search

Talented-Q / POSTER_V2

MIT License
112 stars 20 forks source link

Cannot reproduce the results on AffectNet #12

Open zyh-uaiaaaa opened 1 year ago

zyh-uaiaaaa commented 1 year ago

Hi,

Could you please upload the pre-processing code of AffectNet? It seems that I could not reproduce the results even on my own AffectNet test set using your uploaded trained model. Additionally, training from scratch on my AffectNet dataset cannot achieve over 67% on 7 classes or over 63% on 8 classes. Thanks for your help.

Talented-Q commented 1 year ago

J04VDBJ8R3%WJG10HYNN9}V

Thank you for your interest in this paper. sampler=ImbalancedDatasetSampler(train_dataset)。 The Affectnet dataset should be oversampled during training. sampler=ImbalancedDatasetSampler(train_dataset) should be added to train_loader. We apologize that we did not point out that we had given you the wrong impression. We will make changes to the above issues in the coming work.