Fix bug on the gradients graph computation caused by an Assign node - C++ API (#12397)

* Fix bug on the gradients graph computation caused by an Assign node.
* Adding a reachable_nodes array to the gradient computation graph.
* Adding test cases when the Var has unreachable nodes.

tensorflow/cc/framework/gradients.cc

@@ -77,6 +77,10 @@ class SymbolicGradientBuilder {
   Status CallGradFunction(const Operation& op,
                           const std::vector<Output>& grad_inputs,
                           std::vector<Output>* grad_outputs);
+  
+  // Returns a list mapping whether each node in the graph is reachable
+  // from outputs_. Keyed by node id.
+  std::vector<bool> GetReachableNodes();
 
   const Scope& scope_;
   const ops::GradOpRegistry* registry_;
@@ -143,11 +147,36 @@ Status SymbolicGradientBuilder::BackpropAlongEdge(const Output& dst_grad,
   return Status::OK();
 }
 
+std::vector<bool> SymbolicGradientBuilder::GetReachableNodes() {
+  std::vector<bool> reachable_nodes(scope_.graph()->num_node_ids(), false);
+  std::deque<Node*> queue;
+  for (const Output& out : outputs_) {
+    if (!reachable_nodes[out.node()->id()]) {
+      queue.push_back(out.node());
+      reachable_nodes[out.node()->id()] = true;
+    }
+  }
+  
+  while (!queue.empty()) {
+    Node* n = queue.front();
+    queue.pop_front();
+    for (const Edge* e : n->in_edges()) {
+      if (e->IsControlEdge()) continue;
+      queue.push_back(e->src());
+      reachable_nodes[e->src()->id()] = true;
+    }
+  }
+  return reachable_nodes;
+}
+
 Status SymbolicGradientBuilder::Initialize() {
   if (outputs_.size() != grad_inputs_.size()) {
     return errors::InvalidArgument(
         "Must specify a gradient input for each output.");
   }
+  std::vector<bool> reachable_nodes = GetReachableNodes();
+  // TODO(theflofly) Check that inputs_ are reachable from
+  // outputs_ using reachable_nodes
   grad_outputs_->clear();
   grad_outputs_->resize(inputs_.size());
   // Populate `output_nodes_` from node ids in `outputs_`.
@@ -188,12 +217,15 @@ Status SymbolicGradientBuilder::Initialize() {
       if (output_nodes_.find(n->id()) == output_nodes_.end()) {
         // Internal node: continue BFS along connected outputs.
         for (const Edge* e : n->out_edges()) {
-          if (e->IsControlEdge()) continue;
-          ++num_expected_backprops;
+          // If a node is not reachable from outputs_,
+          // we don't expect it to receive a backpropagated gradient.
+          // It will not be counted in num_expected_backprops.
+          if (e->IsControlEdge() || !reachable_nodes[e->dst()->id()]) continue;
           if (visited.find(e->dst()) == visited.end()) {
             queue.push_back(e->dst());
             visited.insert(e->dst());
           }
+          ++num_expected_backprops;
         }
       } else {
         // Output node: stop BFS and update `num_expected_backprops` for

tensorflow/cc/framework/gradients_test.cc

@@ -364,6 +364,73 @@ TEST_F(GradientsTest, MultipleNodeOutputGrads) {
       test::AsTensor<int>({60, 61, 62, 63, 66, 66, 66, 67}, {4, 2}));
 }
 
+TEST_F(GradientsTest, UnreachableEdgeGradOneOutput) {
+  auto x = Variable(scope_test_, {2, 3}, DT_DOUBLE);
+  auto x_const = Const(scope_test_, {{1.0, 2.0, 3.0}, {4.0, 5.0, 6.0}});
+  auto x_assign = Assign(scope_test_, x, x_const);
+
+  auto y = Variable(scope_test_, {3, 1}, DT_DOUBLE);
+  auto y_const = Const(scope_test_, {{1.0}, {2.0}, {3.0}});
+  auto y_assign = Assign(scope_test_, y, y_const);
+
+  auto m1 = MatMul(scope_test_, x, y);
+
+  auto z = Variable(scope_test_, {1, 3}, DT_DOUBLE);
+  auto z_const = Const(scope_test_, {{9.0, 10.0, 11.0}});
+  auto z_assign = Assign(scope_test_, z, z_const);
+
+  auto m2 = MatMul(scope_test_, y, z);
+
+  auto dm1 = Const(scope_test_, {{0.5}, {0.5}});
+
+  std::vector<Output> grad_outputs;
+  TF_ASSERT_OK(
+      AddSymbolicGradients(scope_test_, {m1}, {y}, {dm1}, &grad_outputs));
+
+  std::vector<Tensor> outputs;
+  test::GetTensors(scope_test_, {x_assign, y_assign, z_assign},
+                   {grad_outputs[0]}, &outputs);
+  // dz/dy = xT * dm1
+  test::ExpectTensorNear<double>(
+      outputs[0], test::AsTensor<double>({2.5, 3.5, 4.5}, {3, 1}), 1e-5);
+}
+
+TEST_F(GradientsTest, UnreachableEdgeGradTwoOutputs) {
+  auto x = Variable(scope_test_, {2, 3}, DT_DOUBLE);
+  auto x_const = Const(scope_test_, {{1.0, 2.0, 3.0}, {4.0, 5.0, 6.0}});
+  auto x_assign = Assign(scope_test_, x, x_const);
+
+  auto y = Variable(scope_test_, {3, 1}, DT_DOUBLE);
+  auto y_const = Const(scope_test_, {{1.0}, {2.0}, {3.0}});
+  auto y_assign = Assign(scope_test_, y, y_const);
+
+  auto m1 = MatMul(scope_test_, x, y);
+
+  auto z = Variable(scope_test_, {1, 3}, DT_DOUBLE);
+  auto z_const = Const(scope_test_, {{9.0, 10.0, 11.0}});
+  auto z_assign = Assign(scope_test_, z, z_const);
+
+  auto m2 = MatMul(scope_test_, y, z);
+
+  auto dm1 = Const(scope_test_, {{0.5}, {0.5}});
+  auto dm2 =
+      Const(scope_test_, {{0.5, 0.5, 0.5}, {0.6, 0.7, 0.8}, {0.6, 0.7, 0.9}});
+
+  std::vector<Output> grad_outputs;
+  TF_ASSERT_OK(AddSymbolicGradients(scope_test_, {m1, m2}, {y}, {dm1, dm2},
+                                    &grad_outputs));
+
+  std::vector<Tensor> outputs;
+  test::GetTensors(scope_test_, {x_assign, y_assign, z_assign},
+                   {grad_outputs[0]}, &outputs);
+
+  // the gradients from m1 and m2 will be summed to compute the gradient
+  // w.r.t y
+  // dz/dy = xT * dm1 + dm2 * zT
+  test::ExpectTensorNear<double>(
+      outputs[0], test::AsTensor<double>({17.5, 24.7, 26.8}, {3, 1}), 1e-5);
+}
+
 // StopGradientSingleOutputMultiEdgeTest tests combinations of valid and
 // 'NoGradient' (induced by StopGradient op) returned along multiple edges from
 // a single nodes output.

tensorflow/cc/framework/testutil.cc

@@ -36,5 +36,19 @@ void GetTensor(const Scope& scope, Output tensor, Tensor* out) {
   *out = outputs[0];
 }
 
+void GetTensors(const Scope& scope, const std::vector<Output>& assign_vars,
+                OutputList tensors, std::vector<Tensor>* out) {
+  ClientSession session(scope);
+  TF_CHECK_OK(session.Run(assign_vars, nullptr));
+  TF_CHECK_OK(session.Run(tensors, out));
+}
+
+void GetTensor(const Scope& scope, const std::vector<Output>& assign_vars,
+               Output tensor, Tensor* out) {
+  std::vector<Tensor> outputs;
+  GetTensors(scope, assign_vars, {std::move(tensor)}, &outputs);
+  *out = outputs[0];
+}
+
 }  // end namespace test
 }  // end namespace tensorflow

tensorflow/cc/framework/testutil.h

@@ -26,9 +26,21 @@ namespace test {
 void GetTensors(const Scope& scope, OutputList tensors,
                 std::vector<Tensor>* out);
 
+// Computes the outputs listed in 'tensors', returns the tensors in 'out'.
+// assign_vars are extra outputs that should be run
+// e.g. to assign values to variables.
+void GetTensors(const Scope& scope, const std::vector<Output>& assign_vars,
+                OutputList tensors, std::vector<Tensor>* out);
+
 /// Computes the output 'tensor', returning the resulting tensor in 'out'.
 void GetTensor(const Scope& scope, Output tensor, Tensor* out);
 
+// Computes the output 'tensor', returning the resulting tensor in 'out'.
+// assign_vars are extra outputs that should be run
+// e.g. to assign values to variables.
+void GetTensor(const Scope& scope, const std::vector<Output>& assign_vars,
+               Output tensor, Tensor* out);
+
 }  // namespace test
 }  // namespace tensorflow