Open Dragon2938734 opened 4 months ago
The results you reproduced are within the acceptable margin of error. After careful checking, it was found that 'torch.cuda.manual_seed_all(seed)' was missing in train.py, so the random seed was not completely fixed, and there would still be differences in each reproduction. I have just added it.
hello, i followed the code , yml, cfg provided and train the model , but i got worse results than the paper described. May be it caused by a different random seed (123 is provided as default)? can you share more train params? thank you !
my training results
paper results