search

wengong-jin / icml18-jtnn

Junction Tree Variational Autoencoder for Molecular Graph Generation (ICML 2018)
MIT License
514 stars 189 forks source link

fast molopt #32

Open redBirdTx opened 5 years ago

redBirdTx commented 5 years ago

Dear Wengong,

Does molopt have a faster implementation as well? (Skipping pretrain and use the vae_train.py with annealing strategy)?

Thank you!

wengong-jin commented 5 years ago

Currently not. In terms of conditional optimization, we developed a new method called graph-to-graph translation that performs much better than running gradient ascent over VAE latent space. The paper and github link is here: https://github.com/wengong-jin/iclr19-graph2graph. We suggest you to use that for conditional optimization instead.