search

coarse-graining / cgnet

learning coarse-grained force fields
BSD 3-Clause "New" or "Revised" License
56 stars 26 forks source link

Model Averaging #133

Closed nec4 closed 4 years ago

nec4 commented 4 years ago

Hey guys. In the vanilla CGnet paper, the models were averaged over 5-folds during for each hyperparameter determined through cross validation. These average models were then used to generate CG simulations for comparison to atomistic simulations (our extended ground truth). I propose to incorporate this functionality into our simulator, wherein users may pass a list of models. The force update will then be an average over all the forces predicted by each model. What do you guys think? The only foreseeable downside I see is that each model will need to occupy space on the GPU, so this approach many not be memory efficient.

brookehus commented 4 years ago

@coarse-graining/developers I just tried to update issue "assignments" to the ones we're actively working on. Is this one @nec4's?

nec4 commented 4 years ago

I'm not sure - but since I brought it up, it can be assigned to me. I was thinking of incorporating it into the simulation tools. I think I can coordinate with @Dom1L on what to do first.

Dom1L commented 4 years ago

I was thinking since it's part of the Simulation class rework we can probably do it together. Once I make the PR for it, we can take a look and both work on it :)

nec4 commented 4 years ago

Closing after merging #179.