Fix crash with the gradient of {inc,set}_subtensor(mat, row) (row get broadcasted inside mat)

theano/tensor/subtensor.py

@@ -1440,6 +1440,25 @@ class IncSubtensor(Op):
         else:
             gx = g_output
         gy = Subtensor(idx_list=self.idx_list)(g_output, *idx_list)
+        if gy.broadcastable != y.broadcastable:
+            y_broad = (True,) * (gy.ndim - y.ndim) + y.broadcastable
+            assert sum(gy.broadcastable) < sum(y_broad)
+            axis_to_sum = []
+            for i in range(gy.ndim):
+                if gy.broadcastable[i] is False and y_broad[i] is True:
+                    axis_to_sum.append(i)
+                elif (gy.broadcastable[i] is True and
+                      y_broad[i] is False):
+                    # This mean that THeano where able to infer that
+                    # gy.shape[i] is 1, so y.shape[i] is 1, but we
+                    # didn't know it. It is fine.
+                    pass
+                else:
+                    assert gy.broadcastable[i] == y_broad[i]
+            gy = gy.sum(axis=axis_to_sum, keepdims=True)
+            if gy.ndim != y.ndim:
+                gy = gy.dimshuffle(*range(y.ndim, gy.ndim))
+            assert gy.broadcastable == y.broadcastable
 
         return [gx, gy] + [DisconnectedType()()] * len(idx_list)
 

theano/tensor/tests/test_subtensor.py

@@ -378,6 +378,26 @@ class T_subtensor(unittest.TestCase, utt.TestOptimizationMixin):
         good[subi:, subi] = numpy.exp(data[subi:, subi])
         self.assertTrue(numpy.allclose(gval, good), (gval, good))
 
+    def test_grad_2d_inc_set_subtensor(self):
+        for n_shape, m_shape in [
+            [(2, 3), (2, 2)],
+            [(3, 2), (2, 2)],
+            [(3, 2), (1, 2)],
+            [(3, 2), (2,)],
+        ]:
+            for op in [inc_subtensor, set_subtensor]:
+                subi = 2
+                data = numpy.asarray(rand(*n_shape), dtype=self.dtype)
+                n = self.shared(data)
+                z = scal.constant(subi)
+                m = matrix('m', dtype=self.dtype)
+                mv = numpy.asarray(rand(*m_shape), dtype=self.dtype)
+
+                t = op(n[:z, :z], m)
+                gn, gm = theano.tensor.grad(theano.tensor.sum(t), [n, m])
+                utt.verify_grad(lambda m: op(n[:z, :z], m), [mv])
+                utt.verify_grad(lambda nn: op(nn[:z, :z], mv), [data])
+
     def test_grad_0d(self):
         data = numpy.asarray(rand(2, 3), dtype=self.dtype)
         n = self.shared(data)