I am trying to port my fastai pipeline to Pytorch for testing some new features. However, I am unable to reproduce my results with FastAI in pytorch.
I have been able to successfully copy the model to PyTorch. However, I feel that there are differences in the way the image data is preprocessed. I create a ImageDataBunch by using the from_df method and keep everything else as default.
Is anyone aware of what preprocessing needs to be done to be comparable to FastAI’s preprocessing?