I was playing around with the example training script for CIFAR-10 and noticed that removing test-time augmentation improved my results from around 92-92.5% to 93% accuracy on the test set. It turns out that torch.fliplr isn't behaving as desired on batches of images, in particular it is causing a vertical rather than horizontal flip. Fixing this bug improves the accuracy with TTA to ~94%.
I was playing around with the example training script for CIFAR-10 and noticed that removing test-time augmentation improved my results from around 92-92.5% to 93% accuracy on the test set. It turns out that
torch.fliplr
isn't behaving as desired on batches of images, in particular it is causing a vertical rather than horizontal flip. Fixing this bug improves the accuracy with TTA to ~94%.Here's a demonstration of the basic problem:
And here are the results of running the training script before/after fixing.
Before fixing:
After fixing:
Thanks for the nice library!