def __init__(
self,
config
):
super(ResNetUNet, self).__init__()
self.n_classes = config.NUM_CLASSES
self.padding = 1
self.up_mode = 'upconv'
assert self.up_mode in ('upconv', 'upsample')
self.encode = ResNet101v2()
prev_channels = 2048
self.up_path = nn.ModuleList()
for i in range(3):
self.up_path.append(
UNetUpBlock(prev_channels, prev_channels // 2, self.up_mode, self.padding)
)
prev_channels //= 2
self.cls_conv_block1 = Block(prev_channels, 32)
self.cls_conv_block2 = Block(32, 16)
self.last = nn.Conv2d(16, self.n_classes, kernel_size=1)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def __init__(self,args
):
super(ResNetUNet, self).__init__()
# 网络参数要跟deeplabv3p一样的参数，是同一个config
self.n_classes = 8
self.padding = 1
self.up_mode = 'upconv'
assert self.up_mode in ('upconv', 'upsample')
# encode改成resnet101v2
self.encode = ResNet101v2()
# 上一层的给出的就是2048
prev_channels = 2048
self.up_path = nn.ModuleList()
self.batch_norm = True
for i in range(3):
self.up_path.append(
UNetUpBlock(prev_channels, prev_channels // 2, self.up_mode, self.padding, self.batch_norm)
)
prev_channels //= 2
self.cls_conv_block1 = Block(prev_channels, 32)
self.cls_conv_block2 = Block(32, 16)
self.last = nn.Conv2d(16, self.n_classes, kernel_size=1)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def __init__(self, config):
super(ResNetUNet, self).__init__()
self.n_classes = config.NUM_CLASSES
self.encode = ResNet101v2()
prev_channels = 2048
self.up_path = nn.ModuleList()
for i in range(3):
self.up_path.append(UNetUpBlock(prev_channels, prev_channels // 2))
prev_channels //= 2
self.cls_conv_block1 = Block(prev_channels, 32)
self.cls_conv_block2 = Block(32, 16)
self.last = nn.Conv2d(16, self.n_classes, kernel_size=1)
self.init_weight()
def __init__(self):
super(ResNet101v2, self).__init__()
self.conv1 = Block(3, 64, 7, 3, 2)
self.pool1 = nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True)
self.conv2_1 = DownBottleneck(64, 256, stride=1)
self.conv2_2 = Bottleneck(256, 256)
self.conv2_3 = Bottleneck(256, 256)
self.layer3 = Layer(256, [512] * 2, "resnet")
self.layer4 = Layer(512, [1024] * 23, "resnet")
self.layer5 = Layer(1024, [2048] * 3, "resnet")
def __init__(
self,
config
):
super(ResNetUNet, self).__init__()
#网络参数要跟deeplabv3p一样的参数，是同一个config
self.n_classes = config.NUM_CLASSES
self.padding = 1
self.up_mode = 'upconv'
assert self.up_mode in ('upconv', 'upsample')
#encode改成resnet101v2
self.encode = ResNet101v2()
#上一层的给出的就是2048
prev_channels = 2048
# 定义self.up_path为nn.Modulelist()，存放三个UnetUpBlock上采样模块，输出channel为输入channel的一半，循环三次
# [UNetUpBlock(2048, 1024),UNetUpBlock(1024, 512),UNetUpBlock(512, 256)]
self.up_path = nn.ModuleList()
for i in range(3):
self.up_path.append(
UNetUpBlock(prev_channels, prev_channels // 2, self.up_mode, self.padding)
)
prev_channels //= 2
# block: input-->conv-->bn-->relu -->out （默认3*3的conv标准模式）
# 定义self.conv_block1,conv_block2，使用3*3常规conv-->bn-->relu -->out模式,下降维度到32维-->16维
self.cls_conv_block1 = Block(prev_channels, 32)
self.cls_conv_block2 = Block(32, 16)
# 定义self.last 1*1conv下降维度到self.n_classes
self.last = nn.Conv2d(16, self.n_classes, kernel_size=1)
# 每层都进行kaiming初始化
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)