search

MolecularAI / Chemformer

Apache License 2.0
211 stars 36 forks source link

Added `model.eval()` and `torch.no_grad()` to disable dropout/gradient calculation during inference #17

Closed salman-moh closed 1 year ago

salman-moh commented 1 year ago

Referencing #13 Simply put, I was getting different predictions for the same SMILE input, turns out this additional resolved that issue.

salman-moh commented 1 year ago

Had to undo the other changes to put a PR for 1 change at a time.

EBjerrum commented 1 year ago

Nice find. How bad did it affect the overall metrics? e.g. R2, RMSE?

salman-moh commented 1 year ago

I did not re-train the model as we changed our focus for now... but during inference there was a +/- difference of 0.3-0.4 on the single ESOL predictions. I'll report back once I get on this.

On Wed, Nov 30, 2022 at 6:55 PM Esben Jannik Bjerrum < @.***> wrote:

Nice find. How bad did it affect the overall metrics? e.g. R2, RMSE?

— Reply to this email directly, view it on GitHub https://github.com/MolecularAI/Chemformer/pull/17#issuecomment-1332150749, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHRKOA7YPUOJSYVO2OUDTK3WK5IVHANCNFSM6AAAAAASIEB77U . You are receiving this because you authored the thread.Message ID: @.***>