Add chunk iterator cache to frame provider (#1367)

* Add chunk iterator cache

* fix

cvat/apps/dataset_manager/bindings.py

@@ -26,7 +26,7 @@ class CvatImagesExtractor(datumaro.Extractor):
         frames = self._frame_provider.get_frames(
             self._frame_provider.Quality.ORIGINAL,
             self._frame_provider.Type.NUMPY_ARRAY)
-        for item_id, image in enumerate(frames):
+        for item_id, (image, _) in enumerate(frames):
             yield datumaro.DatasetItem(
                 id=item_id,
                 image=Image(image),

cvat/apps/engine/frame_provider.py

@@ -2,7 +2,6 @@
 #
 # SPDX-License-Identifier: MIT
 
-import itertools
 import math
 from enum import Enum
 from io import BytesIO
@@ -15,6 +14,33 @@ from cvat.apps.engine.mime_types import mimetypes
 from cvat.apps.engine.models import DataChoice
 
 
+class RandomAccessIterator:
+    def __init__(self, iterable):
+        self.iterable = iterable
+        self.iterator = None
+        self.pos = -1
+
+    def __iter__(self):
+        return self
+
+    def __next__(self):
+        return self[self.pos + 1]
+
+    def __getitem__(self, idx):
+        assert 0 <= idx
+        if self.iterator is None or idx <= self.pos:
+            self.reset()
+        v = None
+        while self.pos < idx:
+            # NOTE: don't keep the last item in self, it can be expensive
+            v = next(self.iterator)
+            self.pos += 1
+        return v
+
+    def reset(self):
+        self.iterator = iter(self.iterable)
+        self.pos = -1
+
 class FrameProvider:
     class Quality(Enum):
         COMPRESSED = 0
@@ -35,7 +61,8 @@ class FrameProvider:
         def load(self, chunk_id):
             if self.chunk_id != chunk_id:
                 self.chunk_id = chunk_id
-                self.chunk_reader = self.reader_class([self.get_chunk_path(chunk_id)])
+                self.chunk_reader = RandomAccessIterator(
+                    self.reader_class([self.get_chunk_path(chunk_id)]))
             return self.chunk_reader
 
     def __init__(self, db_data):
@@ -104,18 +131,18 @@ class FrameProvider:
         chunk_number = self._validate_chunk_number(chunk_number)
         return self._loaders[quality].get_chunk_path(chunk_number)
 
-    def get_frame(self, frame_number, quality=Quality.ORIGINAL):
+    def get_frame(self, frame_number, quality=Quality.ORIGINAL,
+            out_type=Type.BUFFER):
         _, chunk_number, frame_offset = self._validate_frame_number(frame_number)
+        loader = self._loaders[quality]
+        chunk_reader = loader.load(chunk_number)
+        frame, frame_name, _ = chunk_reader[frame_offset]
 
-        chunk_reader = self._loaders[quality].load(chunk_number)
-
-        frame, frame_name, _ = next(itertools.islice(chunk_reader, frame_offset, None))
-        if self._loaders[quality].reader_class is VideoReader:
-            return (self._av_frame_to_png_bytes(frame), 'image/png')
+        frame = self._convert_frame(frame, loader.reader_class, out_type)
+        if loader.reader_class is VideoReader:
+            return (frame, 'image/png')
         return (frame, mimetypes.guess_type(frame_name))
 
     def get_frames(self, quality=Quality.ORIGINAL, out_type=Type.BUFFER):
-        loader = self._loaders[quality]
-        for chunk_idx in range(math.ceil(self._db_data.size / self._db_data.chunk_size)):
-            for frame, _, _ in loader.load(chunk_idx):
-                yield self._convert_frame(frame, loader.reader_class, out_type)
+        for idx in range(self._db_data.size):
+            yield self.get_frame(idx, quality=quality, out_type=out_type)