Old but new molecular generator #RDKit #mishima.syk

In this February, we had opportunity to learn how to use molecular generator in Mishima.syk #13. We used REINVENT for generator.

It works on pytorch ver0.3, but raise error when the code run on new version of pytorch.

Deep learning framework is rapidly upgraded. It is difficult to maintenance OSS code I think.

I demonstrated hands-on code on google colab. But recent version of google colab can’t use old pytorch because cuda version mismatch issue.

So I modified original code of REINVENT and tested it now.

My code is uploaded my repo URL is below.

I modified model.py and train_prior.py. And the code is working now in my PC.

My env is..

CUDA 10.1
GPU GTX 1650
cudatoolkit 10.1.168
cudnn 7.6.0
pytorch 1.3.0
torchvision 0.4.1
rdkit 2019.03.04

This is not enough for efficient learning because my GPU don’t have enough memory but the code work.
Any comment or feedback will be greatly appreciated.