._data_queue.get(timeout=QUEUE_GET_TIMEOUT)
866 867 868 869
            except:
                # NOTE [ avoid handing ] After adding the shared memory mechanism, not only
                # the queue.Empty exception will occur here, but other exceptions will also
                # occur, such as mmap failure. If it is not handled here, it will hang.
870
                self._exit_thread_unexpectedly()
871 872
                logging.error(
                    "DataLoader reader thread failed to read data from the multiprocessing.Queue."
873
                )
874
                six.reraise(*sys.exc_info())
875 876

            if not self._thread_done_event.is_set():
877
                if tensor_list is not None:
878 879
                    try:
                        array = core.LoDTensorArray()
880 881
                        for tensor in tensor_list:
                            array.append(tensor)
882 883 884
                        if not self._blocking_queue.push(array):
                            self._blocking_queue.close()
                    except:
885
                        self._exit_thread_unexpectedly()
886 887
                        six.reraise(*sys.exc_info())
                else:
888
                    self._exit_thread_expectedly()
889

890
    def _reader_thread_loop_for_singleprocess(self):
891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943
        try:
            for sample in self._batch_reader():
                array = core.LoDTensorArray()
                for item in sample:
                    if not isinstance(item, core.LoDTensor):
                        self._check_input_array(item)
                        tmp = core.LoDTensor()
                        tmp.set(item, core.CPUPlace())
                        item = tmp

                    array.append(item)

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

            self._blocking_queue.close()
            self._thread = None
        except Exception:
            self._blocking_queue.kill()
            self._thread = None
            logging.warning(
                "DygraphDataLoader reader thread raised an exception.")
            six.reraise(*sys.exc_info())

    def set_sample_generator(self,
                             reader,
                             batch_size,
                             drop_last=True,
                             places=None):
        assert batch_size > 0, "batch_size must be larger than 0"
        self.set_sample_list_generator(
            paddle.batch(
                reader, batch_size=batch_size, drop_last=drop_last),
            places=places)
        return self

    def set_sample_list_generator(self, reader, places=None):
        def __batch_reader_impl__():
            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

        self.set_batch_generator(__batch_reader_impl__, places)
        return self

    def set_batch_generator(self, reader, places=None):
        self._batch_reader = reader
944 945
        if places is None:
            places = _current_expected_place()
946 947
        self._places = _convert_places(places)
        assert len(self._places) == 1, \
948
            "Number of places must be 1 in imperative mode"
949 950 951
        return self


Z
Zeng Jinle 已提交
952
class GeneratorLoader(DataLoaderBase):
S
sneaxiy 已提交
953
    def __init__(self,
954 955
                 feed_list=None,
                 capacity=None,
S
sneaxiy 已提交
956
                 use_double_buffer=True,
957
                 iterable=True,
958 959
                 return_list=False,
                 drop_last=True):
S
sneaxiy 已提交
960
        self._tensor_reader = None
Z
Zeng Jinle 已提交
961
        self._places = None
S
sneaxiy 已提交
962
        self._thread = None
963
        self._queue = None
964
        self._feed_list = feed_list
965 966 967
        self._exited = False
        self._drop_last = drop_last
        self._keep_order = keep_data_loader_order()
968 969
        if not capacity:
            raise ValueError("Please give value to capacity.")
970 971 972 973
        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
974 975 976 977
        self._use_double_buffer = use_double_buffer
        self._capacity = capacity
        if not self._iterable:
            self._init_non_iterable()
S
sneaxiy 已提交
978

Z
Zeng Jinle 已提交
979
    def _wait_thread_ends(self):
980
        # Get self._thread first to prevent data race, because __thread_main__
Z
Zeng Jinle 已提交
981 982 983 984 985 986 987 988
        # 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()
989 990 991 992 993 994
        self._var_names = [v.name for v in self._feed_list]
        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
        ]
995 996
        self._queue = core.init_lod_tensor_blocking_queue(
            core.Variable(), self._capacity, self._keep_order)
997
        self._reader = None
S
sneaxiy 已提交
998
        self._reader = core.create_py_reader(
999
            self.queue, self._var_names, self._shapes, self._dtypes,
1000 1001
            self._need_check_feed, self._places, self._use_double_buffer,
            self._drop_last)
S
sneaxiy 已提交
1002 1003 1004 1005 1006 1007 1008

    def _init_non_iterable(self):
        lod_levels = []
        dtypes = []
        shape_concat = []
        ranks = []
        shapes = []
1009
        need_check_feed = []
S
sneaxiy 已提交
1010 1011 1012 1013 1014 1015 1016

        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)
1017
            need_check_feed.append(int(feed_data.desc.need_check_feed()))
S
sneaxiy 已提交
1018

Z
Zeng Jinle 已提交
1019 1020 1021 1022
        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
1023

S
sneaxiy 已提交
1024
        var = global_scope().var(queue_name)
1025 1026 1027 1028 1029 1030 1031
        self._queue = core.init_lod_tensor_blocking_queue(var, self._capacity,
                                                          self._keep_order)

        if self._keep_order:
            block = default_main_program().current_block()
        else:
            block = default_startup_program().current_block()
S
sneaxiy 已提交
1032

1033
        reader_var = block.create_var(name=reader_name)
S
sneaxiy 已提交
1034

1035
        dtype_int = [int(t) for t in dtypes]
1036
        block.append_op(
S
sneaxiy 已提交
1037 1038
            type='create_py_reader',
            inputs={'blocking_queue': [queue_name]},
1039
            outputs={'Out': [reader_var]},
S
sneaxiy 已提交
1040 1041 1042
            attrs={
                'shape_concat': shape_concat,
                'lod_levels': lod_levels,
1043 1044
                'dtypes': dtype_int,
                'need_check_feed': need_check_feed,
S
sneaxiy 已提交
1045 1046 1047
                'ranks': ranks
            })

1048 1049 1050
        reader_var.desc.set_dtypes(dtypes)
        reader_var.persistable = True
        reader_var.stop_gradient = True
S
sneaxiy 已提交
1051

1052 1053 1054 1055 1056 1057 1058 1059 1060 1061
        if self._keep_order:
            main_prog_var = reader_var
            reader = main_prog_var
            reader.reset = self._queue.reset
        else:
            main_prog_var = _copy_reader_var_(
                default_main_program().current_block(), reader_var)

            main_prog_var.stop_gradient = True
            main_prog_var.persistable = True
S
sneaxiy 已提交
1062

1063
            reader = monkey_patch_reader_methods(main_prog_var)
S
sneaxiy 已提交
1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077

        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]},
1078 1079
            outputs={'Out': self._feed_list},
            attrs={'drop_last': self._drop_last})
S
sneaxiy 已提交
1080 1081 1082 1083 1084 1085 1086 1087

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

    @property
    def iterable(self):
        return self._iterable
S
sneaxiy 已提交
1088

Z
Zeng Jinle 已提交
1089 1090
    def __iter__(self):
        assert self.iterable, "DataLoader is not iterable"
S
sneaxiy 已提交
1091
        assert self._tensor_reader is not None, \
Z
Zeng Jinle 已提交
1092
            "Data source of DataLoader has not set yet"
S
sneaxiy 已提交
1093

Z
Zeng Jinle 已提交
1094
        self._init_iterable()
S
sneaxiy 已提交
1095
        self._start()
Z
Zeng Jinle 已提交
1096 1097 1098 1099
        return self

    def __next__(self):
        try:
1100 1101
            if self._return_list:
                return self._reader.read_next_list()
1102
            else:
1103
                return self._reader.read_next()
Z
Zeng Jinle 已提交
1104 1105 1106 1107 1108 1109
        except StopIteration:
            self._queue.close()
            self._reset()
            six.reraise(*sys.exc_info())

    def start(self):
1110 1111
        assert not self._iterable, "start() cannot be called when DataLoader is iterable"
        self._start()
Z
Zeng Jinle 已提交
1112 1113

    def reset(self):
1114 1115
        assert not self._iterable, "reset() cannot be called when DataLoader is iterable"
        self._reset()
Z
Zeng Jinle 已提交
1116

1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127
    @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."))

1128 1129
        return arr

Z
Zeng Jinle 已提交
1130 1131 1132
    def _start(self):
        def __thread_main__():
            try:
1133 1134 1135 1136
                while not self._queue.wait_for_inited(1):
                    if self._exited:
                        return

Z
Zeng Jinle 已提交
1137 1138 1139 1140
                for tensors in self._tensor_reader():
                    array = core.LoDTensorArray()
                    for item in tensors:
                        if not isinstance(item, core.LoDTensor):
1141
                            item = self._check_input_array(item)
Z
Zeng Jinle 已提交
1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153
                            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
1154
                self._queue.kill()
Z
Zeng Jinle 已提交
1155 1156 1157 1158 1159 1160 1161
                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
1162

S
sneaxiy 已提交
1163
    def _reset(self):
1164
        self._queue.close()
1165
        self._exited = True
Z
Zeng Jinle 已提交
1166 1167 1168 1169
        thread = self._thread
        if thread is not None:
            thread.join()

1170
        self._exited = False
1171 1172
        self._reader.reset()

Z
Zeng Jinle 已提交
1173 1174 1175 1176 1177 1178
    def set_sample_generator(self,
                             reader,
                             batch_size,
                             drop_last=True,
                             places=None):
        assert batch_size > 0, "batch_size must be larger than 0"
1179 1180 1181 1182 1183 1184 1185
        has_lod = False
        for f in self._feed_list:
            if f.lod_level != 0:
                has_lod = True
                break

        if has_lod:
1186 1187 1188 1189 1190
            self.set_sample_list_generator(
                paddle.batch(
                    reader, batch_size=batch_size, drop_last=drop_last),
                places=places)
        else:
1191 1192 1193 1194 1195 1196 1197
            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)
Z
Zeng Jinle 已提交
1198 1199 1200
        return self

    def set_sample_list_generator(self, reader, places=None):
1201 1202 1203 1204
        with program_guard(Program(), Program()):
            feeder = DataFeeder(
                feed_list=self._feed_list, place=core.CPUPlace())
            paddle_reader = feeder.decorate_reader(reader, multi_devices=False)
Z
Zeng Jinle 已提交
1205

1206 1207 1208
        def __tensor_reader_impl__():
            for slots in paddle_reader():
                yield [slots[var.name] for var in self._feed_list]
Z
Zeng Jinle 已提交
1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247

        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)
        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 
T
tianshuo78520a 已提交
1248
            the name of each fed variables. If return_list=True, the 
Z
Zeng Jinle 已提交
1249 1250 1251 1252 1253
            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
1254 1255 1256 1257
        the created reader object.

    Return type:
        reader(Reader)
Z
Zeng Jinle 已提交
1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276

    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
1277 1278 1279 1280 1281
           
           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
1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292

           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
1293 1294
           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
1295 1296 1297 1298 1299 1300 1301 1302

           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
1303 1304
           loss = network(image, label)
           executor = fluid.Executor(fluid.CPUPlace())
Z
Zeng Jinle 已提交
1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331
           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
1332 1333 1334 1335 1336
           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
1337 1338 1339
           def reader_creator_random_image(height, width):
               def reader():
                   for i in range(ITER_NUM):
G
guofei 已提交
1340 1341 1342
                       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
1343 1344
               return reader

G
guofei 已提交
1345 1346 1347
           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
1348 1349 1350 1351

           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
1352 1353 1354 1355 1356 1357
                   fluid.core.CPUPlace())
           
           loss = network(image, label)
           executor = fluid.Executor(fluid.CPUPlace())
           executor.run(fluid.default_startup_program())
           
Z
Zeng Jinle 已提交
1358 1359
           for _ in range(EPOCH_NUM):
               for data in reader():
G
guofei 已提交
1360
                   executor.run(feed=data, fetch_list=[loss])
Z
Zeng Jinle 已提交
1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414


        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
1415 1416

    def start(self):
S
add doc  
sneaxiy 已提交
1417 1418 1419
        '''
        Start the data feeding thread. 
        Can only call when the reader object is not iterable.  
1420
        
G
guofei 已提交
1421 1422
	Example:
	    .. code-block:: python
Z
Zeng Jinle 已提交
1423
    
H
Huihuang Zheng 已提交
1424 1425 1426 1427
                import paddle
                import paddle.fluid as fluid
                import numpy as np

1428 1429 1430 1431 1432 1433
                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
1434
                image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
1435 1436 1437 1438
                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
1439
                executor = fluid.Executor(fluid.CPUPlace())
1440 1441 1442 1443 1444 1445 1446 1447 1448 1449
                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
1450 1451
	    '''
        self._loader.start()
S
sneaxiy 已提交
1452

S
sneaxiy 已提交
1453
    def reset(self):
S
add doc  
sneaxiy 已提交
1454 1455 1456
        '''
        Reset the reader object when :code:`fluid.core.EOFException` raises. 
        Can only call when the reader object is not iterable.
1457 1458 1459 1460
        
        Example:
            .. code-block:: python

H
Huihuang Zheng 已提交
1461 1462 1463 1464
                import paddle
                import paddle.fluid as fluid
                import numpy as np

1465 1466 1467 1468 1469 1470
                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
1471
                image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
1472 1473 1474 1475
                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
1476
                executor = fluid.Executor(fluid.CPUPlace())
1477 1478 1479 1480 1481 1482 1483 1484 1485 1486
                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
1487
        '''
Z
Zeng Jinle 已提交
1488
        self._loader.reset()
S
sneaxiy 已提交
1489

S
sneaxiy 已提交
1490 1491 1492 1493 1494 1495 1496 1497 1498
    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,
1499
        which yields list(numpy.ndarray)-typed data of each sample.
S
sneaxiy 已提交
1500 1501 1502 1503

        :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
1504
        :code:`decorate_sample_list_generator(paddle.batch(sample_generator, ...))` .
S
sneaxiy 已提交
1505 1506 1507

        Args:
            sample_generator (generator): Python generator that yields
1508
                list(numpy.ndarray)-typed sample data.
S
sneaxiy 已提交
1509 1510 1511 1512 1513
            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.
1514 1515 1516 1517

        Example:
            .. code-block:: python

H
Huihuang Zheng 已提交
1518 1519 1520
                import paddle.fluid as fluid
                import numpy as np

1521 1522 1523
                EPOCH_NUM = 3
                ITER_NUM = 15
                BATCH_SIZE = 3
G
guofei 已提交
1524 1525 1526 1527 1528
        
                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)
1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539

                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
1540 1541
                image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
                label = fluid.data(name='label', shape=[None, 1], dtype='int64')
1542 1543 1544 1545 1546
                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
1547 1548 1549 1550
                                                 places=[fluid.CPUPlace()])
                loss = network(image, label)
                executor = fluid.Executor(fluid.CPUPlace())
                executor.run(fluid.default_startup_program())
1551 1552 1553

                for _ in range(EPOCH_NUM):
                    for data in reader():
G
guofei 已提交
1554
                        executor.run(feed=data, fetch_list=[loss])
1555
    
S
sneaxiy 已提交
1556
        '''
Z
Zeng Jinle 已提交
1557 1558
        self._loader.set_sample_generator(sample_generator, batch_size,
                                          drop_last, places)
S
sneaxiy 已提交
1559

S
sneaxiy 已提交
1560
    def decorate_sample_list_generator(self, reader, places=None):
S
add doc  
sneaxiy 已提交
1561 1562 1563 1564
        '''
        Set the data source of the PyReader object. 

        The provided :code:`reader` should be a Python generator,
S
sneaxiy 已提交
1565
        which yields list(numpy.ndarray) typed batched data. 
S
add doc  
sneaxiy 已提交
1566 1567 1568 1569
        
        :code:`places` must be set when the PyReader object is iterable.

        Args:
S
sneaxiy 已提交
1570 1571 1572 1573
            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.
1574 1575 1576 1577
        
        Example:
            .. code-block:: python

H
Huihuang Zheng 已提交
1578 1579 1580 1581
                import paddle
                import paddle.fluid as fluid
                import numpy as np

1582 1583 1584 1585
                EPOCH_NUM = 3
                ITER_NUM = 15
                BATCH_SIZE = 3

G
guofei 已提交
1586 1587 1588 1589 1590
                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)

1591 1592 1593 1594 1595 1596 1597 1598 1599 1600
                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
1601 1602
                image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
                label = fluid.data(name='label', shape=[None, 1], dtype='int64')
1603 1604 1605 1606 1607
                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
1608 1609 1610 1611 1612
                    fluid.core.CPUPlace())
                
                loss = network(image, label)
                executor = fluid.Executor(fluid.core.CPUPlace())
                executor.run(fluid.default_startup_program())
1613 1614 1615

                for _ in range(EPOCH_NUM):
                    for data in reader():
G
guofei 已提交
1616
                        executor.run(feed=data, fetch_list=[loss])
1617
                 
S
add doc  
sneaxiy 已提交
1618
        '''
Z
Zeng Jinle 已提交
1619
        self._loader.set_sample_list_generator(reader, places)
S
sneaxiy 已提交
1620

S
sneaxiy 已提交
1621
    def decorate_batch_generator(self, reader, places=None):
S
add doc  
sneaxiy 已提交
1622 1623 1624 1625
        '''
        Set the data source of the PyReader object.

        The provided :code:`reader` should be a Python generator,
S
sneaxiy 已提交
1626
        which yields numpy.ndarray-typed or LoDTensor-typed batched data.
S
add doc  
sneaxiy 已提交
1627 1628 1629 1630 1631 1632

        :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
1633
            places (None|list(CUDAPlace)|list(CPUPlace)): place list. Must
S
sneaxiy 已提交
1634
                be provided when PyReader is iterable.
1635 1636 1637 1638

        Example:
            .. code-block:: python

H
Huihuang Zheng 已提交
1639 1640 1641
                import paddle.fluid as fluid
                import numpy as np

1642 1643 1644
                EPOCH_NUM = 3
                ITER_NUM = 15
                BATCH_SIZE = 3
G
guofei 已提交
1645 1646 1647 1648 1649
               
                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)
1650 1651 1652 1653 1654 1655 1656 1657

                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
1658 1659
                            batch_image = batch_image.astype('float32')
                            batch_label = batch_label.astype('int64')
1660 1661 1662
                            yield batch_image, batch_label
                    return generator

G
guofei 已提交
1663 1664
                image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
                label = fluid.data(name='label', shape=[None, 1], dtype='int64')
1665 1666 1667
                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
1668 1669 1670 1671 1672
                reader.decorate_batch_generator(user_defined_generator, fluid.CPUPlace())
                
                loss = network(image, label)
                executor = fluid.Executor(fluid.CPUPlace())
                executor.run(fluid.default_startup_program())
1673 1674 1675

                for _ in range(EPOCH_NUM):
                    for data in reader():
G
guofei 已提交
1676
                        executor.run(feed=data, fetch_list=[loss])
1677

S
add doc  
sneaxiy 已提交
1678
        '''
Z
Zeng Jinle 已提交
1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723
        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()