Fast Data Loading Using Aten


I am writing my data loading pipeline using Aten. I have some observations.

  1. First I just tried to use the normal C++ data structures to do the data processing, but I guess Pybing is doing copy while performing STL conversion i.e from map to dict.
  2. Then I tried to replace my data structures with Pytorch Tensors using Aten. I saw that accesing each element of Tensor is very expensive.

Please suggest me any way in which I can use ATen if I want to interface between Python and C++ using reference. I do not want my tensor to be copied every time I want to access it through python call. The tensor is not of small size.