Improve static output shape of Reshape

pytensor/tensor/shape.py

@@ -669,7 +669,7 @@ def make_node(self, x, shp):
         assert shp.ndim == 1
 
         if isinstance(shp, TensorConstant):
-            out_shape = tuple(int(s) if s >= 0 else None for s in shp.data)
+            out_shape = [int(s) if s >= 0 else None for s in shp.data]
         else:
             out_shape = [None] * self.ndim
             shp_list = shp_orig
@@ -685,6 +685,29 @@ def make_node(self, x, shp):
                 except NotScalarConstantError:
                     pass
 
+        # If we only don't know the size of one output dimension,
+        # but we know all the input dimensions we can deduce it
+        # This happens often when there is -1 as an input of Reshape
+        if None not in x.type.shape and out_shape.count(None) == 1:
+            full_size = np.prod(x.type.shape)
+            known_size = np.prod([s for s in out_shape if s is not None])
+            out_shape[out_shape.index(None)] = int(full_size // known_size)
+
+        out_shape = tuple(out_shape)
+
+        # Run some eager error checks
+        if len(out_shape) != self.ndim:
+            raise ValueError(
+                "Shape argument to Reshape has incorrect length:"
+                f" {len(out_shape)}, should be {self.ndim}"
+            )
+
+        if None not in x.type.shape and None not in out_shape:
+            if np.prod(x.type.shape) != np.prod(out_shape):
+                raise ValueError(
+                    f"Reshape: Input shape {x.type.shape} is incompatible with new shape {out_shape}"
+                )
+
         return Apply(self, [x, shp], [tensor(dtype=x.type.dtype, shape=out_shape)])
 
     def perform(self, node, inp, out_):

tests/tensor/test_shape.py

@@ -1,3 +1,5 @@
+import re
+
 import numpy as np
 import pytest
 
@@ -353,6 +355,29 @@ def test_rebuild(self):
         assert tuple(y_new.shape.eval({i: i_test})) == (4, 25)
         assert y_new.eval({i: i_test}).shape == (4, 25)
 
+    def test_static_shape(self):
+        dim = lscalar("dim")
+        x1 = tensor(shape=(2, 2, None))
+        x2 = specify_shape(x1, (2, 2, 6))
+
+        assert reshape(x1, (6, 2)).type.shape == (6, 2)
+        assert reshape(x1, (6, -1)).type.shape == (6, None)
+        assert reshape(x1, (6, dim)).type.shape == (6, None)
+        assert reshape(x1, (6, dim, 2)).type.shape == (6, None, 2)
+        assert reshape(x1, (6, 3, 99)).type.shape == (6, 3, 99)
+
+        assert reshape(x2, (6, 4)).type.shape == (6, 4)
+        assert reshape(x2, (6, -1)).type.shape == (6, 4)
+        assert reshape(x2, (6, dim)).type.shape == (6, 4)
+        assert reshape(x2, (6, dim, 2)).type.shape == (6, 2, 2)
+        with pytest.raises(
+            ValueError,
+            match=re.escape(
+                "Reshape: Input shape (2, 2, 6) is incompatible with new shape (6, 3, 99)"
+            ),
+        ):
+            reshape(x2, (6, 3, 99))
+
 
 def test_shape_i_hash():
     assert isinstance(Shape_i(np.int64(1)).__hash__(), int)