Cannot reproduce the zero-shot results on UCF101 and HMDB51

[fuyukaiKUN]

Thank you very much for your interesting work and for sharing the code. I have encountered some issues while trying to reproduce your results in the paper, and I hope to resolve them.

Specifically, I use a K400 pretrained ckpt provided by you (DIST_VIT_B16_K400_16+32f_84.4.pyth) to attempt reproducing the zero-shot results on UCF101 and HDMB51. I can reproduce the results on K400 using this ckpt. But I cannot reproduce the zero-shot results using this ckpt on UCF101 and HDMB51. The config is almost the same except some parameters such as datasetname. My reproduced performance are very similar to the raw CLIP.

I wonder if there are any differences in the setup or parameters when runing zero-shot test. Could you please provide the shell scripts and config files for running zero-shot test?

Thank you very much. Looking forward to your reply.