Transition to FastAI Version 2
vc1492a opened this issue · 5 comments
FastAI now has a version 2 available!
See this blog post: https://www.fast.ai/2020/08/21/fastai2-launch/
I'd like to upgrade to FastAI 2 once the rest of the experimental pipeline is in a more mature state.
The difficulty here - in my case - is that the computer cluster I am using for model training is a PowerPC architecture, meaning that Torch and other libraries need to be installed from IBM's repository. Their team has done a good job updating this in the past, but it is never as up to date or as easy as using x86 architectures. While FastAI 2 could easily be installed, we need to ensure that the underlying Torch and related (Numpy etc.) libraries have the appropriate versions available in the IBM catalogue.
I was able to install PyTorch 1.7.0 and FastAI 2.1.8 on my available computing environment for another project. Currently remains untested, but this is good news for us as many of the code examples and newer approaches are best utilized in the newer versions of these packages 🎊
Through this example, can confirm Torch 1.7.0, torchvision, and FastAI version 2 all work in the environment I have access to.
Given that some of the alternative methods (e.g. using an image-based method to anom detection) requires FastAI version 2, it may make sense to explore making the transition in the near future.
Marking this issue as in progress as the image based approach code is based on FastAI 2. @hamlinliu17 let's track your environment setup and progress in running the code here.
@vc1492a I can run the code in image_predict.ipynb with ease so I think I am set on the new dev environment. I have FastAI 2.2.2 with PyTorch 1.7.1. Hopefully the discrepancy in fastai versions will not cause any problems
@hamlinliu17 fantastic! Closing...