[Bug]: Input transforms for deterministic models

[jduerholt]

What happened?

When using input transforms in deterministic models, one always gets a RuntimeWarning with the following message: "Could not update train_inputs with transformed inputs since GenericDeterministicModel does not have a train_inputs attribute. Make sure that the input_transform is applied to both the train inputs and test inputs." It stems from

botorch/botorch/models/model.py

Line 233 in 15a87e1

"Could not update `train_inputs` with transformed inputs "

This warning seems to further propagates into gen_candidates_scipy and produces an OptimizationWarning with "Optimization failed" which yields restarts of gen_candidates_scipy and makes it slower.

I would recommend to not raise this warning for deterministic models. Any idea on how to do this?

Please provide a minimal, reproducible example of the unexpected behavior.

from botorch.models.deterministic import AffineDeterministicModel
from botorch.models.transforms.input import FilterFeatures
import torch

deterministic = AffineDeterministicModel(
    a=torch.tensor([[-1.0, 2.0]]),
    b=0.1
)

deterministic.input_transform = FilterFeatures(
    feature_indices=torch.tensor([0, 2], dtype=torch.long)
)

deterministic.eval()

Please paste any relevant traceback/logs produced by the example provided.

BoTorch Version

recent main

Python Version

No response

Operating System

No response

(Optional) Describe any potential fixes you've considered to the issue outlined above.

No response

Pull Request

Yes

Code of Conduct

• I agree to follow BoTorch's Code of Conduct

[esantorella]

Thanks for reporting this! Could you clarify "This warning seems to further propagates into gen_candidates_scipy and produces an OptimizationWarning with "Optimization failed" which yields restarts of gen_candidates_scipy and makes it slower." ? Do you have a repro of that behavior or a guess of why that is happening? We certainly should fix the warning from deterministic models, but I'm also wondering if the retry logic in the optimization code should change.

[jduerholt]

I think, I was wrong for the propagation into gen_candidates_scipy. Other acqf optimization related warnings are triggering the restart of gen_candidates_scipy. But gen_candidates_scipy is then printing out all warnings that have been acculated so far. And this leads to my misconception, there then also the warning from above is mentioned, despite not triggering the restart. So it is only about fixing the warning for the deterministic model. Any idea how? I can file a PR for this.

[esantorella]

To make the warning go away, DeterministicModel could override Model._set_transformed_inputs -- so there would be DeterministicModel._set_transformed_inputs(self): return None.

[jduerholt]

I will file a PR over the next days for this.