search

rgiordan / zaminfluence

Tools in R for computing and using Z-estimator approximate influence functions.
Apache License 2.0
95 stars 10 forks source link

Hand-code IV and regression derivatives in R so Python is unnecessary #1

Closed rgiordan closed 3 years ago

grantmcdermott commented 3 years ago

I mentioned this to Rachael the other day on Twitter, but might be useful to add here:

Since autograd appears to be the primary reason for the Python dependency, I'd recommend taking a look at the new(ish) torch implementation for R. This has direct bindings to the underlying libtorch C++ libraries, so no Python install required. The syntax is also pretty similar to the PyTorch implementation, which I'm sure you are familiar with.

More importantly, thanks for putting this package together! I still have to sit down and read the actual paper, but looks like it's going to be a big deal.

rgiordan commented 3 years ago

Thank you for the pointer! I didn't know about that, and I agree, it might be a good way to go.

rgiordan commented 3 years ago

In case anyone else is watching this issue, my plan remains to get the IV and OLS derivatives implemented in R first since that's relatively easy (just a matter of finding the time at the moment). I'll look into alternative autodiff frameworks a little later.

rgiordan commented 3 years ago

This is now done in https://github.com/rgiordan/zaminfluence/pull/11. I'm going to make a new issue to look into Torch.