search

irfanICMLL / structure_knowledge_distillation

The official code for the paper 'Structured Knowledge Distillation for Semantic Segmentation'. (CVPR 2019 ORAL) and extension to other tasks.
BSD 2-Clause "Simplified" License
708 stars 103 forks source link

retrain teacher net #42

Closed wl082013 closed 4 years ago

wl082013 commented 4 years ago

Hi, could you provide the information how you trained the teacher network (PSPNet_combine)? I want to retrain PSPNet, so really appreciate if you can provide the link or github where you use the teacher model. Thanks a lot

irfanICMLL commented 4 years ago

I train the code with OCnet two years ago. https://github.com/openseg-group/OCNet.pytorch

I suggest you use the mmsegmentation, which is a good open tool for semantic segmentation

https://github.com/open-mmlab/mmsegmentation/

wl082013 commented 4 years ago

okay thanks a lot I have been stuck by the synchronous BN as used in your tpami paper (later version). As my gpu environment is under pytorch 1.2, cudo 10.0, I have met some difficulty handling this issue.

Could you provide some help replacing the old synchronous BN with the lastest one compiled in pytorch 1.2 or higher ? I have tried to replace it with some, but they all fail when training several epochs (nan loss). I guess it may be caused by the learning rate, but even when I changed learning rate, it still did not work sometimes. PSPNet really caused some problems for me. thanks a lot.