How to get trained GP model from Service API for predictions and plotting

hi,

i m trying to train a basic 1d GP on data from an excel sheet and then continue with .get_next_trial() (and plot after each iteration).

ax_client = AxClient()
ax_client.create_experiment(name=..., parameters=..., objective_name=..., minimize=True)

for _, row in df.iterrows():
    # format parameters and result
    p, trial_index = ax_client.attach_trial(parameters)
    ax_client.complete_trial(trial_index=trial_index, raw_data=result)

However, prediction and plotting is not possible, as ax_client.get_best_parameters() is giving ax.service.utils.best_point: Could not use model predictions to identify best point, will use raw objective values. and feeding ax_client.generation_strategy.model into render(...) gives NotImplementedError: RandomModelBridge does not support prediction.

A look at ax_client.generation_strategy.model shows None.

I can set up a model via model = get_GPEI(ax_client.experiment, ax_client.experiment.lookup_data()), which then can be plotted (and shows correct data).

How can I run the service api on a pre-trained GP and from there do standard bayesian optimization with model update?

best regards

[lena-kashtelyan]

So ax_client.generation_strategy.model gets set after the first call to ax_client.get_next_trial(). If you reverse the order of operations to the following, your example should work:

print('next:', ax_client.get_next_trial())
print('model:', ax_client.generation_strategy.model)
print('best:', ax_client.get_best_parameters())

model = ax_client.generation_strategy.model
render(plot_slice(model, 'x', 'y'))

[btcorgtfo]

hi!

thanks a lot for the A+ clarification. The changed order now has the hoped for / expected behaviour 😃