1#!/usr/bin/env python 

2# -*- coding: utf-8 -*- 

3 

4from collections import OrderedDict 

5 

6import torch 

7import torch.nn 

8 

9from .backbones.vgg import vgg16_bn_for_segmentation 

10from .make_layers import UpsampleCropBlock, conv_with_kaiming_uniform 

11 

12 

13class ConcatFuseBlock(torch.nn.Module): 

14 """ 

15 Takes in four feature maps with 16 channels each, concatenates them 

16 and applies a 1x1 convolution with 1 output channel. 

17 """ 

18 

19 def __init__(self): 

20 super().__init__() 

21 self.conv = torch.nn.Sequential( 

22 conv_with_kaiming_uniform(4 * 16, 1, 1, 1, 0), 

23 torch.nn.BatchNorm2d(1), 

24 ) 

25 

26 def forward(self, x1, x2, x3, x4): 

27 x_cat = torch.cat([x1, x2, x3, x4], dim=1) 

28 x = self.conv(x_cat) 

29 return x 

30 

31 

32class DRIUBN(torch.nn.Module): 

33 """ 

34 DRIU with Batch-Normalization head module 

35 

36 Based on paper by [MANINIS-2016]_. 

37 

38 Parameters 

39 ---------- 

40 in_channels_list : list 

41 number of channels for each feature map that is returned from backbone 

42 """ 

43 

44 def __init__(self, in_channels_list=None): 

45 super(DRIUBN, self).__init__() 

46 ( 

47 in_conv_1_2_16, 

48 in_upsample2, 

49 in_upsample_4, 

50 in_upsample_8, 

51 ) = in_channels_list 

52 

53 self.conv1_2_16 = torch.nn.Conv2d(in_conv_1_2_16, 16, 3, 1, 1) 

54 # Upsample layers 

55 self.upsample2 = UpsampleCropBlock(in_upsample2, 16, 4, 2, 0) 

56 self.upsample4 = UpsampleCropBlock(in_upsample_4, 16, 8, 4, 0) 

57 self.upsample8 = UpsampleCropBlock(in_upsample_8, 16, 16, 8, 0) 

58 

59 # Concat and Fuse 

60 self.concatfuse = ConcatFuseBlock() 

61 

62 def forward(self, x): 

63 """ 

64 Parameters 

65 ---------- 

66 x : list 

67 list of tensors as returned from the backbone network. 

68 First element: height and width of input image. 

69 Remaining elements: feature maps for each feature level. 

70 

71 Returns 

72 ------- 

73 :py:class:`torch.Tensor` 

74 """ 

75 hw = x[0] 

76 conv1_2_16 = self.conv1_2_16(x[1]) # conv1_2_16 

77 upsample2 = self.upsample2(x[2], hw) # side-multi2-up 

78 upsample4 = self.upsample4(x[3], hw) # side-multi3-up 

79 upsample8 = self.upsample8(x[4], hw) # side-multi4-up 

80 out = self.concatfuse(conv1_2_16, upsample2, upsample4, upsample8) 

81 return out 

82 

83 

84def driu_bn(pretrained_backbone=True, progress=True): 

85 """Builds DRIU with batch-normalization by adding backbone and head together 

86 

87 Parameters 

88 ---------- 

89 

90 pretrained_backbone : :py:class:`bool`, Optional 

91 If set to ``True``, then loads a pre-trained version of the backbone 

92 (not the head) for the DRIU network using VGG-16 trained for ImageNet 

93 classification. 

94 

95 progress : :py:class:`bool`, Optional 

96 If set to ``True``, and you decided to use a ``pretrained_backbone``, 

97 then, shows a progress bar of the backbone model downloading if 

98 download is necesssary. 

99 

100 

101 Returns 

102 ------- 

103 

104 module : :py:class:`torch.nn.Module` 

105 Network model for DRIU (vessel segmentation) using batch normalization 

106 

107 """ 

108 

109 backbone = vgg16_bn_for_segmentation( 

110 pretrained=False, return_features=[5, 12, 19, 29] 

111 ) 

112 head = DRIUBN([64, 128, 256, 512]) 

113 

114 order = [("backbone", backbone), ("head", head)] 

115 if pretrained_backbone: 

116 from .normalizer import TorchVisionNormalizer 

117 

118 order = [("normalizer", TorchVisionNormalizer())] + order 

119 

120 model = torch.nn.Sequential(OrderedDict(order)) 

121 model.name = "driu-bn" 

122 return model