duncanwp / esem Goto Github PK

Basically a wrapper around the scikit-learn implementation.

Notes:

• Random Forests have relatively few hyperparameters, but it should be possible to pass these as keywords to the model.
• The predict method will just return a mean, with variance=None, similar to the neural net.

I think the README needs to be updated now as well? Should have thought of this during #11 but nevermind ...

Wow! Great project - thanks for your hard work.

Is there a way to save a trained model, load it into a new notebook, and run inference? Apologies if this is documented somewhere.

e.g.

rf_model = rf_model(X_train, Y_train)
rf_model.train()
rf_model.save("rf_model.pb") # <---- Is there anything like this?

and then in a new notebook

rf_model = esem.open("rf_model.pb") # <---- Is there anything like this?

Create a very simple toy-model (ideally using the existing testing routines), plot it, emulate it and then calibrate some parameters.

I recently came across this inference technique used in astrophysics that might be useful in our setting: https://academic.oup.com/mnras/article/398/4/1601/981502.

This should be easy for GPs where we can just ignore masked values when flattening (although reshaping at the end will need to be done with care). I'm not sure what support keras has for masked values (or NaNs)

Need to be able to specify the kernel(s) and noise variance (or nugget) for GPs. Maybe the NN structure?

It's not clear exactly how far to abstract this. In principle the model class should just be a constructor for the model anyway... Maybe the NN and GP classes are just examples of a model which needs spatial structure and one that doesn't.

Hi

I'm not able to install ESEm via pip and because it cannot find compatible version of tensorflow. I believe that requirement are to loosely specified. However it is essentially impossible for me to guess which version of tensorflow, numpy, and python that is supported.