datamllab / autovideo

AutoVideo: An Automated Video Action Recognition System
https://autoedge.ai/
MIT License
318 stars 36 forks source link

Doubt about TSM temporal shift #8

Open alejandrosatis opened 2 years ago

alejandrosatis commented 2 years ago

Hi,

First of all, I'd like to congratulate about this repo, we've found this very useful. While training TSM, we've discovered that the parameter is_shift is by default false. Also, the import there cannot be resolved since the original make_temporal_shift code is not integrated into this repo.

Without is_shift enabled, does that mean that we're using a vanilla 2D Resnet50 and averaging the output of every input image in the sequence? Am I missing anything? The original contribution of TSM was this special temporal shift in the internal feature maps of any 2D CNN model.

Thanks in advance.

zaidbhat1234 commented 2 years ago

Hey @alejandrosatis,

We are glad that you found this repo useful. Thanks for pointing out the issue. We have integrated the original make_temporal_shift and non_local code in the repo. Please let us know if this resolves your issue.

Thanks!