standardized models

[rkurchin]

Add a new file (presumably just models.jl) that defines some standard model architectures where the user can feed in the hyperparameters and the whole model gets built. Asking @DhairyaLGandhi about the best approach, e.g. just write a function that returns the Chained model, or define a new sort of struct that can act on inputs?

So far I'm only envisioning one, namely the CGCNN archetype from Tian's paper. In that case the necessary parameters would be:

• input feature length
• number of convolutions (optionally, type of activation)
• output feature length
• type of pooling
• pooled crystal feature length
• number of hidden Dense layers (optionally, type of activation)
• length and type of activation for output (i.e. regression or classification)

We can default to the optimized values from the formation energy case or some nearby round numbers to those...

[DhairyaLGandhi]

Are there cases where you have a class of models that you can mix and match pieces from other established models from? I'd use them for easier dispatch if it makes sense and if there are certain properties which are needed to be defined on them. Otherwise, keeping it simple and generic with functions makes most sense.

[rkurchin]

I think for some of the stuff I envision adding later (like graph attentional models) that might be true, but for now probably the function approach makes the most sense. Will start to work on this later today!