Merge pull request #5070 from JesseLivezey/softmax_dnn_grad_opt_fix

theano/sandbox/cuda/dnn.py

@@ -3110,8 +3110,9 @@ if True:
                 return
             ins = []
             for n in node.inputs:
-                if isinstance(n.owner.op, HostFromGpu):
-                    n = n.owner.inputs[0]
+                if n.owner is not None:
+                    if isinstance(n.owner.op, HostFromGpu):
+                        n = n.owner.inputs[0]
                 if n.ndim != 2:
                     return
                 ins.append(n.dimshuffle(0, 'x', 1, 'x'))

theano/sandbox/cuda/tests/test_dnn.py

@@ -582,6 +582,19 @@ class test_DnnSoftMax(test_nnet.test_SoftMax):
         if cuda.dnn.version() != (3000, 3000):
             utt.verify_grad(softmax_op, [x_val2], mode=mode_with_gpu)
 
+    def test_local_softmax_dnn_grad(self):
+        """
+        Check for optimization error when grad of summed
+        softmax is taken over tensor with fixed shape.
+        """
+        x = T.fvector('x')
+        xp = x.reshape((5, 5))
+        y = T.nnet.softmax(xp.flatten()).sum()
+        g = T.grad(y, x)
+        f = theano.function(inputs=[x], outputs=g, mode=self.mode)
+        assert(any(n for n in f.maker.fgraph.toposort() if
+                   isinstance(n.op, dnn.GpuDnnSoftmaxGrad)))
+
     def test_cudnn_softmax_grad_opt(self):
         # Verify that the SoftmaxGrad -> GpuDnnSoftmaxGrad optimization is
         # applied when cudnn is required