// does various checks while doing so. `input_nodes` will contain the same
// information as input_tensors just in a different structure to make
// following processing easier. TODO(iga): Simplify this nested structure.
Status ProcessInputs(
    const TF_Graph* fn_body, const char* fn_name, int ninputs,
    const TF_Output* inputs, std::vector<OutputTensor>* input_tensors,

  TF_Operation* neg = TF_GraphOperationByName(graph, "imported/neg");
  ASSERT_TRUE(scalar != nullptr);
  ASSERT_TRUE(feed != nullptr);
  ASSERT_TRUE(neg != nullptr);

  // Test basic structure of the imported graph.
  EXPECT_EQ(0, TF_OperationNumInputs(scalar));
  EXPECT_EQ(0, TF_OperationNumInputs(feed));
  ASSERT_EQ(1, TF_OperationNumInputs(neg));
  TF_Output neg_input = TF_OperationInput({neg, 0});

      CopyToCore<TF_ReadOnlyMemoryRegionOps>(
          info->ops[index].read_only_memory_region_ops,
          info->ops[index].read_only_memory_region_ops_size);

  // Step 2: Initialize the opaque filesystem structure
  auto filesystem = std::make_unique<TF_Filesystem>();
  TF_Status* c_status = TF_NewStatus();
  Status status = OkStatus();
  core_filesystem_ops->init(filesystem.get(), c_status);
  status = Status(c_status->status);

  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Un-pack the parallel tensor to verify that the operation was
  // successful. The resulting structure should be:
  //   second_device{first_device{1. * 3., 2. * 3.}, 3. * 3.}.
  std::array<TensorHandlePtr, 2> second_components;
  ExtractPerDeviceValues(context.get(), multiply_result.get(),

  // Get theta shape and store in theta_dims.
  int num_dims = TF_NumDims(theta_tensor);
  vector<int64_t> theta_dims(num_dims);
  GetDims(theta_tensor, theta_dims.data());

  // Initialize auxilary data structures.
  vector<float> thetaPlus_data(num_elems);
  vector<float> thetaMinus_data(num_elems);
  AbstractTensorHandle* f_outputs[1];

  // Numerical Grad Check
  for (int i = 0; i < num_elems; i++) {