cwx-worst-one / eat Goto Github PK

Hi! Thanks for your great work!

I wonder when the pre-training and fine-tuning code will be available.

Hi, Great work! I pre-trained the network using the provided script with the full AudioSet (unbalanced + balanced), similar to BEATs and AudioMAE. So my pre-training set is slightly larger than yours (1.93M samples). I trained for 10 epochs (40K updates) using the provided script. However, I am unable to reproduce the reported AS20K performance, achieving only around 35.8 mAP compared to the reported 40.2 mAP. I fine-tuned on AS20K using the shared pre-trained weights(EAT-base_epoch10.pt) and achieved the reported performance, so I don't think there is an issue with the dataset. Am I missing something here?