Spaces:
Running
Running
File size: 6,651 Bytes
278bf2b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 | #!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""Residual Attention Network module - feature extractor in AVRA."""
import torch.nn as nn
import torch.nn.functional as F
class ResidualAttentionNet(nn.Module):
def __init__(self, z=1):
super(ResidualAttentionNet, self).__init__()
num_filters = [8, 16, 32, 64, 128]
k = 0
conv1 = nn.Sequential(
nn.Conv2d(z, num_filters[k], kernel_size=7, stride=2, padding=3, bias=False),
nn.BatchNorm2d(num_filters[k]),
nn.ReLU(inplace=True)
)
maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
resblock1 = ResidualModule(num_filters[k], num_filters[k+1])
k += 1
attention_module1 = AttentionModule(num_filters[k], num_filters[k], stage=1)
resblock2 = ResidualModule(num_filters[k], num_filters[k+1], stride=2)
k += 1
attention_module2 = AttentionModule(num_filters[k], num_filters[k], stage=2)
resblock3 = ResidualModule(num_filters[k], num_filters[k+1], stride=2)
k += 1
attention_module3 = AttentionModule(num_filters[k], num_filters[k], stage=3)
resblock4 = ResidualModule(num_filters[k], num_filters[k+1], stride=2)
k += 1
resblock5 = ResidualModule(num_filters[k], num_filters[k])
resblock6 = ResidualModule(num_filters[k], num_filters[k])
avgpoolblock = nn.Sequential(
nn.BatchNorm2d(num_filters[k]),
nn.ReLU(inplace=True),
nn.AvgPool2d(kernel_size=3, stride=1)
)
self.features = nn.Sequential(
conv1, maxpool,
resblock1, attention_module1,
resblock2, attention_module2,
resblock3, attention_module3,
resblock4, resblock5, resblock6,
avgpoolblock
)
def forward(self, x):
out = self.features(x)
out = out.view(out.size(0), -1)
return out
class ResidualModule(nn.Module):
def __init__(self, inplanes, planes, stride=1):
super(ResidualModule, self).__init__()
planes_4 = int(planes/4)
self.inplanes = inplanes
self.planes = planes
self.stride = stride
self.bn1 = nn.BatchNorm2d(inplanes)
self.relu1 = nn.LeakyReLU()
self.conv1 = nn.Conv2d(inplanes, planes_4, kernel_size=1, stride=1, bias=False)
self.bn2 = nn.BatchNorm2d(planes_4)
self.relu2 = nn.LeakyReLU()
self.conv2 = nn.Conv2d(planes_4, planes_4, kernel_size=3, stride=stride, padding=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes_4)
self.relu3 = nn.LeakyReLU()
self.conv3 = nn.Conv2d(planes_4, planes, kernel_size=1, stride=1, bias=False)
self.conv4 = nn.Conv2d(inplanes, planes, kernel_size=1, stride=stride, bias=False)
self.downsample = (self.inplanes != self.planes) or (self.stride != 1)
def forward(self, x):
residual = x
out = self.bn1(x)
out1 = self.relu1(out)
out = self.conv1(out1)
out = self.bn2(out)
out = self.relu2(out)
out = self.conv2(out)
out = self.bn3(out)
out = self.relu3(out)
out = self.conv3(out)
if self.downsample:
residual = self.conv4(out1)
out += residual
return out
class AttentionModule(nn.Module):
def __init__(self, in_planes, out_planes, stage=1):
super(AttentionModule, self).__init__()
self.stage = stage
self.res1 = ResidualModule(in_planes, out_planes)
self.trunk_branch = nn.Sequential(
ResidualModule(in_planes, out_planes),
ResidualModule(in_planes, out_planes)
)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
if self.stage < 3:
self.block1 = ResidualModule(in_planes, out_planes)
self.skip1 = ResidualModule(in_planes, out_planes)
self.block5 = ResidualModule(in_planes, out_planes)
if self.stage == 1:
self.block2 = ResidualModule(in_planes, out_planes)
self.skip2 = ResidualModule(in_planes, out_planes)
self.block4 = ResidualModule(in_planes, out_planes)
self.block3 = nn.Sequential(
ResidualModule(in_planes, out_planes),
ResidualModule(in_planes, out_planes)
)
self.block_sigmoid = nn.Sequential(
nn.BatchNorm2d(out_planes),
nn.ReLU(inplace=True),
nn.Conv2d(out_planes, out_planes, kernel_size=1, stride=1, bias=False),
nn.BatchNorm2d(out_planes),
nn.ReLU(inplace=True),
nn.Conv2d(out_planes, out_planes, kernel_size=1, stride=1, bias=False),
nn.Sigmoid()
)
self.block6 = ResidualModule(in_planes, out_planes)
def upsample(self, x):
return F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=True)
def forward(self, x):
x = self.res1(x)
trunk_branch = self.trunk_branch(x)
if self.stage < 3:
x = self.maxpool(x)
x = self.block1(x)
skip1 = self.skip1(x)
if self.stage == 1:
x = self.maxpool(x)
x = self.block2(x)
skip2 = self.skip2(x)
x = self.maxpool(x)
x = self.block3(x)
if self.stage == 1:
x = self.upsample(x)
x = x + skip2
x = self.block4(x)
if self.stage < 3:
x = self.upsample(x)
x = x + skip1
x = self.block5(x)
x = self.upsample(x)
mask = self.block_sigmoid(x)
x = (1 + mask) * trunk_branch
x = x + trunk_branch
out_last = self.block6(x)
return out_last
def conv_block(in_planes, out_planes, bigblock, convxd, norm, pooling, fs=3, stride=1, relu=nn.ReLU):
if bigblock:
block = nn.Sequential(
convxd(in_planes, out_planes, fs, 1, int(fs/2)),
relu(True),
norm(out_planes),
convxd(out_planes, out_planes, fs, 1, int(fs/2)),
relu(True),
norm(out_planes),
convxd(out_planes, out_planes, fs, stride, int(fs/2)),
relu(True),
norm(out_planes),
pooling(2, 2)
)
else:
block = nn.Sequential(
convxd(in_planes, out_planes, fs, 1, int(fs/2)),
relu(True),
norm(out_planes),
convxd(out_planes, out_planes, fs, stride, int(fs/2)),
relu(True),
norm(out_planes),
pooling(2, 2)
)
return block
|