Merge pull request #1246 from lamblin/fix_check_incsubtensor

theano/tensor/basic.py

@@ -5052,10 +5052,11 @@ def inc_subtensor(x, y, inplace=False, set_instead_of_inc=False,
         dim_offset = x.ndim - y.ndim
         if (x.broadcastable[dim + dim_offset]
                 and not y.broadcastable[dim]):
-            raise TypeError(("Trying to increment a subtensor with "
-                "broadcastable dimension %d, with a tensor not broadcastable "
-                "on corresponding dimension %d.") % (dim + dim_offset, dim),
-                x.broadcastable, y.broadcastable)
+            # It is acceptable to try to increment a subtensor with a
+            # a broadcastable dim with a tensor that is not broadcastable
+            # on that dimension. However, its length must then be 1.
+            # We insert a Rebroadcast Op to make sure it is the case.
+            y = addbroadcast(y, dim)
 
     if not x.owner:
         raise TypeError('x must be result of a subtensor operation')

theano/tensor/tests/test_inc_subtensor.py

@@ -68,10 +68,23 @@ class Test_inc_subtensor(unittest.TestCase):
         a = tt.col()
         increment = tt.vector()
 
-        self.assertRaises(TypeError, tt.set_subtensor, a[:], increment)
-        self.assertRaises(TypeError, tt.set_subtensor, a[0], increment)
-        self.assertRaises(TypeError, tt.inc_subtensor, a[:], increment)
-        self.assertRaises(TypeError, tt.inc_subtensor, a[0], increment)
+        # These symbolic graphs legitimate, as long as increment has exactly
+        # one element. So it should fail at runtime, not at compile time.
+        rng = numpy.random.RandomState(utt.fetch_seed())
+
+        for op in (tt.set_subtensor, tt.inc_subtensor):
+            for base in (a[:], a[0]):
+                out = op(base, increment)
+                f = theano.function([a, increment], out)
+                # This one should work
+                f(rng.rand(3, 1), rng.rand(1))
+                # These ones should not
+                self.assertRaises(ValueError,
+                        f, rng.rand(3, 1), rng.rand(2))
+                self.assertRaises(ValueError,
+                        f, rng.rand(3, 1), rng.rand(3))
+                self.assertRaises(ValueError,
+                        f, rng.rand(3, 1), rng.rand(0))
 
     def test_simple_3d(self):
         """Increments or sets part of a tensor by a scalar using full slice and