gp_regression_mixed.py

#!/usr/bin/env python3
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

from __future__ import annotations

from typing import Any, Callable, Dict, List, Optional

import torch
from botorch.models.gp_regression import SingleTaskGP
from botorch.models.kernels.categorical import CategoricalKernel
from botorch.models.transforms.input import InputTransform
from botorch.models.transforms.outcome import OutcomeTransform
from botorch.utils.datasets import SupervisedDataset
from botorch.utils.transforms import normalize_indices
from gpytorch.constraints import GreaterThan
from gpytorch.kernels.kernel import Kernel
from gpytorch.kernels.matern_kernel import MaternKernel
from gpytorch.kernels.scale_kernel import ScaleKernel
from gpytorch.likelihoods.gaussian_likelihood import GaussianLikelihood
from gpytorch.likelihoods.likelihood import Likelihood
from gpytorch.priors import GammaPrior
from torch import Tensor


class MixedSingleTaskGP(SingleTaskGP):
    r"""A single-task exact GP model for mixed search spaces.

    This model is similar to `SingleTaskGP`, but supports mixed search spaces,
    which combine discrete and continuous features, as well as solely discrete
    spaces. It uses a kernel that combines a CategoricalKernel (based on
    Hamming distances) and a regular kernel into a kernel of the form

        K((x1, c1), (x2, c2)) =
            K_cont_1(x1, x2) + K_cat_1(c1, c2) +
            K_cont_2(x1, x2) * K_cat_2(c1, c2)

    where `xi` and `ci` are the continuous and categorical features of the
    input, respectively. The suffix `_i` indicates that we fit different
    lengthscales for the kernels in the sum and product terms.

    Since this model does not provide gradients for the categorical features,
    optimization of the acquisition function will need to be performed in
    a mixed fashion, i.e., treating the categorical features properly as
    discrete optimization variables. We recommend using `optimize_acqf_mixed.`

    Example:
        >>> train_X = torch.cat(
                [torch.rand(20, 2), torch.randint(3, (20, 1))], dim=-1)
            )
        >>> train_Y = (
                torch.sin(train_X[..., :-1]).sum(dim=1, keepdim=True)
                + train_X[..., -1:]
            )
        >>> model = MixedSingleTaskGP(train_X, train_Y, cat_dims=[-1])
    """

    def __init__(
        self,
        train_X: Tensor,
        train_Y: Tensor,
        cat_dims: List[int],
        train_Yvar: Optional[Tensor] = None,
        cont_kernel_factory: Optional[
            Callable[[torch.Size, int, List[int]], Kernel]
        ] = None,
        likelihood: Optional[Likelihood] = None,
        outcome_transform: Optional[OutcomeTransform] = None,  # TODO
        input_transform: Optional[InputTransform] = None,  # TODO
    ) -> None:
        r"""A single-task exact GP model supporting categorical parameters.

        Args:
            train_X: A `batch_shape x n x d` tensor of training features.
            train_Y: A `batch_shape x n x m` tensor of training observations.
            cat_dims: A list of indices corresponding to the columns of
                the input `X` that should be considered categorical features.
            train_Yvar: An optional `batch_shape x n x m` tensor of observed
                measurement noise.
            cont_kernel_factory: A method that accepts  `batch_shape`, `ard_num_dims`,
                and `active_dims` arguments and returns an instantiated GPyTorch
                `Kernel` object to be used as the base kernel for the continuous
                dimensions. If omitted, this model uses a Matern-2.5 kernel as
                the kernel for the ordinal parameters.
            likelihood: A likelihood. If omitted, use a standard
                GaussianLikelihood with inferred noise level.
            outcome_transform: An outcome transform that is applied to the
                training data during instantiation and to the posterior during
                inference (that is, the `Posterior` obtained by calling
                `.posterior` on the model will be on the original scale).
            input_transform: An input transform that is applied in the model's
                forward pass. Only input transforms are allowed which do not
                transform the categorical dimensions. If you want to use it
                for example in combination with a `OneHotToNumeric` input transform
                one has to instantiate the transform with `transform_on_train` == False
                and pass in the already transformed input.
        """
        if len(cat_dims) == 0:
            raise ValueError(
                "Must specify categorical dimensions for MixedSingleTaskGP"
            )
        self._ignore_X_dims_scaling_check = cat_dims
        _, aug_batch_shape = self.get_batch_dimensions(train_X=train_X, train_Y=train_Y)

        if cont_kernel_factory is None:

            def cont_kernel_factory(
                batch_shape: torch.Size,
                ard_num_dims: int,
                active_dims: List[int],
            ) -> MaternKernel:
                return MaternKernel(
                    nu=2.5,
                    batch_shape=batch_shape,
                    ard_num_dims=ard_num_dims,
                    active_dims=active_dims,
                    lengthscale_constraint=GreaterThan(1e-04),
                )

        if likelihood is None and train_Yvar is None:
            # This Gamma prior is quite close to the Horseshoe prior
            min_noise = 1e-5 if train_X.dtype == torch.float else 1e-6
            likelihood = GaussianLikelihood(
                batch_shape=aug_batch_shape,
                noise_constraint=GreaterThan(
                    min_noise, transform=None, initial_value=1e-3
                ),
                noise_prior=GammaPrior(0.9, 10.0),
            )

        d = train_X.shape[-1]
        cat_dims = normalize_indices(indices=cat_dims, d=d)
        ord_dims = sorted(set(range(d)) - set(cat_dims))
        if len(ord_dims) == 0:
            covar_module = ScaleKernel(
                CategoricalKernel(
                    batch_shape=aug_batch_shape,
                    ard_num_dims=len(cat_dims),
                    lengthscale_constraint=GreaterThan(1e-06),
                )
            )
        else:
            sum_kernel = ScaleKernel(
                cont_kernel_factory(
                    batch_shape=aug_batch_shape,
                    ard_num_dims=len(ord_dims),
                    active_dims=ord_dims,
                )
                + ScaleKernel(
                    CategoricalKernel(
                        batch_shape=aug_batch_shape,
                        ard_num_dims=len(cat_dims),
                        active_dims=cat_dims,
                        lengthscale_constraint=GreaterThan(1e-06),
                    )
                )
            )
            prod_kernel = ScaleKernel(
                cont_kernel_factory(
                    batch_shape=aug_batch_shape,
                    ard_num_dims=len(ord_dims),
                    active_dims=ord_dims,
                )
                * CategoricalKernel(
                    batch_shape=aug_batch_shape,
                    ard_num_dims=len(cat_dims),
                    active_dims=cat_dims,
                    lengthscale_constraint=GreaterThan(1e-06),
                )
            )
            covar_module = sum_kernel + prod_kernel
        super().__init__(
            train_X=train_X,
            train_Y=train_Y,
            train_Yvar=train_Yvar,
            likelihood=likelihood,
            covar_module=covar_module,
            outcome_transform=outcome_transform,
            input_transform=input_transform,
        )

    @classmethod
    def construct_inputs(
        cls,
        training_data: SupervisedDataset,
        categorical_features: List[int],
        likelihood: Optional[Likelihood] = None,
        **kwargs: Any,
    ) -> Dict[str, Any]:
        r"""Construct `Model` keyword arguments from a dict of `SupervisedDataset`.

        Args:
            training_data: A `SupervisedDataset` containing the training data.
            categorical_features: Column indices of categorical features.
            likelihood: Optional likelihood used to constuct the model.
        """
        base_inputs = super().construct_inputs(training_data=training_data, **kwargs)
        return {
            **base_inputs,
            "cat_dims": categorical_features,
            "likelihood": likelihood,
        }