def __init__(self, pretrained=True, **kwargs):
super().__init__()
self.encoder = resnet50(pretrained=pretrained)
self.conv1 = nn.Sequential(self.encoder.conv1, self.encoder.bn1,
self.encoder.relu)
self.encoder1 = self.encoder.layer1 # 64
self.encoder2 = self.encoder.layer2 # 128
self.encoder3 = self.encoder.layer3 # 256
self.encoder4 = self.encoder.layer4 # 512
self.center_conv = nn.Sequential(
L.ConvBn2d(2048, 512, kernel_size=3, padding=1),
nn.ReLU(inplace=True))
self.FPA = L.FeaturePyramidAttention_v2(512, 64)
self.decoder5 = L.GlobalAttentionUpsample(1024, 64)
self.decoder4 = L.GlobalAttentionUpsample(512, 64)
self.decoder3 = L.GlobalAttentionUpsample(256, 64)
self.upsample = nn.Upsample(scale_factor=4,
mode='bilinear',
align_corners=True)
self.logit = nn.Conv2d(64, 1, kernel_size=1, padding=0)
def __init__(self, pretrained=True, **kwargs):
super().__init__()
self.encoder = se_resnet50(pretrained=pretrained)
self.conv1 = self.encoder.layer0
self.encoder1 = self.encoder.layer1 # 256
self.encoder2 = self.encoder.layer2 # 512
self.encoder3 = self.encoder.layer3 # 1024
self.encoder4 = self.encoder.layer4 # 2048
self.ema = nn.Sequential(
L.ConvBn2d(2048, 512, kernel_size=3, padding=1, act=True),
L.EMAModule(512, 64, lbda=1, alpha=0.1, T=3),
L.ConvBn2d(512, 64, kernel_size=3, padding=1, act=True))
self.decoder5 = L.GlobalAttentionUpsample(1024, 64)
self.decoder4 = L.GlobalAttentionUpsample(512, 64)
self.decoder3 = L.GlobalAttentionUpsample(256, 64)
self.upsample = nn.Upsample(scale_factor=4,
mode='bilinear',
align_corners=True)
self.logit = nn.Conv2d(64, 1, kernel_size=1, padding=0)