TF_RETURN_IF_ERROR(GetValue(theta, &theta_tensor));

  // Get number of elements and fill data.
  int num_elems = TF_TensorElementCount(theta_tensor);
  vector<float> theta_data(num_elems);
  memcpy(theta_data.data(), TF_TensorData(theta_tensor),
         TF_TensorByteSize(theta_tensor));

  // Initialize space for the numerical gradient.
  vector<float> dtheta_approx(num_elems);

TFE_TensorHandle* TestScalarTensorHandle(TFE_Context* ctx, float value) {
  float data[] = {value};
  TF_Status* status = TF_NewStatus();
  TF_Tensor* t = TFE_AllocateHostTensor(ctx, TF_FLOAT, nullptr, 0, status);
  memcpy(TF_TensorData(t), &data[0], TF_TensorByteSize(t));
  TFE_TensorHandle* th = TFE_NewTensorHandleFromTensor(ctx, t, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TF_DeleteTensor(t);

  *len = debug_str.size();
  char* ret = static_cast<char*>(malloc(*len + 1));
  memcpy(ret, debug_str.c_str(), *len + 1);
  return ret;
}

char* TF_FunctionDebugString(TF_Function* func, size_t* len) {
  const auto& debug_str = DebugString(func->record->fdef());
  *len = debug_str.size();
  char* ret = static_cast<char*>(malloc(*len + 1));
  memcpy(ret, debug_str.c_str(), *len + 1);
  return ret;
}

  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());
  TF_Tensor* result_tensor = TFE_TensorHandleResolve(result_t, status.get());

  // Copy Tensor data into an array.
  float result_data[4] = {0};
  memcpy(&result_data[0], TF_TensorData(result_tensor),
         TF_TensorByteSize(result_tensor));

  int data_len = 4;  // length of result_data
  for (int i = 0; i < data_len; i++) {
    EXPECT_EQ(result_data[i], 0);
  }

  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_DeleteTensorHandle(retvals[0]);
  float sum = 0;
  EXPECT_EQ(sizeof(sum), TF_TensorByteSize(t));
  memcpy(&sum, TF_TensorData(t), TF_TensorByteSize(t));
  TF_DeleteTensor(t);
  EXPECT_EQ(sum, 6.0);

  TFE_DeleteTensorHandle(h0);
  TFE_DeleteTensorHandle(h1);
  TFE_DeleteTensorHandle(h2);

  ASSERT_EQ(errors::OK, s.code()) << s.message();
  auto num_elem_numerical = TF_TensorElementCount(numerical_tensor);
  ASSERT_EQ(num_elem_numerical, num_grad);

  float* dnumerical = new float[num_elem_numerical]{0};
  memcpy(&dnumerical[0], TF_TensorData(numerical_tensor),
         TF_TensorByteSize(numerical_tensor));

  for (int j = 0; j < num_grad; j++) {
    ASSERT_NEAR(dnumerical[j], expected_grad[j], abs_error);
  }

  status->status = s->session->PRunSetup(input_names, output_names,
                                         target_oper_names, &new_handle);
  if (status->status.ok()) {
    char* buf = new char[new_handle.size() + 1];
    memcpy(buf, new_handle.c_str(), new_handle.size() + 1);
    *handle = buf;
  }
}

void TF_PRun(TF_DeprecatedSession* s, const char* handle,
             // Input tensors

  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());
  std::unique_ptr<float[]> actual_values(new float[expected_values.size()]);
  EXPECT_EQ(sizeof(float) * expected_values.size(), TF_TensorByteSize(t));
  memcpy(actual_values.get(), TF_TensorData(t), TF_TensorByteSize(t));
  TF_DeleteTensor(t);

  for (int i = 0; i < expected_values.size(); i++) {
    EXPECT_EQ(expected_values[i], actual_values[i])

  }

  TF_Tensor* FloatTensor2x2(const float* values) {
    const int64_t dims[2] = {2, 2};
    TF_Tensor* t = TF_AllocateTensor(TF_FLOAT, dims, 2, sizeof(float) * 4);
    memcpy(TF_TensorData(t), values, sizeof(float) * 4);
    return t;
  }

  TF_Operation* FloatConst2x2(TF_Graph* graph, TF_Status* s,
                              const float* values, const char* name) {