一文搞定U2Net——图像分割（语义分割）

上文中，我们介绍了UNet，今天我们来了解一下U2Net。这个网络是 UNet的加强版。其结构如下图所示：

与UNet相比，U2Net中的每一个小立方体里面都是一个UNet。但是需要注意的是：

U型结构的最下面一层白色的块为空洞卷积。
在最后一层做通道融合的时候，是用类似ResNet的思想，将特征变量进行相加。其他时候仍然用的torch.cat()操作。
与UNet不同，特征矩阵的通道数没有增加。

所以首先我们构建小立方体中的结构，我们称之为Unet_Blockx。和UNet中的一样其包括了卷积层块、下采样、上采样结构。其具体代码如下：

卷积块

"""
卷积块
"""
class Conv_Block(nn.Module):
    def __init__(self, in_c, out_c, dilation=1) -> None:
        super().__init__()
        self.layers = nn.Sequential(
            nn.Conv2d(in_c, out_c,3,1,padding=1*dilation,dilation=1*dilation),
            nn.BatchNorm2d(out_c),
            nn.ReLU(inplace=True)
        )

    def forward(self, x):
        return self.layers(x)

下采样

"""
下采样块
"""
class Down_Sample_Block(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.layers = nn.Sequential(
            nn.MaxPool2d(2,2)
        ) 

    def forward(self, x):
        return self.layers(x)

上采样

"""
上采样块
"""
class Up_Sample_Block(nn.Module):
    def __init__(self, scale_factor=2) -> None:
        super().__init__()
        # 上采样方法1：
        self.upsample1 = nn.UpsamplingBilinear2d(scale_factor=scale_factor) 
        # 上采样方法2：
        self.upsample2 = nn.Upsample(scale_factor=scale_factor, mode='bilinear')

    def forward(self,x,feature):
        # 方法3：x = torch.nn.functional.interpolate(input=x,scale_factor=2, mode="nearest")
        x = self.upsample1(x)
        # 下面两行代码是将
        # resize = Resize((x.shape[2], x.shape[3]))
        # feature = resize(feature)
        res = torch.cat((x,feature),dim=1)
        return res

一、UNet_Block_x 代码实现

所以Unet_Block1的代码如下：

class UNet_Block1(nn.Module):
    def __init__(self, in_c, mid_c, out_c) -> None:
        super().__init__()
        self.down = Down_Sample_Block()
        self.up = Up_Sample_Block()

        self.conv1 = Conv_Block(in_c,out_c)
        self.conv2 = Conv_Block(out_c,mid_c)
        self.conv3 = Conv_Block(mid_c,mid_c)
        self.conv4 = Conv_Block(mid_c,mid_c,dilation=2)
        self.conv5 = Conv_Block(mid_c*2,mid_c)
        self.conv6 = Conv_Block(mid_c*2,out_c)

    def forward(self,x):
        # 下采样过程
        out1 = self.conv1(x) # out_c [1, 5, 224, 224]
        out2 = self.conv2(out1) # mid_c [1, 3, 224, 224]
        out3 = self.conv3(self.down(out2)) # mid_c [1, 3, 112, 112]
        out4 = self.conv3(self.down(out3)) # mid_c [1, 3, 56, 56]
        out5 = self.conv3(self.down(out4)) # mid_c [1, 3, 28, 28]
        out6 = self.conv3(self.down(out5)) # mid_c [1, 3, 14, 14]
        out7 = self.conv3(self.down(out6)) # mid_c [1, 3, 7, 7]

        out8 = self.conv4(out7) # mid_c [1, 3, 7, 7]
        out9 = self.conv5(torch.cat((out7,out8),dim=1)) # mid_c [1, 3, 7, 7]

        # 上采样
        out10 = self.conv5(self.up(out9,out6)) # [1, 3, 14, 14]
        out11 = self.conv5(self.up(out10,out5)) # [1, 3, 28, 28]
        out12 = self.conv5(self.up(out11,out4)) # [1, 3, 56, 56]
        out13 = self.conv5(self.up(out12,out3)) # [1, 3, 112, 112]
        out14 = self.conv6(self.up(out13,out2)) # [1, 3, 224, 224]
        out = out14 + out1 # [1, 5, 224, 224]
        return out

Unet_Block2的代码：

class UNet_Block2(nn.Module):
    def __init__(self, in_c, mid_c, out_c) -> None:
        super().__init__()
        self.down = Down_Sample_Block()
        self.up = Up_Sample_Block()

        self.conv1 = Conv_Block(in_c,out_c)
        self.conv2 = Conv_Block(out_c,mid_c)
        self.conv3 = Conv_Block(mid_c,mid_c)
        self.conv4 = Conv_Block(mid_c,mid_c,dilation=2)
        self.conv5 = Conv_Block(mid_c*2,mid_c)
        self.conv6 = Conv_Block(mid_c*2,out_c)

    def forward(self,x):
        # 下采样过程
        out1 = self.conv1(x) # out_c [1, 5, 112, 112]
        out2 = self.conv2(out1) # mid_c [1, 3, 112, 112]
        out3 = self.conv3(self.down(out2)) # mid_c [1, 3, 56, 56]
        out4 = self.conv3(self.down(out3)) # mid_c [1, 3, 28, 28]
        out5 = self.conv3(self.down(out4)) # mid_c [1, 3, 14, 14]
        out6 = self.conv3(self.down(out5)) # mid_c [1, 3, 7, 7]

        out8 = self.conv4(out6) # mid_c [1, 3, 7, 7]
        out9 = self.conv5(torch.cat((out6,out8),dim=1)) # mid_c [1, 3, 7, 7]

        # 上采样
        out10 = self.conv5(self.up(out9,out5)) # [1, 3, 14, 14]
        out11 = self.conv5(self.up(out10,out4)) # [1, 3, 28, 28]
        out12 = self.conv5(self.up(out11,out3)) # [1, 3, 56, 56]
        out13 = self.conv6(self.up(out12,out2)) # [1, 3, 112, 112]
        out = out13 + out1 # [1, 5, 112, 112]
        return out

Unet_Block3的代码：

class UNet_Block3(nn.Module):
    def __init__(self, in_c, mid_c, out_c) -> None:
        super().__init__()
        self.down = Down_Sample_Block()
        self.up = Up_Sample_Block()

        self.conv1 = Conv_Block(in_c,out_c)
        self.conv2 = Conv_Block(out_c,mid_c)
        self.conv3 = Conv_Block(mid_c,mid_c)
        self.conv4 = Conv_Block(mid_c,mid_c,dilation=2)
        self.conv5 = Conv_Block(mid_c*2,mid_c)
        self.conv6 = Conv_Block(mid_c*2,out_c)

    def forward(self,x):
        # 下采样过程
        out1 = self.conv1(x) # out_c [1, 5, 56, 56]
        out2 = self.conv2(out1) # mid_c [1, 3, 56, 56]
        out3 = self.conv3(self.down(out2)) # mid_c [1, 3, 28, 28]
        out4 = self.conv3(self.down(out3)) # mid_c [1, 3, 14, 14]
        out5 = self.conv3(self.down(out4)) # mid_c [1, 3, 7, 7]

        out8 = self.conv4(out5) # mid_c [1, 3, 7, 7]
        out9 = self.conv5(torch.cat((out5,out8),dim=1)) # mid_c [1, 3, 7, 7]

        # 上采样
        out10 = self.conv5(self.up(out9,out4)) # [1, 3, 14, 14]
        out11 = self.conv5(self.up(out10,out3)) # [1, 3, 28, 28]
        out12 = self.conv6(self.up(out11,out2)) # [1, 3, 56, 56]
        out = out12 + out1 # [1, 5, 56, 56]
        return out

Unet_Block4的代码：

class UNet_Block4(nn.Module):
    def __init__(self, in_c, mid_c, out_c) -> None:
        super().__init__()
        self.down = Down_Sample_Block()
        self.up = Up_Sample_Block()

        self.conv1 = Conv_Block(in_c,out_c)
        self.conv2 = Conv_Block(out_c,mid_c)
        self.conv3 = Conv_Block(mid_c,mid_c)
        self.conv4 = Conv_Block(mid_c,mid_c,dilation=2)
        self.conv5 = Conv_Block(mid_c*2,mid_c)
        self.conv6 = Conv_Block(mid_c*2,out_c)

    def forward(self,x):
        # 下采样过程
        out1 = self.conv1(x) # out_c [1, 5, 28, 28]
        out2 = self.conv2(out1) # mid_c [1, 3, 28, 28]
        out3 = self.conv3(self.down(out2)) # mid_c [1, 3, 14, 14]
        out4 = self.conv3(self.down(out3)) # mid_c [1, 3, 7, 7]

        out8 = self.conv4(out4) # mid_c [1, 3, 7, 7]
        out9 = self.conv5(torch.cat((out4,out8),dim=1)) # mid_c [1, 3, 7, 7]

        # 上采样
        out10 = self.conv5(self.up(out9,out3)) # [1, 3, 14, 14]
        out11 = self.conv6(self.up(out10,out2)) # [1, 3, 28, 28]
        out = out11 + out1 # [1, 5, 28, 28]
        return out

Unet_Block5的代码：

class UNet_Block5(nn.Module):
    def __init__(self, in_c, mid_c, out_c) -> None:
        super().__init__()

        self.conv1 = Conv_Block(in_c,out_c)
        self.conv2 = Conv_Block(out_c,mid_c)
        self.conv3 = Conv_Block(mid_c,mid_c,dilation=2)
        self.conv4 = Conv_Block(mid_c,mid_c,dilation=4)
        self.conv5 = Conv_Block(mid_c,mid_c,dilation=8)
        self.conv6 = Conv_Block(mid_c*2,mid_c,dilation=4)
        self.conv7 = Conv_Block(mid_c*2,mid_c,dilation=2)
        self.conv8 = Conv_Block(mid_c*2,out_c)

    def forward(self,x):
        out1 = self.conv1(x) # out_c [1, 5, 14, 14]
        out2 = self.conv2(out1) 
        out3 = self.conv3(out2) 
        out4 = self.conv4(out3) 

        out5 = self.conv5(out4) 

        out6 = self.conv6(torch.cat((out4,out5),dim=1)) 
        out7 = self.conv7(torch.cat((out3,out6),dim=1)) 
        out8 = self.conv8(torch.cat((out2,out7),dim=1)) 
        out = out8 + out1
        return out

二、U2Net模型代码实现

model2.py

import torch
import torch.nn as nn 

"""
卷积块
"""
class Conv_Block(nn.Module):
    def __init__(self, in_c, out_c, dilation=1) -> None:
        super().__init__()
        self.layers = nn.Sequential(
            nn.Conv2d(in_c, out_c,3,1,padding=1*dilation,dilation=1*dilation),
            nn.BatchNorm2d(out_c),
            nn.ReLU(inplace=True)
        )

    def forward(self, x):
        return self.layers(x) 
    
"""
下采样块
"""
class Down_Sample_Block(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.layers = nn.Sequential(
            nn.MaxPool2d(2,2)
        ) 

    def forward(self, x):
        return self.layers(x)
"""
上采样块
"""
class Up_Sample_Block(nn.Module):
    def __init__(self, scale_factor=2) -> None:
        super().__init__()
        # 上采样方法1：
        self.upsample1 = nn.UpsamplingBilinear2d(scale_factor=scale_factor) 
        # 上采样方法2：
        self.upsample2 = nn.Upsample(scale_factor=scale_factor, mode='bilinear')

    def forward(self,x,feature):
        # 方法3：x = torch.nn.functional.interpolate(input=x,scale_factor=2, mode="nearest")
        x = self.upsample1(x)
        # 下面两行代码是将
        # resize = Resize((x.shape[2], x.shape[3]))
        # feature = resize(feature)
        res = torch.cat((x,feature),dim=1)
        return res
    
"""
输出模块：
"""
class Output(nn.Module):
    def __init__(self,in_c, out_c) -> None:
        super().__init__()
        self.layers = self.layers = nn.Sequential(
            nn.Conv2d(in_c, out_c, 3, 1, 1,bias=False),
            nn.BatchNorm2d(out_c),
            nn.ReLU(),
            nn.Sigmoid()
        )

    def forward(self, x):
        return self.layers(x)

class UNet_Block1(nn.Module):
    def __init__(self, in_c, mid_c, out_c) -> None:
        super().__init__()
        self.down = Down_Sample_Block()
        self.up = Up_Sample_Block()

        self.conv1 = Conv_Block(in_c,out_c)
        self.conv2 = Conv_Block(out_c,mid_c)
        self.conv3 = Conv_Block(mid_c,mid_c)
        self.conv4 = Conv_Block(mid_c,mid_c,dilation=2)
        self.conv5 = Conv_Block(mid_c*2,mid_c)
        self.conv6 = Conv_Block(mid_c*2,out_c)

    def forward(self,x):
        # 下采样过程
        out1 = self.conv1(x) # out_c [1, 5, 224, 224]
        out2 = self.conv2(out1) # mid_c [1, 3, 224, 224]
        out3 = self.conv3(self.down(out2)) # mid_c [1, 3, 112, 112]
        out4 = self.conv3(self.down(out3)) # mid_c [1, 3, 56, 56]
        out5 = self.conv3(self.down(out4)) # mid_c [1, 3, 28, 28]
        out6 = self.conv3(self.down(out5)) # mid_c [1, 3, 14, 14]
        out7 = self.conv3(self.down(out6)) # mid_c [1, 3, 7, 7]

        out8 = self.conv4(out7) # mid_c [1, 3, 7, 7]
        out9 = self.conv5(torch.cat((out7,out8),dim=1)) # mid_c [1, 3, 7, 7]

        # 上采样
        out10 = self.conv5(self.up(out9,out6)) # [1, 3, 14, 14]
        out11 = self.conv5(self.up(out10,out5)) # [1, 3, 28, 28]
        out12 = self.conv5(self.up(out11,out4)) # [1, 3, 56, 56]
        out13 = self.conv5(self.up(out12,out3)) # [1, 3, 112, 112]
        out14 = self.conv6(self.up(out13,out2)) # [1, 3, 224, 224]
        out = out14 + out1 # [1, 5, 224, 224]
        return out
    
class UNet_Block2(nn.Module):
    def __init__(self, in_c, mid_c, out_c) -> None:
        super().__init__()
        self.down = Down_Sample_Block()
        self.up = Up_Sample_Block()

        self.conv1 = Conv_Block(in_c,out_c)
        self.conv2 = Conv_Block(out_c,mid_c)
        self.conv3 = Conv_Block(mid_c,mid_c)
        self.conv4 = Conv_Block(mid_c,mid_c,dilation=2)
        self.conv5 = Conv_Block(mid_c*2,mid_c)
        self.conv6 = Conv_Block(mid_c*2,out_c)

    def forward(self,x):
        # 下采样过程
        out1 = self.conv1(x) # out_c [1, 5, 112, 112]
        out2 = self.conv2(out1) # mid_c [1, 3, 112, 112]
        out3 = self.conv3(self.down(out2)) # mid_c [1, 3, 56, 56]
        out4 = self.conv3(self.down(out3)) # mid_c [1, 3, 28, 28]
        out5 = self.conv3(self.down(out4)) # mid_c [1, 3, 14, 14]
        out6 = self.conv3(self.down(out5)) # mid_c [1, 3, 7, 7]

        out8 = self.conv4(out6) # mid_c [1, 3, 7, 7]
        out9 = self.conv5(torch.cat((out6,out8),dim=1)) # mid_c [1, 3, 7, 7]

        # 上采样
        out10 = self.conv5(self.up(out9,out5)) # [1, 3, 14, 14]
        out11 = self.conv5(self.up(out10,out4)) # [1, 3, 28, 28]
        out12 = self.conv5(self.up(out11,out3)) # [1, 3, 56, 56]
        out13 = self.conv6(self.up(out12,out2)) # [1, 3, 112, 112]
        out = out13 + out1 # [1, 5, 112, 112]
        return out
    
class UNet_Block3(nn.Module):
    def __init__(self, in_c, mid_c, out_c) -> None:
        super().__init__()
        self.down = Down_Sample_Block()
        self.up = Up_Sample_Block()

        self.conv1 = Conv_Block(in_c,out_c)
        self.conv2 = Conv_Block(out_c,mid_c)
        self.conv3 = Conv_Block(mid_c,mid_c)
        self.conv4 = Conv_Block(mid_c,mid_c,dilation=2)
        self.conv5 = Conv_Block(mid_c*2,mid_c)
        self.conv6 = Conv_Block(mid_c*2,out_c)

    def forward(self,x):
        # 下采样过程
        out1 = self.conv1(x) # out_c [1, 5, 56, 56]
        out2 = self.conv2(out1) # mid_c [1, 3, 56, 56]
        out3 = self.conv3(self.down(out2)) # mid_c [1, 3, 28, 28]
        out4 = self.conv3(self.down(out3)) # mid_c [1, 3, 14, 14]
        out5 = self.conv3(self.down(out4)) # mid_c [1, 3, 7, 7]

        out8 = self.conv4(out5) # mid_c [1, 3, 7, 7]
        out9 = self.conv5(torch.cat((out5,out8),dim=1)) # mid_c [1, 3, 7, 7]

        # 上采样
        out10 = self.conv5(self.up(out9,out4)) # [1, 3, 14, 14]
        out11 = self.conv5(self.up(out10,out3)) # [1, 3, 28, 28]
        out12 = self.conv6(self.up(out11,out2)) # [1, 3, 56, 56]
        out = out12 + out1 # [1, 5, 56, 56]
        return out
    
class UNet_Block4(nn.Module):
    def __init__(self, in_c, mid_c, out_c) -> None:
        super().__init__()
        self.down = Down_Sample_Block()
        self.up = Up_Sample_Block()

        self.conv1 = Conv_Block(in_c,out_c)
        self.conv2 = Conv_Block(out_c,mid_c)
        self.conv3 = Conv_Block(mid_c,mid_c)
        self.conv4 = Conv_Block(mid_c,mid_c,dilation=2)
        self.conv5 = Conv_Block(mid_c*2,mid_c)
        self.conv6 = Conv_Block(mid_c*2,out_c)

    def forward(self,x):
        # 下采样过程
        out1 = self.conv1(x) # out_c [1, 5, 28, 28]
        out2 = self.conv2(out1) # mid_c [1, 3, 28, 28]
        out3 = self.conv3(self.down(out2)) # mid_c [1, 3, 14, 14]
        out4 = self.conv3(self.down(out3)) # mid_c [1, 3, 7, 7]

        out8 = self.conv4(out4) # mid_c [1, 3, 7, 7]
        out9 = self.conv5(torch.cat((out4,out8),dim=1)) # mid_c [1, 3, 7, 7]

        # 上采样
        out10 = self.conv5(self.up(out9,out3)) # [1, 3, 14, 14]
        out11 = self.conv6(self.up(out10,out2)) # [1, 3, 28, 28]
        out = out11 + out1 # [1, 5, 28, 28]
        return out
    
class UNet_Block5(nn.Module):
    def __init__(self, in_c, mid_c, out_c) -> None:
        super().__init__()

        self.conv1 = Conv_Block(in_c,out_c)
        self.conv2 = Conv_Block(out_c,mid_c)
        self.conv3 = Conv_Block(mid_c,mid_c,dilation=2)
        self.conv4 = Conv_Block(mid_c,mid_c,dilation=4)
        self.conv5 = Conv_Block(mid_c,mid_c,dilation=8)
        self.conv6 = Conv_Block(mid_c*2,mid_c,dilation=4)
        self.conv7 = Conv_Block(mid_c*2,mid_c,dilation=2)
        self.conv8 = Conv_Block(mid_c*2,out_c)

    def forward(self,x):
        out1 = self.conv1(x) # out_c [1, 5, 14, 14]
        out2 = self.conv2(out1) 
        out3 = self.conv3(out2) 
        out4 = self.conv4(out3) 

        out5 = self.conv5(out4) 

        out6 = self.conv6(torch.cat((out4,out5),dim=1)) 
        out7 = self.conv7(torch.cat((out3,out6),dim=1)) 
        out8 = self.conv8(torch.cat((out2,out7),dim=1)) 
        out = out8 + out1
        return out
    
class U2NET(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.down = Down_Sample_Block()
        self.up1 = nn.UpsamplingNearest2d(scale_factor=2)
        self.up2 = nn.UpsamplingNearest2d(scale_factor=4)
        self.up3 = nn.UpsamplingNearest2d(scale_factor=8)
        self.up4 = nn.UpsamplingNearest2d(scale_factor=16)
        self.up5 = nn.UpsamplingNearest2d(scale_factor=32)

        self.unet1 = UNet_Block1(1,32,64) 
        self.unet2 = UNet_Block2(64,32,128) 
        self.unet3 = UNet_Block3(128,64,256) 
        self.unet4 = UNet_Block4(256,128,512) 
        self.unet5 = UNet_Block5(512,256,512) 
        
        self.unet6 = UNet_Block5(512,256,512) 

        self.de_unet1 = UNet_Block1(128,16,64) 
        self.de_unet2 = UNet_Block2(256,32,64) 
        self.de_unet3 = UNet_Block3(512,64,128) 
        self.de_unet4 = UNet_Block4(1024,128,256) 
        self.de_unet5 = UNet_Block5(1024,256,512) 

        self.out1 = Output(64,1)
        self.out2 = Output(64,1)
        self.out3 = Output(128,1)
        self.out4 = Output(256,1)
        self.out5 = Output(512,1)

    def forward(self, x):
        # 下采样，编码
        conv1 = self.unet1(x)
        en1 = self.down(conv1)
        conv2 = self.unet2(en1)
        en2 = self.down(conv2)
        conv3 = self.unet3(en2)
        en3 = self.down(conv3)
        conv4 = self.unet4(en3)
        en4 = self.down(conv4)
        conv5 = self.unet5(en4)
        en5 = self.down(conv5)

        conv6 = self.unet6(en5)

        # 上采样，解码
        de1 = self.up1(conv6) # [1, 512, 14, 14]
        conv7 = self.de_unet5(torch.cat((conv5,de1),dim=1)) # [1, 512, 14, 14]
        de2 = self.up1(conv7) # [1, 512, 28, 28]
        conv8 = self.de_unet4(torch.cat((conv4,de2),dim=1)) # [1, 256, 28, 28]
        de3 = self.up1(conv8) # [1, 256, 56, 56]
        conv9 = self.de_unet3(torch.cat((conv3,de3),dim=1)) # [1, 128, 56, 56]
        de4 = self.up1(conv9) # [1, 128, 112, 112]
        conv10 = self.de_unet2(torch.cat((conv2,de4),dim=1)) # [1, 64, 112, 112]
        de5 = self.up1(conv10) # [1, 64, 224, 224]

        # 输出
        out1 = self.up5(self.out5(conv6)) # [1, 1, 224, 224]
        out2 = self.up4(self.out5(conv7)) # [1, 1, 224, 224]
        out3 = self.up3(self.out4(conv8)) # [1, 1, 224, 224]
        out4 = self.up2(self.out3(conv9)) # [1, 1, 224, 224]
        out5 = self.up1(self.out2(conv10)) # [1, 1, 224, 224]
        out6 = self.out1(de5) # [1, 1, 224, 224]

        out = (out1 + out2 + out3 + out4 + out5 + out6)/6

        return out
    
if __name__ == "__main__":
    x = torch.randn((1,1,224,224))
    conv = U2NET()
    y = conv(x)
    print(y.shape)

三、数据集和实验结果

（1）数据集：

VOC数据集：里面包含了图像分割数据，但训练效果并不好。
眼球血管分割数据集：可以使用这个开源数据集练练手，这里也给出链接：眼底血管分割 - 飞桨AI Studio (baidu.com)

(2)训练代码：

train2.py

import PIL.Image as Image
import numpy as np 
import torch 
import torch.nn as nn
from torch.utils.data import DataLoader, Dataset
import os 
from model2 import U2NET
from torchvision.utils import save_image

IMG_PATH = "./data/FundusVessels/JPEGImages/"
TARGET_PATH = "./data/FundusVessels/Annotations/"
DST_DIR = "./img2"

class EYE_Dataset(Dataset):
    def __init__(self) -> None:
        super().__init__()
        self.img_list =  os.listdir(IMG_PATH)
        self.target_list = os.listdir(TARGET_PATH)

    def __len__(self) -> int:
        return len(self.img_list)
    
    def __getitem__(self, index):
        name = self.img_list[index][0:-4]
        img = Image.open(IMG_PATH+f"{name}.jpg").convert("L").resize((224,224))
        lable = Image.open(TARGET_PATH+f"{name}.png").resize((224,224))
        img = np.array(img, dtype=np.float32)/255
        img = img[np.newaxis,:]
        lable = np.array(lable, dtype=np.float32)
        return img, lable
    


if __name__ == "__main__":
    device = "cuda" if torch.cuda.is_available() else "cpu"
    # device = "cpu"
    train_dataset = EYE_Dataset()
    train_loader = DataLoader(train_dataset,batch_size=1,shuffle=True)

    net = U2NET()
    if os.path.exists("./U2Net.pt"):
        params = torch.load("./U2Net.pt")
        net.load_state_dict(params)

    net.to(device)
    optim = torch.optim.Adam(net.parameters())
    loss_fn = nn.BCELoss()
    epoch=1
    net.train()
    while True:
        for i,(img,target) in enumerate(train_loader):
            img, target = img.to(device), target.to(device)
            y = net(img)
            loss = loss_fn(y, target.unsqueeze(dim=0))
            optim.zero_grad()
            loss.backward()
            optim.step()

            if i%1 == 0:
                # 保存测试结果
                img2 =  (y[0]>0.6).float() *255
                res = torch.stack([img[0],img2],dim=0)
                save_image(res.cpu(), DST_DIR + f"/epoch{epoch}_{i}.jpg", nrow=2)
                print(f"epoch {epoch},loss: {loss.item()}")
                torch.save(net.state_dict(),"./U2Net.pt")
            epoch += 1

(3)训练结果

大约训练了1500轮，多训练一会儿，效果还可以更好。

人面猴
序言：七十年代末，一起剥皮案震惊了整个滨河市，随后出现的几起案子，更是在滨河造成了极大的恐慌，老刑警刘岩，带你破解...
沈念sama阅读 205,236评论 6赞 478
死咒
序言：滨河连续发生了三起死亡事件，死亡现场离奇诡异，居然都是意外死亡，警方通过查阅死者的电脑和手机，发现死者居然都...
沈念sama阅读 87,867评论 2赞 381
救了他两次的神仙让他今天三更去死
文/潘晓璐我一进店门，熙熙楼的掌柜王于贵愁眉苦脸地迎上来，“玉大人，你说我怎么就摊上这事。” “怎么了？”我有些...
开封第一讲书人阅读 151,715评论 0赞 340
道士缉凶录：失踪的卖姜人
文/不坏的土叔我叫张陵，是天一观的道长。经常有香客问我，道长，这世上最难降的妖魔是什么？我笑而不...
开封第一讲书人阅读 54,899评论 1赞 278
港岛之恋（遗憾婚礼）
正文为了忘掉前任，我火速办了婚礼，结果婚礼上，老公的妹妹穿的比我还像新娘。我一直安慰自己，他们只是感情好，可当我...
茶点故事阅读 63,895评论 5赞 368
恶毒庶女顶嫁案：这布局不是一般人想出来的
文/花漫我一把揭开白布。她就那样静静地躺着，像睡着了一般。火红的嫁衣衬着肌肤如雪。梳的纹丝不乱的头发上，一...
开封第一讲书人阅读 48,733评论 1赞 283
城市分裂传说
那天，我揣着相机与录音，去河边找鬼。笑死，一个胖子当着我的面吹牛，可吹牛的内容都是我干的。我是一名探鬼主播，决...
沈念sama阅读 38,085评论 3赞 399
双鸳鸯连环套：你想象不到人心有多黑
文/苍兰香墨我猛地睁开眼，长吁一口气：“原来是场噩梦啊……” “哼！你这毒妇竟也来了？” 一声冷哼从身侧响起，我...
开封第一讲书人阅读 36,722评论 0赞 258
万荣杀人案实录
序言：老挝万荣一对情侣失踪，失踪者是张志新（化名）和其女友刘颖，没想到半个月后，有当地人在树林里发现了一具尸体，经...
沈念sama阅读 43,025评论 1赞 300
护林员之死
正文独居荒郊野岭守林人离奇死亡，尸身上长有42处带血的脓包…… 初始之章·张勋以下内容为张勋视角年9月15日...
茶点故事阅读 35,696评论 2赞 323
白月光启示录
正文我和宋清朗相恋三年，在试婚纱的时候发现自己被绿了。大学时的朋友给我发了我未婚夫和他白月光在一起吃饭的照片。...
茶点故事阅读 37,816评论 1赞 333
活死人
序言：一个原本活蹦乱跳的男人离奇死亡，死状恐怖，灵堂内的尸体忽然破棺而出，到底是诈尸还是另有隐情，我是刑警宁泽，带...
沈念sama阅读 33,447评论 4赞 322
日本核电站爆炸内幕
正文年R本政府宣布，位于F岛的核电站，受9级特大地震影响，放射性物质发生泄漏。R本人自食恶果不足惜，却给世界环境...
茶点故事阅读 39,057评论 3赞 307
男人毒药：我在死后第九天来索命
文/蒙蒙一、第九天我趴在偏房一处隐蔽的房顶上张望。院中可真热闹，春花似锦、人声如沸。这庄子的主人今日做“春日...
开封第一讲书人阅读 30,009评论 0赞 19
一桩弑父案，背后竟有这般阴谋
文/苍兰香墨我抬头看了看天上的太阳。三九已至，却和暖如春，着一层夹袄步出监牢的瞬间，已是汗流浃背。一阵脚步声响...
开封第一讲书人阅读 31,254评论 1赞 260
情欲美人皮
我被黑心中介骗来泰国打工，没想到刚下飞机就差点儿被人妖公主榨干…… 1. 我叫王不留，地道东北人。一个月前我还...
沈念sama阅读 45,204评论 2赞 352
代替公主和亲
正文我出身青楼，却偏偏与公主长得像，于是被迫代替她去往敌国和亲。传闻我的和亲对象是个残疾皇子，可洞房花烛夜当晚...
茶点故事阅读 42,561评论 2赞 343