和FCN类似，UNet是另一个做语义分割的网络，网络从输入到输出中间呈一个U型而得名。
相比于FCN，UNet增加了更多的中间连接，能够更好处理不同尺度上的特征。
网络结构如下：
下面代码是用UNet对VOC数据集做的语义分割。
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset,DataLoader
from torchvision import transforms
import os
from PIL import Image
import numpy as np
transform = transforms.Compose([
transforms.Resize((256, 256)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
colormap = [[0,0,0],[128,0,0],[0,128,0], [128,128,0], [0,0,128],
[128,0,128],[0,128,128],[128,128,128],[64,0,0],[192,0,0],
[64,128,0],[192,128,0],[64,0,128],[192,0,128],
[64,128,128],[192,128,128],[0,64,0],[128,64,0],
[0,192,0],[128,192,0],[0,64,128]]
class VOCData(Dataset):
def __init__(self, root):
super(VOCData, self).__init__()
self.lab_path = root + 'VOC2012/SegmentationClass/'
self.img_path = root + 'VOC2012/JPEGImages/'
self.lab_names = self.get_file_names(self.lab_path)
self.img_names=[]
for file in self.lab_names:
self.img_names.append(file.replace('.png', '.jpg'))
self.cm2lbl = np.zeros(256**3)
for i,cm in enumerate(colormap):
self.cm2lbl[cm[0]*256*256+cm[1]*256+cm[2]] = i
self.image = []
self.label = []
for i in range(len(self.lab_names)):
image = Image.open(self.img_path+self.img_names[i]).convert('RGB')
image = transform(image)
label = Image.open(self.lab_path+self.lab_names[i]).convert('RGB').resize((256,256))
label = torch.from_numpy(self.image2label(label))
self.image.append(image)
self.label.append(label)
def __len__(self):
return len(self.image)
def __getitem__(self, idx):
return self.image[idx], self.label[idx]
def get_file_names(self,directory):
file_names = []
for file_name in os.listdir(directory):
if os.path.isfile(os.path.join(directory, file_name)):
file_names.append(file_name)
return file_names
def image2label(self,im):
data = np.array(im, dtype='int32')
idx = data[:, :, 0] * 256 * 256 + data[:, :, 1] * 256 + data[:, :, 2]
return np.array(self.cm2lbl[idx], dtype='int64')
class convblock(nn.Module):
def __init__(self, in_channels, out_channels):
super(convblock, self).__init__()
self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
self.bn1 = nn.BatchNorm2d(out_channels)
self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1)
self.bn2 = nn.BatchNorm2d(out_channels)
self.relu = nn.ReLU(inplace=True)
def forward(self, x):
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
x = self.conv2(x)
x = self.bn2(x)
x = self.relu(x)
return x
class Unet(nn.Module):
def __init__(self, num_classes):
super(Unet, self).__init__()
self.conv_block1 = convblock(3,64)
self.conv_block2 = convblock(64,128)
self.conv_block3 = convblock(128,256)
self.conv_block4 = convblock(256,512)
self.conv_block5 = convblock(512,1024)
self.upsample1 = nn.ConvTranspose2d(1024, 512, kernel_size=2, stride=2, padding=0)
self.upsample2 = nn.ConvTranspose2d(512, 256, kernel_size=2, stride=2, padding=0)
self.upsample3 = nn.ConvTranspose2d(256, 128, kernel_size=2, stride=2, padding=0)
self.upsample4 = nn.ConvTranspose2d(128, 64, kernel_size=2, stride=2, padding=0)
self.conv_block6 = convblock(1024,512)
self.conv_block7 = convblock(512,256)
self.conv_block8 = convblock(256,128)
self.conv_block9 = convblock(128,64)
self.maxpool = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv_out = convblock(64,num_classes)
def forward(self, x):
x1 = self.conv_block1(x)
x = self.maxpool(x1)
x2 = self.conv_block2(x)
x = self.maxpool(x2)
x3 = self.conv_block3(x)
x = self.maxpool(x3)
x4 = self.conv_block4(x)
x = self.maxpool(x4)
x = self.conv_block5(x)
x = self.upsample1(x)
x = torch.cat([x4,x],dim=1)
x = self.conv_block6(x)
x = self.upsample2(x)
x = torch.cat([x3,x],dim=1)
x = self.conv_block7(x)
x = self.upsample3(x)
x = torch.cat([x2,x],dim=1)
x = self.conv_block8(x)
x = self.upsample4(x)
x = torch.cat([x1,x],dim=1)
x = self.conv_block9(x)
x = self.conv_out(x)
return x
def train():
train_dataset = VOCData(root='./VOCdevkit/')
train_loader = DataLoader(train_dataset, batch_size=4, shuffle=True)
net = Unet(21)
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
criterion = nn.CrossEntropyLoss()
net.to(device)
net.train()
num_epochs = 100
for epoch in range(num_epochs):
loss_sum = 0
img_sum = 0
for inputs, labels in train_loader:
inputs = inputs.to(device)
labels = labels.to(device)
outputs = net(inputs)
loss = criterion(outputs, labels)
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_sum += loss.item()
img_sum += inputs.shape[0]
print('epochs:',epoch,loss_sum / img_sum )
torch.save(net.state_dict(), 'unet.pth')
def val():
net = Unet(21)
net.load_state_dict(torch.load('unet.pth'))
net.to(device)
net.eval()
image = Image.open('./VOCdevkit/VOC2012/JPEGImages/2012_001064.jpg').convert('RGB')
image = transform(image).unsqueeze(0).to(device)
out = net(image).squeeze(0)
ToPIL= transforms.ToPILImage()
maxind = torch.argmax(out,dim=0)
outimg = torch.zeros([3,256,256])
for y in range(256):
for x in range(256):
outimg[:,x,y] = torch.from_numpy(np.array(colormap[maxind[x,y]]))
re = ToPIL(outimg)
re.show()
if __name__ == "__main__":
train()
#val()
标签：__,conv,self,torch,学习,UNet,深度,128,256
From： https://www.cnblogs.com/tiandsp/p/17922362.html