properly handle the cases where video decoding fails in compute_clips of class VideoClips

Author: zyan3

torchvision/datasets/video_utils.py

@@ -168,9 +168,15 @@ def compute_clips(self, num_frames, step, frame_rate=None):
         self.clips = []
         self.resampling_idxs = []
         for video_pts, info in zip(self.video_pts, self.info):
-            clips, idxs = self.compute_clips_for_video(video_pts, num_frames, step, info["video_fps"], frame_rate)
-            self.clips.append(clips)
-            self.resampling_idxs.append(idxs)
+            if "video_fps" in info:
+                clips, idxs = self.compute_clips_for_video(video_pts, num_frames, step, info["video_fps"], frame_rate)
+                self.clips.append(clips)
+                self.resampling_idxs.append(idxs)
+            else:
+                # properly handle the cases where video decoding fails
+                self.clips.append(torch.zeros(0, num_frames, dtype=torch.int64))
+                self.resampling_idxs.append(torch.zeros(0, dtype=torch.int64))
+
         clip_lengths = torch.as_tensor([len(v) for v in self.clips])
         self.cumulative_sizes = clip_lengths.cumsum(0).tolist()