Closed yuanqing-wang closed 2 years ago
I generally agree with the structure @jchodera has proposed but I would be careful trying to provide abstractions for training / dataset partitioning / evaluating. These components might look drastically different from experient to experiment and I think it would make sense to leave these in the scirpts to avoid over-energineering.
There's no need to refactor everything at once, either! Simplifying/streamlining the components that are most critical right now is all that is needed, and the approach can be gradual and iterative.
@jchodera @yuanqing-wang I think I've implemented the structure you're talking about, and I've updated the README to show how the new interface is used. I haven't done anything with data yet, but I've written a torch
Dataset
subclass for the SPICE DES370K dataset elsewhere that I can add in eg an mtenn.datasets
submodule.
Looks great. Thanks!
I think this looks great!
Description
Provide a brief description of the PR's purpose here.
Todos
Notable points that this PR has either accomplished or will accomplish.
Questions
Status