const int num_dims = 2;
  int64_t* dims = new int64_t[num_dims];
  int64_t num_elements = 1;
  dims[0] = batch_size;
  dims[1] = 1;
  for (int64_t i = 0; i < num_dims; ++i) {
    num_elements *= dims[i];
  }
  TF_Tensor* t = TF_AllocateTensor(TF_STRING, dims, num_dims,
                                   sizeof(TF_TString) * num_elements);
  delete[] dims;

    *   Changing `tf.data.experimental.save` to store the type specification of
        the dataset elements. This avoids the need for explicitly specifying the
        `element_spec` argument of `tf.data.experimental.load` when loading the
        previously saved dataset.
    *   Add `.element_spec` property to `tf.data.DatasetSpec` to access the
        inner spec. This can be used to extract the structure of nested

  shapes.resize(num_shapes);
  for (int i = 0; i < num_shapes; ++i) {
    if (proto_lens[i] > std::numeric_limits<int>::max()) {
      status->status = InvalidArgument(
          "length of element ", i, " in the list (", proto_lens[i],
          " bytes) is too large to be parsed by the protocol buffer library");
      return;
    }
    if (!shapes[i].ParseFromArray(protos[i], static_cast<int>(proto_lens[i]))) {

  std::vector<Tensor> all_input_tensors;
  // Update the vector with information from `input_tensors` if provided.
  if (input_tensors != nullptr) {
    // Note that we take the address of the elements in `all_input_tensors`
    // below. Allocate enough space so that no reallocation happens, which will
    // make the pointers invalid.
    all_input_tensors.reserve(num_inputs);
    for (int i = 0; i < num_inputs; ++i) {