Merge remote-tracking branch 'remotes/origin/master' into adaptiveLR

ovotools/pytorch/__init__.py

+from .threading_dataloader import BatchThreadingDataLoader, ThreadingDataLoader
\ No newline at end of file

ovotools/pytorch/threading_dataloader.py

+'''
+Multithread pytorch dataloaders.
+Usefull for due to improper work of standard DataLoader in a multiprocess (num_workers > 0) mode.
+See issue https://github.com/pytorch/pytorch/issues/12831 etc.
+
+Based on https://github.com/lopuhin/kaggle-imet-2019/blob/f1ec0827149a8218430a6884acf49c27ba6fcb1f/imet/utils.py#L35-L56
+'''
+import torch
+from multiprocessing.pool import ThreadPool
+
+
+class BatchThreadingDataLoader(torch.utils.data.DataLoader):
+    '''
+    Prepares each batch in a separate thread, including collate, so no processing is done in the caller thread.
+    Requires map-style dataset (i.e. dataset with indexing) and automatic batching.
+    Ignores worker_init_fn parameter.
+    '''
+    def __iter__(self):
+        sample_iter = iter(self.batch_sampler)
+        if self.num_workers == 0:
+            for indices in sample_iter:
+                yield self._get_batch(indices)
+        else:
+            prefetch = self.num_workers
+            with ThreadPool(processes=self.num_workers) as pool:
+                futures = []
+                for indices in sample_iter:
+                    futures.append(pool.apply_async(self._get_batch, args=(indices,)))
+                    if len(futures) > prefetch:
+                        yield futures.pop(0).get()
+                for batch_futures in futures:
+                    yield batch_futures.get()
+
+    def _get_item(self, i):
+        return self.dataset[i]
+
+    def _get_batch(self, indices):
+        return self.collate_fn([self._get_item(i) for i in indices])
+
+
+class ThreadingDataLoader(torch.utils.data.DataLoader):
+    '''
+    Original dataset from https://github.com/lopuhin/kaggle-imet-2019
+    Prepares each dataset item in a separate thread, but collation is processed in the caller thread.
+    Requires map-style dataset (i.e. dataset with indexing) and automatic batching.
+    Ignores worker_init_fn parameter.
+    '''
+    def __iter__(self):
+        sample_iter = iter(self.batch_sampler)
+        if self.num_workers == 0:
+            for indices in sample_iter:
+                yield self.collate_fn([self._get_item(i) for i in indices])
+        else:
+            prefetch = 1
+            with ThreadPool(processes=self.num_workers) as pool:
+                futures = []
+                for indices in sample_iter:
+                    futures.append([pool.apply_async(self._get_item, args=(i,))
+                                    for i in indices])
+                    if len(futures) > prefetch:
+                        yield self.collate_fn([f.get() for f in futures.pop(0)])
+                for batch_futures in futures:
+                    yield self.collate_fn([f.get() for f in batch_futures])
+
+    def _get_item(self, i):
+        return self.dataset[i]