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

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

  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);
  }
  delete[] dnumerical;

    TF_DeleteTensor(grad_tensor);
    dtheta_approx[i] = grad_data[0];
  }

  // Populate *numerical_grad with the data from dtheta_approx.
  TF_RETURN_IF_ERROR(TestTensorHandleWithDims<float, TF_FLOAT>(
      ctx, dtheta_approx.data(), theta_dims.data(), num_dims, numerical_grad));
  TF_DeleteTensor(theta_tensor);
  return absl::OkStatus();
}

}  // namespace gradients

namespace tensorflow {
namespace gradients {

/* Returns numerical grad inside `dtheta_approx` given `forward` model and
 * parameter specified by `input_index`.
 *
 * I.e. if y = <output of the forward model> and w = inputs[input_index],
 * this will calculate dy/dw numerically.
 *
 * `use_function` indicates whether to use graph mode(true) or eager(false).
 *
 * `numerical_grad` is the pointer to the AbstractTensorHandle* which will

   *
   * <p>This is guaranteed to return zero if the dataset contains a single pair of finite values. It
   * is not guaranteed to return zero when the dataset consists of the same pair of values multiple
   * times, due to numerical errors.
   *
   * <h3>Non-finite values</h3>
   *
   * <p>If the dataset contains any non-finite values ({@link Double#POSITIVE_INFINITY}, {@link

   *
   * <p>This is guaranteed to return zero if the dataset contains only exactly one finite value. It
   * is not guaranteed to return zero when the dataset consists of the same value multiple times,
   * due to numerical errors. However, it is guaranteed never to return a negative result.
   *
   * <h3>Non-finite values</h3>
   *
   * <p>If the dataset contains any non-finite values ({@link Double#POSITIVE_INFINITY}, {@link

   *
   * <p>This is guaranteed to return zero if the dataset contains a single pair of finite values. It
   * is not guaranteed to return zero when the dataset consists of the same pair of values multiple
   * times, due to numerical errors.
   *
   * <h3>Non-finite values</h3>
   *
   * <p>If the dataset contains any non-finite values ({@link Double#POSITIVE_INFINITY}, {@link

   *
   * <p>This is guaranteed to return zero if the dataset contains only exactly one finite value. It
   * is not guaranteed to return zero when the dataset consists of the same value multiple times,
   * due to numerical errors. However, it is guaranteed never to return a negative result.
   *
   * <h3>Non-finite values</h3>
   *
   * <p>If the dataset contains any non-finite values ({@link Double#POSITIVE_INFINITY}, {@link

   * itself. In all other cases, the inverse is a transformation such that applying both the
   * original transformation and its inverse to a value gives you the original value give-or-take
   * numerical errors. Calling this method multiple times on the same instance will always return
   * the same instance. Calling this method on the result of calling this method on an instance will
   * always return that original instance.

          BuildImmediateExecutionContext(std::get<1>(GetParam()), &ctx_raw);
      ASSERT_EQ(errors::OK, status_.code()) << status_.message();
      immediate_execution_ctx_.reset(ctx_raw);
    }

    // Computing numerical gradients with TensorFloat-32 is numerically
    // unstable. Some forward pass tests also fail with TensorFloat-32 due to
    // low tolerances
    enable_tensor_float_32_execution(false);
  }