search

Ramprasad-Group / polygnn

polyGNN is a Python library to automate ML model training for polymer informatics.
Other
32 stars 5 forks source link

Add example_predict.py file. #15

Closed rishigurnani closed 1 year ago

rishigurnani commented 1 year ago

The goal of this PR is to create an example_predict.py file to illustrate how to just do prediction using previously-trained models.

rishigurnani commented 1 year ago

So far, I have written a partial example_predict.py file. While writing it, I realized that example.py was broken. It appears that empty graph_feats tensors are not getting added to Data objects when users provide data without a "graph_feats" column. I am writing a fix locally that updates the prepare.py file in polygnn_trainer. Fixed by https://github.com/rishigurnani/polygnn_trainer/commit/1ab50cfec8345d6c2fa84d7c638d77b575e2d7ae.

EDIT: After "fixing" the graph_feats issue, I am seeing that the ensemble predictions in example.py and example2.py are low R^2. However, the models train with high R^2. Fixed by https://github.com/rishigurnani/polygnn_trainer/commit/f62c9010e909520b4b4423b0d3bef7acf0ac0fce and https://github.com/rishigurnani/polygnn_trainer/commit/f3eb5ba3ef73a746aed2121ddac2926cc942d83d.