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

ESPNet training #32

Closed sungsooo closed 4 years ago

sungsooo commented 4 years ago

Hi, I want to reproduce the ESPNet KD performance. So, I just want to know how did you train the ESPNet?

Did you use the same res101-pspnet as the teacher net?

Which features were used for the pi, pa, ho loss? Thank you

irfanICMLL commented 4 years ago

Yes. I use re101 as the teacher net. For ESP-C it is the same as present 18. I use the feature at 8x resolution. For ESP, I add at the 8x feature in the middle. Therefore, the improvements are not as good as ECP-C.

betogulliver commented 4 years ago

I'd also like to reproduce the results using ESPNet, any ideas on how to proceed? I guess I need to adapt "networks/pspnet_combine.py" for ESPNet. Am I'm on the right track? (can you share your steps?)

thanks for any/all help.