[Release 0.7 Cherry-Pick] Let get_frames_at and get_frames_played_at accept tensor indices … (#882)

Author: NicolasHug

src/torchcodec/decoders/_video_decoder.py

@@ -226,6 +226,12 @@ def get_frames_at(self, indices: list[int]) -> FrameBatch:
         Returns:
             FrameBatch: The frames at the given indices.
         """
+        if isinstance(indices, torch.Tensor):
+            # TODO we should avoid converting tensors to lists and just let the
+            # core ops and C++ code natively accept tensors.  See
+            # https://github.com/pytorch/torchcodec/issues/879
+            indices = indices.to(torch.int).tolist()
+
         data, pts_seconds, duration_seconds = core.get_frames_at_indices(
             self._decoder, frame_indices=indices
         )
@@ -301,6 +307,12 @@ def get_frames_played_at(self, seconds: list[float]) -> FrameBatch:
         Returns:
             FrameBatch: The frames that are played at ``seconds``.
         """
+        if isinstance(seconds, torch.Tensor):
+            # TODO we should avoid converting tensors to lists and just let the
+            # core ops and C++ code natively accept tensors.  See
+            # https://github.com/pytorch/torchcodec/issues/879
+            seconds = seconds.to(torch.float).tolist()
+
         data, pts_seconds, duration_seconds = core.get_frames_by_pts(
             self._decoder, timestamps=seconds
         )

test/test_decoders.py

@@ -1389,6 +1389,17 @@ def test_10bit_videos_cpu(self, asset):
     #                     custom_frame_mappings=custom_frame_mappings,
     #                 )
 
+    def test_get_frames_at_tensor_indices(self):
+        # Non-regression test for tensor support in get_frames_at() and
+        # get_frames_played_at()
+        decoder = VideoDecoder(NASA_VIDEO.path)
+
+        decoder.get_frames_at(torch.tensor([0, 10], dtype=torch.int))
+        decoder.get_frames_at(torch.tensor([0, 10], dtype=torch.float))
+
+        decoder.get_frames_played_at(torch.tensor([0, 1], dtype=torch.int))
+        decoder.get_frames_played_at(torch.tensor([0, 1], dtype=torch.float))
+
 
 class TestAudioDecoder:
     @pytest.mark.parametrize("asset", (NASA_AUDIO, NASA_AUDIO_MP3, SINE_MONO_S32))