Fix in advanced_inc_subtensor1 if index is broadcastable

theano/tensor/subtensor.py

@@ -1157,7 +1157,11 @@ def inc_subtensor(x, y, inplace=False, set_instead_of_inc=False,
             inplace=inplace,
             set_instead_of_inc=set_instead_of_inc,
             tolerate_inplace_aliasing=tolerate_inplace_aliasing)
-        return x.owner.op(inner_incsubtensor, *x.owner.inputs[1:])
+        # The broadcastable pattern of inner_x may not be the same as
+        # the one of x, so we have to build a new dimshuffle here,
+        # instead of reusing x.owner.op().
+        return inner_incsubtensor.dimshuffle(x.owner.op.new_order)
+
     elif isinstance(x.owner.op, theano.tensor.Reshape):
         # This case happens when the indices are not arranged as a vector, but
         # as a higher-dimensional array. This is handled by the subtensor

theano/tensor/tests/test_subtensor.py

@@ -1325,6 +1325,20 @@ class TestIncSubtensor1(unittest.TestCase):
         utt.assert_allclose(a2val[2], mval[2] * 3)
         utt.assert_allclose(a2val[3], mval[3] * 2)
 
+    def test_inc_bcastableidx(self):
+        idx = tensor.constant([0])
+        c_inc = tensor.col()
+        m_inc = tensor.matrix()
+        out1 = inc_subtensor(self.m[:, idx], c_inc)
+        out2 = inc_subtensor(self.m[:, idx], m_inc)
+
+        f = theano.function([self.m, c_inc, m_inc], [out1, out2])
+        mval = self.rng.random_sample((10, 5))
+        incval = self.rng.random_sample((10, 1)).astype(config.floatX)
+
+        out1val, out2val = f(mval, incval, incval)
+        utt.assert_allclose(out1val, out2val)
+
 
 inplace_increment_missing = SkipTest(
     "inc_subtensor with advanced indexing not enabled. "