Monotonic rejection model and generator (facebookresearch#458)

Summary:

monotonic rejection model GPU support, since they're tied to the generator, we also ensure the generators are gpu ready as well.

Differential Revision: D65638150

Author: JasonKChow

aepsych/generators/monotonic_rejection_generator.py

@@ -101,7 +101,7 @@ def gen(
         )
 
         # Augment bounds with deriv indicator
-        bounds = torch.cat((model.bounds_, torch.zeros(2, 1)), dim=1)
+        bounds = torch.cat((model.bounds_, torch.zeros(2, 1).to(model.device)), dim=1)
         # Fix deriv indicator to 0 during optimization
         fixed_features = {(bounds.shape[1] - 1): 0.0}
         # Fix explore features to random values

aepsych/means/constant_partial_grad.py

@@ -26,6 +26,6 @@ def forward(self, input: torch.Tensor) -> torch.Tensor:
         idx = input[..., -1].to(dtype=torch.long) > 0
         mean_fit = super(ConstantMeanPartialObsGrad, self).forward(input[..., ~idx, :])
         sz = mean_fit.shape[:-1] + torch.Size([input.shape[-2]])
-        mean = torch.zeros(sz)
+        mean = torch.zeros(sz).to(input)
         mean[~idx] = mean_fit
         return mean

aepsych/models/base.py

@@ -415,7 +415,12 @@ def set_train_data(self, inputs=None, targets=None, strict=False):
     def device(self) -> torch.device:
         # We assume all models have some parameters and all models will only use one device
         # notice that this has no setting, don't let users set device, use .to().
-        return next(self.parameters()).device
+        try:
+            return next(self.parameters()).device
+        except (
+            AttributeError
+        ):  # Fallback for cases where we need device before we have params
+            return torch.device("cpu")
 
     @property
     def train_inputs(self) -> Optional[Tuple[torch.Tensor]]:

aepsych/models/derivative_gp.py

@@ -13,6 +13,7 @@
 import torch
 from aepsych.kernels.rbf_partial_grad import RBFKernelPartialObsGrad
 from aepsych.means.constant_partial_grad import ConstantMeanPartialObsGrad
+from aepsych.models.base import AEPsychModelDeviceMixin
 from botorch.models.gpytorch import GPyTorchModel
 from gpytorch.distributions import MultivariateNormal
 from gpytorch.kernels import Kernel
@@ -22,7 +23,7 @@
 from gpytorch.variational import CholeskyVariationalDistribution, VariationalStrategy
 
 
-class MixedDerivativeVariationalGP(gpytorch.models.ApproximateGP, GPyTorchModel):
+class MixedDerivativeVariationalGP(gpytorch.models.ApproximateGP, AEPsychModelDeviceMixin, GPyTorchModel):
     """A variational GP with mixed derivative observations.
 
     For more on GPs with derivative observations, see e.g. Riihimaki & Vehtari 2010.
@@ -99,6 +100,7 @@ def __init__(
         self._num_outputs = 1
         self.train_inputs = (train_x,)
         self.train_targets = train_y
+        self.to(self.device) # Needed to prep for below
         self(train_x)  # Necessary for CholeskyVariationalDistribution
 
     def forward(self, x: torch.Tensor) -> MultivariateNormal:

aepsych/models/monotonic_projection_gp.py

@@ -136,14 +136,17 @@ def posterior(
             # using numpy because torch doesn't support vectorized linspace,
             # pytorch/issues/61292
             grid: Union[np.ndarray, torch.Tensor] = np.linspace(
-                self.lb[dim],
-                X[:, dim].numpy(),
+                self.lb[dim].cpu().numpy(),
+                X[:, dim].cpu().numpy(),
                 s + 1,
             )  # (s+1 x n)
             grid = torch.tensor(grid[:-1, :], dtype=X.dtype)  # Drop x; (s x n)
             X_aug[(1 + i * s) : (1 + (i + 1) * s), :, dim] = grid
         # X_aug[0, :, :] is X, and then subsequent indices are points in the grids
         # Predict marginal distributions on X_aug
+
+        X = X.to(self.device)
+        X_aug = X_aug.to(self.device)
         with torch.no_grad():
             post_aug = super().posterior(X=X_aug)
         mu_aug = post_aug.mean.squeeze()  # (m*s+1 x n)
@@ -158,12 +161,13 @@ def posterior(
         # Adjust the whole covariance matrix to accomadate the projected marginals
         with torch.no_grad():
             post = super().posterior(X=X)
-            R = cov2corr(post.distribution.covariance_matrix.squeeze().numpy())
-            S_proj = torch.tensor(corr2cov(R, sigma_proj.numpy()), dtype=X.dtype)
+            R = cov2corr(post.distribution.covariance_matrix.squeeze().cpu().numpy())
+            S_proj = torch.tensor(corr2cov(R, sigma_proj.cpu().numpy()), dtype=X.dtype)
         mvn_proj = gpytorch.distributions.MultivariateNormal(
-            mu_proj.unsqueeze(0),
-            S_proj.unsqueeze(0),
+            mu_proj.unsqueeze(0).to(self.device),
+            S_proj.unsqueeze(0).to(self.device),
         )
+
         return GPyTorchPosterior(mvn_proj)
 
     def sample(self, x: torch.Tensor, num_samples: int) -> torch.Tensor:

aepsych/models/monotonic_rejection_gp.py

@@ -18,7 +18,7 @@
 from aepsych.factory.monotonic import monotonic_mean_covar_factory
 from aepsych.kernels.rbf_partial_grad import RBFKernelPartialObsGrad
 from aepsych.means.constant_partial_grad import ConstantMeanPartialObsGrad
-from aepsych.models.base import AEPsychMixin
+from aepsych.models.base import AEPsychModelDeviceMixin
 from aepsych.models.utils import select_inducing_points
 from aepsych.utils import _process_bounds, promote_0d
 from botorch.fit import fit_gpytorch_mll
@@ -32,7 +32,7 @@
 from torch import Tensor
 
 
-class MonotonicRejectionGP(AEPsychMixin, ApproximateGP):
+class MonotonicRejectionGP(AEPsychModelDeviceMixin, ApproximateGP):
     """A monotonic GP using rejection sampling.
 
     This takes the same insight as in e.g. Riihimäki & Vehtari 2010 (that the derivative of a GP
@@ -83,15 +83,15 @@ def __init__(
             objective (Optional[MCAcquisitionObjective], optional): Transformation of GP to apply before computing acquisition function. Defaults to identity transform for gaussian likelihood, probit transform for probit-bernoulli.
             extra_acqf_args (Optional[Dict[str, object]], optional): Additional arguments to pass into the acquisition function. Defaults to None.
         """
-        self.lb, self.ub, self.dim = _process_bounds(lb, ub, dim)
+        lb, ub, self.dim = _process_bounds(lb, ub, dim)
         if likelihood is None:
             likelihood = BernoulliLikelihood()
 
         self.inducing_size = num_induc
         self.inducing_point_method = inducing_point_method
         inducing_points = select_inducing_points(
             inducing_size=self.inducing_size,
-            bounds=self.bounds,
+            bounds=torch.stack((lb, ub)),
             method="sobol",
         )
 
@@ -134,7 +134,9 @@ def __init__(
 
         super().__init__(variational_strategy)
 
-        self.bounds_ = torch.stack([self.lb, self.ub])
+        self.register_buffer("lb", lb)
+        self.register_buffer("ub", ub)
+        self.register_buffer("bounds_", torch.stack([self.lb, self.ub]))
         self.mean_module = mean_module
         self.covar_module = covar_module
         self.likelihood = likelihood
@@ -144,7 +146,7 @@ def __init__(
         self.num_samples = num_samples
         self.num_rejection_samples = num_rejection_samples
         self.fixed_prior_mean = fixed_prior_mean
-        self.inducing_points = inducing_points
+        self.register_buffer("inducing_points", inducing_points)
 
     def fit(self, train_x: Tensor, train_y: Tensor, **kwargs) -> None:
         """Fit the model
@@ -161,7 +163,7 @@ def fit(self, train_x: Tensor, train_y: Tensor, **kwargs) -> None:
             X=self.train_inputs[0],
             bounds=self.bounds,
             method=self.inducing_point_method,
-        )
+        ).to(self.device)
         self._set_model(train_x, train_y)
 
     def _set_model(
@@ -284,13 +286,14 @@ def predict_probability(
         return self.predict(x, probability_space=True)
 
     def _augment_with_deriv_index(self, x: Tensor, indx) -> Tensor:
+        x = x.to(self.device)
         return torch.cat(
-            (x, indx * torch.ones(x.shape[0], 1)),
+            (x, indx * torch.ones(x.shape[0], 1).to(self.device)),
             dim=1,
         )
 
     def _get_deriv_constraint_points(self) -> Tensor:
-        deriv_cp = torch.tensor([])
+        deriv_cp = torch.tensor([]).to(self.device)
         for i in self.monotonic_idxs:
             induc_i = self._augment_with_deriv_index(self.inducing_points, i + 1)
             deriv_cp = torch.cat((deriv_cp, induc_i), dim=0)
@@ -299,8 +302,8 @@ def _get_deriv_constraint_points(self) -> Tensor:
     @classmethod
     def from_config(cls, config: Config) -> MonotonicRejectionGP:
         classname = cls.__name__
-        num_induc = config.gettensor(classname, "num_induc", fallback=25)
-        num_samples = config.gettensor(classname, "num_samples", fallback=250)
+        num_induc = config.getint(classname, "num_induc", fallback=25)
+        num_samples = config.getint(classname, "num_samples", fallback=250)
         num_rejection_samples = config.getint(
             classname, "num_rejection_samples", fallback=5000
         )

aepsych/strategy.py

@@ -71,13 +71,6 @@ class Strategy(object):
 
     _n_eval_points: int = 1000
 
-    no_gpu_acqfs = (
-        MonotonicMCAcquisition,
-        MonotonicBernoulliMCMutualInformation,
-        MonotonicMCPosteriorVariance,
-        MonotonicMCLSE,
-    )
-
     def __init__(
         self,
         generator: Union[AEPsychGenerator, ParameterTransformedGenerator],
@@ -182,13 +175,7 @@ def __init__(
                 )
                 self.generator_device = torch.device("cpu")
             else:
-                if hasattr(generator, "acqf") and generator.acqf in self.no_gpu_acqfs:
-                    warnings.warn(
-                        f"GPU requested for acquistion function {type(generator.acqf).__name__}, but this acquisiton function does not support GPU! Using CPU instead.",
-                        UserWarning,
-                    )
-                    self.generator_device = torch.device("cpu")
-                elif not torch.cuda.is_available():
+                if not torch.cuda.is_available():
                     warnings.warn(
                         f"GPU requested for generator {type(generator).__name__}, but no GPU found! Using CPU instead.",
                         UserWarning,
@@ -283,9 +270,11 @@ def normalize_inputs(
             x = x[None, :]
 
         if self.x is not None:
+            x = x.to(self.x)
             x = torch.cat((self.x, x), dim=0)
 
         if self.y is not None:
+            y = y.to(self.y)
             y = torch.cat((self.y, y), dim=0)
 
         # Ensure the correct dtype

tests_gpu/acquisition/__init__.py

@@ -0,0 +1,6 @@
+#!/usr/bin/env python3
+# Copyright (c) Meta, Inc. and its affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

tests_gpu/acquisition/test_monotonic.py

@@ -0,0 +1,51 @@
+#!/usr/bin/env python3
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from aepsych.acquisition.monotonic_rejection import MonotonicMCLSE
+from aepsych.acquisition.objective import ProbitObjective
+from aepsych.models.derivative_gp import MixedDerivativeVariationalGP
+from botorch.acquisition.objective import IdentityMCObjective
+from botorch.utils.testing import BotorchTestCase
+
+
+class TestMonotonicAcq(BotorchTestCase):
+    def test_monotonic_acq_gpu(self):
+        # Init
+        train_X_aug = torch.tensor(
+            [[0.0, 0.0, 0.0], [1.0, 1.0, 0.0], [2.0, 2.0, 0.0]]
+        ).cuda()
+        deriv_constraint_points = torch.tensor(
+            [[0.0, 0.0, 1.0], [1.0, 1.0, 1.0], [2.0, 2.0, 1.0]]
+        ).cuda()
+        train_Y = torch.tensor([[1.0], [2.0], [3.0]]).cuda()
+
+        m = MixedDerivativeVariationalGP(
+            train_x=train_X_aug, train_y=train_Y, inducing_points=train_X_aug
+        ).cuda()
+        acq = MonotonicMCLSE(
+            model=m,
+            deriv_constraint_points=deriv_constraint_points,
+            num_samples=5,
+            num_rejection_samples=8,
+            target=1.9,
+        )
+        self.assertTrue(isinstance(acq.objective, IdentityMCObjective))
+        acq = MonotonicMCLSE(
+            model=m,
+            deriv_constraint_points=deriv_constraint_points,
+            num_samples=5,
+            num_rejection_samples=8,
+            target=1.9,
+            objective=ProbitObjective(),
+        ).cuda()
+        # forward
+        acq(train_X_aug)
+        Xfull = torch.cat((train_X_aug, acq.deriv_constraint_points), dim=0)
+        posterior = m.posterior(Xfull)
+        samples = acq.sampler(posterior)
+        self.assertEqual(samples.shape, torch.Size([5, 6, 1]))