Reduce of accuracy after update to version 0.7

[sh1ng]

Hi,

noticed that after upgrade to version 0.7 accuracy of trained models from 0.63 reduces on a few percentage. And it would be not so critical, but retraining from scratch on version 0.7 can't get previous results. Unfortunately can't provide some code as efficient-net is only small part of the system. We use b3 network and crop size 288.

Could you provide an example how to change image size? It's a single part that differs our code from your examples.

[lukemelas]

Thanks for the issue. This seems important and I will investigate.

For the image size, it's just changed with the normal PyTirch data transforms (i.e. Resize(300))