Returning predictions

[PatWalters]

I've been hacking on the code and trying to modify model.train_and_test to return the predictions. I thought I could just put

pred = trainer.predict(fastprop_model, test_dataloader)

in place of

test_results = trainer.test(fastprop_model, test_dataloader, verbose=False)

However, the correlations I calculate from the values returned in pred are much lower than those returned in test_results. What am I missing?

[JacksonBurns]

Ah, I've built a foot gun here - the predict method in fastprop automatically applies the feature scaling to its inputs, which is needed for inference when loading from a checkpoint but breaks the workflow in your case.

You can comment out the part of the predict method responsible for this here to get it to work now:
https://github.com/JacksonBurns/fastprop/blob/main/fastprop/model.py#L174-L175

There's a clever solution to make this work more generally, I am sure.