Can we make train_X and train_Y optional for SingleTaskGP?

[dynobo]

The two init args train_X, train_Y of SingleTaskGP are typed as Tensor (none-optional). It's parent class gpytorch.models.ExactGP actually handles the case of None as input data and SingleTaskGP also seems to work well when instantiated with train_X=None, train_Y=None.

As it is sometimes useful to initialize without training data and the implementation of the parent ExactGP is explicitly implemented to support that, does anything speak against typing them as Optional[Tensor] in botorch?

(If not, I'd open a PR accordingly)

[saitcakmak]

Using None inputs will error out if you use input / outcome transforms. You can initialize it with empty tensors though.

model = SingleTaskGP(torch.ones(0, 2), torch.ones(0, 2), outcome_transform=Standardize(m=2))

(swapping for None above will error out)

[ptighin]

The other option is to forbid the passing of input and output transforms to a model with empty train data. This might give more meaningful feedback to the users that standardisation of an empty array makes little sense.

The third option is to adapt the input and output transforms such that they can handle this corner case under the hood. For instance, Standardize() would take a "trivial" state with mean zero and variance one. This option, however, requires a little more surgery I think, since all normalisers would need to be adjusted to this corner case.

What do you think, @saitcakmak?

[ptighin]

@saitcakmak I managed to mess up the fork to the above link and your tests disappeared. Please see this commit, saitcakmak@310246d, for the correct version. Note that I also expanded the tests for empty inputs a little.

[saitcakmak]

if empty data is passed, no transforms should be applied to the data since it makes no sense to normalise an empty vector.

This makes sense on a conceptual level. One concern with this is that we apply transforms to both train and test data. In many cases, we also use the train data to set the parameters of the transforms (this is pretty much why the tests were failing). Suppose you use a prior that is meant to work with standardized data (botorch defaults are designed for standardized data). If you use a transform and pass no training data, your transforms are not properly initialized, and we get around this by not applying them (at least in the proposed design). If we then evaluate some test inputs on this model, these will be in untransformed space. When we use the same prior with non-empty train data, then the test data will be on standardized space. So, we end up with a bit of inconsistency there. To me this just says if you want to use empty inputs with your model, you shouldn't ever use transforms with it - which agrees with your suggestion.

See the fix at saitcakmak@dcee8bb. What do you think?

I updated #1229 with your changes, and also added the model training step to the unit test, which fails. It makes sense that you should not be able to train a model without any training data. We could probably add a check somewhere in fit_gpytorch_model that skips training if there's no data.

A main concern about that design is overwriting the transforms to None without warning the user (and the need to do this in each model). Silently overwriting the user inputs is generally a bad practice and we should avoid that. I'd rather let it error out while trying to transform things.

The third option is to adapt the input and output transforms such that they can handle this corner case under the hood.

This would require lots of changes in transforms, so ideally we wouldn't do that.

All that said, one cleanish solution I see is just making the change in transform_inputs to skip over empty inputs and adding something in fit_gpytorch_model to skip training if there are no inputs. This would be rather minimal and would support using empty inputs as long as you don't use any transforms (you can always manually transform the data in your code). Would that work for you? I can look into what needs to be done in fit_gpytorch_model to make this work.

[saitcakmak]

Actually, you don't even need the change in transform_inputs if the input_transform is None. So, we would only need to skip model fitting

[saitcakmak]

Updated #1229 with the proposed change to fit_gpytorch_model.