const auto &init_values_ty = op.getInitValues().getType();
  int n_init_values = init_values_ty.size();
  if (n_init_values != n_inputs) {
    return op.emitOpError() << "Number of inputs (" << n_inputs
                            << ") is different than number of init_values ("
                            << n_init_values << ")";
  }

  auto input_ty_0 = inputs_ty[0].cast<ShapedType>();

  auto y = Elu(scope_, x);
  Tensor x_init_value = test::AsTensor<float>(
      {-0.9f, -0.7f, -0.5f, -0.3f, -0.1f, 0.1f, 0.3f, 0.5f, 0.7f, 0.9f},
      {5, 2});
  RunTest(x, x_init_value, y, shape);
}

TEST_F(NNGradTest, SeluGrad) {
  TensorShape shape({5, 2});
  auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(shape));
  auto y = Selu(scope_, x);
  Tensor x_init_value = test::AsTensor<float>(

    EXPECT_LT(max_error, 1e-3);
  }

  void RunTest(const Output& x, const Tensor& x_init_value, const Output& y,
               const TensorShape& y_shape) {
    TF_ASSERT_OK(scope_.status());
    float max_error;
    TF_ASSERT_OK((ComputeGradientError<float, float, float>(
        scope_, x, x_init_value, y, y_shape, &max_error)));
    EXPECT_LT(max_error, 1e-3);
  }

  Scope scope_;
};

Status ComputeGradientError(const Scope& scope, const Output& x,
                            const Tensor& x_init_value, const Output& y,
                            const TensorShape& y_shape, JAC_T* max_error) {
  // Initialize 'x_data' from 'x_init_value'.
  std::vector<Tensor> x_datas(1, Tensor(x_init_value));
  // Compute gradient error.
  return ComputeGradientErrorInternal<X_T, Y_T, JAC_T>(

template <typename X_T, typename Y_T, typename JAC_T>
Status ComputeGradientError(const Scope& scope, const Output& x,
                            const Tensor& x_init_value, const Output& y,
                            const TensorShape& y_shape, JAC_T* max_error);

}  // namespace tensorflow

    # updated to these new ssa values before it enters the scope of the loop.
    out_types = []
    init_values = []
    for symbol, ssa_value in zip(loop_carried, ret_ssa_values):
      init, ty = self.symbol_table.lookup(symbol)
      self.symbol_table.insert_symbol(symbol, ssa_value, ty)
      out_types.append(str(ty))
      init_values.append((init, ty))

    # Create a new scope in case the local variables are leaked.

// set of values. Returns the final set of values via `final_values`. The
// initial set of values is passed in via `init_values`.
//
// This effectively does:
//
// ```c++
// SmallVector<Values, 4> old_values = init_values;
// SmallVector<Values, 4> new_values;
// for (int i = 0; i < num_iterations; ++i) {
//   body_fn(old_values, &new_values, ...);
//   old_values = new_values;

  }];

  let arguments = (ins
    Arg<Variadic<TF_Tensor>, [{the input tensor(s)}]>:$inputs,
    Arg<Variadic<TF_Tensor>, [{scalar initial value(s) for the reduction}]>:$init_values,

    I64ArrayAttr:$dimensions_to_reduce,
    SymbolRefAttr:$reducer
  );

  let results = (outs
    Variadic<TF_Tensor>:$outputs
  );