search

insilicomedicine / GENTRL

Generative Tensorial Reinforcement Learning (GENTRL) model
596 stars 215 forks source link

What can I do after training the model (VAE and RL)? #6

Closed dakoner closed 4 years ago

dakoner commented 4 years ago

Hi, I have successfully completed both steps of training. It's not clear what the next steps are. I am familiar with previous autoenc/dec models, so I want to (for example), take two SMILES strings, draw a vector between them in encoded space, and decode valid SMILES strings along their interpolation, or, explore around a starting SMILES strings.

If you explain, I can write up some example and send a PR.

dakoner commented 4 years ago

I missed that there is already a notebook called 'examples/sample.ipynb' that does this.

xuzhang5788 commented 4 years ago

@dakoner I ran sample.ipynb after I successfully completed both steps of training, but I got 100 same molecules. I am not sure if my training or RL have problems.

dakoner commented 4 years ago

There's a couple bug in the script, I will send a PR. I

xuzhang5788 commented 4 years ago

@dakoner Did you know if the bugs have been fixed? Many thanks

Bibyutatsu commented 4 years ago

Hi xuzhang5788, If everything is correct and you are sampling correctly then I think this mainly happens when your RL step is overfitted. Please try with less number of num_iterations.