Problem with general_beat_saber_dataset

[chaonan99]

Thanks for the great project! I am trying to go through the training process and encountered a problem in general_beat_saber_dataset.py. Specifically, from L169 to L172, y is padded on both sides, which makes sequence_length in L178 equals the padded length. This subsequently makes indices in L193 possibly larger than the length of blocks_reduced_classes, and triggers an error in retrieving a subsequence of blocks_reduced_classes if so (L235).

I am not entirely sure how to fix this, especially how features and targets are aligned in the presence of time_shift. Could you please point out how this can be fixed? Thanks!

[deyituo]

indices - time_shifts // 2