Add SNAPER HMC from TFP to mcmc algorithms.

See https://arxiv.org/abs/2110.11576 for details.

PiperOrigin-RevId: 606680099

bayeux/_src/mcmc/tfp.py

@@ -0,0 +1,116 @@
+# Copyright 2024 The bayeux Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Copyright 2024 The bayeux Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""NumPyro specific code."""
+
+import arviz as az
+from bayeux._src import shared
+import jax
+import numpy as np
+import tensorflow_probability.substrates.jax as tfp
+
+
+class SnaperHMC(shared.Base):
+  """Implements SNAPER HMC [1] with step size adaptation.
+
+  [1]: Sountsov, P. & Hoffman, M. (2021). Focusing on Difficult Directions for
+  Learning HMC Trajectory Lengths. <https://arxiv.org/abs/2110.11576>
+  """
+  name = "tfp_snaper_hmc"
+
+  def get_kwargs(self, **kwargs):
+    kwargs_with_defaults = {
+        "num_results": 1_000,
+        "num_chains": 8,
+    } | kwargs
+    snaper = tfp.experimental.mcmc.sample_snaper_hmc
+    snaper_kwargs, snaper_required = shared.get_default_signature(snaper)
+    snaper_kwargs.update({k: kwargs_with_defaults[k] for k in snaper_required
+                          if k in kwargs_with_defaults})
+    snaper_required.remove("model")
+    # Initial state is handled internally
+    snaper_kwargs.pop("init_state")
+    # Seed set later
+    snaper_kwargs.pop("seed")
+
+    snaper_required = snaper_required - snaper_kwargs.keys()
+
+    if snaper_required:
+      raise ValueError(f"Unexpected required arguments: "
+                       f"{','.join(snaper_required)}. Probably file a bug, but "
+                       "you can try to manually supply them as keywords.")
+    snaper_kwargs.update({k: kwargs_with_defaults[k] for k in snaper_kwargs
+                          if k in kwargs_with_defaults})
+    return {
+        snaper: snaper_kwargs,
+        "extra_parameters": {
+            "return_pytree": kwargs.get("return_pytree", False)
+        },
+    }
+
+  def __call__(self, seed, **kwargs):
+    snaper = tfp.experimental.mcmc.sample_snaper_hmc
+    init_key, sample_key = jax.random.split(seed)
+    kwargs = self.get_kwargs(**kwargs)
+    initial_state = self.get_initial_state(
+        init_key, num_chains=kwargs[snaper]["num_chains"])
+
+    vmapped_constrained_log_prob = jax.vmap(self.constrained_log_density())
+
+    def tlp(*args, **kwargs):
+      if args:
+        return vmapped_constrained_log_prob(args)
+      else:
+        return vmapped_constrained_log_prob(kwargs)
+
+    (draws, trace), *_ = snaper(
+        model=tlp, init_state=initial_state, seed=sample_key, **kwargs[snaper]
+    )
+    draws = self.transform_fn(draws)
+    if kwargs["extra_parameters"]["return_pytree"]:
+      return draws
+
+    if hasattr(draws, "_asdict"):
+      draws = draws._asdict()
+    elif not isinstance(draws, dict):
+      draws = {"var0": draws}
+
+    draws = {x: np.swapaxes(v, 0, 1) for x, v in draws.items()}
+    return az.from_dict(posterior=draws, sample_stats=_tfp_stats_to_dict(trace))
+
+
+def _tfp_stats_to_dict(stats):
+  new_stats = {}
+  for k, v in stats.items():
+    if k == "variance_scaling":
+      continue
+    if np.ndim(v) > 1:
+      new_stats[k] = np.swapaxes(v, 0, 1)
+    else:
+      new_stats[k] = v
+  return new_stats

bayeux/mcmc/__init__.py

@@ -15,9 +15,13 @@
 """Imports from submodules."""
 # pylint: disable=g-importing-member
 # pylint: disable=g-import-not-at-top
+# pylint: disable=g-bad-import-order
 import importlib
 
-__all__ = []
+# TFP-on-JAX always installed
+from bayeux._src.mcmc.tfp import SnaperHMC as SNAPER_HMC_TFP
+__all__ = ["SNAPER_HMC_TFP"]
+
 if importlib.util.find_spec("blackjax") is not None:
   from bayeux._src.mcmc.blackjax import CheesHMC as CheesHMCblackjax
   from bayeux._src.mcmc.blackjax import HMC as HMCblackjax

bayeux/tests/mcmc_test.py

@@ -63,6 +63,20 @@ def test_return_pytree_numpyro():
   assert pytree["x"]["y"].shape == (4, 10)
 
 
+def test_return_pytree_tfp():
+  model = bx.Model(log_density=lambda pt: -pt["x"]["y"]**2,
+                   test_point={"x": {"y": jnp.array(1.)}})
+  seed = jax.random.PRNGKey(0)
+  pytree = model.mcmc.tfp_snaper_hmc(
+      seed=seed,
+      return_pytree=True,
+      num_chains=4,
+      num_results=10,
+      num_burnin_steps=10,
+  )
+  assert pytree["x"]["y"].shape == (10, 4)
+
+
 @pytest.mark.parametrize("method", METHODS)
 def test_samplers(method):
   # flowMC samplers are broken for 0 or 1 dimensions, so just test