reader.py 54.2 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
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
Z
Zeng Jinle 已提交
21
from .framework import Program, Variable, program_guard, default_main_program, default_startup_program, in_dygraph_mode, cpu_places
S
sneaxiy 已提交
22
from .executor import global_scope
23
from .data_feeder import DataFeeder, BatchedTensorProvider
S
sneaxiy 已提交
24
from .layers.io import monkey_patch_reader_methods, _copy_reader_var_, double_buffer
S
sneaxiy 已提交
25
from .unique_name import UniqueNameGenerator
26
import logging
Z
Zeng Jinle 已提交
27
from .dataset import DatasetBase, InMemoryDataset
S
sneaxiy 已提交
28

29 30 31 32 33 34 35 36
### Dygraph DataLoader configs ###
import multiprocessing
import signal
# NOTE: queue has a different name in python2 and python3
if sys.version_info[0] == 2:
    import Queue as queue
else:
    import queue
37 38 39
# NOTE: [ avoid hanging ] These value is used in getting data from another process
QUEUE_GET_TIMEOUT = 5
MAX_GET_FAILED_TIME = 12
40

Z
Zeng Jinle 已提交
41 42 43
__all__ = ['PyReader', 'DataLoader']

data_loader_unique_name_generator = UniqueNameGenerator()
S
sneaxiy 已提交
44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


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
61 62 63
class DataLoaderBase(object):
    def __init__(self):
        self._places = None
S
sneaxiy 已提交
64

Z
Zeng Jinle 已提交
65 66
    def __call__(self):
        return self
S
sneaxiy 已提交
67

Z
Zeng Jinle 已提交
68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89
    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,
90 91
                       return_list=False,
                       use_multiprocess=False):
Z
Zeng Jinle 已提交
92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113
        """
        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.
114
                The variables should be created by :code:`fluid.data()`.
Z
Zeng Jinle 已提交
115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130
            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
131 132 133 134 135 136
                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.
Z
Zeng Jinle 已提交
137 138 139 140 141 142 143

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

        Examples:
            
            .. code-block:: python
S
sneaxiy 已提交
144

Z
Zeng Jinle 已提交
145 146
                import paddle.fluid as fluid
                import numpy as np
147

Z
Zeng Jinle 已提交
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
                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
203

Z
Zeng Jinle 已提交
204
                    return __reader__
205

Z
Zeng Jinle 已提交
206 207 208 209 210
                # 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])
211

Z
Zeng Jinle 已提交
212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230
                # 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')
231

232 233
                image = fluid.data(name='image', shape=[None, 784], dtype='float32')
                label = fluid.data(name='label', shape=[None, 1], dtype='int64')
234

Z
Zeng Jinle 已提交
235 236
                # Define DataLoader 
                loader = fluid.io.DataLoader.from_generator(feed_list=[image, label], capacity=16, iterable=ITERABLE)
237

Z
Zeng Jinle 已提交
238 239
                # Define network
                loss = simple_net(image, label)
S
sneaxiy 已提交
240

Z
Zeng Jinle 已提交
241 242 243 244 245 246 247 248 249
                # 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
250

Z
Zeng Jinle 已提交
251 252
                exe = fluid.Executor(places[0])
                exe.run(fluid.default_startup_program())
H
Huihuang Zheng 已提交
253

Z
Zeng Jinle 已提交
254
                prog = fluid.CompiledProgram(fluid.default_main_program()).with_data_parallel(loss_name=loss.name)
255

Z
Zeng Jinle 已提交
256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273
                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]
        """
274 275 276 277 278 279 280
        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,
                                   iterable, return_list)
Z
Zeng Jinle 已提交
281 282 283 284 285 286

    @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.
287

Z
Zeng Jinle 已提交
288 289 290 291 292 293 294
        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. 
295

Z
Zeng Jinle 已提交
296 297 298
        Returns:
            loader (DataLoader): the created DataLoader object, which can be 
                treated as a Python generator.   
299

Z
Zeng Jinle 已提交
300 301 302
        Examples:

            .. code-block:: python
303

Z
Zeng Jinle 已提交
304
                import paddle.fluid as fluid
305

306 307
                image = fluid.data(name='image', shape=[None, 784], dtype='float32')
                label = fluid.data(name='label', shape=[None, 1], dtype='int64')
308

Z
Zeng Jinle 已提交
309 310 311 312 313
                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') 
314

Z
Zeng Jinle 已提交
315 316 317
                loader = fluid.io.DataLoader.from_dataset(dataset, fluid.cpu_places())
        """
        return DatasetLoader(dataset, places, drop_last)
S
sneaxiy 已提交
318

S
sneaxiy 已提交
319

320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488
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(
                "Please NOTE: dygraph can support iterable mode only. Change to iterable mode."
            )
        self._iterable = True
        if not return_list:
            logging.warning(
                "Please NOTE: dygraph can support return as list only. Change to return as list."
            )
        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

    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:
            self._data_queue.cancel_join_thread()
            self._data_queue.close()
            process.join()
            # erase process id
            core._erase_process_pid(id(self))

    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(
            core.Variable(), self._capacity)
        self._reader = core.create_py_reader(
            self.queue, self._var_names, self._shapes, self._dtypes,
            self._need_check_feed, self._places, self._use_double_buffer)

    def _start(self):
        if self._use_multiprocess:
            # Set data_queue and process
            self._data_queue = multiprocessing.Queue(self._capacity)
            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.
            self._set_child_signal_handler()

            # Set reader_thread
            self._thread_done_event = threading.Event()
            self._thread = threading.Thread(
                target=self._reader_thread_loop_with_process)
            self._thread.daemon = True
            self._thread.start()
        else:
            self._thread = threading.Thread(target=self._reader_thread_loop)
            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.")

    def _set_child_signal_handler(self):
        core._set_process_pid(id(self), self._process.pid)
        current_handler = signal.getsignal(signal.SIGCHLD)
        if not callable(current_handler):
            current_handler = None

        def __handler__(signum, frame):
            core._throw_error_if_process_failed()
            if current_handler is not None:
                current_handler(signum, frame)

        signal.signal(signal.SIGCHLD, __handler__)

489 490 491 492 493 494 495 496 497 498 499
    def _exit_thread_expectedly(self):
        self._thread_done_event.set()
        self._blocking_queue.close()
        self._data_queue.close()

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

500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520
    def _reader_process_loop(self):
        try:
            # set signal handler
            core._set_process_signal_handler()

            for sample in self._batch_reader():
                if sample is None:
                    raise ValueError(
                        "Sample in reader is None. Please check whether your dataset is valid."
                    )
                self._data_queue.put(sample)
            self._data_queue.put(None)
        except KeyboardInterrupt:
            # NOTE: Main process will raise KeyboardInterrupt anyways, ignore it in child process
            pass
        except:
            self._data_queue.cancel_join_thread()
            self._data_queue.close()
            six.reraise(*sys.exc_info())

    def _reader_thread_loop_with_process(self):
521
        get_sample_try_time = 0
522 523 524 525 526 527 528
        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`
529 530
                sample = self._data_queue.get(timeout=QUEUE_GET_TIMEOUT)
                get_sample_try_time = 0
531
            except queue.Empty:
532 533 534 535 536 537 538 539 540 541 542 543
                get_sample_try_time += 1
                if get_sample_try_time > MAX_GET_FAILED_TIME:
                    self._exit_thread_unexpectedly()
                    raise RuntimeError(
                        "DataLoader reader thread has not read data for a long time (60s)."
                    )
                else:
                    # NOTE: [ avoid failed quickly ] Sometimes if the reader child process has 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, failure to read data should not be
                    # counted as a fatal error, there should be a certain number of attempts.
                    continue
544 545 546 547 548 549 550 551 552 553 554 555 556 557 558

            if not self._thread_done_event.is_set():
                if sample is not None:
                    try:
                        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):
                            self._blocking_queue.close()
                    except:
559
                        self._exit_thread_unexpectedly()
560 561
                        six.reraise(*sys.exc_info())
                else:
562
                    self._exit_thread_expectedly()
563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624

    def _reader_thread_loop(self):
        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
        assert places is not None, "Places cannot be None when DataLoader is iterable"
        self._places = _convert_places(places)
        assert len(self._places) == 1, \
            "Number of places must be 1 in dygraph mode"
        return self


Z
Zeng Jinle 已提交
625
class GeneratorLoader(DataLoaderBase):
S
sneaxiy 已提交
626
    def __init__(self,
627 628
                 feed_list=None,
                 capacity=None,
S
sneaxiy 已提交
629
                 use_double_buffer=True,
630 631
                 iterable=True,
                 return_list=False):
S
sneaxiy 已提交
632
        self._tensor_reader = None
Z
Zeng Jinle 已提交
633
        self._places = None
S
sneaxiy 已提交
634
        self._thread = None
635
        self._queue = None
636 637 638
        self._feed_list = feed_list
        if not capacity:
            raise ValueError("Please give value to capacity.")
639 640 641 642
        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
643 644 645 646
        self._use_double_buffer = use_double_buffer
        self._capacity = capacity
        if not self._iterable:
            self._init_non_iterable()
S
sneaxiy 已提交
647

Z
Zeng Jinle 已提交
648
    def _wait_thread_ends(self):
649
        # Get self._thread first to prevent data race, because __thread_main__
Z
Zeng Jinle 已提交
650 651 652 653 654 655 656 657
        # 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()
658 659 660 661 662 663
        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
        ]
S
sneaxiy 已提交
664 665 666
        self._queue = core.init_lod_tensor_blocking_queue(core.Variable(),
                                                          self._capacity)
        self._reader = core.create_py_reader(
667 668
            self.queue, self._var_names, self._shapes, self._dtypes,
            self._need_check_feed, self._places, self._use_double_buffer)
S
sneaxiy 已提交
669 670 671 672 673 674 675

    def _init_non_iterable(self):
        lod_levels = []
        dtypes = []
        shape_concat = []
        ranks = []
        shapes = []
676
        need_check_feed = []
S
sneaxiy 已提交
677 678 679 680 681 682 683

        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)
684
            need_check_feed.append(int(feed_data.desc.need_check_feed()))
S
sneaxiy 已提交
685

Z
Zeng Jinle 已提交
686 687 688 689
        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
690

S
sneaxiy 已提交
691
        var = global_scope().var(queue_name)
S
sneaxiy 已提交
692 693 694 695 696
        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)

697
        dtype_int = [int(t) for t in dtypes]
S
sneaxiy 已提交
698 699 700 701 702 703 704
        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,
705 706
                'dtypes': dtype_int,
                'need_check_feed': need_check_feed,
S
sneaxiy 已提交
707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741
                '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
742

Z
Zeng Jinle 已提交
743 744
    def __iter__(self):
        assert self.iterable, "DataLoader is not iterable"
S
sneaxiy 已提交
745
        assert self._tensor_reader is not None, \
Z
Zeng Jinle 已提交
746
            "Data source of DataLoader has not set yet"
S
sneaxiy 已提交
747

Z
Zeng Jinle 已提交
748
        self._init_iterable()
S
sneaxiy 已提交
749
        self._start()
Z
Zeng Jinle 已提交
750 751 752 753
        return self

    def __next__(self):
        try:
754 755
            if self._return_list:
                return self._reader.read_next_list()
756
            else:
757
                return self._reader.read_next()
Z
Zeng Jinle 已提交
758 759 760 761 762 763
        except StopIteration:
            self._queue.close()
            self._reset()
            six.reraise(*sys.exc_info())

    def start(self):
764 765
        assert not self._iterable, "start() cannot be called when DataLoader is iterable"
        self._start()
Z
Zeng Jinle 已提交
766 767

    def reset(self):
768 769
        assert not self._iterable, "reset() cannot be called when DataLoader is iterable"
        self._reset()
Z
Zeng Jinle 已提交
770

771 772 773 774 775 776 777 778 779 780 781
    @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
782 783 784 785 786 787 788
    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):
789
                            self._check_input_array(item)
Z
Zeng Jinle 已提交
790 791 792 793 794 795 796 797 798 799 800 801
                            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
802
                self._queue.kill()
Z
Zeng Jinle 已提交
803 804 805 806 807 808 809
                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
810

S
sneaxiy 已提交
811
    def _reset(self):
812
        self._queue.close()
Z
Zeng Jinle 已提交
813 814 815 816
        thread = self._thread
        if thread is not None:
            thread.join()

817 818
        self._reader.reset()

Z
Zeng Jinle 已提交
819 820 821 822 823 824
    def set_sample_generator(self,
                             reader,
                             batch_size,
                             drop_last=True,
                             places=None):
        assert batch_size > 0, "batch_size must be larger than 0"
825 826 827 828 829 830 831
        has_lod = False
        for f in self._feed_list:
            if f.lod_level != 0:
                has_lod = True
                break

        if has_lod:
832 833 834 835 836
            self.set_sample_list_generator(
                paddle.batch(
                    reader, batch_size=batch_size, drop_last=drop_last),
                places=places)
        else:
837 838 839 840 841 842 843
            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
844 845 846
        return self

    def set_sample_list_generator(self, reader, places=None):
847 848 849 850
        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
851

852 853 854
        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
855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899

        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 
            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
900 901 902 903
        the created reader object.

    Return type:
        reader(Reader)
Z
Zeng Jinle 已提交
904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922

    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
923 924 925 926 927
           
           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
928 929 930 931 932 933 934 935 936 937 938

           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
939 940
           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
941 942 943 944 945 946 947 948

           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
949 950
           loss = network(image, label)
           executor = fluid.Executor(fluid.CPUPlace())
Z
Zeng Jinle 已提交
951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977
           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
978 979 980 981 982
           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
983 984 985
           def reader_creator_random_image(height, width):
               def reader():
                   for i in range(ITER_NUM):
G
guofei 已提交
986 987 988
                       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
989 990
               return reader

G
guofei 已提交
991 992 993
           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
994 995 996 997

           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
998 999 1000 1001 1002 1003
                   fluid.core.CPUPlace())
           
           loss = network(image, label)
           executor = fluid.Executor(fluid.CPUPlace())
           executor.run(fluid.default_startup_program())
           
Z
Zeng Jinle 已提交
1004 1005
           for _ in range(EPOCH_NUM):
               for data in reader():
G
guofei 已提交
1006
                   executor.run(feed=data, fetch_list=[loss])
Z
Zeng Jinle 已提交
1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060


        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
1061 1062

    def start(self):
S
add doc  
sneaxiy 已提交
1063 1064 1065
        '''
        Start the data feeding thread. 
        Can only call when the reader object is not iterable.  
1066
        
G
guofei 已提交
1067 1068
	Example:
	    .. code-block:: python
Z
Zeng Jinle 已提交
1069
    
H
Huihuang Zheng 已提交
1070 1071 1072 1073
                import paddle
                import paddle.fluid as fluid
                import numpy as np

1074 1075 1076 1077 1078 1079
                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
1080
                image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
1081 1082 1083 1084
                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
1085
                executor = fluid.Executor(fluid.CPUPlace())
1086 1087 1088 1089 1090 1091 1092 1093 1094 1095
                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
1096 1097
	    '''
        self._loader.start()
S
sneaxiy 已提交
1098

S
sneaxiy 已提交
1099
    def reset(self):
S
add doc  
sneaxiy 已提交
1100 1101 1102
        '''
        Reset the reader object when :code:`fluid.core.EOFException` raises. 
        Can only call when the reader object is not iterable.
1103 1104 1105 1106
        
        Example:
            .. code-block:: python

H
Huihuang Zheng 已提交
1107 1108 1109 1110
                import paddle
                import paddle.fluid as fluid
                import numpy as np

1111 1112 1113 1114 1115 1116
                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
1117
                image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
1118 1119 1120 1121
                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
1122
                executor = fluid.Executor(fluid.CPUPlace())
1123 1124 1125 1126 1127 1128 1129 1130 1131 1132
                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
1133
        '''
Z
Zeng Jinle 已提交
1134
        self._loader.reset()
S
sneaxiy 已提交
1135

S
sneaxiy 已提交
1136 1137 1138 1139 1140 1141 1142 1143 1144
    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,
1145
        which yields list(numpy.ndarray)-typed data of each sample.
S
sneaxiy 已提交
1146 1147 1148 1149

        :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
1150
        :code:`decorate_sample_list_generator(paddle.batch(sample_generator, ...))` .
S
sneaxiy 已提交
1151 1152 1153

        Args:
            sample_generator (generator): Python generator that yields
1154
                list(numpy.ndarray)-typed sample data.
S
sneaxiy 已提交
1155 1156 1157 1158 1159
            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.
1160 1161 1162 1163

        Example:
            .. code-block:: python

H
Huihuang Zheng 已提交
1164 1165 1166
                import paddle.fluid as fluid
                import numpy as np

1167 1168 1169
                EPOCH_NUM = 3
                ITER_NUM = 15
                BATCH_SIZE = 3
G
guofei 已提交
1170 1171 1172 1173 1174
        
                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)
1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185

                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
1186 1187
                image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
                label = fluid.data(name='label', shape=[None, 1], dtype='int64')
1188 1189 1190 1191 1192
                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
1193 1194 1195 1196
                                                 places=[fluid.CPUPlace()])
                loss = network(image, label)
                executor = fluid.Executor(fluid.CPUPlace())
                executor.run(fluid.default_startup_program())
1197 1198 1199

                for _ in range(EPOCH_NUM):
                    for data in reader():
G
guofei 已提交
1200
                        executor.run(feed=data, fetch_list=[loss])
1201
    
S
sneaxiy 已提交
1202
        '''
Z
Zeng Jinle 已提交
1203 1204
        self._loader.set_sample_generator(sample_generator, batch_size,
                                          drop_last, places)
S
sneaxiy 已提交
1205

S
sneaxiy 已提交
1206
    def decorate_sample_list_generator(self, reader, places=None):
S
add doc  
sneaxiy 已提交
1207 1208 1209 1210
        '''
        Set the data source of the PyReader object. 

        The provided :code:`reader` should be a Python generator,
S
sneaxiy 已提交
1211
        which yields list(numpy.ndarray) typed batched data. 
S
add doc  
sneaxiy 已提交
1212 1213 1214 1215
        
        :code:`places` must be set when the PyReader object is iterable.

        Args:
S
sneaxiy 已提交
1216 1217 1218 1219
            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.
1220 1221 1222 1223
        
        Example:
            .. code-block:: python

H
Huihuang Zheng 已提交
1224 1225 1226 1227
                import paddle
                import paddle.fluid as fluid
                import numpy as np

1228 1229 1230 1231
                EPOCH_NUM = 3
                ITER_NUM = 15
                BATCH_SIZE = 3

G
guofei 已提交
1232 1233 1234 1235 1236
                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)

1237 1238 1239 1240 1241 1242 1243 1244 1245 1246
                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
1247 1248
                image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
                label = fluid.data(name='label', shape=[None, 1], dtype='int64')
1249 1250 1251 1252 1253
                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
1254 1255 1256 1257 1258
                    fluid.core.CPUPlace())
                
                loss = network(image, label)
                executor = fluid.Executor(fluid.core.CPUPlace())
                executor.run(fluid.default_startup_program())
1259 1260 1261

                for _ in range(EPOCH_NUM):
                    for data in reader():
G
guofei 已提交
1262
                        executor.run(feed=data, fetch_list=[loss])
1263
                 
S
add doc  
sneaxiy 已提交
1264
        '''
Z
Zeng Jinle 已提交
1265
        self._loader.set_sample_list_generator(reader, places)
S
sneaxiy 已提交
1266

S
sneaxiy 已提交
1267
    def decorate_batch_generator(self, reader, places=None):
S
add doc  
sneaxiy 已提交
1268 1269 1270 1271
        '''
        Set the data source of the PyReader object.

        The provided :code:`reader` should be a Python generator,
S
sneaxiy 已提交
1272
        which yields numpy.ndarray-typed or LoDTensor-typed batched data.
S
add doc  
sneaxiy 已提交
1273 1274 1275 1276 1277 1278

        :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
1279
            places (None|list(CUDAPlace)|list(CPUPlace)): place list. Must
S
sneaxiy 已提交
1280
                be provided when PyReader is iterable.
1281 1282 1283 1284

        Example:
            .. code-block:: python

H
Huihuang Zheng 已提交
1285 1286 1287
                import paddle.fluid as fluid
                import numpy as np

1288 1289 1290
                EPOCH_NUM = 3
                ITER_NUM = 15
                BATCH_SIZE = 3
G
guofei 已提交
1291 1292 1293 1294 1295
               
                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)
1296 1297 1298 1299 1300 1301 1302 1303

                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
1304 1305
                            batch_image = batch_image.astype('float32')
                            batch_label = batch_label.astype('int64')
1306 1307 1308
                            yield batch_image, batch_label
                    return generator

G
guofei 已提交
1309 1310
                image = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
                label = fluid.data(name='label', shape=[None, 1], dtype='int64')
1311 1312 1313
                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
1314 1315 1316 1317 1318
                reader.decorate_batch_generator(user_defined_generator, fluid.CPUPlace())
                
                loss = network(image, label)
                executor = fluid.Executor(fluid.CPUPlace())
                executor.run(fluid.default_startup_program())
1319 1320 1321

                for _ in range(EPOCH_NUM):
                    for data in reader():
G
guofei 已提交
1322
                        executor.run(feed=data, fetch_list=[loss])
1323

S
add doc  
sneaxiy 已提交
1324
        '''
Z
Zeng Jinle 已提交
1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369
        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()