Source code for megengine.data.dataloader

# -*- coding: utf-8 -*-
import collections
import gc
import itertools
import multiprocessing
import os
import pickle
import platform
import queue
import random
import sys
import threading
import time
import traceback
from multiprocessing import Array, Value

import numpy as np

from ..device import _sh, get_default_device
from ..functional.tensor import copy
from ..logger import get_logger
from ..random.rng import _random_seed_generator
from ..tensor import Tensor
from .collator import Collator
from .dataset import Dataset, StreamDataset
from .sampler import MapSampler, Sampler, SequentialSampler, StreamSampler
from .signal_handling import (
    _reset_worker_pids,
    _set_sigchld_handler,
    _set_worker_pids,
    _set_worker_signal_handlers,
)
from .transform import PseudoTransform, Transform

try:
    import thread
except:
    import _thread as thread

logger = get_logger(__name__)


GLOBAL_TIMEOUT = 5

# Use to enable DataLoader monitoring
data_monitor = int(os.environ.get("MGE_DATA_MONITOR", "0"))

# This variable controls the interval of iterations for printing the monitored time. Default: 10.
# Currently only supported when DataLoader num_workers is greater than 0.
data_monitor_frequency = int(os.environ.get("MGE_DATA_MONITOR_FREQUENCY", "10"))

if data_monitor:
    # The number of processes to be monitored, greater than or equal to num_workers of dataloader
    monitor_num_workers = Value("i", 16)

    # Each monitor_workers represents the process id of a dataloader process
    monitor_workers = Array("i", [0] * monitor_num_workers.value)

    # When Parallel, the main process needs to obtain the time to communicate with the child process,
    # and IPC_worker is the process id of the main process
    IPC_worker = Value("i", 0)

    # When Parallel, get the time when each child process put data
    put_time = Array("d", [0.0] * monitor_num_workers.value)

    # The time to get the dataset for each step
    dataset_time = 0.0
    # The time to transform data for each step
    transform_time = 0.0
    # The time to collate data for each step
    collate_time = 0.0


def get_monitor_time_stats(func, *args, **kwargs):
    tik = time.perf_counter()
    ret = func(*args, **kwargs)
    tok = time.perf_counter()
    return ret, tok - tik


def raise_timeout_error():
    raise RuntimeError("dataloader timeout")


[docs]class DataLoader: r"""Data loader. Combines a dataset and a sampler, and provides a convenient way to iterate on a given dataset. The process is as follows: .. mermaid:: :align: center flowchart LR Dataset.__len__ -- Sampler --> Indices batch_size -- Sampler --> Indices Indices -- Dataset.__getitem__ --> Samples Samples -- Transform + Collator --> mini-batch See :ref:`data-guide` for more details. Args: dataset: dataset from which to load the minibatch. sampler: defines the strategy to sample data from the dataset. If ``None``, it will sequentially sample from the dataset one by one. transform: defined the transforming strategy for a sampled batch. collator: defined the merging strategy for a transformed batch. num_workers: the number of sub-process to load, transform and collate the batch. ``0`` means using single-process. Default: 0 timeout: if positive, means the timeout value(second) for collecting a batch from workers. Default: 0 preload: whether to enable the preloading strategy of the dataloader. When enabling, the dataloader will preload one batch to the device memory to speed up the whole training process. parallel_stream: whether to splitting workload across all workers when dataset is streamdataset and num_workers > 0. When enabling, each worker will collect data from different dataset in order to speed up the whole loading process. See ref:`streamdataset-example` for more details Examples: >>> import megengine.data as data >>> dataset = CustomDataset() # doctest: +SKIP >>> dataloader = DataLoader(dataset) # doctest: +SKIP >>> for batch_data in DataLoader: # doctest: +SKIP >>> print(batch_data.shape) # doctest: +SKIP .. admonition:: The effect of enabling preload :class: warning * All elements in :class:`map`, :class:`list`, and :class:`tuple` will be converted to :class:`~.Tensor` by preloading, and you will get :class:`~.Tensor` instead of the original Numpy array or Python built-in data structrure. * Tensors' host2device copy and device kernel execution will be overlapped, which will improve the training speed at the cost of **higher device memory usage** (due to one more batch data on device memory). This feature saves more time when your NN training time is short or your machine's host PCIe bandwidth for each device is low. """ def __init__( self, dataset: Dataset, sampler: Sampler = None, transform: Transform = None, collator: Collator = None, num_workers: int = 0, timeout: int = 0, preload: bool = False, parallel_stream: bool = False, ): if num_workers < 0: raise ValueError("num_workers should not be negative") if timeout < 0: raise ValueError("timeout should not be negative") self.dataset = dataset self.num_workers = num_workers self.timeout = timeout self.preload = preload self.parallel_stream = parallel_stream if isinstance(dataset, StreamDataset): self.sampler = sampler if sampler else StreamSampler(batch_size=1) assert isinstance( self.sampler, StreamSampler ), "types of dataset and sampler do not match" if parallel_stream is False and self.num_workers > 1: logger.warning( "Data time will be affected by getting origin-data, please set parallel_stream in order to speed up dataloader!" ) self.datakind = "stream" else: assert isinstance( dataset, Dataset ), "Can not recognize this kind of dataset: %s" % type(dataset) self.sampler = ( sampler if sampler else SequentialSampler(dataset, batch_size=1, drop_last=False) ) assert isinstance( self.sampler, MapSampler ), "types of dataset and sampler do not match" self.datakind = "map" if transform is None: self.transform = PseudoTransform() else: self.transform = transform if collator is None: self.collator = Collator() else: self.collator = collator if platform.system() == "Linux" and self.num_workers > 0: self.check_memory_rationality() def __iter__(self): if platform.system() == "Windows" and self.num_workers > 0: print( "pyarrow.plasma does not support ParallelDataLoader on windows, changing num_workers to be zero" ) self.num_workers = 0 if os.getenv("TERMUX_VERSION"): # FIXME: termux install pyarrow will build error now # remove this logic after pyarrow fix this issue print( "pyarrow do not support on termux env now, changing num_workers to be zero" ) self.num_workers = 0 if isinstance(self.dataset, StreamDataset): if not self.num_workers: return _SerialStreamDataLoaderIter(self, self.preload) else: return _ParallelStreamDataLoaderIter(self, self.preload) else: assert isinstance( self.dataset, Dataset ), "Can not recognize this kind of dataset: %s" % type(self.dataset) if not self.num_workers: return _SerialMapDataLoaderIter(self, self.preload) else: return _ParallelMapDataLoaderIter(self, self.preload) def __len__(self): return len(self.sampler) def check_memory_rationality(self): import psutil main_memory = psutil.Process(os.getpid()).memory_info().rss / 1024 / 1024 / 1024 total_memory = (self.num_workers + 1) * main_memory limit_in_bytes = os.popen("cat /sys/fs/cgroup/memory/memory.limit_in_bytes") current_memory = int(limit_in_bytes.read()) / 1024 / 1024 / 1024 limit_in_bytes.close() if current_memory < total_memory: logger.warning( "Each worker need to read the shared meta-data, which will be increasing the reference count." "Copy-On-Write propety will lead to 'memory leak', the memory usage will end up being " + str(total_memory) + " GB, " "However the current requested memory is " + str(current_memory) + " GB, " "Maybe you can request more memory or uesd np-array to save meta-data rather than List or Tuple" )
class PreLoader: def __init__(self, loader, preload): self.dataset = loader.dataset self.sampler = loader.sampler self.seed = _random_seed_generator().__next__() self.transform = loader.transform self.collator = loader.collator self.num_workers = loader.num_workers self.timeout = loader.timeout self.num_processed = 0 self.datakind = loader.datakind self.parallel_stream = loader.parallel_stream if preload: self.default_device = get_default_device() self.pre_load_device = self.default_device + ":" + str(_sh.get_next()) self.pre_load_device_cache = None self.preload = preload if data_monitor: global monitor_num_workers, monitor_workers, put_time monitor_num_workers.value = self.num_workers monitor_workers = Array("i", [0] * monitor_num_workers.value) put_time = Array("d", [0.0] * monitor_num_workers.value) def __iter__(self): return self """ strategy one: load from numpy data, and generate dtype tensor """ def _load_tensor(self, batch, cached=True): if isinstance(batch, np.ndarray): device = self.pre_load_device if cached else self.default_device return Tensor(batch, device=device) elif isinstance(batch, collections.abc.Mapping): return {k: self._load_tensor(v, cached) for k, v in batch.items()} elif isinstance(batch, tuple) and hasattr(batch, "_fields"): # namedtuple return type(batch)(*(self._load_tensor(value, cached) for value in batch)) elif isinstance(batch, collections.abc.Sequence): return [self._load_tensor(value, cached) for value in batch] else: return batch """ strategy two: load from cache that is already tensor just do d2d copy """ def _load_cache(self, data): if isinstance(data, Tensor): if data.device == self.default_device: return data return copy(data, device=self.default_device) elif isinstance(data, collections.abc.Mapping): return {k: self._load_cache(v) for k, v in data.items()} elif isinstance(data, tuple) and hasattr(data, "_fields"): # namedtuple return type(data)(*(self._load_cache(value) for value in data)) elif isinstance(data, collections.abc.Sequence): return [self._load_cache(value) for value in data] else: return data def _swap_out_cache(self): out = self._load_cache(self.pre_load_device_cache) self.pre_load_device_cache = None # clean cache return out class _ParallelDataLoaderIter: def __init__(self): self._worker_queue_idx_cycle = itertools.cycle(range(self.num_workers)) from .tools._queue import PlasmaShmQueue self._worker_result_queue = PlasmaShmQueue() self._shutdown = False self._workers_done_event = multiprocessing.Event() self._index_queues = [] self._workers = [] for i in range(self.num_workers): index_queue = multiprocessing.Queue() index_queue.cancel_join_thread() w = multiprocessing.Process( target=_worker_loop, args=( self.dataset, index_queue, self._worker_result_queue, self._workers_done_event, self.transform, self.collator, self.sampler.batch_size, self.seed + i, i, self.num_workers, self.datakind, self.parallel_stream, ), daemon=True, ) gc.collect() w.start() self._index_queues.append(index_queue) self._workers.append(w) self._data_queue = self._worker_result_queue self._worker_pids_set = False _set_worker_pids(id(self), tuple(w.pid for w in self._workers)) _set_sigchld_handler() self._worker_pids_set = True self._reset() def _try_put_index(self): raise NotImplementedError def _reset(self): self._sampler_iter = iter(self.sampler) self._send_idx = 0 self._rcvd_idx = 0 self._task_info = {} self._workers_status = [True for _ in range(self.num_workers)] for _ in range(2 * self.num_workers): self._try_put_index() def _process_data(self, data): self._rcvd_idx += 1 self._try_put_index() if isinstance(data, bytes): data = pickle.loads(data) data.reraise() return data def _get_data(self): if self.timeout > 0: success, data = self._try_get_data(self.timeout) if success: return data else: raise_timeout_error() else: while True: success, data = self._try_get_data() if success: return data def _get_next_batch(self): while True: while self._rcvd_idx < self._send_idx: info = self._task_info[self._rcvd_idx] worker_id = info[0] if ( len(info) == 2 or self._workers_status[worker_id] ): # has data or work is still active break del self._task_info[self._rcvd_idx] self._rcvd_idx += 1 else: self._shutdown_workers() raise StopIteration if len(self._task_info[self._rcvd_idx]) == 2: data = self._task_info.pop(self._rcvd_idx)[1] return self._process_data(data) _get_data = self._get_data() if len(_get_data) == 1: assert isinstance(_get_data[0], bytes) exception = pickle.loads(_get_data[0]) exception.reraise() self._try_put_index() continue else: idx, data = _get_data if isinstance(data, int): # Check if StopIteration in StreamDataset self._mark_worker_as_unavailable(data) self._try_put_index() continue if idx != self._rcvd_idx: self._task_info[idx] += (data,) else: del self._task_info[idx] return self._process_data(data) def _try_get_data(self, timeout=GLOBAL_TIMEOUT): try: data = self._data_queue.get(timeout=timeout) if data_monitor: pid = os.getpid() idx = data[0] if IPC_worker.value == 0: IPC_worker.value = pid if pid == IPC_worker.value and idx % data_monitor_frequency == 0: data_tok = time.perf_counter() worker_idx = idx % monitor_num_workers.value IPC_time = data_tok - put_time[worker_idx] print(f"pid: {pid}, idx: {idx}, IPC_time: {IPC_time:.6f}") return (True, data) except Exception as e: failed_workers = [] for worker_id, w in enumerate(self._workers): if self._workers_status[worker_id] and not w.is_alive(): failed_workers.append((worker_id, w)) self._mark_worker_as_unavailable(worker_id) if w.exitcode == -9: logger.debug( "Maybe memory is not enough, please request for more memory!" ) if len(failed_workers) > 0: pids_str = ", ".join(str(w_info[1].pid) for w_info in failed_workers) w_ids_str = ", ".join(str(w_info[0]) for w_info in failed_workers) exitcode_str = ", ".join( str(w_info[1].exitcode) for w_info in failed_workers ) raise RuntimeError( "DataLoader worker (worker(s): {} , pid(s): {}) exited unexpectedly, exitcode(s): {}".format( w_ids_str, pids_str, exitcode_str ) ) if isinstance(e, queue.Empty): return (False, None) def _mark_worker_as_unavailable(self, worker_id, shutdown=False): q = self._index_queues[worker_id] q.put(None) self._workers_status[worker_id] = False assert self._workers_done_event.is_set() == shutdown def _shutdown_workers(self): if not self._shutdown: self._shutdown = True try: self._workers_done_event.set() for worker_id in range(len(self._workers)): if self._workers_status[worker_id]: self._mark_worker_as_unavailable(worker_id, shutdown=True) for w in self._workers: if w.is_alive(): w.terminate() w.join(timeout=GLOBAL_TIMEOUT) for q in self._index_queues: q.cancel_join_thread() q.close() self._data_queue.cancel_join_thread() self._data_queue.close() finally: if self._worker_pids_set: _reset_worker_pids(id(self)) self._worker_pids_set = False def __del__(self): self._shutdown_workers() class _BaseMapDataLoaderIter(PreLoader): def __init__(self, loader, preload): super().__init__(loader, preload) def __len__(self): return len(self.sampler) def __next__(self): if self.preload: cached = self.pre_load_device_cache if cached is None: # first and last if self.num_processed >= len(self): # last raise StopIteration elif self.num_processed == 0: # first self._try_load_tensor(cached=False) # first do the h2d out = self._swap_out_cache() self._try_load_tensor() return out else: data = self._get_next_batch() return data def _try_load_tensor(self, cached=True): if self.num_processed >= len(self): return else: self.num_processed += 1 batch = self._get_next_batch() self.pre_load_device_cache = self._load_tensor(batch, cached) class _SerialMapDataLoaderIter(_BaseMapDataLoaderIter): def __init__(self, loader, preload): super(_SerialMapDataLoaderIter, self).__init__(loader, preload) self._sampler_iter = iter(self.sampler) def _get_next_batch(self): if data_monitor: global monitor_workers pid = os.getpid() if monitor_workers[0] == 0: monitor_workers[0] = pid data_tik = time.perf_counter() indices = next(self._sampler_iter) items = [self.dataset[idx] for idx in indices] data_tok = time.perf_counter() dataset_time = data_tok - data_tik trans_items, transform_time = get_monitor_time_stats( self.transform.apply_batch, items ) collate_ret, collate_time = get_monitor_time_stats( self.collator.apply, trans_items ) print( f"pid: {pid}, dataset_time: {dataset_time:.6f}, transform_time: {transform_time:.6f}, collate_time: {collate_time:.6f}" ) return collate_ret else: indices = next(self._sampler_iter) items = [self.dataset[idx] for idx in indices] trans_items = self.transform.apply_batch(items) return self.collator.apply(trans_items) class _ParallelMapDataLoaderIter(_BaseMapDataLoaderIter, _ParallelDataLoaderIter): def __init__(self, loader, preload): _BaseMapDataLoaderIter.__init__(self, loader, preload) _ParallelDataLoaderIter.__init__(self) def _try_put_index(self): try: index = next(self._sampler_iter) except StopIteration: return for _ in range(self.num_workers): # find the next active worker, if any worker_queue_idx = next(self._worker_queue_idx_cycle) if self._workers_status[worker_queue_idx]: break self._index_queues[worker_queue_idx].put((self._send_idx, index)) self._task_info[self._send_idx] = (worker_queue_idx,) self._send_idx += 1 _worker_info = None class WorkerInfo(object): __initialized = False def __init__(self, **kwargs): for k, v in kwargs.items(): setattr(self, k, v) self.__keys = tuple(kwargs.keys()) self.__initialized = True def __setattr__(self, key, val): if self.__initialized: raise RuntimeError( "Cannot assign attributes to {} objects".format(self.__class__.__name__) ) return super(WorkerInfo, self).__setattr__(key, val) def __repr__(self): items = [] for k in self.__keys: items.append("{}={}".format(k, getattr(self, k))) return "{}({})".format(self.__class__.__name__, ", ".join(items)) def get_worker_info(): return _worker_info class _BaseStreamDataLoaderIter(PreLoader): def __init__(self, loader, preload): super().__init__(loader, preload) self.dataset_iter = iter(self.dataset) def __next__(self): if self.preload: if self.pre_load_device_cache is None: self._try_load_tensor(cached=False) # load in current out = self._swap_out_cache() self._try_load_tensor() # load in cached return out else: return self._get_next_batch() def _try_load_tensor(self, cached=True): batch = self._get_next_batch() self.pre_load_device_cache = self._load_tensor(batch, cached) class _SerialStreamDataLoaderIter(_BaseStreamDataLoaderIter): def __init__(self, loader, preload): super().__init__(loader, preload) self.dataset_iter = iter(self.dataset) def _try_get_raw_data(self, start_time): raw_data = None while not raw_data: try: if self.timeout > 0: timer = threading.Timer(self.timeout, thread.interrupt_main) timer.start() raw_data = next(self.dataset_iter) if self.timeout > 0: timer.cancel() except AttributeError as error: raise error except: if self.timeout > 0: timer.cancel() waited_time = time.time() - start_time if waited_time > self.timeout: raise_timeout_error() return raw_data def _get_next_batch(self): ret = [] start_time = time.time() if data_monitor: global monitor_workers pid = os.getpid() if monitor_workers[0] == 0: monitor_workers[0] = pid while len(ret) < self.sampler.batch_size: if pid == monitor_workers[0]: raw_data, dataset_time = get_monitor_time_stats( self._try_get_raw_data, start_time ) raw_data, transform_time = get_monitor_time_stats( self.transform.apply, raw_data ) ret.append(raw_data) else: raw_data = self._try_get_raw_data(start_time) ret.append(self.transform.apply(raw_data)) if pid == monitor_workers[0]: collate_ret, collate_time = get_monitor_time_stats( self.collator.apply, ret ) print( f"pid: {pid}, dataset_time: {dataset_time:.6f}, transform_time: {transform_time:.6f}, collate_time: {collate_time:.6f}" ) else: collate_ret = self.collator.apply(ret) return collate_ret else: while len(ret) < self.sampler.batch_size: raw_data = self._try_get_raw_data(start_time) ret.append(self.transform.apply(raw_data)) return self.collator.apply(ret) class _ParallelStreamDataLoaderIter(_BaseStreamDataLoaderIter, _ParallelDataLoaderIter): def __init__(self, loader, preload): _BaseStreamDataLoaderIter.__init__(self, loader, preload) _ParallelDataLoaderIter.__init__(self) def _get_remaind_data(self, place_holder): num = self.sampler.batch_size for _ in range(num - 1): place_holder.append(next(self.dataset_iter)) return place_holder def _try_put_index(self): try: if self.parallel_stream is False: start_time = time.time() place_holder = [next(self.dataset_iter)] waited_time = time.time() - start_time if self.timeout > 0 and waited_time > self.timeout: raise_timeout_error() place_holder = self._get_remaind_data(place_holder) else: place_holder = next(self._sampler_iter) except StopIteration: return for _ in range(self.num_workers): worker_queue_idx = next(self._worker_queue_idx_cycle) if self._workers_status[worker_queue_idx]: break else: return self._index_queues[worker_queue_idx].put((self._send_idx, place_holder)) self._task_info[self._send_idx] = (worker_queue_idx,) self._send_idx += 1 class ManagerWatchdog(object): def __init__(self): self.manager_pid = os.getppid() self.manager_dead = False def is_alive(self): if not self.manager_dead: self.manager_dead = os.getppid() != self.manager_pid return not self.manager_dead def stream_fetcher( dataset_iter, place_holder, transform, collate, parallel_stream, batch_size ): pid = os.getpid() global dataset_time, transform_time, collate_time data = [] for idx in place_holder: try: if parallel_stream is False: raw_data = idx else: if data_monitor: raw_data, dataset_time = get_monitor_time_stats(next, dataset_iter) else: raw_data = next(dataset_iter) if data_monitor: trans_items, transform_time = get_monitor_time_stats( transform.apply, raw_data ) else: trans_items = transform.apply(raw_data) data.append(trans_items) except StopIteration: break if len(data) == 0: raise StopIteration if data_monitor: data, collate_time = get_monitor_time_stats(collate.apply, data) return data, pid, dataset_time, transform_time, collate_time else: data = collate.apply(data) return data def map_fetcher(dataset, place_holder, transform, collate, parallel_stream, batch_size): if data_monitor: pid = os.getpid() global dataset_time, transform_time, collate_time data_tik = time.perf_counter() items = [dataset[idx] for idx in place_holder] data_tok = time.perf_counter() dataset_time = data_tok - data_tik trans_items, transform_time = get_monitor_time_stats( transform.apply_batch, items ) data, collate_time = get_monitor_time_stats(collate.apply, trans_items) return data, pid, dataset_time, transform_time, collate_time else: items = [dataset[idx] for idx in place_holder] trans_items = transform.apply_batch(items) data = collate.apply(trans_items) return data def _worker_loop( dataset, index_queue, data_queue, done_event, transform, collate, batch_size, seed, worker_id, num_workers, datakind, parallel_stream, ): _set_worker_signal_handlers() random.seed(seed) np.random.seed(seed) watchdog = ManagerWatchdog() iteration_end = False fetcher = map_fetcher if datakind == "stream": global _worker_info _worker_info = WorkerInfo(idx=worker_id, worker=num_workers, seed=seed) dataset = iter(dataset) fetcher = stream_fetcher while watchdog.is_alive(): try: r = index_queue.get(timeout=GLOBAL_TIMEOUT) except queue.Empty: continue if r is None: assert done_event.is_set() or iteration_end break elif done_event.is_set() or iteration_end: continue idx, place_holder = r try: if data_monitor: data, pid, dataset_time, transform_time, collate_time = fetcher( dataset, place_holder, transform, collate, parallel_stream, batch_size, ) if idx < monitor_num_workers.value: if monitor_workers[idx] == 0: monitor_workers[idx] = pid worker_idx = idx % monitor_num_workers.value if ( pid == monitor_workers[worker_idx] and idx % data_monitor_frequency == 0 ): print( f"pid: {pid}, idx: {idx}, dataset_time: {dataset_time:.6f}, transform_time: {transform_time:.6f}, collate_time: {collate_time:.6f}" ) else: data = fetcher( dataset, place_holder, transform, collate, parallel_stream, batch_size, ) except Exception as e: if isinstance(e, StopIteration) and datakind == "stream": data = worker_id iteration_end = True else: from .tools._queue import _ExceptionWrapper exc_info = sys.exc_info() where = "in DataLoader worker process {}".format(worker_id) exc_msg = "".join(traceback.format_exception(*exc_info)) data = _ExceptionWrapper(exc_info[0].__name__, exc_msg, where) data = pickle.dumps(data) if data_monitor: pid = os.getpid() worker_idx = idx % monitor_num_workers.value if pid == monitor_workers[worker_idx]: data_tik = time.perf_counter() put_time[worker_idx] = data_tik data_queue.put((idx, data)) del data, idx, place_holder, r if done_event.is_set(): data_queue.disconnect_client() data_queue.close()