python/mxnet/gluon/data/dataloader.py - master - githubmirror/incubator-mxnet

python/mxnet/gluon/data/dataloader.py
816 строк34 KB
Przemyslaw Tredak
WAR the dataloader issue with forked processes holding stale references (#19925)
21 фев 2021, 06:46
21 фев 2021, 06:46fee8bf4
# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements.  See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership.  The ASF licenses this file
# to you 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.

# coding: utf-8
# pylint: disable=ungrouped-imports
"""Dataset generator."""
__all__ = ['DataLoader']

import pickle
import logging
import io
import sys
import signal
import multiprocessing
import multiprocessing.queues
from multiprocessing.reduction import ForkingPickler
from multiprocessing.pool import ThreadPool
import threading
import numpy as np

try:
    import multiprocessing.resource_sharer
except ImportError:
    pass

from . import sampler as _sampler
from . import batchify as _batchify
from ... import ndarray as nd, context
from ...util import is_np_shape, is_np_array, set_np
from ... import numpy as _mx_np  # pylint: disable=reimported

if sys.platform == 'darwin' or sys.platform == 'win32':
    def rebuild_ndarray(*args):
        """Rebuild ndarray from pickled shared memory"""
        # pylint: disable=no-value-for-parameter
        return nd.NDArray(nd.ndarray._new_from_shared_mem(*args))

    def reduce_ndarray(data):
        """Reduce ndarray to shared memory handle"""
        return rebuild_ndarray, data._to_shared_mem()
else:
    def rebuild_ndarray(pid, fd, shape, dtype):
        """Rebuild ndarray from pickled shared memory"""
        # pylint: disable=no-value-for-parameter
        fd = fd.detach()
        return nd.NDArray(nd.ndarray._new_from_shared_mem(pid, fd, shape, dtype))

    def reduce_ndarray(data):
        """Reduce ndarray to shared memory handle"""
        # keep a local ref before duplicating fd
        data = data.as_in_context(context.Context('cpu_shared', 0))
        pid, fd, shape, dtype = data._to_shared_mem()
        fd = multiprocessing.reduction.DupFd(fd)
        return rebuild_ndarray, (pid, fd, shape, dtype)

ForkingPickler.register(nd.NDArray, reduce_ndarray)

if sys.platform == 'darwin' or sys.platform == 'win32':
    def rebuild_np_ndarray(*args):
        """Rebuild ndarray from pickled shared memory"""
        # pylint: disable=no-value-for-parameter
        return _mx_np.ndarray(nd.ndarray._new_from_shared_mem(*args))

    def reduce_np_ndarray(data):
        """Reduce ndarray to shared memory handle"""
        return rebuild_np_ndarray, data._to_shared_mem()
else:
    def rebuild_np_ndarray(pid, fd, shape, dtype):
        """Rebuild ndarray from pickled shared memory"""
        # pylint: disable=no-value-for-parameter
        fd = fd.detach()
        return _mx_np.ndarray(nd.ndarray._new_from_shared_mem(pid, fd, shape, dtype))

    def reduce_np_ndarray(data):
        """Reduce ndarray to shared memory handle"""
        # keep a local ref before duplicating fd
        data = data.as_in_context(context.Context('cpu_shared', 0))
        pid, fd, shape, dtype = data._to_shared_mem()
        fd = multiprocessing.reduction.DupFd(fd)
        return rebuild_np_ndarray, (pid, fd, shape, dtype)

ForkingPickler.register(_mx_np.ndarray, reduce_np_ndarray)


class ConnectionWrapper(object):
    """Connection wrapper for multiprocessing that supports sending
    NDArray via shared memory."""

    def __init__(self, conn):
        self._conn = conn

    def send(self, obj):
        """Send object"""
        buf = io.BytesIO()
        ForkingPickler(buf, pickle.HIGHEST_PROTOCOL).dump(obj)
        self.send_bytes(buf.getvalue())

    def recv(self):
        """Receive object"""
        buf = self.recv_bytes()
        return pickle.loads(buf)

    def __getattr__(self, name):
        """Emmulate conn"""
        attr = self.__dict__.get('_conn', None)
        return getattr(attr, name)


class Queue(multiprocessing.queues.Queue):
    """Wrapper for multiprocessing queue that dumps NDArray with shared memory."""
    def __init__(self, *args, **kwargs):
        super().__init__(*args, ctx=multiprocessing.get_context(), **kwargs)
        self._reader = ConnectionWrapper(self._reader)
        self._writer = ConnectionWrapper(self._writer)
        self._send = self._writer.send
        self._recv = self._reader.recv


class SimpleQueue(multiprocessing.queues.SimpleQueue):
    """Wrapper for multiprocessing SimpleQueue that dumps NDArray with shared memory.
       SimpleQueue don't use threading internally.
    """
    def __init__(self, *args, **kwargs):
        super().__init__(*args, ctx=multiprocessing.get_context(), **kwargs)
        self._reader = ConnectionWrapper(self._reader)
        self._writer = ConnectionWrapper(self._writer)
        self._send = self._writer.send
        self._recv = self._reader.recv

def default_batchify_fn(data):
    """Collate data into batch."""
    if isinstance(data[0], nd.NDArray):
        return _mx_np.stack(data) if is_np_array() else nd.stack(*data)
    elif isinstance(data[0], tuple):
        data = zip(*data)
        return [default_batchify_fn(i) for i in data]
    else:
        data = np.asarray(data)
        array_fn = _mx_np.array if is_np_array() else nd.array
        return array_fn(data, dtype=data.dtype)


def default_mp_batchify_fn(data):
    """Collate data into batch. Use shared memory for stacking."""
    if isinstance(data[0], nd.NDArray):
        empty_fn = _mx_np.empty if is_np_array() else nd.empty
        out = empty_fn((len(data),) + data[0].shape, dtype=data[0].dtype,
                       ctx=context.Context('cpu_shared', 0))
        if is_np_array():
            return _mx_np.stack(data, out=out)
        else:
            return nd.stack(*data, out=out)
    elif isinstance(data[0], tuple):
        data = zip(*data)
        return [default_mp_batchify_fn(i) for i in data]
    else:
        data = np.asarray(data)
        array_fn = _mx_np.array if is_np_array() else nd.array
        return array_fn(data, dtype=data.dtype,
                        ctx=context.Context('cpu_shared', 0))


def _as_in_context(data, ctx):
    """Move data into new context."""
    if isinstance(data, nd.NDArray):
        return data.as_in_context(ctx)
    elif isinstance(data, (list, tuple)):
        return [_as_in_context(d, ctx) for d in data]
    return data


def worker_loop_v1(dataset, key_queue, data_queue, batchify_fn):
    """Worker loop for multiprocessing DataLoader."""
    while True:
        idx, samples = key_queue.get()
        if idx is None:
            break
        batch = batchify_fn([dataset[i] for i in samples])
        data_queue.put((idx, batch))

def fetcher_loop_v1(data_queue, data_buffer, pin_memory=False,
                    pin_device_id=0, data_buffer_lock=None):
    """Fetcher loop for fetching data from queue and put in reorder dict."""
    while True:
        idx, batch = data_queue.get()
        if idx is None:
            break
        if pin_memory:
            batch = _as_in_context(batch, context.cpu_pinned(pin_device_id))
        else:
            batch = _as_in_context(batch, context.cpu())
        if data_buffer_lock is not None:
            with data_buffer_lock:
                data_buffer[idx] = batch
        else:
            data_buffer[idx] = batch


class _MultiWorkerIterV1(object):
    """Internal multi-worker iterator for DataLoader."""
    def __init__(self, num_workers, dataset, batchify_fn, batch_sampler,
                 pin_memory=False, pin_device_id=0, worker_fn=worker_loop_v1):
        assert num_workers > 0, "_MultiWorkerIter is not for {} workers".format(num_workers)
        self._num_workers = num_workers
        self._dataset = dataset
        self._batchify_fn = batchify_fn
        self._batch_sampler = batch_sampler
        self._key_queue = Queue()
        self._data_queue = SimpleQueue()

        self._data_buffer = {}
        self._data_buffer_lock = threading.Lock()

        self._rcvd_idx = 0
        self._sent_idx = 0
        self._iter = iter(self._batch_sampler)
        self._shutdown = False

        workers = []
        for _ in range(self._num_workers):
            worker = multiprocessing.Process(
                target=worker_fn,
                args=(self._dataset, self._key_queue, self._data_queue, self._batchify_fn))
            worker.daemon = True
            worker.start()
            workers.append(worker)
        self._workers = workers

        self._fetcher = threading.Thread(
            target=fetcher_loop_v1,
            args=(self._data_queue, self._data_buffer, pin_memory,
                  pin_device_id, self._data_buffer_lock))
        self._fetcher.daemon = True
        self._fetcher.start()

        # pre-fetch
        for _ in range(2 * self._num_workers):
            self._push_next()

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

    def __del__(self):
        self.shutdown()

    def _push_next(self):
        """Assign next batch workload to workers."""
        r = next(self._iter, None)
        if r is None:
            return
        self._key_queue.put((self._sent_idx, r))
        self._sent_idx += 1

    def __next__(self):
        assert not self._shutdown, "call __next__ after shutdown is forbidden"
        if self._rcvd_idx == self._sent_idx:
            assert not self._data_buffer, "Data buffer should be empty at this moment"
            self.shutdown()
            raise StopIteration

        while True:
            if self._rcvd_idx in self._data_buffer:
                with self._data_buffer_lock:
                    batch = self._data_buffer.pop(self._rcvd_idx)
                self._rcvd_idx += 1
                self._push_next()
                return batch

    def next(self):
        return self.__next__()

    def __iter__(self):
        return self

    def shutdown(self):
        """Shutdown internal workers by pushing terminate signals."""
        if not self._shutdown:
            # send shutdown signal to the fetcher and join data queue first
            # Remark:   loop_fetcher need to be joined prior to the workers.
            #           otherwise, the fetcher may fail at getting data
            self._data_queue.put((None, None))
            self._fetcher.join()
            # send shutdown signal to all worker processes
            for _ in range(self._num_workers):
                self._key_queue.put((None, None))
            # force shut down any alive worker processes
            for w in self._workers:
                if w.is_alive():
                    w.terminate()
            self._shutdown = True


class DataLoaderV1(object):
    """Loads data from a dataset and returns mini-batches of data.

    Parameters
    ----------
    dataset : Dataset
        Source dataset. Note that numpy and mxnet arrays can be directly used
        as a Dataset.
    batch_size : int
        Size of mini-batch.
    shuffle : bool
        Whether to shuffle the samples.
    sampler : Sampler
        The sampler to use. Either specify sampler or shuffle, not both.
    last_batch : {'keep', 'discard', 'rollover'}
        How to handle the last batch if batch_size does not evenly divide
        `len(dataset)`:
        - ``keep`` - A batch with less samples than previous batches is returned.
        - ``discard`` - The last batch is discarded if its incomplete.
        - ``rollover`` - The remaining samples are rolled over to the next epoch.
    batch_sampler : Sampler
        A sampler that returns mini-batches. Do not specify batch_size,
        shuffle, sampler, and last_batch if batch_sampler is specified.
    batchify_fn : callable
        Callback function to allow users to specify how to merge samples
        into a batch. Defaults to ``default_batchify_fn``.

        .. code-block:: python

            def default_batchify_fn(data):
                if isinstance(data[0], nd.NDArray):
                    return nd.stack(*data)
                elif isinstance(data[0], tuple):
                    data = zip(*data)
                    return [default_batchify_fn(i) for i in data]
                else:
                    data = np.asarray(data)
                    return nd.array(data, dtype=data.dtype)

    num_workers : int, default 0
        The number of multiprocessing workers to use for data preprocessing.
    pin_memory : boolean, default False
        If ``True``, the dataloader will copy NDArrays into pinned memory
        before returning them. Copying from CPU pinned memory to GPU is faster
        than from normal CPU memory.
    pin_device_id : int, default 0
        The device id to use for allocating pinned memory if pin_memory is ``True``
    """
    def __init__(self, dataset, batch_size=None, shuffle=False, sampler=None,
                 last_batch=None, batch_sampler=None, batchify_fn=None,
                 num_workers=0, pin_memory=False, pin_device_id=0):
        self._dataset = dataset
        self._pin_memory = pin_memory
        self._pin_device_id = pin_device_id

        if batch_sampler is None:
            if batch_size is None:
                raise ValueError("batch_size must be specified unless " \
                                 "batch_sampler is specified")
            if sampler is None:
                if shuffle:
                    sampler = _sampler.RandomSampler(len(dataset))
                else:
                    sampler = _sampler.SequentialSampler(len(dataset))
            elif shuffle:
                raise ValueError("shuffle must not be specified if sampler is specified")

            batch_sampler = _sampler.BatchSampler(
                sampler, batch_size, last_batch if last_batch else 'keep')
        elif batch_size is not None or shuffle or sampler is not None or \
                last_batch is not None:
            raise ValueError("batch_size, shuffle, sampler and last_batch must " \
                             "not be specified if batch_sampler is specified.")

        self._batch_sampler = batch_sampler
        self._num_workers = num_workers if num_workers >= 0 else 0
        if batchify_fn is None:
            if num_workers > 0:
                self._batchify_fn = _batchify.Stack(use_shared_mem=True)
            else:
                self._batchify_fn = _batchify.Stack()
        else:
            self._batchify_fn = batchify_fn

    def __iter__(self):
        if self._num_workers == 0:
            def same_process_iter():
                for batch in self._batch_sampler:
                    ret = self._batchify_fn([self._dataset[idx] for idx in batch])
                    if self._pin_memory:
                        ret = _as_in_context(ret, context.cpu_pinned(self._pin_device_id))
                    yield ret
            return same_process_iter()

        # multi-worker
        return _MultiWorkerIterV1(self._num_workers, self._dataset,
                                  self._batchify_fn, self._batch_sampler,
                                  self._pin_memory, self._pin_device_id)

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


def _thread_worker_initializer(active_shape, active_array):
    """Initializer for ThreadPool."""
    set_np(shape=active_shape, array=active_array)


_worker_dataset = None
def _worker_initializer(dataset, active_shape, active_array):
    """Initialier for processing pool."""
    # global dataset is per-process based and only available in worker processes
    # this is only necessary to handle MXIndexedRecordIO because otherwise dataset
    # can be passed as argument
    global _worker_dataset
    _worker_dataset = dataset
    set_np(shape=active_shape, array=active_array)

def _worker_fn(samples, batchify_fn, dataset=None):
    """Function for processing data in worker process."""
    # pylint: disable=unused-argument
    # it is required that each worker process has to fork a new MXIndexedRecordIO handle
    # preserving dataset as global variable can save tons of overhead and is safe in new process
    global _worker_dataset
    batch = batchify_fn([_worker_dataset[i] for i in samples])
    buf = io.BytesIO()
    ForkingPickler(buf, pickle.HIGHEST_PROTOCOL).dump(batch)
    return buf.getvalue()

def _thread_worker_fn(samples, batchify_fn, dataset):
    """Threadpool worker function for processing data."""
    return batchify_fn([dataset[i] for i in samples])

class _MultiWorkerIter(object):
    """Internal multi-worker iterator for DataLoader."""
    def __init__(self, worker_pool, batchify_fn, batch_sampler, pin_memory=False,
                 pin_device_id=0, worker_fn=_worker_fn, prefetch=0, dataset=None,
                 data_loader=None, timeout=120):
        self._worker_pool = worker_pool
        self._batchify_fn = batchify_fn
        self._batch_sampler = batch_sampler
        self._data_buffer = {}
        self._rcvd_idx = 0
        self._sent_idx = 0
        self._iter = iter(self._batch_sampler)
        self._worker_fn = worker_fn
        self._pin_memory = pin_memory
        self._pin_device_id = pin_device_id
        self._dataset = dataset
        self._data_loader = data_loader
        self._timeout = timeout
        # pre-fetch
        for _ in range(prefetch):
            self._push_next()

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

    def _push_next(self):
        """Assign next batch workload to workers."""
        r = next(self._iter, None)
        if r is None:
            return
        async_ret = self._worker_pool.apply_async(
            self._worker_fn, (r, self._batchify_fn, self._dataset))
        self._data_buffer[self._sent_idx] = async_ret
        self._sent_idx += 1

    def __next__(self):
        self._push_next()
        if self._rcvd_idx == self._sent_idx:
            assert not self._data_buffer, "Data buffer should be empty at this moment"
            raise StopIteration

        assert self._rcvd_idx < self._sent_idx, "rcvd_idx must be smaller than sent_idx"
        assert self._rcvd_idx in self._data_buffer, "fatal error with _push_next, rcvd_idx missing"
        ret = self._data_buffer.pop(self._rcvd_idx)
        try:
            if self._dataset is None:
                batch = pickle.loads(ret.get(self._timeout))
            else:
                batch = ret.get(self._timeout)
            if self._pin_memory:
                batch = _as_in_context(batch, context.cpu_pinned(self._pin_device_id))
            self._rcvd_idx += 1
            return batch
        except multiprocessing.context.TimeoutError:
            msg = '''Worker timed out after {} seconds. This might be caused by \n
            - Slow transform. Please increase timeout to allow slower data loading in each worker.
            '''.format(self._timeout)
            if not isinstance(self._worker_pool, multiprocessing.pool.ThreadPool):
                msg += '''- Insufficient shared_memory if `timeout` is large enough.
            Please consider reduce `num_workers` or increase shared_memory in system.
            '''
            print(msg)
            raise
        except Exception:
            self._worker_pool.terminate()
            raise

    def next(self):
        return self.__next__()

    def __iter__(self):
        return self


class DataLoader(object):
    """Loads data from a dataset and returns mini-batches of data.

    Parameters
    ----------
    dataset : Dataset
        Source dataset. Note that numpy and mxnet arrays can be directly used
        as a Dataset.
    batch_size : int
        Size of mini-batch.
    shuffle : bool
        Whether to shuffle the samples.
    sampler : Sampler
        The sampler to use. Either specify sampler or shuffle, not both.
    last_batch : {'keep', 'discard', 'rollover'}
        How to handle the last batch if batch_size does not evenly divide
        ``len(dataset)``.

        keep - A batch with less samples than previous batches is returned.
        discard - The last batch is discarded if its incomplete.
        rollover - The remaining samples are rolled over to the next epoch.
    batch_sampler : Sampler
        A sampler that returns mini-batches. Do not specify batch_size,
        shuffle, sampler, and last_batch if batch_sampler is specified.
    batchify_fn : callable
        Callback function to allow users to specify how to merge samples
        into a batch. Defaults to `gluon.data.batchify.Stack()`.

        .. code-block:: python

            def default_batchify_fn(data):
                if isinstance(data[0], nd.NDArray):
                    return nd.stack(*data)
                elif isinstance(data[0], np.ndarray):
                    return np.stack(*data)
                elif isinstance(data[0], tuple):
                    data = zip(*data)
                    return [default_batchify_fn(i) for i in data]
                else:
                    data = np.asarray(data)
                    return np.ndarray(data, dtype=data.dtype)

    num_workers : int, default 0
        The number of multiprocessing workers to use for data preprocessing.
    pin_memory : boolean, default False
        If ``True``, the dataloader will copy NDArrays into pinned memory
        before returning them. Copying from CPU pinned memory to GPU is faster
        than from normal CPU memory.
    pin_device_id : int, default 0
        The device id to use for allocating pinned memory if pin_memory is ``True``
    prefetch : int, default is `num_workers * 2`
        The number of prefetching batches only works if `num_workers` > 0.
        If `prefetch` > 0, it allow worker process to prefetch certain batches before
        acquiring data from iterators.
        Note that using large prefetching batch will provide smoother bootstrapping performance,
        but will consume more shared_memory. Using smaller number may forfeit the purpose of using
        multiple worker processes, try reduce `num_workers` in this case.
        By default it defaults to `num_workers * 2`.
    thread_pool : bool, default False
        If ``True``, use threading pool instead of multiprocessing pool. Using threadpool
        can avoid shared memory usage. If `DataLoader` is more IO bounded or GIL is not a killing
        problem, threadpool version may achieve better performance than multiprocessing.
    timeout : int, default is 120
        The timeout in seconds for each worker to fetch a batch data. Only modify this number
        unless you are experiencing timeout and you know it's due to slow data loading.
        Sometimes full `shared_memory` will cause all workers to hang and causes timeout. In these
        cases please reduce `num_workers` or increase system `shared_memory` size instead.
    try_nopython : bool or None, default is None
        Try compile python dataloading pipeline into pure MXNet c++ implementation. The benefit is
        potentially faster iteration, no `shared_memory` usage, and less processes managed by python.
        The compilation is not gauranteed to support all use cases, but it will fallback to python in
        case of failure. You can set `try_nopython` to `False` to disable auto-detection of the
        compilation feature or leave it to `None` to allow MXNet to determine it automatically.
        If you request `try_nopython` to `True` and the compilation fails, it will raise a
        RuntimeError with the failure reason.

    """
    def __init__(self, dataset, batch_size=None, shuffle=False, sampler=None,
                 last_batch=None, batch_sampler=None, batchify_fn=None,
                 num_workers=0, pin_memory=False, pin_device_id=0,
                 prefetch=None, thread_pool=False, timeout=120, try_nopython=None):
        self._dataset = dataset
        self._pin_memory = pin_memory
        self._pin_device_id = pin_device_id
        self._thread_pool = thread_pool
        self._timeout = timeout
        self._mx_iter = None
        assert timeout > 0, "timeout must be positive, given {}".format(timeout)

        if batch_sampler is None:
            if batch_size is None:
                raise ValueError("batch_size must be specified unless " \
                                 "batch_sampler is specified")
            if sampler is None:
                if shuffle:
                    sampler = _sampler.RandomSampler(len(dataset))
                else:
                    sampler = _sampler.SequentialSampler(len(dataset))
            elif shuffle:
                raise ValueError("shuffle must not be specified if sampler is specified")

            batch_sampler = _sampler.BatchSampler(
                sampler, batch_size, last_batch if last_batch else 'keep')
        elif batch_size is not None or shuffle or sampler is not None or \
                last_batch is not None:
            raise ValueError("batch_size, shuffle, sampler and last_batch must " \
                             "not be specified if batch_sampler is specified.")

        self._batch_sampler = batch_sampler
        self._num_workers = num_workers if num_workers >= 0 else 0
        self._worker_pool = None
        self._prefetch = max(0, int(prefetch) if prefetch is not None else 2 * self._num_workers)
        if batchify_fn is None:
            if num_workers > 0:
                self._batchify_fn = _batchify.Stack(use_shared_mem=True)
            else:
                self._batchify_fn = _batchify.Stack()
        else:
            self._batchify_fn = batchify_fn

        if num_workers > 0 and (try_nopython or try_nopython is None):
            # check for capability to use mx backend threadedLoader
            use_mx_iter, mx_iter_args = _check_mx_loader_capability(
                self._dataset, self._batch_sampler, self._batchify_fn)
            if not use_mx_iter:
                if try_nopython:
                    raise RuntimeError(mx_iter_args)
        else:
            use_mx_iter = False

        if use_mx_iter:
            logging.info("Using MXNet backend ThreadedDataLoader with %s workers "
                         "instead of python dataloader.", self._num_workers)
            self._mx_iter = _MXThreadedDataLoader(
                num_workers=self._num_workers,
                pin_memory=self._pin_memory,
                pin_device_id=self._pin_device_id,
                prefetch=self._prefetch, **mx_iter_args)
        else:
            nd.waitall()
            import gc
            gc.collect()
            nd.waitall()
            if self._num_workers > 0:
                if self._thread_pool:
                    self._worker_pool = ThreadPool(self._num_workers,
                                                   initializer=_thread_worker_initializer,
                                                   initargs=(is_np_shape(), is_np_array()))
                else:
                    # set ignore keyboard interupt signal before forking processes
                    original_sigint_handler = signal.signal(signal.SIGINT, signal.SIG_IGN)
                    self._worker_pool = multiprocessing.Pool(
                        self._num_workers, initializer=_worker_initializer,
                        initargs=[self._dataset, is_np_shape(), is_np_array()])
                    # resume keyboard interupt signal in main process
                    signal.signal(signal.SIGINT, original_sigint_handler)

    def __iter__(self):
        if self._mx_iter is not None:
            return iter(self._mx_iter)

        if self._num_workers == 0:
            def same_process_iter():
                for batch in self._batch_sampler:
                    ret = self._batchify_fn([self._dataset[idx] for idx in batch])
                    if self._pin_memory:
                        ret = _as_in_context(ret, context.cpu_pinned(self._pin_device_id))
                    yield ret
            return same_process_iter()

        # multi-worker
        return _MultiWorkerIter(self._worker_pool, self._batchify_fn, self._batch_sampler,
                                pin_memory=self._pin_memory, pin_device_id=self._pin_device_id,
                                worker_fn=_thread_worker_fn if self._thread_pool else _worker_fn,
                                prefetch=self._prefetch,
                                dataset=self._dataset if self._thread_pool else None,
                                data_loader=self, timeout=self._timeout)

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

    def __del__(self):
        if self._worker_pool:
            # manually terminate due to a bug that pool is not automatically terminated
            # https://bugs.python.org/issue34172
            assert isinstance(self._worker_pool, multiprocessing.pool.Pool)
            self._worker_pool.terminate()

def _check_mx_loader_capability(dataset, batch_sampler, batchify_fn):
    from ._internal import MXDataset, MXSampler
    from ._internal import MXBatchifyFunction
    mx_loader_args = {}
    error_template = "MXNet backend loader compatibility: " \
        "[dataset - {}][batchify_fn - {}][batch sampler - {}]"

    # supported dataset
    if isinstance(dataset, MXDataset):
        mx_loader_args['dataset'] = dataset
    elif hasattr(dataset, '__mx_handle__'):
        try:
            mx_loader_args['dataset'] = dataset.__mx_handle__()
        except NotImplementedError:
            return False, error_template.format('fail', 'unknown', 'unknown')
    else:
        return False, error_template.format('fail', 'unknown', 'unknown')

    # supported batchify functions
    if hasattr(batchify_fn, '__mx_handle__'):
        mx_loader_args['batchify_fn'] = batchify_fn.__mx_handle__()
    elif isinstance(batchify_fn, MXBatchifyFunction):
        mx_loader_args['batchify_fn'] = batchify_fn
    else:
        return False, error_template.format('pass', 'fail', 'unknown')

    # supported sampler
    if isinstance(batch_sampler, _sampler.BatchSampler):
        if isinstance(batch_sampler._sampler, _sampler.SequentialSampler):
            mx_loader_args['batch_sampler'] = MXSampler(
                'SequentialSampler', length=batch_sampler._sampler._length,
                start=batch_sampler._sampler._start,
                batch_size=batch_sampler._batch_size,
                last_batch=batch_sampler._last_batch)
        elif isinstance(batch_sampler._sampler, _sampler.RandomSampler):
            mx_loader_args['batch_sampler'] = MXSampler(
                'RandomSampler', length=batch_sampler._sampler._length,
                batch_size=batch_sampler._batch_size,
                last_batch=batch_sampler._last_batch)
        else:
            return False, error_template.format('pass', 'pass', 'fail')
    elif isinstance(batch_sampler, MXSampler):
        mx_loader_args['batch_sampler'] = batch_sampler
    else:
        return False, error_template.format('pass', 'pass', 'fail')
    # all good
    return True, mx_loader_args


class _MXThreadedDataLoader(object):
    """MXNet internal C++ threaded Data Iterator in form of DataLoader

    parameters
    ----------
    dataset : Dataset
        Source dataset. Note that numpy and mxnet arrays can be directly used
        as a Dataset.
    batch_sampler : Sampler
        A sampler that returns mini-batches.
    batchify_fn : callable
        Callback function to allow users to specify how to merge samples
        into a batch. Defaults to `gluon.data.batchify.Stack()`::
    num_workers : int, default 0
        The number of multiprocessing workers to use for data preprocessing.
    pin_memory : boolean, default False
        If ``True``, the dataloader will copy NDArrays into pinned memory
        before returning them. Copying from CPU pinned memory to GPU is faster
        than from normal CPU memory.
    pin_device_id : int, default 0
        The device id to use for allocating pinned memory if pin_memory is ``True``
    prefetch : int, default is `num_workers * 2`
        The number of prefetching batches only works if `num_workers` > 0.
        If `prefetch` > 0, it allow worker process to prefetch certain batches before
        acquiring data from iterators.
        Note that using large prefetching batch will provide smoother bootstrapping performance,
        but will consume more shared_memory. Using smaller number may forfeit the purpose of using
        multiple worker processes, try reduce `num_workers` in this case.
        By default it defaults to `num_workers * 2`, maximum prefetch size is `16`.
    """
    def __init__(self, dataset, batch_sampler, batchify_fn,
                 num_workers=0, pin_memory=False, pin_device_id=0,
                 prefetch=4):
        from ._internal import MXDataset, MXSampler, MXBatchifyFunction
        from ...io.io import ThreadedDataLoader
        assert isinstance(dataset, MXDataset)
        assert isinstance(batch_sampler, MXSampler)
        assert isinstance(batchify_fn, MXBatchifyFunction)
        self._dataset = dataset
        self._batch_sampler = batch_sampler
        self._batchify_fn = batchify_fn
        if num_workers == 0:
            num_workers = 1  # different convention for single thread
        if prefetch == 0:
            prefetch = 1  # at least one buffer required
        pin_device_id = pin_device_id if pin_memory else -1
        ctx = 'cpu_pinned' if pin_memory else 'cpu'
        self._iter = ThreadedDataLoader(num_workers=num_workers, dataset=dataset,
                                        sampler=batch_sampler, batchify_fn=batchify_fn,
                                        prefetch_buffer=prefetch, ctx=ctx,
                                        device_id=pin_device_id)

    def __iter__(self):
        while self._iter.iter_next():
            self._iter.first_batch = None
            items = self._iter.getitems()
            pad = self._iter.getpad()
            if pad > 0:
                items = tuple([x[:-pad] for x in items])
            if len(items) < 2:
                items = items[0]
            yield items
        self._iter.reset()

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