allow for learning of single observation modality depending on the co…

…ntent of pA list

pymdp/jax/agent.py

@@ -104,8 +104,8 @@ def __init__(
         policy_len=1,
         control_fac_idx=None,
         policies=None,
-        gamma=16.0,
-        alpha=16.0,
+        gamma=1.0,
+        alpha=1.0,
         inductive_depth=1,
         inductive_threshold=0.1,
         inductive_epsilon=1e-3,
@@ -156,7 +156,7 @@ def __init__(
             factor_dims = tuple([self.num_states[f] for f in self.A_dependencies[m]])
             assert self.A[m].shape[2:] == factor_dims, f"Please input an `A_dependencies` whose {m}-th indices correspond to the hidden state factors that line up with lagging dimensions of A[{m}]..." 
             if self.pA != None:
-                assert self.pA[m].shape[2:] == factor_dims, f"Please input an `A_dependencies` whose {m}-th indices correspond to the hidden state factors that line up with lagging dimensions of pA[{m}]..." 
+                assert self.pA[m].shape[2:] == factor_dims if self.pA[m] is not None else True, f"Please input an `A_dependencies` whose {m}-th indices correspond to the hidden state factors that line up with lagging dimensions of pA[{m}]..." 
             assert max(self.A_dependencies[m]) <= (self.num_factors - 1), f"Check modality {m} of `A_dependencies` - must be consistent with `num_states` and `num_factors`..."
 
         # Ensure consistency of B_dependencies with num_states and num_factors
@@ -280,6 +280,7 @@ def infer_parameters(self, beliefs_A, outcomes, actions, beliefs_B=None, lr_pA=1
             lr = jnp.broadcast_to(lr_pA, (self.batch_size,))
             qA, E_qA = vmap(update_A)(
                 self.pA,
+                self.A,
                 outcomes,
                 marginal_beliefs,
                 lr=lr,

pymdp/jax/learning.py

@@ -26,10 +26,11 @@ def update_obs_likelihood_dirichlet_m(pA_m, obs_m, qs, dependencies_m, lr=1.0):
     dfda = vmap(multidimensional_outer)([obs_m] + relevant_factors).sum(axis=0)
 
     new_pA_m = pA_m + lr * dfda
+    A_m = dirichlet_expected_value(new_pA_m)
 
-    return new_pA_m, dirichlet_expected_value(new_pA_m)
+    return new_pA_m, A_m
 
-def update_obs_likelihood_dirichlet(pA, obs, qs, *, A_dependencies, onehot_obs, num_obs, lr):
+def update_obs_likelihood_dirichlet(pA, A, obs, qs, *, A_dependencies, onehot_obs, num_obs, lr):
     """ JAX version of ``pymdp.learning.update_obs_likelihood_dirichlet`` """
 
     obs_m = lambda o, dim: nn.one_hot(o, dim) if not onehot_obs else o
@@ -39,9 +40,13 @@ def update_obs_likelihood_dirichlet(pA, obs, qs, *, A_dependencies, onehot_obs,
     result = tree_map(update_A_fn, pA, obs, num_obs, A_dependencies)
     qA = []
     E_qA = []
-    for r in result:
-        qA.append(r[0])
-        E_qA.append(r[1])
+    for i, r in enumerate(result):
+        if r is None:
+            qA.append(r)
+            E_qA.append(A[i])
+        else:
+            qA.append(r[0])
+            E_qA.append(r[1])
 
     return qA, E_qA