A scikit-learn compatible neural network library that wraps pytorch

2 Likes

Yeah…Interesting project. In a lot of ways fastai also adopts the keras and sklearn API of fit, predict. But sticks with the PyTorch way of DataLoaders and Transformers. One key difference is, fastai adds functionality to PyTorch (lr_find, cycles etc), whereas this repo, provides a wrapper to make it easier with fewer lines of code. But definitely great to see so many projects inspired by Scikit-learn. Infact, Keras has a scikit-learn wrapper - https://keras.io/scikit-learn-api/

One thing missing in fastai is GridSearch of Model Parameters that this wrapper and keras-scikit-learn has, would be good to have a function also available in fastai.

7 Likes

That would be an easy addition I think, if anyone feels like having a go at it… :slight_smile:

1 Like

One thing missing in fastai is GridSearch of Model Parameters that this wrapper and keras-scikit-learn has, would be good to have a function also available in fastai.

That would be an easy addition I think, if anyone feels like having a go at it… :slight_smile:

I wonder if it’s worth adding a scikit-learn wrapper for fastai?

I did this walkthrough on a parameter tuning / stacking tool called Xcessiv yesterday and thought it was really cool. I feel its organized visual interface would save me time down the road and help eke out better results.

The only catch is you need to give it a scikit-learn compatible wrapper to your model, which I don’t think fastai has yet. So, I added an item to the “fastai features wishlist”

I suppose it’s possible that people don’t do much parameter tuning for deep learning due to the time needed for training, and so stacking/tuning tools aren’t considered critical processes. But, we always do some, and any amount of this in my opinion is tricky to keep track of. I like the way Xcessiv handles this visually.