[Chapter 3: Image Data Augmentations] Xavi's FastAI example works better than the TF version. Investigate! #3
Description
@XanaGA wrote...
I have been working on the PyTorch version of the notebook. I am using Fastai, which is a library that has a workflow very similar to Keras. In the first notebook I tried to imitate the training procedure of your implementation. But Fastai has a "special" way to train the models (using lr_finder and the fit_one_cycle policy). So I made a second one following those recommendations. The second one gets better performance but is not as similar as the codelab from the book. It also doesn't get so much improvement by using the data augmentation. If you think it is correct maybe I can upload to the book's repo the one that is more similar and publish in my blog the other one.