Dataset issues

[QiuZhiZhe123]

We got top1 accuracy of 50.25 by running the training python file of the sthv1 dataset you shared, we suspect that there is something wrong with the dataset we collected, do you share your dataset?

[yzfly]

Sorry for the very late reply. We cannot redistribute the Something-Something V1 dataset — it is licensed by TwentyBN (now hosted by Qualcomm), so you need to download it from the official source: https://developer.qualcomm.com/software/ai-datasets/something-something

A top-1 accuracy of ~50% on SSv1 usually indicates a data preparation problem rather than a model problem. Common causes:

1. Frame extraction: frames must be extracted exactly as in the TSM pipeline (see https://github.com/mit-han-lab/temporal-shift-module/tree/master/tools) — wrong FPS or resizing during extraction hurts accuracy noticeably.
2. Annotation/list mismatch: make sure train_videofolder.txt / val_videofolder_latest.txt were regenerated against your extracted frames (frame counts must match actual files).
3. Learning rate vs batch size: as noted in the README, lr should scale linearly with batch size (0.01 for b=32).

Closing due to age; feel free to reopen if you still work on this.