Multimodels using partial or full data in optimization

[An-xu]

Hi,
I would like to ask about an error I encounter during using botorch.
Below here is my code.
def initialize_model(x, y):
models = []
for i in range(y.shape[-1]):
train_objective = y[:, i].unsqueeze(-1)
if i == 0:
input_data = x[:, 0:2]
models.append(
SingleTaskGP(input_data, train_objective)
)
else:
models.append(
SingleTaskGP(x, train_objective)
)
model = ModelListGP(*models)
mll = SumMarginalLogLikelihood(model.likelihood, model)
return mll, model

#%%
def generate_next_candidate(x, y, n_candidates=1):
mll, model = initialize_model(x, y)
fit_gpytorch_model(mll)

pred_list = []
for i, sub_model in enumerate(model.models):
    input_data = x[:, :2] if i == 0 else x
    with torch.no_grad():
        pred = sub_model.posterior(input_data).mean
    pred_list.append(pred)
pred = torch.cat(pred_list, dim=1)

acq_fun_list = []
for _ in range(n_candidates):
    partitioning = FastNondominatedPartitioning(ref_point=ref_point, Y=pred)
    acq_fun = ExpectedHypervolumeImprovement(
        model=model, ref_point=ref_point, partitioning=partitioning
    )
    acq_fun_list.append(acq_fun)

# modify bounds based on input dimensionality
bounds_list = []
for i in range(len(model.models)):
    d = 2 if i == 0 else 4
    bounds = torch.zeros(2, d)
    bounds[1] = 1.0
    bounds_list.append(bounds)

candidates_list = []
for i, acq_fun in enumerate(acq_fun_list):
    candidates, _ = optimize_acqf(
        acq_function=acq_fun,
        bounds=bounds_list[i],
        q=n_candidates,
        num_restarts=1000,
        batch_initial_conditions=None,
        raw_samples=11000,
    )
    candidates_list.append(candidates)

return candidates_list, pred, model

x = torch.tensor(data[['ratio1', 'ratio2','ratio3','ratio4']].values, dtype=torch.float)
y = torch.tensor(data[['PF', 'Printing']].values, dtype=torch.float)

for i in range(n_iter):
print(f"Iteration {i}")
candidates, pred, model = generate_next_candidate(train_x_norm, train_y_norm, n_candidates=1)

When I iterate the function, I have an error :RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 4 but got size 2 for tensor number 1 in the list.
What approach I should take to fix the error?

Thanks a lot for your help!!

[An-xu]

Hi, Since the model need to remain the same input dimension, I set the non-related variables to 0. The function is written below. Let me know if there are better solutions.
Thank you so much for your help!

def initialize_model(train_x, train_y):
models = []
for i in range(train_y.shape[1]):
train_objective = train_y[:, i].unsqueeze(-1)
if i == 0:
input_data = torch.cat((train_x[:,0:2],torch.zeros(train_x.shape[0],2)),1)
models.append(
SingleTaskGP(input_data, train_objective)
)
else:
models.append(
SingleTaskGP(train_x, train_objective)
)
model = ModelListGP(*models)
mll = SumMarginalLogLikelihood(model.likelihood, model)
return mll, model

[Balandat]

So one option you have is to use the FilterFeatures input transform that just picks out the desired features for the model. That way you don't have to introduce irrelevant zero features for one of your models.

[An-xu]

Thank you! That is really helpful!

[An-xu]

Hi Balandat,
Thanks again for your reply!
After filtering the feature, I encounter a problem during the optimization process, I couldn't set the bound separately. It has the error:
RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 4 but got size 3 for tensor number 1 in the list.

def initialize_model(x, y, var=None):
models = []
for i in range(y.shape[1]):
if i == 2:
train_x = x
else:
feature_indices = torch.tensor([0, 1, 2]) # indices of the features we want to keep
transform = FilterFeatures(feature_indices)
train_x = transform(x)
train_objective = y[:, i].unsqueeze(-1)
models.append(
HeteroskedasticSingleTaskGP(train_x, train_objective, var[:, i].unsqueeze(-1))
)

model = ModelListGP(*models)
mll = SumMarginalLogLikelihood(model.likelihood, model)
return mll, model

def optimize_qehvi_and_get_observation(model, x, y, ref_point, n_candidates, sampler):
with torch.no_grad():
feature_indices = torch.tensor([0, 1, 2]) # indices of the features we want to keep
transform = FilterFeatures(feature_indices)
x_val_transformed = transform(x)
preds_0 = model.models[0].posterior(x_val_transformed)
preds_1 = model.models[1].posterior(x_val_transformed)
preds_2 = model.models[2].posterior(x)
preds = torch.stack((preds_0.mean, preds_1.mean, preds_2.mean), dim=1).squeeze(-1)
print("preds shape:", preds.shape)
acq_fun_list = []

partitioning = FastNondominatedPartitioning(ref_point=ref_point,Y=preds)

acq_fun =qExpectedHypervolumeImprovement(model=model, ref_point=ref_point, partitioning=partitioning, sampler=sampler)
# acq_fun =ExpectedHypervolumeImprovement(model=model, ref_point=ref_point, partitioning=partitioning)

bounds = torch.zeros(2, 3)
bounds[1] = 1
candidates, _ = optimize_acqf(
    acq_function=acq_fun,
    bounds= bounds,
    q=n_candidates,
    num_restarts=100,
    batch_initial_conditions=None,
    raw_samples=1100,
)
return candidates, preds