About the training of Camvid.

Reagan1311 / DABNet

Depth-wise Asymmetric Bottleneck for Real-time Semantic Segmentation (BMVC2019)

https://github.com/Reagan1311/DABNet

MIT License

140 stars 36 forks source link

About the training of Camvid. #13

Closed sunpeng1996 closed 5 years ago

sunpeng1996 commented 5 years ago

Hi, thanks for your excellent work. I have a question, the Camvid dataset is small (i,e. the images are few), do you share your training tricks or tips for Camvid?

And I see you that your max-epoch is 1000 on Camvid, don't you? Wow, it's too many epochs. Thanks!!!

Reagan1311 commented 5 years ago

Adam is more stable in the low-resolution image like CamVid dataset and class weights are employed in the loss function just as ENet. Actually, it does not need to set the epoch to 1000 on camvid, the model can saturate about 500~800 epochs, I just set it the same as cityscapes, you can try different epoch.