Basic implementation

Author: jessegrabowski

pymc/distributions/multivariate.py

@@ -1510,7 +1510,7 @@ def helper_deterministics(cls, n, packed_chol):
 
 class LKJCorrRV(RandomVariable):
     name = "lkjcorr"
-    signature = "(),()->(n)"
+    signature = "(),()->(n,n)"
     dtype = "floatX"
     _print_name = ("LKJCorrRV", "\\operatorname{LKJCorrRV}")
 
@@ -1527,8 +1527,8 @@ def make_node(self, rng, size, n, eta):
 
     def _supp_shape_from_params(self, dist_params, **kwargs):
         n = dist_params[0].squeeze()
-        dist_shape = ((n * (n - 1)) // 2,)
-        return dist_shape
+        # dist_shape = ((n * (n - 1)) // 2,)
+        return (n, n)
 
     @classmethod
     def rng_fn(cls, rng, n, eta, size):
@@ -1609,23 +1609,26 @@ def logp(value, n, eta):
         -------
         TensorVariable
         """
-        if value.ndim > 1:
-            raise NotImplementedError("LKJCorr logp is only implemented for vector values (ndim=1)")
-
-        # TODO: PyTensor does not have a `triu_indices`, so we can only work with constant
-        #  n (or else find a different expression)
+        # if value.ndim > 1:
+        #     raise NotImplementedError("LKJCorr logp is only implemented for vector values (ndim=1)")
+        #
         try:
             n = int(get_underlying_scalar_constant_value(n))
         except NotScalarConstantError:
             raise NotImplementedError("logp only implemented for constant `n`")
 
-        shape = n * (n - 1) // 2
-        tri_index = np.zeros((n, n), dtype="int32")
-        tri_index[np.triu_indices(n, k=1)] = np.arange(shape)
-        tri_index[np.triu_indices(n, k=1)[::-1]] = np.arange(shape)
+        # shape = n * (n - 1) // 2
+        # tri_index = np.zeros((n, n), dtype="int32")
+        # tri_index[np.triu_indices(n, k=1)] = np.arange(shape)
+        # tri_index[np.triu_indices(n, k=1)[::-1]] = np.arange(shape)
+
+        # value = pt.take(value, tri_index)
+        # value = pt.fill_diagonal(value, 1)
 
-        value = pt.take(value, tri_index)
-        value = pt.fill_diagonal(value, 1)
+        # print(n, type(n))
+        # print(value.type.shape)
+        # value = value @ value.T
+        # print(value.type.shape)
 
         # TODO: _lkj_normalizing_constant currently requires `eta` and `n` to be constants
         try:

pymc/distributions/transforms.py

@@ -164,6 +164,56 @@ def log_jac_det(self, value, *inputs):
         return pt.sum(value[..., self.diag_idxs], axis=-1)
 
 
+class CholeskyCorr(Transform):
+    """Get a Cholesky Corr from a packed vector."""
+
+    name = "cholesky-corr-packed"
+
+    def __init__(self, n):
+        """Create a CholeskyCorrPack object.
+
+        Parameters
+        ----------
+        n: int
+            Number of diagonal entries in the LKJCholeskyCov distribution
+        """
+        self.n = n
+
+    def _compute_L_and_logdet(self, value, *inputs):
+        n = self.n
+        counter = 0
+        L = pt.eye(n)
+        log_det = 0
+
+        for i in range(1, n):
+            y_star = value[counter : counter + i]
+            dsy = y_star.dot(y_star)
+            alpha_r = 1 / (dsy + 1)
+            gamma = pt.sqrt(dsy + 2) * alpha_r
+
+            x = pt.join(0, gamma * y_star, pt.atleast_1d(alpha_r))
+            L = L[i, : i + 1].set(x)
+            log_det += pt.log(2) + 0.5 * (i - 2) * pt.log(dsy + 2) - i * pt.log(1 + dsy)
+
+            counter += i
+
+        # Return whole matrix? Or just lower triangle?
+        return L, log_det
+
+    def backward(self, value, *inputs):
+        L, _ = self._compute_L_and_logdet(value, *inputs)
+        return L
+
+    def forward(self, value, *inputs):
+        # TODO: This is a placeholder
+        n = self.n
+        return pt.as_tensor_variable(np.random.normal(size=(n,)))
+
+    def log_jac_det(self, value, *inputs):
+        _, log_det = self._compute_L_and_logdet(value, *inputs)
+        return log_det
+
+
 Chain = ChainedTransform
 
 simplex = SimplexTransform()