Fix the gradient for functions when its output has no dependencies on

its inputs.
Change: 120128592

tensorflow/core/graph/gradients.cc

@@ -156,9 +156,8 @@ class SymbolicGradientBuilder {
   // add dy as an input of the gradient function.
   std::deque<Node*> ready_;
 
-  // The set of nodes at which to stop backprop.
+  // The set of node ids at which to stop backprop.
-  // Maps from node.id -> index of 'x_node_outputs_'
+  std::unordered_set<int> stop_nodes_;
-  std::unordered_map<int, int> stop_nodes_;
 
   // Initialize pending_ and ready_.
   void InitBackprop();
@@ -190,7 +189,7 @@ SymbolicGradientBuilder::SymbolicGradientBuilder(
   x_grad_node_outputs_->resize(x_node_outputs_.size());
   stop_nodes_.reserve(x_node_outputs_.size());
   for (int i = 0; i < x_node_outputs_.size(); ++i) {
-    stop_nodes_.insert(std::make_pair(x_node_outputs_[i].node->id(), i));
+    stop_nodes_.insert(x_node_outputs_[i].node->id());
   }
 }
 
@@ -319,11 +318,9 @@ Status SymbolicGradientBuilder::Compute() {
 
     auto iter = stop_nodes_.find(n->id());
     if (iter != stop_nodes_.end()) {
-      // Stop backprop and add gradient sum to 'x_grad_node_outputs_'.
+      // Stop backprop.
       // TODO(andydavis) Support stop nodes with more than one output.
       CHECK_EQ(1, num_y);
-      const int index = iter->second;
-      (*x_grad_node_outputs_)[index] = SumGradients(x_node_outputs_[index]);
       continue;
     }
 
@@ -362,6 +359,10 @@ Status SymbolicGradientBuilder::Compute() {
     }
   }
 
+  for (int i = 0; i < x_node_outputs_.size(); ++i) {
+    (*x_grad_node_outputs_)[i] = SumGradients(x_node_outputs_[i]);
+  }
+
   return Status::OK();
 }
 

tensorflow/python/framework/function_test.py

@@ -232,6 +232,21 @@ class FunctionTest(tf.test.TestCase):
       self.assertEquals(x.get_shape(), dx.get_shape())
       self.assertEquals(y.get_shape(), dy.get_shape())
 
+  def testZNoDepOnY(self):
+    with tf.Graph().as_default():
+      # z = Foo(x, y). z doe
+      @function.Defun(tf.float32, tf.float32)
+      def Foo(x, y):
+        return x * 2
+      x = tf.constant(1.0)
+      y = tf.constant(2.0)
+      z = Foo(x, y)
+      dx, dy = tf.gradients([z], [x, y])
+      with tf.Session() as sess:
+        dx_val, dy_val = sess.run([dx, dy])
+        self.assertEquals([2.0], dx_val)
+        self.assertEquals([0.0], dy_val)
+
   def testDefineFunctionNoArgs(self):
 
     def AConstant():