For tinyimagenet, the test acc is very low, 0.009 about moon

[QinbinLi]

Hi @rambo-coder ,

I can achieve a normal test acc by running your commands with torch 1.0.0 on a V100 GPU. Can you double-check your dataset (i.e., tiny-imagenet-200) and run again? You can also try the other algorithms such as FedAvg to see whether the acc is normal or not. Btw, there is a typo in your commands, "noiid"->"noniid".

[rambo-coder]

Hi @QinbinLi,

Thank you so much for your reply.

I am wondering whether the dataset from kaggle is the same as the dataset from http://cs231n.stanford.edu/tiny-imagenet-200.zip. It is not available to download the dataset from kaggle now. I used tiny-imagenet-200 from stanford-cs231 for training.

[QinbinLi]

Hi @rambo-coder ,

What you used is the correct dataset. Did you preprocess the validation folder to move the images to subfolders by the classes (here is an example)?

[rambo-coder]

Hi @rambo-coder ,

What you used is the correct dataset. Did you preprocess the validation folder to move the images to subfolders by the classes (here is an example)?

Thanks, @QinbinLi , It works now. That is the reason !