pytorch/vision

Consistent argument names for video datasets and VideoClips class

Open

#1227 aperta il 10 ago 2019

Vedi su GitHub
 (5 commenti) (0 reazioni) (0 assegnatari)Python (6858 fork)batch import
enhancementhelp wantedmodule: iomodule: video

Metriche repository

Star
 (15.050 star)
Metriche merge PR
 (Merge medio 12g 8h) (14 PR mergiate in 30 g)

Descrizione

First of all, the new support for Video data is amazing. Really useful. 😃

Just when I was going through the code of the video datasets and utilities, I found this a bit confusing.

Sample Dataset: https://github.com/pytorch/vision/blob/8635be94d1216f10fb8302da89233bd86445e449/torchvision/datasets/ucf101.py#L27-L36

VideoClips class: https://github.com/pytorch/vision/blob/8635be94d1216f10fb8302da89233bd86445e449/torchvision/datasets/video_utils.py#L43-L49

From what I understand, the argument frames_per_clip from the dataset is the same as clip_length_in_frames from the VideoClips class.

Similarly, step_between_clips <-> frames_between_clips.

Is there a reason for this difference. Or did I understand it wrong?

Would it be better if both, the video datasets and the VideoClips class had consistent arguments?

In my opinion, frames_per_clip looks better than clip_length_in_frames. And frames_between_clips can be favored over step_between_clips since the former feels more intuitive.

Any suggestions? @fmassa

Guida contributor