std::vector<TF_Output> tf_outputs(inputs.size());
    for (int i = 0; i < inputs.size(); i++) {
      GraphTensor* t = dyn_cast<GraphTensor>(inputs[i]);
      if (!t) {
        return tensorflow::errors::InvalidArgument(
            "Unable to cast input to GraphTensor");
      }
      tf_outputs[i] = t->output_;
    }
    TF_AddInputList(op_.get(), tf_outputs.data(), tf_outputs.size());

  cond_graph->parent_inputs = inputs;
  body_graph->parent = g;
  body_graph->parent_inputs = inputs;

  TF_Output* cond_inputs = new TF_Output[ninputs];
  TF_Output cond_output = {nullptr, -1};
  TF_Output* body_inputs = new TF_Output[ninputs];
  TF_Output* body_outputs = new TF_Output[ninputs];
  for (int i = 0; i < ninputs; ++i) body_outputs[i] = {nullptr, -1};
  const char* name = nullptr;

  TF_Graph* const cond_graph;
  const TF_Output* const cond_inputs;
  TF_Output cond_output;

  // The loop body graph. The inputs are the current values of the loop
  // variables. The outputs are the updated values of the loop variables.
  TF_Graph* const body_graph;
  const TF_Output* const body_inputs;
  TF_Output* const body_outputs;

    TF_RETURN_IF_ERROR(
        RunAndMaybeSum(ctx, forward, theta_inputs, f_outputs, use_function));
    AbstractTensorHandlePtr fPlus(f_outputs[0]);

    // Get f(theta - eps):
    theta_inputs[input_index] = thetaMinus.get();
    TF_RETURN_IF_ERROR(
        RunAndMaybeSum(ctx, forward, theta_inputs, f_outputs, use_function));
    AbstractTensorHandlePtr fMinus(f_outputs[0]);

                                    TF_Graph* graph, TF_Operation* ctrl_op,
                                    TF_Status* s, const char* name = "add");

TF_Operation* Add(TF_Output l, TF_Output r, TF_Graph* graph, TF_Status* s,
                  const char* name = "add");

TF_Operation* Min(TF_Operation* l, TF_Operation* r, TF_Graph* graph,
                  TF_Status* s, const char* name = "min");

  TF_SetRequestedDevice(graph, op, device);
}

void UpdateEdge(TF_Graph* graph, TF_Output new_src, TF_Input dst,
                TF_Status* status) {
  TF_UpdateEdge(graph, new_src, dst, status);
}

void ExtendSession(TF_Session* session, TF_Status* status) {
  TF_ExtendSession(session, status);
}

std::string GetHandleShapeAndType(TF_Graph* graph, TF_Output output) {
  Node* node = &output.oper->node;

  TF_SessionOptions* opts = TF_NewSessionOptions();
  TF_Session* sess = TF_NewSession(graph, opts, s);
  TF_DeleteSessionOptions(opts);

  TF_Output feeds[] = {TF_Output{a, 0}, TF_Output{b, 0}};
  TF_Output fetches[] = {TF_Output{plus2, 0}, TF_Output{plusB, 0}};

  const char* handle = nullptr;
  TF_SessionPRunSetup(sess, feeds, TF_ARRAYSIZE(feeds), fetches,

  TF_Operation* op;
  BinaryOpHelper("Mul", l, r, graph, s, name, &op, "", true);
  return op;
}

TF_Operation* Add(TF_Output l, TF_Output r, TF_Graph* graph, TF_Status* s,
                  const char* name) {
  TF_OperationDescription* desc = TF_NewOperation(graph, "AddN", name);
  TF_Output inputs[2] = {l, r};
  TF_AddInputList(desc, inputs, 2);
  return TF_FinishOperation(desc, s);
}

std::string GetHandleShapeAndType(TF_Graph* graph, TF_Output output);

// Sets `output` based on `proto`, which should be a serialized
// CppShapeInferenceResult::HandleData proto. `output` should be a resource
// or variant tensor.
// NOTE(skyewm): `proto` is passed a void*/size_t pair instead of a std::string
// because I couldn't get SWIG to work otherwise.
void SetHandleShapeAndType(TF_Graph* graph, TF_Output output, const void* proto,

namespace tensorflow {
namespace gradients {
namespace {

class ReluGradientFunction : public GradientFunction {
 public:
  explicit ReluGradientFunction(vector<AbstractTensorHandle*> f_outputs)
      : forward_outputs_(f_outputs) {
    for (auto output : forward_outputs_) {
      if (output) {
        output->Ref();
      }
    }
  }

  Status Compute(AbstractContext* ctx,