How to access lengthscales from default GP kernel (Matern or RBF) in BoTorch 0.16.1?

[AM4691]

Hello,

I’m using BoTorch 0.16.1 (with Ax 1.2.1) to perform Bayesian Optimization. My goal is to extract the ARD lengthscales from the default GP kernel to analyze parameter importance (as 1/lengthscale).

Questions:
1. What is the default kernel type in BoTorch 0.16.1 for SingleTaskGP? Is it Matern 5/2 or RBF?
2. How can I access the lengthscale values from the trained GP model?

I’ve tried accessing the model via client.experiment, generator_run, and generation_strategy, but none have a _model or model attribute.
I checked this discussion, which explains how to access lengthscale in pure BoTorch, but I’m using Ax.

Thank you for your help!
Best regards

Relevant Code:

import torch

# Setup AxClient
client = AxClient(random_seed=1, verbose_logging=False)
client.create_experiment(
    name="optim_bo",
    parameters=[
        {"name": "param1", "type": "range", "bounds": [0.0, 1.0], "value_type": "float"},
        # ... (7 parameters total)
    ],
    objectives={"objective": ObjectiveProperties(minimize=False)},
)

# Run 3 BO iterations (minimum to fit GP)
for _ in range(3):
    params, trial_idx = client.get_next_trial()
    client.complete_trial(trial_idx, raw_data={"objective": (1.0, 1e-3)})

# Attempt to access lengthscales (FAILS)
try:
    last_trial = list(client.experiment.trials.values())[-1]
    gp = last_trial.generator_run._model.model.model  # Fails: no '_model' in generator_run
    ls = gp.covar_module.base_kernel.lengthscale
except AttributeError as e:
    print(f"Error: {e}")  # 'GeneratorRun' object has no attribute '_model'

[hvarfner]

Hi @AM4691 ,
Thanks for reaching out!

Two issues:

1. The Ax attributes have changed
2. The BoTorch attributes have changed

(1) We cleaned up the Ax sequence of attributes. Previously, a lot of things used to be called model. Now, the sequence is
Adapter --> Generator --> Surrogate --> Model
where the Model is the actual BoTorch object you are interested in. So, in your case you should go with
gp = client._generation_strategy.adapter.generator.surrogate.model

(2) Since we dropped the ScaleKernel, the SingleTaskGP no longer has a base_kernel by default. Thus, you should just subtract that attribute:
ls = gp.covar_module.base_kernel.lengthscale --> ls = gp.covar_module.lengthscale

Unfortunately, your coding agent of choice still has a pretty difficult time understanding this, but we'll get there!

[AM4691]

Thank you, @hvarfner!
This clarification was exactly what I needed.
I love that the solution keeps improving, even if it occasionally means chasing down breaking changes.
Thanks again for the help!

[hvarfner]

Happy to help!