search

cuihaoleo / kaggle-dfdc

2nd place solution for Kaggle Deepfake Detection Challenge
MIT License
160 stars 43 forks source link

Some issues with kaggle-dfdc/train-wsdan.py #15

Open thaondc opened 2 years ago

thaondc commented 2 years ago

Hello @cuihaoleo,

I have read your project. It is so great. But I had some problems when I read it. In kaggle-dfdc/train-wsdan.py, lines 177-198, I think you computed epoch_loss, epoch_raw_acc ... to save to log, but I don't understand what did dist do in lines 181-184. Can you help me to answer?

Thank you.

cuihaoleo commented 2 years ago

Not sure. It looks like it obtains the sum of losses from all parallel workers for logging, and doesn't really affect the training.

@yoctta wrote this part.

thaondc commented 2 years ago

Thank you.