sparse variational GP with get_fantasy_model and qNegIntegratedPosteriorVariance

[cagoc]

Hi,
I am trying todo an active learning task with SingletaskVariationalGP and use qNegIntegratedPosteriorVariance as acqusition function to decide which point to sample next. And use condition_on_observations to update the SingletaskvariationalGP model then use updated model to decide next sampling point. But I ran into

'SingleTaskVariationalGP' object has no attribute 'fantasize' 

I have noticed SVGP Model Updating in Gpytorch so I tried but I found that the origin model and updated_model cond_model both have no attribute 'fantasize', which means it is impossible to use optimize_acqf.
Thanks for your reply in advance!

[saitcakmak]

Hi @cagoc. qNegIntegratedPosteriorVariance utilizes fantasy models to predict what the posterior variance would look like after evaluating the current candidate. Since SingletaskVariationalGP does not support fantasizing, this doesn't work.

For ExactGP models, fantasizing adds a fantasy observation at the given point to the training inputs & targets (train_X/Y)of the model. This concept does not trivially extend to approximate / variational GPs, since they utilize inducing points rather than relying on the training data directly. The GPyTorch documentation you linked shows one option that could be used to fantasize using these models. Approximate GP support in BoTorch is rather limited and we do not have an implementation of this fantasize method.

You could create a subclass of SingleTaskVariationalGP that implements this method and use it with qNegIntegratedPosteriorVariance. Here's the implementation for fantasize method for ExactGP models in BoTorch: https://github.com/pytorch/botorch/blob/main/botorch/models/model.py#L338

[cagoc]

Thanks!
To implement SingleTaskVariationalGP with qNegIntergratedPosteriorVariance, class SingleTaskVariationalGP should have function 'fantasize'. I have tried to add fantasize function in botorch class SingleTaskVariantionalGP like botorch.models.model FantasizeMixin do :

def fantasize(
            self,
            X,
            sampler,
            observation_noise,
            **kwargs,
    ):
        r"""Construct a fantasy model.

        Constructs a fantasy model in the following fashion:
        (1) compute the model posterior at `X`, including observation noise.
        If `observation_noise` is a Tensor, use it directly as the observation
        noise to add.
        (2) sample from this posterior (using `sampler`) to generate "fake"
        observations.
        (3) condition the model on the new fake observations.

        Args:
            X: A `batch_shape x n' x d`-dim Tensor, where `d` is the dimension of
                the feature space, `n'` is the number of points per batch, and
                `batch_shape` is the batch shape (must be compatible with the
                batch shape of the model).
            sampler: The sampler used for sampling from the posterior at `X`.
            observation_noise: A `model_batch_shape x 1 x m`-dim tensor or
                a `model_batch_shape x n' x m`-dim tensor containing the average
                noise for each batch and output, where `m` is the number of outputs.
                `noise` must be in the outcome-transformed space if an outcome
                transform is used.
                If None and using an inferred noise likelihood, the noise will be the
                inferred noise level. If using a fixed noise likelihood, the mean across
                the observation noise in the training data is used as observation noise.
            kwargs: Will be passed to `model.condition_on_observations`

        Returns:
            The constructed fantasy model.
        """
        if not isinstance(observation_noise, Tensor) and observation_noise is not None:
            raise DeprecationError(
                "`fantasize` no longer accepts a boolean for `observation_noise`."
            )
        elif observation_noise is None and isinstance(
                self.likelihood, FNGaussianLikelihood
        ):
            if self.num_outputs > 1:
                # make noise ... x n x m
                observation_noise = self.likelihood.noise.transpose(-1, -2)
            else:
                observation_noise = self.likelihood.noise.unsqueeze(-1)
            observation_noise = observation_noise.mean(dim=-2, keepdim=True)
        
        # if the inputs are empty, expand the inputs
        if X.shape[-2] == 0:
            output_shape = (
                    sampler.sample_shape
                    + X.shape[:-2]
                    + self.batch_shape
                    + torch.Size([0, self.num_outputs])
            )
            Y = torch.empty(output_shape, dtype=X.dtype, device=X.device)
            if observation_noise is not None:
                kwargs["noise"] = observation_noise.expand(Y.shape[1:])
            return self.condition_on_observations(
                X=self.transform_inputs(X),
                Y=Y,
                **kwargs,
            )
        propagate_grads = kwargs.pop("propagate_grads", False)
        with fantasize_flag():
            with settings.propagate_grads(propagate_grads):
                post_X = self.posterior(
                    X,
                    observation_noise=(
                        True if observation_noise is None else observation_noise
                    ),
                )
            Y_fantasized = sampler(post_X)  # num_fantasies x batch_shape x n' x m
            if observation_noise is not None:
                kwargs["noise"] = observation_noise.expand(Y_fantasized.shape[1:])
            return self.condition_on_observations(
                X=self.transform_inputs(X), Y=Y_fantasized, **kwargs
            )

but I ran into to bugs

otorch.exceptions.errors.BotorchTensorDimensionError: Expected X and Y to have the same number of dimensions (got X with dimension 3 and Y with dimension 2).

So I reshape observation_noise to deal with BotorchDimensionError . But I run into new errors:

  File "c:\code\online_vargp-main\volatilitygp\models\single_task_variational_gp.py", line 597, in get_fantasy_model
    return super().get_fantasy_model(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "c:\code\online_vargp-main\volatilitygp\models\single_task_variational_gp.py", line 333, in get_fantasy_model
    inducing_exact_model = self.inducing_model()
                           ^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\.conda\envs\env01\Lib\site-packages\gpytorch\utils\memoize.py", line 59, in g
    return _add_to_cache(self, cache_name, method(self, *args, **kwargs), *args, kwargs_pkl=kwargs_pkl)
                                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "c:\code\online_vargp-main\volatilitygp\models\single_task_variational_gp.py", line 226, in inducing_model
    inducing_noise_covar, inducing_mean = self.pseudo_points()
                                          ^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\.conda\envs\env01\Lib\site-packages\gpytorch\utils\memoize.py", line 59, in g
    return _add_to_cache(self, cache_name, method(self, *args, **kwargs), *args, kwargs_pkl=kwargs_pkl)
                                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "c:\code\online_vargp-main\volatilitygp\models\single_task_variational_gp.py", line 169, in pseudo_points
    res = var_cov.add_jitter(-1.0)._mul_constant(-1.0)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\.conda\envs\env01\Lib\site-packages\linear_operator\operators\sum_linear_operator.py", line 55, in _mul_constant
    return self.__class__(*[lt._mul_constant(other) for lt in self.linear_ops])
                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\.conda\envs\env01\Lib\site-packages\linear_operator\operators\sum_linear_operator.py", line 55, in <listcomp>
    return self.__class__(*[lt._mul_constant(other) for lt in self.linear_ops])
                            ^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\.conda\envs\env01\Lib\site-packages\linear_operator\operators\root_linear_operator.py", line 75, in _mul_constant
    if (other > 0).all():
       ^^^^^^^^^^^^^^^
AttributeError: 'bool' object has no attribute 'all'

I have no idea to fix this. Could you give me more hints about how to add fantasize function into SingleTaskVariationalGP? Thank you!

[saitcakmak]

That implementation of fantasize method is mainly intended for ExactGP models. I am not familiar enough with approximate models to tell what a fantasize method would look like. Your best guide is likely the GPyTorch document you linked, but getting it working would likely still require a decent bit of debugging.