Can sahi do slice predictions without torch dependencies?

[zhaoruibing]

I am trying to deploy a YOLOv8 model with get_sliced_prediction(), pack everything into a small container.
get_sliced_prediction() calls postprocess functions, and these postprocess functions in sahi.postprocess.combine depends on torch.

sahi is said to be 'lightweight' and in this discussion independent of torch.

We wanted sahi to work independent from torch thats why all detectors accept numpy or str as image input 👍

The question is, could sahi do slice predictions without torch dependencies?
Or this is something sahi plans to do in the future?

[HowcanoeWang]

I have the same requirements, so I manually modify the source code to change all torch based to numpy based (just for yolo detection, I also removed several unnecessary codes), you can check the source code here (based on v0.11.20):

https://github.com/UTokyo-FieldPhenomics-Lab/EasyAMS/tree/0e74099e311052943b280eb402f61caa6ad752d9/easyams/sahi_onnx

The modification can be checked at this commit:

UTokyo-FieldPhenomics-Lab/EasyAMS@0e74099

[onuralpszr]

Next release will be not gonna include torch and all of the process can be done via "numpy" but it can be use via torch or numba as well. So I make sure it is now optional as well. But please try and let us know, thank you.