search

sacmehta / ESPNet

ESPNet: Efficient Spatial Pyramid of Dilated Convolutions for Semantic Segmentation
https://sacmehta.github.io/ESPNet/
MIT License
541 stars 112 forks source link

training on 2 classes #36

Closed Jee-King closed 5 years ago

Jee-King commented 5 years ago

Hi, I'm trying to use this model to segment drivable regions and only have 2 classes. I did not get good results by changing the model.py as you said in #13. The smallest train loss I got was 0.036. So do I need to change softmax function to sigmoid? Or do I need to make other changes? Thank you very much.

sacmehta commented 5 years ago

Training loss itself is too small and you should get good results. Did you try to measure dice score or mIOU or visualize results on val/test images?

Jee-King commented 5 years ago

Thank you for your reply. It has worked.