Cast output of local_func_inv and local_exp_log to float when needed

aesara/tensor/math_opt.py

@@ -225,7 +225,12 @@ def local_func_inv(fgraph, node):
         if is_inverse_pair(node_op, prev_op, inv_pair):
             # We don't need to copy stack trace, because the optimization
             # is trivial and maintains the earlier stack trace
-            return x.owner.inputs
+            ottype = node.out.dtype
+            inp = x.owner.inputs[0]
+            # Functions may have casted integer input to float
+            if inp.dtype != ottype:
+                inp = cast(inp, ottype)
+            return [inp]
 
     return
 
@@ -246,7 +251,12 @@ def local_exp_log(fgraph, node):
 
     # Case for log(exp(x))
     if isinstance(prev_op, aes.Exp) and isinstance(node_op, aes.Log):
-        return x.owner.inputs
+        new_out = x.owner.inputs[0]
+        old_out = node.outputs[0]
+        # Exp may have casted integer input to float
+        if new_out.dtype != old_out.dtype:
+            new_out = cast(new_out, old_out.dtype)
+        return [new_out]
 
     # Case for exp(softplus(x)) aka exp(log1pexp)
     if isinstance(prev_op, aes_math.Softplus) and isinstance(node_op, aes.Exp):

tests/tensor/test_math_opt.py

@@ -2488,6 +2488,16 @@ def test(self):
         self.assert_func_pair_optimized(rad2deg, rad2deg, dx, should_copy=False)
         self.assert_func_pair_optimized(rad2deg, cosh, dx, should_copy=False)
 
+    def test_integer_upcast(self):
+        """
+        All invertible methods (except for `Neg`) can upgrade their input to float.
+        Here we test that the rewrite works with just one pair of methods
+        """
+        x = ivector("x")
+        f = function([x], deg2rad(rad2deg(x)), mode=self.mode)
+        topo = f.maker.fgraph.toposort()
+        assert len(topo) == 1
+
 
 class TestExpLog:
     def setup_method(self):
@@ -2512,6 +2522,17 @@ def test_log_exp(self):
         assert len(ops_graph) == 0
         np.testing.assert_array_equal(f(data), data)
 
+    def test_log_exp_integer_upcast(self):
+        x = ivector("x")
+        f = function([x], log(exp(x)), mode=self.mode)
+        ops_graph = [
+            node
+            for node in f.maker.fgraph.toposort()
+            if isinstance(node.op, Elemwise)
+            and isinstance(node.op.scalar_op, (aes.Log, aes.Exp))
+        ]
+        assert len(ops_graph) == 0
+
     def test_exp_log(self):
         # exp(log(x)) -> switch(x >= 0, x, nan)
         data_valid = np.random.random((4, 3)).astype("float32")