Coverage for src/deepdraw/configs/models/resunet.py: 100%

1# SPDX-FileCopyrightText: Copyright © 2023 Idiap Research Institute <contact@idiap.ch> 

2# 

3# SPDX-License-Identifier: GPL-3.0-or-later 

4 

5"""Residual U-Net for image segmentation. 

6 

7A semantic segmentation neural network which combines the strengths of residual 

8learning and U-Net is proposed for road area extraction. The network is built 

9with residual units and has similar architecture to that of U-Net. The benefits 

10of this model is two-fold: first, residual units ease training of deep 

11networks. Second, the rich skip connections within the network could facilitate 

12information propagation, allowing us to design networks with fewer parameters 

13however better performance. 

14 

15Reference: [ZHANG-2017]_ 

16""" 

17 

18from torch.optim.lr_scheduler import MultiStepLR 

19 

20from deepdraw.engine.adabound import AdaBound 

21from deepdraw.models.losses import SoftJaccardBCELogitsLoss 

22from deepdraw.models.resunet import resunet50 

23 

24# config 

25lr = 0.001 

26betas = (0.9, 0.999) 

27eps = 1e-08 

28weight_decay = 0 

29final_lr = 0.1 

30gamma = 1e-3 

31eps = 1e-8 

32amsbound = False 

33 

34scheduler_milestones = [900] 

35scheduler_gamma = 0.1 

36 

37model = resunet50() 

38 

39# optimizer 

40optimizer = AdaBound( 

41 model.parameters(), 

42 lr=lr, 

43 betas=betas, 

44 final_lr=final_lr, 

45 gamma=gamma, 

46 eps=eps, 

47 weight_decay=weight_decay, 

48 amsbound=amsbound, 

49) 

50 

51# criterion 

52criterion = SoftJaccardBCELogitsLoss(alpha=0.7) 

53 

54# scheduler 

55scheduler = MultiStepLR( 

56 optimizer, milestones=scheduler_milestones, gamma=scheduler_gamma 

57)