Confusion about the baseline

[wei666666]

Hi, I'm very interested in your project. So, I'm coming back! In your ablation study, I found you mentioned baseline without S_TDM or L-TDM, only with shift operation in latter stages. So I reimplemented it. WOW~ It's amazing that the top1 accuracy reached 47.2% (45.2% in paper). I don't know what's wrong with the code. I would appreciate it if you could provide the code of baseline! (1845848350@qq.com)

Only two changes I made in your code:
1.base_module.py

2.tdn_net.py

[yztongzhan]

Hi @wei666666, thanks so much for sharing your interest in our TDN!
We have updated our camera ready version with stronger baseline 46.6%. You can check the new baseline and any other result in arxiv and openaccess.
Your implementation seems no difference with ours. We may attribute this phenomenon to the randomness introduced by the experiment settings such as GPU number, batch size and etc. You could try again and check the new result. If you have more questions, feel free to comment below.