this is a mess

autofit/messages/abstract.py

@@ -52,10 +52,21 @@ def __init__(
 
         self.id = next(self.ids) if id_ is None else id_
         self.log_norm = log_norm
-        self._broadcast = np.broadcast(*parameters)
+
+        self._broadcast = None
+        self._broadcast_jnp = None
+
+        if isinstance(parameters[0], (np.float64, float, int)):
+            self._broadcast = np.broadcast(*parameters)
+        else:
+            import jax.numpy as jnp
+            self._broadcast_jnp = jnp.broadcast_arrays(*parameters)
 
         if self.shape:
-            self.parameters = tuple(np.asanyarray(p) for p in parameters)
+            if isinstance(parameters[0], (np.float64, float, int)):
+                self.parameters = tuple(np.asanyarray(p) for p in parameters)
+            else:
+                self.parameters = tuple(jnp.asarray(p) for p in parameters)
         else:
             self.parameters = tuple(parameters)
 

autofit/messages/interface.py

@@ -23,7 +23,11 @@ def broadcast(self):
 
     @property
     def shape(self) -> Tuple[int, ...]:
-        return self.broadcast.shape
+
+        if self.broadcast is not None:
+            return self.broadcast.shape
+
+        return ()
 
     @property
     def size(self) -> int:
@@ -47,6 +51,7 @@ def logpdf(self, x: Union[np.ndarray, float]) -> np.ndarray:
 
     def _broadcast_natural_parameters(self, x):
         shape = np.shape(x)
+        print(shape, self.shape)
         if shape == self.shape:
             return self.natural_parameters
         elif shape[1:] == self.shape:
@@ -78,8 +83,15 @@ def log_partition(self) -> np.ndarray:
 
     @classmethod
     def natural_logpdf(cls, eta, t, log_base, log_partition):
-        eta_t = np.multiply(eta, t).sum(0)
-        return np.nan_to_num(log_base + eta_t - log_partition, nan=-np.inf)
+
+        if isinstance(eta, (np.ndarray, np.float64)):
+            eta_t = np.multiply(eta, t).sum(0)
+            return np.nan_to_num(log_base + eta_t - log_partition, nan=-np.inf)
+
+        import jax.numpy as jnp
+
+        eta_t = jnp.multiply(eta, t).sum(0)
+        return jnp.nan_to_num(log_base + eta_t - log_partition, nan=-jnp.inf)
 
     def numerical_logpdf_gradient(
         self, x: np.ndarray, eps: float = 1e-6

autofit/messages/normal.py

@@ -40,7 +40,13 @@ def log_partition(self):
         This ensures normalization of the exponential-family distribution.
         """
         eta1, eta2 = self.natural_parameters
-        return -(eta1**2) / 4 / eta2 - np.log(-2 * eta2) / 2
+
+        if isinstance(eta1, (np.ndarray, np.float64)):
+            return -(eta1**2) / 4 / eta2 - np.log(-2 * eta2) / 2
+
+        import jax.numpy as jnp
+
+        return -(eta1**2) / 4 / eta2 - jnp.log(-2 * eta2) / 2
 
     log_base_measure = -0.5 * np.log(2 * np.pi)
     _support = ((-np.inf, np.inf),)
@@ -73,8 +79,9 @@ def __init__(
         id_
             An optional unique identifier used to track the message in larger probabilistic graphs or models.
         """
-        if (np.array(sigma) < 0).any():
-            raise exc.MessageException("Sigma cannot be negative")
+        if isinstance(sigma, (float, int, np.ndarray)):
+            if (np.array(sigma) < 0).any():
+                raise exc.MessageException("Sigma cannot be negative")
 
         super().__init__(
             mean,
@@ -158,7 +165,13 @@ def calc_natural_parameters(mu : Union[float, np.ndarray], sigma : Union[float,
             η₂ = -1 / (2σ²)
         """
         precision = 1 / sigma**2
-        return np.array([mu * precision, -precision / 2])
+
+        if isinstance(mu, (np.ndarray, np.float64)):
+            return np.array([mu * precision, -precision / 2])
+
+        import jax.numpy as jnp
+
+        return jnp.array([mu * precision, -precision / 2])
 
     @staticmethod
     def invert_natural_parameters(natural_parameters : np.ndarray) -> Tuple[float, float]:
@@ -197,7 +210,13 @@ def to_canonical_form(x : Union[float, np.ndarray]) -> np.ndarray:
         -------
         The sufficient statistics [x, x²].
         """
-        return np.array([x, x**2])
+
+        if isinstance(x, (np.ndarray, np.float64)):
+            return np.array([x, x**2])
+
+        import jax.numpy as jnp
+
+        return jnp.array([x, x**2])
 
     @classmethod
     def invert_sufficient_statistics(cls, suff_stats: Tuple[float, float]) -> np.ndarray: