Bingkun Huang

Sorry, I cannot find the script. Maybe you can print the keys of the original model and the converted model, and then compare the difference. The key value should not...

> Would you please update the data preprocessing part, so it would be easier to check the correctness of the codes. The preprocessing script hasn't been implemented yet. I use...

kindly ping :p @dreamerlin

Sorry, Please remove `from petrel_client.client import Client`, which is used to read data on CephBackend.

This is the script file we put up by mistake for processing videos, it has nothing to do with VideoMAE and we will delete the file later.

Because the target of reconstruction task is the normalized pixel instead of pixel

too small batch size may degrade the model performance, maybe you can use checkpoint.checkpoint to save memory.

with deepspeed enabled, only support '--auto_resume'. you can copy the `checkpoint-45` to the work dir.

I've done similar experiments and achieved similar results to mae. Based on my limited experimental results:predictive features are easier to train than predictive pixels; the potential of this training method...

I think the training is normal. How's it going back then?