[BUG] a bug in the dataset/tiktok_video_dataset.py

[ljh0v0]

When sampling a segment of video to train the temporal module, the length of the dataset is divided by the train_sample_interval = 4.

But in this line:

DisCo/dataset/tiktok_video_dataset.py

Line 295 in 8538889

idx = min(idx, len(self) - 1)

the code limits the start frame id to be min(idx, len(self) - 1). The len(self) = num_frames / train_sample_interval, so it means only the first quarter of the images in the dataset can be the first frame. Any video segment whose first frame id is larger than num_frames / train_sample_interval will be replaced by the segment starting with the id num_frames / train_sample_interval-1. This causes only the first quarter of datasets will be used in the training.