| 73 | |
| 74 | |
| 75 | class _PreprocessSingleProcessDataLoaderIter(_BaseDataLoaderIter): |
| 76 | |
| 77 | def __init__(self, loader, preprocess: Preprocess=None, queue_max_size=2): |
| 78 | self._preprocess = preprocess |
| 79 | if self._preprocess is not None and self._preprocess.has_cpu_preprocess(): |
| 80 | org_create_fetcher = _DatasetKind.create_fetcher |
| 81 | _DatasetKind.create_fetcher = _create_fetcher_proxy(org_create_fetcher, self._preprocess) |
| 82 | |
| 83 | super(_PreprocessSingleProcessDataLoaderIter, self).__init__(loader) |
| 84 | |
| 85 | self._dataset_fetcher = _DatasetKind.create_fetcher( |
| 86 | self._dataset_kind, self._dataset, self._auto_collation, self._collate_fn, self._drop_last) |
| 87 | |
| 88 | if self._preprocess is not None and self._preprocess.has_cpu_preprocess(): |
| 89 | _DatasetKind.create_fetcher = org_create_fetcher |
| 90 | |
| 91 | self._timeout = self._timeout if self._timeout > 0 else (12 * _utils.MP_STATUS_CHECK_INTERVAL) |
| 92 | |
| 93 | if self._preprocess is not None and self._preprocess.has_gpu_preprocess(): |
| 94 | self._stopped = False |
| 95 | self._stream = torch.cuda.Stream() |
| 96 | print(f"xxxxxx _PreprocessSingleProcessDataLoaderIter queue_max_size: {queue_max_size}") |
| 97 | self._data_queue = queue.Queue(maxsize=queue_max_size) |
| 98 | self._device_id = torch.cuda.current_device() |
| 99 | self._preprocess_thread_done_event = threading.Event() |
| 100 | self._preprocess_thread = threading.Thread(target=self._preprocess_loop) |
| 101 | self._preprocess_thread.daemon = True |
| 102 | self._preprocess_thread.start() |
| 103 | |
| 104 | def _preprocess_loop(self): |
| 105 | # torch.set_num_threads(1) |
| 106 | torch.cuda.set_device(self._device_id) |
| 107 | while not self._preprocess_thread_done_event.is_set(): |
| 108 | while not self._stopped: |
| 109 | index = None |
| 110 | data = None |
| 111 | try: |
| 112 | index = self._next_index() |
| 113 | data = self._dataset_fetcher.fetch(index) |
| 114 | if self._pin_memory: |
| 115 | data = _utils.pin_memory.pin_memory(data) |
| 116 | # print(f"_preprocess_loop data is {type(data)}") |
| 117 | # print(f"************** _PreprocessSingleProcessDataLoaderIter _preprocess_loop **************") |
| 118 | if self._preprocess is not None and self._preprocess.has_gpu_preprocess(): |
| 119 | with torch.no_grad(), torch.cuda.stream(self._stream): |
| 120 | data = self._preprocess.gpu_preprocess(data, self._stream) |
| 121 | # print(f"xxxxxx data: {data.keys()})") |
| 122 | self._stream.synchronize() |
| 123 | # print(f"xxxxxx _stream.synchronize") |
| 124 | # self._data_queue.put((index, data)) |
| 125 | except StopIteration as e: |
| 126 | # print("xxxxxx StopIteration") |
| 127 | self._stopped = True |
| 128 | index = None |
| 129 | data = None |
| 130 | except Exception as e: |
| 131 | data = _utils.ExceptionWrapper(where=f"in _preprocess_loop thread for device {self._device_id}") |
| 132 | r = (index, data) |
no outgoing calls
no test coverage detected