[GENERAL SUPPORT]: Improving run time of BO iterations

[RoeyYadgar]

Question

Hi, I'm trying to use BO on a problem in which the sampling time (time it takes to evaluate the function to be optimized at a specific arm) is relatively short. In this case the BO iteration time can become the computational bottleneck, and I'd like to use Ax's framework without some of the additional computations (that aren't strictly necessary for the BO) it performs

More specifically, I've seen that considerable part of the run time per iteration is spent on get_fit_and_std_quality_and_generalization_dict (https://github.com/facebook/Ax/blob/main/ax/modelbridge/cross_validation.py#L409) which is called by ModelSpec.gen (https://github.com/facebook/Ax/blob/main/ax/modelbridge/model_spec.py#L239).
I'd like to skip the computation of the cross validation each iteration as I'm not using it anyway, I tried looking for some flag in ModelBridge (or anywhere else for that matter) that allows me to skip that but wasn't able to find one. Is there a way to do that or should I approach this in a different manner (Like inhereting from ModelSpec and overwriting this method)?

Additionally, I've seen that TorchModelBridge._gen also computes best_point at each iteration (https://github.com/facebook/Ax/blob/main/ax/modelbridge/torch.py#L729) which I would also like to skip, but not sure if there's a simple flag that allows me to do so (This computation is very fast when using best in sample point so that's of less importance to me, however I would sometimes to use a custom TorchModelBridge that computes best_point by optimizing the posterior mean, and I'd still want to have a flag that allows me to skip that computation in _gen)

I've also seen that Experiment.fetch_data (https://github.com/facebook/Ax/blob/main/ax/core/experiment.py#L575) takes quite a bit of time on each iteration. but I wasn't able to understand what it really does and what makes it computationally expensive?

Below is a profiler on a single BO iteration using Models.BOTORCH_MODULAR with 250 samples.

Thanks!

Please provide any relevant code snippet if applicable.

No response

Code of Conduct

• I agree to follow this Ax's Code of Conduct