CE loss on CUB-200-2011 does not reach reported accuracy

[mouhamadoufall2023-gif]

Dear Binguan,

I hope you are doing well.

I am currently using the official MbLS GitHub code, but I am struggling to reproduce the CE results reported in the paper on the CUB-200-2011 dataset. Specifically, for CE loss, I only achieve around 62% accuracy, whereas the paper reports 73%.

I may have missed an important parameter or setting, and I would greatly appreciate your guidance on how to reproduce the reported performance.

Here is exactly what I am doing:

Dataset: CUB-200-2011 with 5,994 training and 5,794 test images, 200 bird species.

Training augmentation: resize images to 256×256, then randomly crop 224×224 patches or their horizontal flips.

Model: ResNet pre-trained on ImageNet.

Training: 200 epochs, batch size 16, SGD with momentum 0.9.

Learning rate: 0.1, decayed by a factor of 0.1 every 80 epochs.

Margin m: optimal m found on Tiny-ImageNet (no validation set for CUB).

Could you please advise if there are any additional settings, preprocessing, or tricks required to reproduce the CE accuracy reported in the paper?

Thank you very much for your time and help.

Best regards,
Mouhamadou Fall