Memory Issue torchvision.datasets.video_utils import VideoClips

[haoranD]

Hi, thank you for your excellent works.

I am trying to extract the features myself. But I found that when I ran the feature_extractor, it cost me all the memory(32GB) and my pc crashed.

Can I ask if I should make any changes or can you please give me any suggestion?

[ekosman]

Can you provide more details, like when does it crash?

[haoranD]

Can you provide more details, like when does it crash?

Much appreciated for your reply.

Run the Feature_Extractor.py ---> main() function ---> train_loader = VideoIterTrain ---> class VideoIterTrain(data.Dataset) ---> self.video_clips = VideoClips() ---> VideoClips() function initialization ---> _compute_frame_pts() ---> located at for batch in dl:

I didn't get any errors but my pc just crashed after 5 minutes. I monitored the GPU, CPU, Memory. I found this step didn't use GPU too much. But cost massive memory(at least 48GB memory and need more or PC crashing).

Thanks and have a nice day.

[ekosman]

That's probably has to do with torchvision's VideoClip class.

The quickest solution I suggest is to rewrite data_loader so that it loads one video at at time, instead of loading the whole dataset which is what I've done.

[haoranD]

That's probably has to do with torchvision's VideoClip class.

The quickest solution I suggest is to rewrite data_loader so that it loads one video at at time, instead of loading the whole dataset which is what I've done.

Perfect! I will try it and come back. Thx

[PHDPeter]

Ok so just in case you are slow like me (and my Pc is would seem) Here is code to add to data_loader (around line 30) to get it to load one video at a time.

with tqdm(total=len(self.video_list[1:])+1,desc=' total % of videos loaded') as pbar1: for video_list_used in self.video_list[1:]: #length of load?) pbar1.update(1) video_clips_out = VideoClips(video_paths=[video_list_used], clip_length_in_frames=self.total_clip_length_in_frames, frames_between_clips=self.total_clip_length_in_frames) self.video_clips.clips.append(video_clips_out.clips[0]) self.video_clips.cumulative_sizes.append(video_clips_out.cumulative_sizes[0]) self.video_clips.resampling_idxs.append(video_clips_out.resampling_idxs[0]) self.video_clips.video_fps.append(video_clips_out.video_fps[0]) self.video_clips.video_paths.append(video_clips_out.video_paths[0]) self.video_clips.video_pts.append(video_clips_out.video_pts[0]) You dont have to add the with tqdm(total=len(self.video_list[1:])+1,desc=' total % of videos loaded') as pbar1: line as I was just plaing about with tqdm to gave a "tatal count". Without this you will still have a tqdm bar for ever video loaded, inless you want to mess with the code in torchvision's VideoClip class. Hope this helps some one.

[PHDPeter]

Ok so just in case you are slow like me (and my Pc is would seem) Here is code to add to data_loader (around line 30) to get it to load one video at a time.

with tqdm(total=len(self.video_list[1:])+1,desc=' total % of videos loaded') as pbar1: for video_list_used in self.video_list[1:]: #length of load?) pbar1.update(1) video_clips_out = VideoClips(video_paths=[video_list_used], clip_length_in_frames=self.total_clip_length_in_frames, frames_between_clips=self.total_clip_length_in_frames) self.video_clips.clips.append(video_clips_out.clips[0]) self.video_clips.cumulative_sizes.append(video_clips_out.cumulative_sizes[0]) self.video_clips.resampling_idxs.append(video_clips_out.resampling_idxs[0]) self.video_clips.video_fps.append(video_clips_out.video_fps[0]) self.video_clips.video_paths.append(video_clips_out.video_paths[0]) self.video_clips.video_pts.append(video_clips_out.video_pts[0]) You dont have to add the with tqdm(total=len(self.video_list[1:])+1,desc=' total % of videos loaded') as pbar1: line as I was just plaing about with tqdm to gave a "tatal count". Without this you will still have a tqdm bar for ever video loaded, inless you want to mess with the code in torchvision's VideoClip class. Hope this helps some one.

Hey, sorry I screwed up. The cumulative_sizes update I put here does not add the running total meaning, i.e. it will just list the size of that video clips. you need to add self.video_clips.cumulative_sizes.append(self.video_clips.cumulative_sizes[-1]+video_clips_out.cumulative_sizes[0]) This should then keep the running total. Hope that is now alright, please let me know if anyone spots any other problems. Thanks