search

GRAPH-0 / JODO

Learning Joint 2D & 3D Diffusion Models for Complete Molecule Generation
MIT License
40 stars 8 forks source link

Colab GPU #2

Closed SFN1373 closed 11 months ago

SFN1373 commented 1 year ago

Excuse me, how much memory does this code need and how many hours does it take? I mean in Colab.

GRAPH-0 commented 1 year ago

I haven't tried it on Colab. The version I released does require a lot of memory and training time. You can train a smaller model on QM9 to adapt for your GPU, such as reducing the number of layers or hidden dimensions. To speed up inference, you can reduce sampling steps.

GRAPH-0 commented 1 year ago

I have provided the generated molecules at https://github.com/GRAPH-0/JODO/tree/5c3e37d7ae6928a17cccff07861a9c9e15efbfb0/rdkit_mols . For evaluation of the generated molecules, you can refer to https://github.com/GRAPH-0/JODO#evaluation .

SFN1373 commented 12 months ago

What kind of GPU is needed to train diffusion model?

GRAPH-0 commented 12 months ago

What kind of GPU is needed to train diffusion model?

RTX3090 for QM9 Geom-Drugs need larger GPU memory, e.g. V100, A40

SFN1373 commented 12 months ago

Thank you.

How long does it take?

GRAPH-0 commented 12 months ago

It depends on the number of epochs. I think within 2 days, the model could converge in QM9