joos2010kj / clip-tsa Goto Github PK

Hi @joos2010kj,

Is it possible for you to share the code/repositories used for feature extraction and pre-processing? Essentially, the code that could help in generating the features that are to be fed during training/evaluation.

Thank you

Would you share the features of the dataset you extracted with vit? Is there any reference code in the extraction process?

Congrats on the nice work!

I have a small question on the extracted CLIP features. It seems that in your provided CLIP features, the testing set of UCF Crime dataset is not complete? Could you provide the complete version? Thanks so much.

I would appreciate it very much if you could reply.

Hello

in the xdv train set , there a video missing its npy:
v=8cTqh9tMz_I__#1_label_A

which match the error i got when extracting the actual available files
ERROR: Data Error : v=8cTqh9tMz_I__#1_label_A.mp4

maybe be useful for others,

if in the middletime i find the original mp4, i can upload

cheers

@joos2010kj hello great work ！ I have encountered an issue with insufficient GPU memory and I have tried more gpus to train，such as CUDA_VISIBLE_DEVICES= 0，1，2 python. But it doesn't work,why? Why does using multiple GPUs for training still encounter this problem？

Hello,

Thanks for the code.
Could you please upload the model's checkpoints?

Nice works!
Have you considered releasing the Clip visual features extracted from UCF Crime and XD violence so that people can follow them better :)