Thread by @TheZachMueller

🎉🎉New Blog Alert!🎉🎉

Come learn how to integrate your @PyTorch into @fastdotai with the most minimal fastai possible!

What does this entail?

- Pytorch Datasets
- Pytorch DataLoaders
- Pytorch Models

But, is it really so easy as advertized?

muellerzr.github.io/fastblog/2021/…

Let's find out. In total, this article used four @fastdotai imports to have full access to the training loop and pretty fit outputs:

Optimizer, DataLoaders, Learner, and ProgressCallback

From there, it's really just as simple as passing it all to `Learner` and having it go from there!

It should be noted that if you want @fastdotai's `fit_one_cycle`, `fit_flat_cos`, etc methods, then the import should be adjusted to:

I hope this helps alleviate some confusion and blocking for folks wanting to quickly utilize @fastdotai's training loop, this was honestly the most minimal I could get.

Some folks were likely hoping for me to show how to use a raw Pytorch optimizer in the framework, and sadly it's not quite as simple as plug and play, this is due to @fastdotai's new optimizer setup, which Jeremy discusses in this video:

However what their new optimizer framework has done instead is make Optimizers much more relatable to each other. They noticed where similarities existed, so now SGD's is simply a few lines of code, vs Pytorch's 30 or so

@threadreaderapp unroll