Support UnpackTransform.inv via pack_fn (pyro-ppl#1824)

* add pack_fn to UnpackTransform for inverse

* better pack_fn

numpyro/distributions/transforms.py

@@ -9,7 +9,6 @@
 
 import jax
 from jax import lax, vmap
-from jax.flatten_util import ravel_pytree
 from jax.nn import log_sigmoid, softplus
 import jax.numpy as jnp
 from jax.scipy.linalg import solve_triangular
@@ -1102,13 +1101,15 @@ class UnpackTransform(Transform):
     Transforms a contiguous array to a pytree of subarrays.
 
     :param unpack_fn: callable used to unpack a contiguous array.
+    :param pack_fn: callable used to pack a pytree into a contiguous array.
     """
 
     domain = constraints.real_vector
     codomain = constraints.dependent
 
-    def __init__(self, unpack_fn):
+    def __init__(self, unpack_fn, pack_fn=None):
         self.unpack_fn = unpack_fn
+        self.pack_fn = pack_fn
 
     def __call__(self, x):
         batch_shape = x.shape[:-1]
@@ -1121,9 +1122,15 @@ def __call__(self, x):
             return self.unpack_fn(x)
 
     def _inverse(self, y):
+        if self.pack_fn is None:
+            raise NotImplementedError(
+                "pack_fn needs to be provided to perform UnpackTransform.inv."
+            )
         leading_dims = [
             v.shape[0] if jnp.ndim(v) > 0 else 0 for v in jax.tree.flatten(y)[0]
         ]
+        if not leading_dims:
+            return jnp.array([])
         d0 = leading_dims[0]
         not_scalar = d0 > 0 or len(leading_dims) > 1
         if not_scalar and all(d == d0 for d in leading_dims[1:]):
@@ -1132,7 +1139,7 @@ def _inverse(self, y):
                 " cannot transform a batch of unpacked arrays.",
                 stacklevel=find_stack_level(),
             )
-        return ravel_pytree(y)[0]
+        return self.pack_fn(y)
 
     def log_abs_det_jacobian(self, x, y, intermediates=None):
         return jnp.zeros(jnp.shape(x)[:-1])
@@ -1145,10 +1152,14 @@ def inverse_shape(self, shape):
 
     def tree_flatten(self):
         # XXX: what if unpack_fn is a parametrized callable pytree?
-        return (), ((), {"unpack_fn": self.unpack_fn})
+        return (), ((), {"unpack_fn": self.unpack_fn, "pack_fn": self.pack_fn})
 
     def __eq__(self, other):
-        return isinstance(other, UnpackTransform) and self.unpack_fn is other.unpack_fn
+        return (
+            isinstance(other, UnpackTransform)
+            and (self.unpack_fn is other.unpack_fn)
+            and (self.pack_fn is other.pack_fn)
+        )
 
 
 def _get_target_shape(shape, forward_shape, inverse_shape):

numpyro/infer/autoguide.py

@@ -624,13 +624,20 @@ def _unravel_dict(x_flat, shape_dict):
 def _ravel_dict(x):
     """Return the flatten version of `x` and shapes of each item in `x`."""
     assert isinstance(x, dict)
-    shape_dict = {}
+    shape_dict = {name: jnp.shape(value) for name, value in x.items()}
+    x_flat = _ravel_dict_with_shape_dict(x, shape_dict)
+    return x_flat, shape_dict
+
+
+def _ravel_dict_with_shape_dict(x, shape_dict):
+    assert set(x.keys()) == set(shape_dict.keys())
     x_flat = []
-    for name, value in x.items():
-        shape_dict[name] = jnp.shape(value)
+    for name, shape in shape_dict.items():
+        value = x[name]
+        assert shape == jnp.shape(value)
         x_flat.append(jnp.reshape(value, -1))
     x_flat = jnp.concatenate(x_flat) if x_flat else jnp.zeros((0,))
-    return x_flat, shape_dict
+    return x_flat
 
 
 class AutoContinuous(AutoGuide):
@@ -661,7 +668,9 @@ def _setup_prototype(self, *args, **kwargs):
         unpack_latent = partial(_unravel_dict, shape_dict=shape_dict)
         # this is to match the behavior of Pyro, where we can apply
         # unpack_latent for a batch of samples
-        self._unpack_latent = UnpackTransform(unpack_latent)
+        self._unpack_latent = UnpackTransform(
+            unpack_latent, _ravel_dict_with_shape_dict
+        )
         self.latent_dim = jnp.size(self._init_latent)
         if self.latent_dim == 0:
             raise RuntimeError(

test/test_distributions.py

@@ -2703,7 +2703,8 @@ def test_compose_transform_with_intermediates(ts):
 def test_unpack_transform(x_dim, y_dim):
     xy = np.random.randn(x_dim + y_dim)
     unpack_fn = lambda xy: {"x": xy[:x_dim], "y": xy[x_dim:]}  # noqa: E731
-    transform = transforms.UnpackTransform(unpack_fn)
+    pack_fn = lambda d: jnp.concatenate([d["x"], d["y"]], axis=-1)  # noqa: E731
+    transform = transforms.UnpackTransform(unpack_fn, pack_fn)
     z = transform(xy)
     if x_dim == y_dim:
         with pytest.warns(UserWarning, match="UnpackTransform.inv"):