// 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);

  // Get theta shape and store in theta_dims.
  int num_dims = TF_NumDims(theta_tensor);

  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);
  TF_DeleteStatus(status);
  return th;
}

    TFE_DeleteTensorHandle(hcopy);
    if (TF_GetCode(status.get()) != TF_OK) {
      ADD_FAILURE() << tag;
      continue;
    }
    EXPECT_EQ(TF_TensorByteSize(t), TF_TensorByteSize(tcopy)) << tag;
    EXPECT_EQ(
        0, memcmp(TF_TensorData(t), TF_TensorData(tcopy), TF_TensorByteSize(t)))
        << tag;
    TF_DeleteTensor(tcopy);
  }

  TF_DeleteDeviceList(devices);
  TF_DeleteTensor(t);

  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);

  EXPECT_NE(b, nullptr);

  EXPECT_EQ(6, TF_TensorElementCount(a));
  EXPECT_EQ(1, TF_TensorElementCount(b));
  EXPECT_EQ(6 * TF_DataTypeSize(TF_UINT64), TF_TensorByteSize(a));
  EXPECT_EQ(TF_DataTypeSize(TF_UINT64), TF_TensorByteSize(b));

  int64_t new_dims[] = {3, 2};
  TF_Status* status = TF_NewStatus();
  TF_TensorBitcastFrom(a, TF_UINT64, b, new_dims, 2, status);

  TF_Status* status = TF_NewStatus();
  TF_Tensor* t = TFE_AllocateHostTensor(ctx, datatype, dims, num_dims, status);
  memcpy(TF_TensorData(t), data, TF_TensorByteSize(t));
  TFE_TensorHandle* th = TFE_NewTensorHandleFromTensor(ctx, t, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TF_DeleteTensor(t);
  TF_DeleteStatus(status);
  return th;
}

  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);
  }

  TF_DeleteTensor(result_tensor);

  TFE_DeleteExecutor(old_executor);
  TFE_DeleteContext(ctx);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  float product[4] = {0};
  EXPECT_EQ(sizeof(product), TF_TensorByteSize(t));
  memcpy(&product[0], TF_TensorData(t), TF_TensorByteSize(t));
  TF_DeleteTensor(t);
  EXPECT_EQ(7, product[0]);
  EXPECT_EQ(10, product[1]);
  EXPECT_EQ(15, product[2]);
  EXPECT_EQ(22, product[3]);
  TF_DeleteStatus(status);
}

  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);
  }
  delete[] dnumerical;
  TF_DeleteTensor(numerical_tensor);
}

  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])