Based on some of the initial results, the hyperparameter optimization doesn't seem to

<a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url="/us

Maybe restrict the parameter search to fewer parameters about crabnet-hyperparameter HOT 1 CLOSED

sparks-baird commented on September 23, 2024

Maybe restrict the parameter search to fewer parameters

from crabnet-hyperparameter.

Comments (1)

sgbaird commented on September 23, 2024

@mliu7051 we can use the following to figure out which parameters are most important in addition to using our own intuition.

ax.plot.feature_importances.plot_feature_importance_by_feature(model)

facebook/Ax#426 (comment)

I think we can load the experiment and then "reload" the model from the saved experiment. The full workflow is something like the following (haven't tested this):

from os.path import join
from ax.storage.json_store.load import load_experiment
from ax.modelbridge.factory import get_GPEI
from ax.plot.feature_importances import plot_feature_importance_by_feature

fpath = join("experiments", "experiments0.json")
experiment = load_experiment(fpath)
model = get_GPEI(experiment, experiment.fetch_data())
plot_feature_importance_by_feature(model)