search

chemreps / chemreps

Implementation of various machine learning representations for molecules
MIT License
24 stars 12 forks source link

Bidirectional #72

Open hypnopump opened 4 years ago

hypnopump commented 4 years ago

It's okay to have encoding methods, but decoding ones are also needed for more complex stuff than just classification/regression. Reversible transformations would be awesome.

shivupa commented 3 years ago

Sorry for the delay (feel like I am saying this on every message I write these days..)

This sounds interesting and would definitely been in the scope of this project. Do you have a reference to such methods particularly for molecular applications? So far all we have needed is molecule->encoded representation, but it'd be nice to see it grow.

hypnopump commented 3 years ago

Hi there,

For most of the encodings that are available on this repository, I think it's not possible to find the inverse transform, but there are some end2end autoencoders in which the idea of encoding-decoding is explored. Example here: https://github.com/jrwnter/cddd