Update sample_all_priors to support wider set of priors (#2371)

Summary:
Pull Request resolved: #2371

Addresses #780

Previously, this would pass in `closure(module).shape` as the `sample_shape`, which only worked if the prior was a univariate distribution. `Distribution.sample` produces samples of shape `Distribution._extended_shape(sample_shape) = sample_shape + Distribution._extended_shape()`, so we can calculate the `sample_shape` required to support both univariate and multivariate / batched priors.

Reviewed By: dme65

Differential Revision: D58377495

fbshipit-source-id: 17510505012838a3fe670492656be4d13bc0db5e

botorch/optim/utils/model_utils.py

@@ -161,7 +161,12 @@ def sample_all_priors(model: GPyTorchModel, max_retries: int = 100) -> None:
             )
         for i in range(max_retries):
             try:
-                setting_closure(module, prior.sample(closure(module).shape))
+                # Set sample shape, so that the prior samples have the same shape
+                # as `closure(module)` without having to be repeated.
+                closure_shape = closure(module).shape
+                prior_shape = prior._extended_shape()
+                sample_shape = closure_shape[: -len(prior_shape)]
+                setting_closure(module, prior.sample(sample_shape=sample_shape))
                 break
             except NotImplementedError:
                 warn(

test/optim/utils/test_model_utils.py

@@ -25,13 +25,14 @@
 )
 from botorch.utils.testing import BotorchTestCase
 from gpytorch.constraints import GreaterThan
+from gpytorch.kernels import RBFKernel
 from gpytorch.kernels.matern_kernel import MaternKernel
 from gpytorch.kernels.scale_kernel import ScaleKernel
 from gpytorch.likelihoods import GaussianLikelihood
 from gpytorch.mlls.exact_marginal_log_likelihood import ExactMarginalLogLikelihood
 from gpytorch.priors import UniformPrior
 from gpytorch.priors.prior import Prior
-from gpytorch.priors.torch_priors import GammaPrior
+from gpytorch.priors.torch_priors import GammaPrior, NormalPrior
 
 
 class DummyPrior(Prior):
@@ -244,3 +245,26 @@ def test_sample_all_priors(self):
             original_state_dict = dict(deepcopy(mll.model.state_dict()))
             with self.assertRaises(RuntimeError):
                 sample_all_priors(model)
+
+    def test_with_multivariate_prior(self) -> None:
+        # This is modified from https://github.com/pytorch/botorch/issues/780.
+        for batch in (torch.Size([]), torch.Size([3])):
+            model = SingleTaskGP(
+                train_X=torch.randn(*batch, 2, 2),
+                train_Y=torch.randn(*batch, 2, 1),
+                covar_module=RBFKernel(
+                    ard_num_dims=2,
+                    batch_shape=batch,
+                    lengthscale_prior=NormalPrior(
+                        # Make this almost singular for easy comparison below.
+                        torch.tensor([[1.0, 1.0]]),
+                        torch.tensor(1e-10),
+                    ),
+                ),
+            )
+            # Check that the lengthscale is replaced with the sampled values.
+            original_lengthscale = model.covar_module.lengthscale
+            sample_all_priors(model)
+            new_lengthscale = model.covar_module.lengthscale
+            self.assertFalse(torch.allclose(original_lengthscale, new_lengthscale))
+            self.assertAllClose(new_lengthscale, torch.ones(*batch, 1, 2))