Skip to content
This repository was archived by the owner on Apr 22, 2020. It is now read-only.

MODEL.md

Latest commit

 

History

History
14 lines (11 loc) · 685 Bytes

MODEL.md

File metadata and controls

14 lines (11 loc) · 685 Bytes

MODEL import from PYTHON

This project's purpose is 'train in python, inference in cpp'

Overall Process

  1. Train weight file in maskrcnn-benchmark
  2. Using to_jit.py script, convert model's weight file from python to pytorch jit.
    This script only takes weight except logics(eg. nms..)
  3. Using jit_to_cpp.cpp, convert pytorch jit weight to libtorch weight.
    If you are using new type of backbone(not resnet or vovnet), you have to declare and define new backbone mapper.
    It makes hash map python weight name <-> cpp weight name
  4. Saved as /models/new_pth_from_python_cpp.pth
  5. Rename and use