Fix dataclass pytree registration bug (#34)

* Fix dataclass pytree registration

* Simplify

pgmax/fg/graph.py

@@ -117,30 +117,29 @@ def run_bp(
             msgs = init_msgs
         else:
             msgs = self.get_init_msgs(msgs_context)
+
+        wiring = jax.device_put(self.wiring)
         evidence = self.get_evidence(evidence_data, evidence_context)
-        edges_num_states = jax.device_put(self.wiring.edges_num_states)
-        var_states_for_edges = jax.device_put(self.wiring.var_states_for_edges)
-        factor_configs_edge_states = jax.device_put(
-            self.wiring.factor_configs_edge_states
-        )
-        max_msg_size = int(jnp.max(edges_num_states))
+        max_msg_size = int(jnp.max(wiring.edges_num_states))
 
         # Normalize the messages to ensure the maximum value is 0.
-        msgs = infer.normalize_and_clip_msgs(msgs, edges_num_states, max_msg_size)
-        num_val_configs = int(factor_configs_edge_states[-1, 0])
+        msgs = infer.normalize_and_clip_msgs(
+            msgs, wiring.edges_num_states, max_msg_size
+        )
+        num_val_configs = int(wiring.factor_configs_edge_states[-1, 0])
 
         @jax.jit
         def message_passing_step(msgs, _):
             # Compute new variable to factor messages by message passing
             vtof_msgs = infer.pass_var_to_fac_messages(
                 msgs,
                 evidence,
-                var_states_for_edges,
+                wiring.var_states_for_edges,
             )
             # Compute new factor to variable messages by message passing
             ftov_msgs = infer.pass_fac_to_var_messages(
                 vtof_msgs,
-                factor_configs_edge_states,
+                wiring.factor_configs_edge_states,
                 num_val_configs,
             )
             # Use the results of message passing to perform damping and
@@ -151,7 +150,7 @@ def message_passing_step(msgs, _):
             # them.
             msgs = infer.normalize_and_clip_msgs(
                 msgs,
-                edges_num_states,
+                wiring.edges_num_states,
                 max_msg_size,
             )
             return msgs, None
@@ -180,8 +179,6 @@ def decode_map_states(
         # NOTE: Having to regenerate the evidence here is annoying - there must be a better way to handle evidence and
         # message initialization.
         evidence = self.get_evidence(evidence_data, evidence_context)
-        # TODO: Once issue #20 is resolved, just grab this from self.wiring instead of
-        # casting it to a jnp.array
         var_states_for_edges = jax.device_put(self.wiring.var_states_for_edges)
         final_var_states = evidence.at[var_states_for_edges].add(msgs)
         var_to_map_dict = {}

pgmax/fg/nodes.py

@@ -1,8 +1,9 @@
 """A module containing classes that specify the components of a Factor Graph."""
 
-from dataclasses import dataclass
+from dataclasses import asdict, dataclass
 from typing import Mapping, Tuple, Union
 
+import jax
 import jax.numpy as jnp
 import numpy as np
 
@@ -18,7 +19,7 @@ class Variable:
     num_states: int
 
 
-@utils.register_pytree_node_dataclass
+@jax.tree_util.register_pytree_node_class
 @dataclass(frozen=True, eq=False)
 class EnumerationWiring:
     """Wiring for enumeration factors.
@@ -40,7 +41,15 @@ class EnumerationWiring:
 
     def __post_init__(self):
         for field in self.__dataclass_fields__:
-            getattr(self, field).flags.writeable = False
+            if isinstance(getattr(self, field), np.ndarray):
+                getattr(self, field).flags.writeable = False
+
+    def tree_flatten(self):
+        return jax.tree_util.tree_flatten(asdict(self))
+
+    @classmethod
+    def tree_unflatten(cls, aux_data, children):
+        return cls(**aux_data.unflatten(children))
 
 
 @dataclass(frozen=True, eq=False)

pgmax/utils.py

@@ -1,10 +1,7 @@
 """A module containing helper functions useful while constructing Factor Graphs."""
 
-import dataclasses
 import functools
-from typing import Any, Callable
-
-import jax
+from typing import Callable
 
 
 def cached_property(func: Callable) -> property:
@@ -17,23 +14,3 @@ def cached_property(func: Callable) -> property:
         Decorated cached property
     """
     return property(functools.lru_cache(None)(func))
-
-
-def register_pytree_node_dataclass(cls: Any) -> Any:
-    """Decorator to register a dataclass as a pytree
-
-    Args:
-        cls: A dataclass to be registered as a pytree
-
-    Returns:
-        The registered dataclass
-    """
-
-    def _flatten(obj):
-        jax.tree_flatten(dataclasses.asdict(obj))
-
-    def _unflatten(d, children):
-        cls(**d.unflatten(children))
-
-    jax.tree_util.register_pytree_node(cls, _flatten, _unflatten)
-    return cls