fix: Replace unsafe under vectorize pt.zeros(value.shape) with zeros_like

Author: ferrine

pymc/logprob/transforms.py

@@ -961,7 +961,7 @@ def log_jac_det(self, value, *inputs):
         N = N.astype(value.dtype)
         sum_value = pt.sum(value, -1, keepdims=True)
         value_sum_expanded = value + sum_value
-        value_sum_expanded = pt.concatenate([value_sum_expanded, pt.zeros(sum_value.shape)], -1)
+        value_sum_expanded = pt.concatenate([value_sum_expanded, pt.zeros_like(sum_value)], -1)
         logsumexp_value_expanded = pt.logsumexp(value_sum_expanded, -1, keepdims=True)
         res = pt.log(N) + (N * sum_value) - (N * logsumexp_value_expanded)
         return pt.sum(res, -1)
@@ -977,7 +977,7 @@ def forward(self, value, *inputs):
         return pt.as_tensor_variable(value)
 
     def log_jac_det(self, value, *inputs):
-        return pt.zeros(value.shape)
+        return pt.zeros_like(value)
 
 
 class ChainedTransform(Transform):