visinf / lnfmm Goto Github PK

Due to a limit of 11GB GPU resource, I am forced to reduce the batch size for image2text to 8 instead of 128.
After training for a while, I was unable to get a converged trained model that replicated results in the paper.

Hence, my question is:

1. How much GPU memory did you use?
2. Does using smaller batch size leads to failure in training?

Hi @sroth-visinf ,

Thanks for the implementation of the excellent work. I ran into the error RuntimeError: unexpected EOF, expected 170260432 more bytes. The file might be corrupted. while loading the released pre-trained model weight. It seems that the checkpoint file is corrupted and torch.load() raises the error. Could you please check or re-upload the model weights. Thank you very much!

Regards,
Chang