the problem of accuracy

[yinzi6666666]

Hello author, I noticed that when using the weights of emonet_8.pth, the test accuracy of the affectnet dataset is 38%, which significantly deviates from the ACC=0.75. Could you please let me know if there is any missing operation?

[larryfans]

me too

[timlac]

I also get very counterintuitive results when I validate the emonet_8.pth model on other datasets. It would be very appreciated if the author if this repository could reconfirm that the correct weights are uploaded and specify if there's some particular pre-processing required to reproduce the results...