std::vector<int64_t> transpose_order;
      transpose_order.resize(num_spatial_dims + 2);
      // the output always in b, 0, 1, f format
      transpose_order[dnums.getOutputBatchDimension()] = 0;
      transpose_order[dnums.getOutputFeatureDimension()] = num_spatial_dims + 1;
      for (size_t i = 0; i < num_spatial_dims; ++i) {
        transpose_order[dnums.getOutputSpatialDimensions().data()[i]] = i + 1;