CIFAR-100 dataset in train_cifar.py has only 20 classes.

[yimiandai]

CIFAR-100 dataset in train_cifar.py has only 20 classes.

[YimianDai]

My god... Many thanks for this important information. It is a terrible mistake. I didn't notice this fine_label=True thing.

I will soon delete my thing on paperswithcode and modify the paper, and I will re-run the CIFAR-100 experiment as soon as possible.

Many thanks for your information.

[takedarts]

It's my pleasure if I can help with your work.
I am looking forward to new results.

[YimianDai]

Many thanks!

I will try to report the new CIFAR-100 results training from scratch in 24 hours.

Luckily, the ablation study results still hold since they are all compared in 20 superclasses (I will explain it in the paper that the accuracies are on 20 superclasses), and the ImageNet experiments has no such problems.

[YimianDai]

Emmm... I have submitted the Machine Request to the HPC for nearly 24 hours, but it is still in queue...

I will keep updating the progress.

Edit: I have applied one machine, and the new experiment is running.

[YimianDai]

Hi. The accuracy of AFF-ResNeXt-47 training from scratch on the fine-grained CIFAR-100 is 85.4%. The trained models and log file is uploaded.

[takedarts]

I trained ResNet-32 (not AFF-ResNet-32) in your implementation with fine_label=True on my computer with a 1080Ti GPU.
The validation accuracy is 82.6% which is close to accuracy of WideResNet-34k4 in my own implementation.

Your AFF-ResNeXt-47 result (85.4%) also seems to be reasonable.