Dataloader to load video with neighboring windows (without nvvl)?

Hi again,

I am looking for help in writing pytorch dataloader for my dataset, which can load small video clips ( loading sequence of frames using neighbouring window) from 1 big video.

I have some dataset which contains lots of videos and each video can have multiple labels.

Now I would like to train my network with 16 frames with neighbouring video. Also, each video is of different length. I strictly don’t want to use nvvl because of some package dependancy issue which I have already tried.