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;

  // once) and produces the gradient of the target tensors with respect to the
  // source tensors. The output gradients are used if not empty and not
  // null. The result is populated with one tensor per target element.
  // When running backward functions, builds zeros-like tensors for
  // incoming grads which are nullptrs, unless `build_default_zeros_grads`
  // is set to false.
  Status ComputeGradient(

    *   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]))) {

class TapeVSpace
    : public eager::VSpace<AbstractTensorHandle, GradientFunction, TapeTensor> {
 public:
  explicit TapeVSpace(AbstractContext* ctx) : ctx_(ctx) {}
  ~TapeVSpace() override {}

  // Returns the number of elements in the gradient tensor.
  int64_t NumElements(AbstractTensorHandle* tensor) const override;

  // Consumes references to the tensors in the gradient_tensors list and returns
  // a tensor with the result.

  return th;
}

TFE_TensorHandle* TestMatrixTensorHandle100x100(TFE_Context* ctx) {
  constexpr int64_t dims[] = {100, 100};
  constexpr int num_elements = dims[0] * dims[1];
  float data[num_elements];
  for (int i = 0; i < num_elements; ++i) {
    data[i] = 1.0f;
  }
  TF_Status* status = TF_NewStatus();
  TF_Tensor* t = TFE_AllocateHostTensor(ctx, TF_FLOAT, &dims[0],

  if (h == nullptr) {
    status->status = tensorflow::errors::InvalidArgument("Invalid handle");
    return -1;
  }

  int64_t num_elements = -1;
  status->status = tensorflow::unwrap(h)->NumElements(&num_elements);
  return num_elements;
}

int64_t TFE_TensorHandleDim(TFE_TensorHandle* h, int dim_index,
                            TF_Status* status) {
  if (h == nullptr) {

  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) {

  // Convert from AbstractTensor to TF_Tensor.
  TF_Tensor* theta_tensor;
  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));