Fabian Pedregosa

taking this on during the sprint

Update: I wrote an alternative to flax_resnet.py that uses GroupNorm instead of BatchNorm: https://gist.github.com/fabianp/7c6c2da33156010d5d822bdad997c942 It doesn't seem to do better on the test set accuracy, which is stuck around 70%,...

CC @fllinares who is facing similar issues

This is the train/test accuracy on CIFAR10 with the latest script with GroupNorm ( https://gist.github.com/fabianp/7c6c2da33156010d5d822bdad997c942) 

thanks @ita9naiwa for looking into it. I think you might be running on MNIST, where accuracy is indeed at 99%. Can you run it in the cifar dataset, for this...

BTW by validation accuracy I mean the "Test acc" reported

thanks @ita9naiwa , I'll try this, this is helpful!

There are still some things I don't understand. For example, this code (https://github.com/fabianp/pytorch-cifar) achieves 80% test set accuracy after just 5 epochs in Pytorch using constant-learning rate Adam and without...

Thanks a lot @ita9naiwa , this is extremely helpful. I've been busy for the last couple of days but will get back to this and try these suggestions out as...

Coming back to this issue. I don't think it's a dataset loading issue. Here I adapted the current example to load the dataset using pytorch, including data augmentation that yields...