Closed ChunyunShen closed 4 years ago
all you need to do is change the backbone in CUB-200-2011.py, and change the output channel of the backbone to 600.
Hi @dongliangchang According to your advise, I have trained ResNet18 from scratch based on this repo. The performance has improved! However the results I obtained (29.89 with CE loss, 51.97 with MC loss) are still much worse than reported in the paper (45.70 with CE loss, 59.41 with MC loss). Maybe the hyperparameters are not optimal... I am very confused.
I will release the related codes within one month. Thanks for your attention.
I will release the related codes within one month. Thanks for your attention.
Hi, can you please release the related codes? I am very interested in your research and need the codes for researching.
Sorry for this. As I need to pass the Duolingo English Test this month, I don't have enough time to update this code. But the only difference between the MC_loss with VGG16 and the MC_loss with ResNet18 is the backbone, just change it for test. Maybe you can refer to this code (https://github.com/Kurumi233/Mutual-Channel-Loss).
After I pass the Duolingo English Test, I will update this code.
Sorry for this. As I need to pass the Duolingo English Test this month, I don't have enough time to update this code. But the only difference between the MC_loss with VGG16 and the MC_loss with ResNet18 is the backbone, just change it for test. Maybe you can refer to this code (https://github.com/Kurumi233/Mutual-Channel-Loss).
After I pass the Duolingo English Test, I will update this code.
Great! Thank you for your help. I Sincerely wish you pass the Duolingo English Test.
@ChunyunShen @zhangchuanyi96 Hi, I have update the related code. The result of this code is 59.66 ± 0.54 (59.41, 59.16, 60.41).
Hi authors: Could you release codes for training with ResNet18 (trained from scratch)? I find it is hard to obtain the reported scores (45.7 with CE loss) training with the hyperparameters provided in your paper. Maybe I miss some critical issues.