Upgrade deprecated APIs in PyTorch 1.11 & clean out some legacy codes (…

…#1378)

Summary:
Pull Request resolved: #1378

PyTorch 1.11 was released last Thursday, which introduces a few Deprecation Warnings that were [caught by our CI](https://github.com/facebookresearch/beanmachine/runs/5507917348?check_suite_focus=true). This diff addresses those warnings by updating the relevant function calls.

It also caught my attention that some of the warnings come from our legacy inference module, which was kept around when we did the migration just in case "something happens." Now that the new inference has been out for a while, we can begin to clean out the legacy modules. In this diff I still kept some algorithms around because there are still some modules that depends on them -- in theory it should be safe to swap all of them out with the new implementation, though it's probably do that in a separate diff.

Reviewed By: jpchen

Differential Revision: D34833587

fbshipit-source-id: 6ef0456766236c9851ee079304dfd324b112f5a2

src/beanmachine/ppl/experimental/tests/gp/inference_test.py

@@ -14,7 +14,6 @@
 from beanmachine.ppl.experimental.gp import likelihoods
 from beanmachine.ppl.experimental.gp.kernels import PeriodicKernel, ScaleKernel
 from beanmachine.ppl.experimental.gp.models import SimpleGP
-from beanmachine.ppl.legacy.inference import SingleSiteNoUTurnSampler
 from gpytorch.distributions import MultivariateNormal
 
 
@@ -78,7 +77,7 @@ def test_simple_regression(self):
             self.lengthscale_prior(),
             self.period_length_prior(),
         ]
-        samples = SingleSiteNoUTurnSampler().infer(
+        samples = bm.SingleSiteNoUTurnSampler().infer(
             queries, obs, n_samples, num_chains=1
         )
 

src/beanmachine/ppl/experimental/tests/gp/models_test.py

@@ -10,7 +10,6 @@
 import torch.distributions as dist
 from beanmachine.ppl.experimental.gp import kernels, likelihoods
 from beanmachine.ppl.experimental.gp.models import BoTorchGP, SimpleGP
-from beanmachine.ppl.legacy.inference import SingleSiteNoUTurnSampler
 from botorch.posteriors.gpytorch import GPyTorchPosterior
 
 
@@ -34,7 +33,7 @@ def mean():
 
     def test_infer(self):
         self.model.train()
-        SingleSiteNoUTurnSampler().infer([self.p()], {}, num_samples=2, num_chains=1)
+        bm.SingleSiteNoUTurnSampler().infer([self.p()], {}, num_samples=2, num_chains=1)
 
     def test_load_and_predict(self):
         self.model.eval()

src/beanmachine/ppl/experimental/tests/vi/vi_test.py

@@ -450,7 +450,7 @@ def w():
         def q_y():
             weights = w()
             data = x()
-            p = torch.sigmoid(data @ weights.T)
+            p = torch.sigmoid(data @ weights)
             return dist.Bernoulli(p)
 
         opt_params = VariationalInference().infer(

src/beanmachine/ppl/inference/proposer/nmc/single_site_real_space_nmc_proposer.py

@@ -137,11 +137,11 @@ def get_proposal_distribution(self, world: World) -> dist.Distribution:
             # flip(flip(H^-1)) = flip((L^-1)') @ flip(L^-1)
             # H^-1 = flip(L^-1)' @ flip(L^-1)
             # flip(L^-1)' is the lower triangular cholesky factor for H^-1.
-            L_inv = torch.triangular_solve(
-                torch.eye(L.size(-1)).to(dtype=neg_hessian.dtype, device=node_device),
+            L_inv = torch.linalg.solve_triangular(
                 L,
+                torch.eye(L.size(-1)).to(dtype=neg_hessian.dtype, device=node_device),
                 upper=False,
-            ).solution
+            )
             L_chol = L_inv.flip([0, 1]).T
             distance = torch.cholesky_solve(first_gradient.unsqueeze(1), L).t()
             proposal_args = _ProposalArgs(

src/beanmachine/ppl/legacy/inference/__init__.py

@@ -8,33 +8,18 @@
 from beanmachine.ppl.legacy.inference.single_site_ancestral_mh import (
     SingleSiteAncestralMetropolisHastings,
 )
-from beanmachine.ppl.legacy.inference.single_site_hamiltonian_monte_carlo import (
-    SingleSiteHamiltonianMonteCarlo,
-)
 from beanmachine.ppl.legacy.inference.single_site_newtonian_monte_carlo import (
     SingleSiteNewtonianMonteCarlo,
 )
-from beanmachine.ppl.legacy.inference.single_site_no_u_turn_sampler import (
-    SingleSiteNoUTurnSampler,
-)
 from beanmachine.ppl.legacy.inference.single_site_random_walk import (
     SingleSiteRandomWalk,
 )
-from beanmachine.ppl.legacy.inference.single_site_uniform_mh import (
-    SingleSiteUniformMetropolisHastings,
-)
 
 
 __all__ = [
     "CompositionalInference",
     "RejectionSampling",
     "SingleSiteAncestralMetropolisHastings",
-    "SingleSiteHamiltonianMonteCarlo",
     "SingleSiteNewtonianMonteCarlo",
-    "SingleSiteNoUTurnSampler",
     "SingleSiteRandomWalk",
-    "SingleSiteUniformMetropolisHastings",
-    "Predictive",
-    "empirical",
-    "simulate",
 ]

src/beanmachine/ppl/legacy/inference/proposer/__init__.py

@@ -9,17 +9,9 @@
 from beanmachine.ppl.legacy.inference.proposer.single_site_newtonian_monte_carlo_proposer import (
     SingleSiteNewtonianMonteCarloProposer,
 )
-from beanmachine.ppl.legacy.inference.proposer.single_site_no_u_turn_sampler_proposer import (
-    SingleSiteNoUTurnSamplerProposer,
-)
-from beanmachine.ppl.legacy.inference.proposer.single_site_uniform_proposer import (
-    SingleSiteUniformProposer,
-)
 
 
 __all__ = [
     "SingleSiteAncestralProposer",
     "SingleSiteNewtonianMonteCarloProposer",
-    "SingleSiteNoUTurnSamplerProposer",
-    "SingleSiteUniformProposer",
 ]