[Question] Issue with `ScalarizedPosteriorTransform`

[r-ashwin]

I am running into an issue with using ScalarizedPosteriorTransform with a multioutput GP. I am not doing anything fancy, just the following:

weights = torch.tensor(s / s.sum())
post_transf = ScalarizedPosteriorTransform(weights=weights)

mfacq = qMultiFidelityKnowledgeGradient(model=model, num_fantasies=32,
                                        posterior_transform=post_transf,)

Then upon using it within the optimize_acqf call I see the following error:

---------------------------------------------------------------------------
UnsupportedError                          Traceback (most recent call last)
<ipython-input-10-52db2f5a269d> in <module>
     27     return new_x
     28 
---> 29 optimize_acqf_(mfacq)

<ipython-input-10-52db2f5a269d> in optimize_acqf_(acqf, fixed_features)
      7     _num_restarts = 10
      8 
----> 9     X_init = gen_one_shot_kg_initial_conditions(
     10                 acq_function=acqf,
     11                 bounds=bounds,

~/opt/anaconda3/lib/python3.8/site-packages/botorch/optim/initializers.py in gen_one_shot_kg_initial_conditions(acq_function, bounds, q, num_restarts, raw_samples, fixed_features, options, inequality_constraints, equality_constraints)
    291 
    292     # compute maximizer of the value function
--> 293     value_function = _get_value_function(
    294         model=acq_function.model,
    295         objective=acq_function.objective,

~/opt/anaconda3/lib/python3.8/site-packages/botorch/acquisition/knowledge_gradient.py in _get_value_function(model, objective, posterior_transform, sampler, project, valfunc_cls, valfunc_argfac)
    548         else:
    549             print(posterior_transform)
--> 550             base_value_function = PosteriorMean(
    551                 model=model, posterior_transform=posterior_transform
    552             )

~/opt/anaconda3/lib/python3.8/site-packages/botorch/acquisition/analytic.py in __init__(self, model, posterior_transform, maximize)
    177                 does actually return -1 * minimum of the posterior mean.
    178         """
--> 179         super().__init__(model=model, posterior_transform=posterior_transform)
    180         self.maximize = maximize
    181 

~/opt/anaconda3/lib/python3.8/site-packages/botorch/acquisition/analytic.py in __init__(self, model, posterior_transform, **kwargs)
     53         if posterior_transform is None:
     54             if model.num_outputs != 1:
---> 55                 raise UnsupportedError(
     56                     "Must specify a posterior transform when using a "
     57                     "multi-output model."

UnsupportedError: Must specify a posterior transform when using a multi-output model.

Happy to provide full code if necessary. While it is clear that the PosteriorMean value function call within KnowledgeGradient sees None for the Posterior_transform, printing the attribute for debugging revealed that it becomes None only on the second call (the first call seems to be fine).

Can someone shed some light on this? thanks.

[saitcakmak]

Thanks for flagging this! It's an actual bug due to a missing line in gen_one_shot_kg_initial_conditions. I'll put up a PR shortly.

[r-ashwin]

Happy to try on my own if you give me an outline of what needs to be done.

[saitcakmak]

Happy to try on my own if you give me an outline of what needs to be done.

Saw this a little too late. The fix is in #1187 :)