Pardon me, but as of my actual understanding of Fast.ai library (not good not terrible) that’s impossible.
pretrained=False means random initialization of the model’s weights (instead of using the ones obtained with several days of training of ImageNet) so it’s totally unrealistc that such a network can converge in just one epoch.
I’m writing this because, no matter what I try, I can’t go past 90% accuracy with a ResNet-50. So your strategy lured me (and I admit it could also pay off if done properly) but not in one epoch. Moreover in your notebook you train for one epoch (with
bs=16) and then you do an
unfreeze() on an already unfrozen model (because of the
pretrained=False) and then train again for just two epochs with a very low
max_lr. So it’s totally impossible that the notebook you linked yields 96% accuracy, sorry. I’m writing this also to prevent other people from wasting time on this.