Did you fine tune every model?

[xuzhang5788]

Thank you so much for your great repo. From your demos, you always set epochs=100 for training. If we want to use some of the models, do we need to fine tune the hyperparameters and retrain them?

[kexinhuang12345]

Hi, ideally, i found depending on different dataset and task, the models are somewhat parameters sensitive. I would suggest do hyperparameter searching for individual use case. We also provide a tutorial on Bayesian Hyperparameter search in the demo: https://github.com/kexinhuang12345/DeepPurpose/blob/master/DEMO/Drug_Property_Pred-Ax-Hyperparam-Tune.ipynb

[kexinhuang12345]

the 100 epochs are following DeepDTA's implementation. But for small dataset, I usually find convergence in 10-20 epochs

[xuzhang5788]

Thanks a lot

[xuzhang5788]

I want to try https://github.com/kexinhuang12345/DeepPurpose/blob/master/DEMO/Drug_Property_Pred-Ax-Hyperparam-Tune.ipynb, but I can't find the data

[kexinhuang12345]

Hi, the data is from mit ai cures, you have to send an email to get it: https://www.aicures.mit.edu/forum

Checkout the open data section in https://www.aicures.mit.edu/data

[pykao]

Hi Kexin, do you have any reference paper for Bayesian hyper-parameter search?

[kexinhuang12345]

Hi, this is a good description of the BO using by Ax platform: https://ax.dev/docs/bayesopt.html