BiSeNet
BiSeNet copied to clipboard
Any suggestions for optimization?
trafficstars
any suggestions for optimization?
- Speed up the reasoning
- improve the accuracy of segmentation details thx!
This is what the researchers are doing everyday, isn't it? Maybe you can read paper of your interest and try the proposed method.
For part 2: I've added Diceloss for 2-class seg(for unbalanced labels) and multiplied the different losses(pre, aux, dice) with weights([1,0.5,1]) but get worse result. ===> now, I fix the weight equally and train again. BTW the training process is so slow...
Good job!!! I am closing this.