1#!/usr/bin/env python 

2# -*- coding: utf-8 -*- 

3 

4"""Residual U-Net for image segmentation 

5 

6A semantic segmentation neural network which combines the strengths of residual 

7learning and U-Net is proposed for road area extraction. The network is built 

8with residual units and has similar architecture to that of U-Net. The benefits 

9of this model is two-fold: first, residual units ease training of deep 

10networks. Second, the rich skip connections within the network could facilitate 

11information propagation, allowing us to design networks with fewer parameters 

12however better performance. 

13 

14Reference: [ZHANG-2017]_ 

15""" 

16 

17from torch.optim.lr_scheduler import MultiStepLR 

18 

19from bob.ip.binseg.engine.adabound import AdaBound 

20from bob.ip.binseg.models.losses import SoftJaccardBCELogitsLoss 

21from bob.ip.binseg.models.resunet import resunet50 

22 

23# config 

24lr = 0.001 

25betas = (0.9, 0.999) 

26eps = 1e-08 

27weight_decay = 0 

28final_lr = 0.1 

29gamma = 1e-3 

30eps = 1e-8 

31amsbound = False 

32 

33scheduler_milestones = [900] 

34scheduler_gamma = 0.1 

35 

36model = resunet50() 

37 

38# optimizer 

39optimizer = AdaBound( 

40 model.parameters(), 

41 lr=lr, 

42 betas=betas, 

43 final_lr=final_lr, 

44 gamma=gamma, 

45 eps=eps, 

46 weight_decay=weight_decay, 

47 amsbound=amsbound, 

48) 

49 

50# criterion 

51criterion = SoftJaccardBCELogitsLoss(alpha=0.7) 

52 

53# scheduler 

54scheduler = MultiStepLR( 

55 optimizer, milestones=scheduler_milestones, gamma=scheduler_gamma 

56)