Runyu-Zhang
commented
3 months ago I reformatted the code.
Code
def create_gp_model_kwargs():
return {'surrogate' : Surrogate(botorch_model_class = train_default_gp_model()),
'botorch_acqf_class': qNoisyExpectedHypervolumeImprovement}
def train_default_gp_model(self):
train_x, train_y = initialize_training_data()
train_x = normalize(X=train_x_motors, bounds=params_bounds_tensor)
models = []
for i in range(train_y.shape[-1]):
models.append(SingleTaskGP(train_X = train_x,
train_Y = train_y[:, i].unsqueeze(-1),
outcome_transform = Standardize(m = 1)))
model = ModelListGP(*models)
mll = SumMarginalLogLikelihood(model.likelihood, model)
fit_gpytorch_mll(mll)
return model
parameters_list = [dict(name = param,
type = 'range',
bounds = range_values,
value_type = 'float')
for param in all_parameters]
objectives = {obj: ObjectiveProperties(minimize=True, threshold=hypervolume_reference_point)
for obj in all_objectives}
steps = [GenerationStep(model = Models.SOBOL, # Add 1 sobol to avoid "ax.exceptions.core.DataRequiredError: StandardizeY transform requires non-empty data."
num_trials = 1),
GenerationStep(model = Models.BOTORCH_MODULAR,
num_trials. = num_bo_trials,
model_kwargs = create_gp_model_kwargs())]
# Create Ax Client and optimize:
ax_client = AxClient(generation_strategy = GenerationStrategy(steps=steps),
random_seed = fixed_seed,
verbose_logging = True)
ax_client.create_experiment(name = 'Ax_Modular',
parameters = parameters_list,
objectives = objectives,
overwrite_existing_experiment = True)
for i in range(BO_trials):
suggested_values, trial_index = ax_client.get_next_trial()
# Throw the error after one Sobol trial
loss_values = evaluate_suggested_values(suggested_values=suggested_values)
ax_client.complete_trial(trial_index=trial_index, raw_data=loss_values.copy())
Cesar-Cardoso
saitcakmak
RyanSaat
Hello, thank you all for creating and maintaining this wonderful Ax library for BO. I have a question regarding pre-fitting a GP model and pass it to Ax Client for multi-objective optimization. Similar questions can be found: https://github.com/pytorch/botorch/issues/2299, https://github.com/pytorch/botorch/issues/1750. I followed @Balandat suggestion https://github.com/facebook/Ax/issues/1647#issuecomment-1581428813, to use the Modular BoTorch Model interface to pass in a new custom BoTorch class together with a pre-fitted model (Surrogate?). I create one Sobol step in GenerationStrategy just to avoid "ax.exceptions.core.datarequirederror:
standardizeytransform requires non-empty data.". However, I still got error:which is likely due related to botorch_model_class in Surrogate() cannot be a ModelList. Is there a way to do this? Basically pass a pre-fitted multi-objective GP model to Ax and continue the MOBO process? My code implementation:
`def create_gp_model_kwargs(self): return {'surrogate' : Surrogate(botorch_model_class = train_default_gp_model()), 'botorch_acqf_class': qNoisyExpectedHypervolumeImprovement}
def train_default_gp_model(self): train_x, train_y = initialize_training_data() train_x = normalize(X=train_x_motors, bounds=params_bounds_tensor) models = [] for i in range(train_y.shape[-1]): models.append(SingleTaskGP(train_X = train_x, train_Y = train_y[:, i].unsqueeze(-1), outcome_transform = Standardize(m = 1))) model = ModelListGP(*models) mll = SumMarginalLogLikelihood(model.likelihood, model) fit_gpytorch_mll(mll) return model
parameters_list = [dict(name = param, type = 'range', bounds = range_values, value_type = 'float') for param in all_parameters]
objectives = {obj: ObjectiveProperties(minimize=True, threshold=hypervolume_reference_point) for obj in all_objectives}
steps = [GenerationStep(model = Models.SOBOL, # Add 1 sobol to avoid "ax.exceptions.core.DataRequiredError:
StandardizeYtransform requires non-empty data." num_trials = 1), GenerationStep(model. = Models.BOTORCH_MODULAR, num_trials. = num_bo_trials, model_kwargs = create_gp_model_kwargs())]ax_client = AxClient(generation_strategy = GenerationStrategy(steps=steps_bom), random_seed = fixed_seed, verbose_logging = True) ax_client.create_experiment(name = 'Ax_Modular', parameters = parameters_list, objectives = objectives, overwrite_existing_experiment = True)
for i in range(BO_trials): suggested_values, trial_index = ax_client.get_next_trial() loss_values = evaluate_suggested_values(suggested_values=suggested_values) ax_client.complete_trial(trial_index=trial_index, raw_data=loss_values.copy())`