大家好，我又好久没有给大家更新这个系列了，但是我内心一直没有忘记要更新pytorch初学者系列文章，今天给大家分享一下Pytorch如何构建UNet网络并实现模型训练与测试，实现一个道路裂纹检测!
数据集
CrackForest数据集，包括118张标注数据，37张验证与测试数据。数据集的目录有groundtruth、image、seg三个子目录，分别是标注数据、原始图像、分割信息。其中标注信息是matlab格式的文件，通过字典方式实现数据存储与读写，seg文件本质是text文件，按行来组织信息，前面几行是图像属性与格式化信息，data部分的格式如下：
代码语言：javascript复制Seg_num 空格 row_index 空格 column1 column2- 空格表示space，
- seg_num值为0或者1
- row_index表示当前行
- column1表示开始列位置
- column2 表示结束列位置
假设seg中描述的图像宽度为480，高度为320，表示第一行的分割信息表示如下：
代码语言：javascript复制0 0 0 479 表示图像第一行从列0到列479为0，黑色
1 200 141 151 表示图像中第200行中列141到151为1，白色最终解释上述数据集生成的mask数据显示如下：大小均为（480x320）
Pytorch中定义对应数据集类的代码实现如下：
代码语言：javascript复制class SegmentationDataset(Dataset):
def __init__(self, image_dir, mask_dir):
self.images = []
self.masks = []
files = os.listdir(image_dir)
sfiles = os.listdir(mask_dir)
for i in range(len(sfiles)):
img_file = os.path.join(image_dir, files[i])
mask_file = os.path.join(mask_dir, sfiles[i])
# print(img_file, mask_file)
self.images.append(img_file)
self.masks.append(mask_file)
def __len__(self):
return len(self.images)
def num_of_samples(self):
return len(self.images)
def __getitem__(self, idx):
if torch.is_tensor(idx):
idx = idx.tolist()
image_path = self.images[idx]
mask_path = self.masks[idx]
else:
image_path = self.images[idx]
mask_path = self.masks[idx]
img = cv.imread(image_path, cv.IMREAD_GRAYSCALE) # BGR order
mask = cv.imread(mask_path, cv.IMREAD_GRAYSCALE)
# 输入图像
img = np.float32(img) / 255.0
img = np.expand_dims(img, 0)
# 目标标签0 ~ 1， 对于
mask[mask <= 128] = 0
mask[mask > 128] = 1
mask = np.expand_dims(mask, 0)
sample = {'image': torch.from_numpy(img), 'mask': torch.from_numpy(mask),}
return sample模型构建
UNet网络是图像语义分割网络，整个网络可以分为两个部分来解释。第一部分是编码网络，不断的降低分辨率，实现图像特征提取；第二部分是解码网络，不断提升分辨率同时尝试重建图像有用信息，最终输出结果。网络模型结构如下：
代码实现如下：
代码语言：javascript复制
代码实现如下class UNetModel(torch.nn.Module):
def __init__(self, in_features=1, out_features=2, init_features=32):
super(UNetModel, self).__init__()
features = init_features
self.encode_layer1 = torch.nn.Sequential(
torch.nn.Conv2d(in_channels=in_features, out_channels=features, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features),
torch.nn.ReLU(),
torch.nn.Conv2d(in_channels=features, out_channels=features, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features),
torch.nn.ReLU()
)
self.pool1 = torch.nn.MaxPool2d(kernel_size=2, stride=2)
self.encode_layer2 = torch.nn.Sequential(
torch.nn.Conv2d(in_channels=features, out_channels=features*2, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features*2),
torch.nn.ReLU(),
torch.nn.Conv2d(in_channels=features*2, out_channels=features*2, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features * 2),
torch.nn.ReLU()
)
self.pool2 = torch.nn.MaxPool2d(kernel_size=2, stride=2)
self.encode_layer3 = torch.nn.Sequential(
torch.nn.Conv2d(in_channels=features*2, out_channels=features*4, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features * 4),
torch.nn.ReLU(),
torch.nn.Conv2d(in_channels=features*4, out_channels=features*4, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features * 4),
torch.nn.ReLU()
)
self.pool3 = torch.nn.MaxPool2d(kernel_size=2, stride=2)
self.encode_layer4 = torch.nn.Sequential(
torch.nn.Conv2d(in_channels=features*4, out_channels=features*8, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features * 8),
torch.nn.ReLU(),
torch.nn.Conv2d(in_channels=features*8, out_channels=features*8, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features * 8),
torch.nn.ReLU(),
)
self.pool4 = torch.nn.MaxPool2d(kernel_size=2, stride=2)
self.encode_decode_layer = torch.nn.Sequential(
torch.nn.Conv2d(in_channels=features*8, out_channels=features*16, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features * 16),
torch.nn.ReLU(),
torch.nn.Conv2d(in_channels=features*16, out_channels=features*16, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features * 16),
torch.nn.ReLU()
)
self.upconv4 = torch.nn.ConvTranspose2d(
features * 16, features * 8, kernel_size=2, stride=2
)
self.decode_layer4 = torch.nn.Sequential(
torch.nn.Conv2d(in_channels=features*16, out_channels=features*8, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features*8),
torch.nn.ReLU(),
torch.nn.Conv2d(in_channels=features*8, out_channels=features*8, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features * 8),
torch.nn.ReLU(),
)
self.upconv3 = torch.nn.ConvTranspose2d(
features * 8, features * 4, kernel_size=2, stride=2
)
self.decode_layer3 = torch.nn.Sequential(
torch.nn.Conv2d(in_channels=features*8, out_channels=features*4, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features * 4),
torch.nn.ReLU(),
torch.nn.Conv2d(in_channels=features*4, out_channels=features*4, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features * 4),
torch.nn.ReLU()
)
self.upconv2 = torch.nn.ConvTranspose2d(
features * 4, features * 2, kernel_size=2, stride=2
)
self.decode_layer2 = torch.nn.Sequential(
torch.nn.Conv2d(in_channels=features*4, out_channels=features*2, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features * 2),
torch.nn.ReLU(),
torch.nn.Conv2d(in_channels=features*2, out_channels=features*2, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features * 2),
torch.nn.ReLU()
)
self.upconv1 = torch.nn.ConvTranspose2d(
features * 2, features, kernel_size=2, stride=2
)
self.decode_layer1 = torch.nn.Sequential(
torch.nn.Conv2d(in_channels=features*2, out_channels=features, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features),
torch.nn.ReLU(),
torch.nn.Conv2d(in_channels=features, out_channels=features, kernel_size=3, padding=1, stride=1),
torch.nn.BatchNorm2d(num_features=features),
torch.nn.ReLU()
)
self.out_layer = torch.nn.Sequential(
torch.nn.Conv2d(in_channels=features, out_channels=out_features, kernel_size=1, padding=0, stride=1),
)
def forward(self, x):
enc1 = self.encode_layer1(x)
enc2 = self.encode_layer2(self.pool1(enc1))
enc3 = self.encode_layer3(self.pool2(enc2))
enc4 = self.encode_layer4(self.pool3(enc3))
bottleneck = self.encode_decode_layer(self.pool4(enc4))
dec4 = self.upconv4(bottleneck)
dec4 = torch.cat((dec4, enc4), dim=1)
dec4 = self.decode_layer4(dec4)
dec3 = self.upconv3(dec4)
dec3 = torch.cat((dec3, enc3), dim=1)
dec3 = self.decode_layer3(dec3)
dec2 = self.upconv2(dec3)
dec2 = torch.cat((dec2, enc2), dim=1)
dec2 = self.decode_layer2(dec2)
dec1 = self.upconv1(dec2)
dec1 = torch.cat((dec1, enc1), dim=1)
dec1 = self.decode_layer1(dec1)
out = self.out_layer(dec1)
return out训练过程
基于像素的交叉熵损失与Adam优化器实现模型训练，输入图像格式为：
代码语言：javascript复制NCHW=2x1x320x480如果硬件条件允许，建议把batchSize可以开4或者8、16尝试做对比测试。这里我训练了15个epoch，训练部分的代码如下：
代码语言：javascript复制index = 0
for epoch in range(num_epochs):
train_loss = 0.0
for i_batch, sample_batched in enumerate(dataloader):
images_batch, target_labels =
sample_batched['image'], sample_batched['mask']
if train_on_gpu:
images_batch, target_labels = images_batch.cuda(), target_labels.cuda()
optimizer.zero_grad()
# forward pass: compute predicted outputs by passing inputs to the model
m_label_out_ = unet(images_batch)
# calculate the batch loss
target_labels = target_labels.contiguous().view(-1)
m_label_out_ = m_label_out_.transpose(1,3).transpose(1, 2).contiguous().view(-1, 2)
target_labels = target_labels.long()
loss = cross_loss(m_label_out_, target_labels)
# backward pass: compute gradient of the loss with respect to model parameters
loss.backward()
# perform a single optimization step (parameter update)
optimizer.step()
# update training loss
train_loss = loss.item()
if index % 100 == 0:
print('step: {} tcurrent Loss: {:.6f} '.format(index, loss.item()))
index = 1
# 计算平均损失
train_loss = train_loss / num_train_samples
# 显示训练集与验证集的损失函数
print('Epoch: {} tTraining Loss: {:.6f} '.format(epoch, train_loss))
# save model
unet.eval()
torch.save(unet, 'unet_road_model.pt')模型测试
对训练生成的UNet模型，使用下面的代码进行测试与验证。测试运行代码如下：
代码语言：javascript复制cnn_model = torch.load("./unet_road_model.pt")
root_dir = "D:/pytorch/CrackForest-dataset/test"
fileNames = os.listdir(root_dir)
for f in fileNames:
image = cv.imread(os.path.join(root_dir, f), cv.IMREAD_GRAYSCALE)
h, w = image.shape
img = np.float32(image) /255.0
img = np.expand_dims(img, 0)
x_input = torch.from_numpy(img).view( 1, 1, h, w)
probs = cnn_model(x_input.cuda())
m_label_out_ = probs.transpose(1, 3).transpose(1, 2).contiguous().view(-1, 2)
_, output = m_label_out_.data.max(dim=1)
output[output > 0] = 255
predic_ = output.view(h, w).cpu().detach().numpy()
print(predic_.shape)
cv.imshow("input", image)
result = cv.resize(np.uint8(predic_), (w, h))
cv.imshow("unet-segmentation-demo", result)
cv.waitKey(0)
cv.destroyAllWindows()运行结果如下：
无裂纹道路
有裂纹道路
君子藏器于身，待时而动
