Merge pull request #342 from calebweinreb/parallel_hmm_posterior_sample

Parallel hmm posterior sample

dynamax/hidden_markov_model/__init__.py

@@ -23,4 +23,5 @@
 from dynamax.hidden_markov_model.inference import compute_transition_probs
 
 from dynamax.hidden_markov_model.parallel_inference import hmm_filter as parallel_hmm_filter
-from dynamax.hidden_markov_model.parallel_inference import hmm_smoother as parallel_hmm_smoother
+from dynamax.hidden_markov_model.parallel_inference import hmm_smoother as parallel_hmm_smoother
+from dynamax.hidden_markov_model.parallel_inference import hmm_posterior_sample as parallel_hmm_posterior_sample

dynamax/hidden_markov_model/inference_test.py

@@ -2,6 +2,7 @@
 import itertools as it
 import jax.numpy as jnp
 import jax.random as jr
+from jax import vmap
 import dynamax.hidden_markov_model.inference as core
 import dynamax.hidden_markov_model.parallel_inference as parallel
 
@@ -285,3 +286,32 @@ def test_parallel_smoother(key=0, num_timesteps=100, num_states=3):
     posterior = core.hmm_smoother(initial_probs, transition_matrix, log_likelihoods)
     posterior2 = parallel.hmm_smoother(initial_probs, transition_matrix, log_likelihoods)
     assert jnp.allclose(posterior.smoothed_probs, posterior2.smoothed_probs, atol=1e-1)
+
+
+def test_parallel_posterior_sample(
+        key=0, num_timesteps=5, num_states=2, eps=1e-3, 
+        num_samples=1000000, num_iterations=5
+):
+    if isinstance(key, int):
+        key = jr.PRNGKey(key)
+
+    max_unique_size = 1 << num_timesteps
+
+    def iterate_test(key_iter):
+        keys_iter = jr.split(key_iter, num_samples)
+        args = random_hmm_args(key_iter, num_timesteps, num_states)
+
+        # Sample sequences from posterior
+        state_seqs = vmap(parallel.hmm_posterior_sample, (0, None, None, None), (0, 0))(keys_iter, *args)[1]
+        unique_seqs, counts = jnp.unique(state_seqs, axis=0, size=max_unique_size, return_counts=True)
+        blj_sample = counts / counts.sum()
+
+        # Compute joint probabilities
+        blj = jnp.exp(big_log_joint(*args))
+        blj = jnp.ravel(blj / blj.sum())
+
+        # Compare the joint distributions
+        return jnp.allclose(blj_sample, blj, rtol=0, atol=eps)
+
+    keys = jr.split(key, num_iterations)
+    assert iterate_test(keys[0])

dynamax/hidden_markov_model/parallel_inference.py

@@ -1,11 +1,23 @@
 import jax.numpy as jnp
+import jax.random as jr
 from jax import lax, vmap, value_and_grad
-from jaxtyping import Array, Float
+from jaxtyping import Array, Float, Int
 from typing import NamedTuple, Union
+from functools import partial
 
 from dynamax.hidden_markov_model.inference import HMMPosterior, HMMPosteriorFiltered
 
-class Message(NamedTuple):
+#---------------------------------------------------------------------------#
+#                                Filtering                                  #
+#---------------------------------------------------------------------------#
+
+class FilterMessage(NamedTuple):
+    """Filtering associative scan elements.
+
+    Attributes:
+        A: $p(z_j \mid z_i)$
+        log_b: $\log P(y_{i+1}, ..., y_j \mid z_i)$
+    """    
     A: Float[Array, "num_timesteps num_states num_states"]
     log_b: Float[Array, "num_timesteps num_states"]
 
@@ -43,15 +55,15 @@ def marginalize(m_ij, m_jk):
         A_ij_cond, lognorm = _condition_on(m_ij.A, m_jk.log_b)
         A_ik = A_ij_cond @ m_jk.A
         log_b_ik = m_ij.log_b + lognorm
-        return Message(A=A_ik, log_b=log_b_ik)
+        return FilterMessage(A=A_ik, log_b=log_b_ik)
 
 
     # Initialize the messages
     A0, log_b0 = _condition_on(initial_probs, log_likelihoods[0])
     A0 *= jnp.ones((K, K))
     log_b0 *= jnp.ones(K)
     A1T, log_b1T = vmap(_condition_on, in_axes=(None, 0))(transition_matrix, log_likelihoods[1:])
-    initial_messages = Message(
+    initial_messages = FilterMessage(
         A=jnp.concatenate([A0[None, :, :], A1T]),
         log_b=jnp.vstack([log_b0, log_b1T])
     )
@@ -72,6 +84,11 @@ def marginalize(m_ij, m_jk):
                                 predicted_probs=predicted_probs)
 
 
+#---------------------------------------------------------------------------#
+#                                 Smoothing                                 #
+#---------------------------------------------------------------------------#
+
+
 def hmm_smoother(initial_probs: Float[Array, "num_states"],
                  transition_matrix: Float[Array, "num_states num_states"],
                  log_likelihoods: Float[Array, "num_timesteps num_states"]
@@ -109,4 +126,69 @@ def log_normalizer(log_initial_probs, log_transition_matrix, log_likelihoods):
         initial_probs=smoothed_probs[0],
         smoothed_probs=smoothed_probs,
         trans_probs=trans_probs
-    )
+    )
+
+
+#---------------------------------------------------------------------------#
+#                                  Sampling                                 #
+#---------------------------------------------------------------------------#
+"""Associative scan elements $E_ij$ are vectors specifying a sample::
+
+    $z_j ~ p(z_j \mid z_i)$ 
+    
+for each possible value of $z_i$.
+"""
+
+def _initialize_sampling_messages(rng, transition_matrix, filtered_probs):
+    """Preprocess filtering output to construct input for sampling assocative scan."""
+
+    T, K = filtered_probs.shape
+    rngs = jr.split(rng, T)
+
+    def _last_message(rng, probs):
+        state = jr.choice(rng, K, p=probs)
+        return jnp.repeat(state, K)
+
+    @vmap
+    def _generic_message(rng, probs):
+        smoothed_probs = probs * transition_matrix.T
+        smoothed_probs = smoothed_probs / smoothed_probs.sum(1).reshape(K,1)
+        return vmap(lambda p: jr.choice(rng, K, p=p))(smoothed_probs)
+
+    En = _last_message(rngs[-1], filtered_probs[-1])
+    Et = _generic_message(rngs[:-1], filtered_probs[:-1])
+    return jnp.concatenate([Et, En[None]])
+
+
+def hmm_posterior_sample(rng: jr.PRNGKey,
+                         initial_distribution: Float[Array, "num_states"],
+                         transition_matrix: Float[Array, "num_states num_states"],
+                         log_likelihoods: Float[Array, "num_timesteps num_states"]
+) -> Int[Array, "num_timesteps"]:
+    r"""Sample a sequence of hidden states from the posterior.
+
+    Args:
+        rng: random number generator
+        initial_distribution: $p(z_1 \mid u_1, \theta)$
+        transition_matrix: $p(z_{t+1} \mid z_t, u_t, \theta)$
+        log_likelihoods: $p(y_t \mid z_t, u_t, \theta)$ for $t=1,\ldots, T$.
+ 
+    Returns:
+        log_normalizer: $\log P(y_{1:T} \mid u_{1:T}, \theta)$
+        states: sequence of hidden states $z_{1:T}$
+    """
+    T, K = log_likelihoods.shape
+
+    # Run the HMM filter
+    post = hmm_filter(initial_distribution, transition_matrix, log_likelihoods)
+    log_normalizer = post.marginal_loglik
+    filtered_probs = post.filtered_probs
+
+    @vmap
+    def _operator(E_jk, E_ij):
+        return jnp.take(E_ij, E_jk)
+
+    initial_messages = _initialize_sampling_messages(rng, transition_matrix, filtered_probs)
+    final_messages = lax.associative_scan(_operator, initial_messages, reverse=True)
+    states = final_messages[:,0]
+    return log_normalizer, states