Closed epistoteles closed 1 year ago
Hi @epistoteles ! Thanks for this contribution. Just as a sanity check, can you make sure the CIFAR accuracy is still roughly the same as before using the default training script? I want to make sure the hyperparameters there aren't overfit to the wrong normalization.
Ran with old: test accuracy: 93.8% With new/correct mean and std: test accuracy: 94.0%
Varies a bit over runs of course, but seems fine @andrewilyas
Thanks so much! Merging this now
For some reason, the CIFAR10_STD used in the example differed significantly from the actual standard deviation of the CIFAR10 train dataset. I corrected both the MEAN and STD with 3 decimal places accuracy.
I calculated the values as seen here: https://gist.github.com/epistoteles/c35bd5154a036748651d8caca11a7efe