Add JAX implementation for InvGammaRV

aesara/link/jax/dispatch/random.py

@@ -453,3 +453,24 @@ def sample_fn(rng, size, dtype, *parameters):
         return (rng, samples)
 
     return sample_fn
+
+
+@jax_sample_fn.register(aer.InvGammaRV)
+def jax_sample_fn_invgamma(op):
+    """JAX implementation of `InvGammaRV`."""
+
+    def sample_fn(rng, size, dtype, *parameters):
+        rng_key = rng["jax_state"]
+        rng_key, sampling_key = jax.random.split(rng_key, 2)
+
+        (
+            shape,
+            scale,
+        ) = parameters
+        # InvGamma[shape, scale] <-> 1 / Gamma[shape, 1 / scale]
+        samples = 1 / (jax.random.gamma(sampling_key, shape, size, dtype) / scale)
+
+        rng["jax_state"] = rng_key
+        return (rng, samples)
+
+    return sample_fn

tests/link/jax/test_random.py

@@ -467,6 +467,19 @@ def test_random_dirichlet(parameter, size):
     np.testing.assert_allclose(samples.mean(axis=0), 0.5, 1)
 
 
+@pytest.mark.parametrize(
+    "shape, scale",
+    [(3, 3), (2, 1), (2, 5)],
+)
+def test_random_invgamma(shape, scale):
+    rng = shared(np.random.RandomState(123))
+    g = at.random.invgamma(shape, scale, size=(100000,), rng=rng)
+    g_fn = function([], g, mode=jax_mode)
+    samples = g_fn()
+    # mean = scale / (shape - 1) only exists for shape > 1
+    np.testing.assert_allclose(samples.mean(), scale / (shape - 1), rtol=1e-01)
+
+
 def test_random_choice():
     # Elements are picked at equal frequency
     num_samples = 10000