pytorch用voc分割数据集训练FCN

语义分割是对图像中的每一个像素进行分类，从而完成图像分割的过程。分割主要用于医学图像领域和无人驾驶领域。

和其他算法一样，图像分割发展过程也经历了传统算法到深度学习算法的转变，传统的分割算法包括阈值分割、分水岭、边缘检测等等，面临的问题也跟其他传统图像处理算法一样，就是鲁棒性不够，但在一些场景单一不变的场合，传统图像处理依旧用的较多。

FCN是2014年的一篇论文，深度学习语义分割的开山之作，从思想上奠定了语义分割的基础。

Fully Convolutional Networks for Semantic Segmentation
Submitted on 14 Nov 2014
https://arxiv.org/abs/1411.4038

一、FCN理论介绍

 上图是原论文中的截图，从整体架构上描绘了FCN的网络架构。其实就是图像经过一系列卷积运算，然后再上采样成原图大小，输出每一个像素的类别概率。

 上图更加细致的描述了FCN的网络。backbone采用VGG16，把VGG的fully-connect层用卷积来表示，即conv6-7（一个大小和feature_map同样size的卷积核，就相当于全连接）。总的来说，网络有下列几个关键点：

1. Fully Convolution: 用于解决像素的预测问题。通过将基础网络（如VGG16）最后全连接层替换为卷积层，可实现任意大小的图像输入，并且输出图像大小与输入相对应；

2. Transpose Convolution: 上采样过程，用于恢复图片尺寸，方便后续进行逐个像素的预测；

3. Skip Architecture : 用于融合高底层特征信息。因为卷积是个下采样操作，而转置卷积虽然恢复了图像尺寸，但毕竟不是卷积的逆操作，所以信息肯定有丢失，而skip architecture可以融合千层的细粒度信息和深层的粗粒度信息，提高分割的精细程度。

 

FCN-32s:  没有跳连接，按照每层转置卷积放大2倍的速度放大，经过五层后放大32倍复原原图大小。

FCN-16s: 一个skip-connect，（1/32）放大为（1/16）后，再与vgg的（1/16）相加，然后继续放大，直到原图大小。 

FCN-8s: 两个skip-connect，一个是（1/32）放大为（1/16）后，再与vgg的（1/16）相加；另外一个是（1/16）放大为（1/8）之后，再与vgg的（1/8）相加，然后继续放大，直到原图大小。

二、训练过程

pytorch训练深度学习模型主要实现三个文件即可，分别为data.py, model.py, train.py。其中data.py里实现数据批量处理功能，model.py定义网络模型，train.py实现训练步骤。

2.1 voc数据集介绍

 

下载地址：Pascal VOC Dataset Mirror

图片的名称在/ImageSets/Segmentation/train.txt ans val.txt里
图片都在./data/VOC2012/JPEGImages文件夹下面，需要在train.txt读取的每一行后面加上.jpg
标签都在./data/VOC2012/SegmentationClass文件夹下面，需要在读取的每一行后面加上.png

voc_seg_data.py

import torch
import torch.nn as nn
import torchvision.transforms as T
from torch.utils.data import DataLoader,Dataset
import numpy as np
import os
from PIL import Image
from datetime import datetime



class VOC_SEG(Dataset):
    def __init__(self, root, width, height, train=True, transforms=None):
        # 图像统一剪切尺寸（width, height）
        self.width = width
        self.height = height
        # VOC数据集中对应的标签
        self.classes = ['background','aeroplane','bicycle','bird','boat',
           'bottle','bus','car','cat','chair','cow','diningtable',
           'dog','horse','motorbike','person','potted plant',
           'sheep','sofa','train','tv/monitor']
        # 各种标签所对应的颜色
        self.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]]
        # 辅助变量
        self.fnum = 0
        if transforms is None:
            normalize = T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
            self.transforms = T.Compose([
                T.ToTensor(),
                normalize
            ])
        # 像素值(RGB)与类别label(0,1,3...)一一对应
        self.cm2lbl = np.zeros(256**3)
        for i, cm in enumerate(self.colormap):
            self.cm2lbl[(cm[0]*256+cm[1])*256+cm[2]] = i

        
        if train:
            txt_fname = root+"/ImageSets/Segmentation/train.txt"
        else:
            txt_fname = root+"/ImageSets/Segmentation/val.txt"
        with open(txt_fname, 'r') as f:
            images = f.read().split()
        imgs = [os.path.join(root, "JPEGImages", item+".jpg") for item in images]
        labels = [os.path.join(root, "SegmentationClass", item+".png") for item in images]
        self.imgs = self._filter(imgs)
        self.labels = self._filter(labels)
        if train:
            print("训练集：加载了 " + str(len(self.imgs)) + " 张图片和标签" + ",过滤了" + str(self.fnum) + "张图片")
        else:
            print("测试集：加载了 " + str(len(self.imgs)) + " 张图片和标签" + ",过滤了" + str(self.fnum) + "张图片")

    def _crop(self, data, label):
        """
        切割函数，默认都是从图片的左上角开始切割。切割后的图片宽是width,高是height
        data和label都是Image对象
        """
        box = (0,0,self.width,self.height)
        data = data.crop(box)
        label = label.crop(box)
        return data, label

    def _image2label(self, im):
        data = np.array(im, dtype="int32")
        idx = (data[:,:,0]*256+data[:,:,1])*256+data[:,:,2]
        return np.array(self.cm2lbl[idx], dtype="int64")
        
    def _image_transforms(self, data, label):
        data, label = self._crop(data,label)
        data = self.transforms(data)
        label = self._image2label(label)
        label = torch.from_numpy(label)
        return data, label

    def _filter(self, imgs): 
        img = []
        for im in imgs:
            if (Image.open(im).size[1] >= self.height and 
               Image.open(im).size[0] >= self.width):
                img.append(im)
            else:
                self.fnum  = self.fnum+1
        return img

    def __getitem__(self, index: int):
        img_path = self.imgs[index]
        label_path = self.labels[index]
        img = Image.open(img_path)
        label = Image.open(label_path).convert("RGB")
        img, label = self._image_transforms(img, label)
        return img, label

    def __len__(self) :
        return len(self.imgs)



if __name__=="__main__":
    root = "./VOCdevkit/VOC2012"
    height = 224
    width = 224
    voc_train = VOC_SEG(root, width, height, train=True)
    voc_test = VOC_SEG(root, width, height, train=False)

    # train_data = DataLoader(voc_train, batch_size=8, shuffle=True)
    # valid_data = DataLoader(voc_test, batch_size=8)
    for data, label in voc_train:
        print(data.shape)
        print(label.shape)
        break

• 我这里为了省事把一些辅助函数，如_crop(), _filter()，还是有变量colormap等都写到类里面了。实际上脱离出来另外写一个数据预处理的文件比较好，这样在训练结束后，推理测试时可以直接调用相应的处理函数。
• 数据处理的结果是得到data, label。data是tensor格式的图像，label也是tensor，且已经把像素(RGB)替换为了int类别号。这样在训练时候，交叉熵函数直接会实现one-hot处理，就跟训练分类网络一样。
   

2.2  网络定义

fcn8s_net.py

import torch
import torch.nn as nn
from torch.autograd import Variable
import torch.nn.functional as F
from torchsummary import summary
from torchvision import models


class FCN8s(nn.Module):
    def __init__(self, num_classes=21):
        super(FCN8s,self).__init__()
        net = models.vgg16(pretrained=True)   # 从预训练模型加载VGG16网络参数
        self.premodel = net.features          # 只使用Vgg16的五层卷积层（特征提取层）（3，224，224）----->（512，7，7）

        # self.conv6 = nn.Conv2d(512,512,kernel_size=1,stride=1,padding=0,dilation=1) 
        # self.conv7 = nn.Conv2d(512,512,kernel_size=1,stride=1,padding=0,dilation=1)
        # (512,7,7)
        self.relu = nn.ReLU(inplace=True)
        self.deconv1 = nn.ConvTranspose2d(512,512,kernel_size=3,stride=2,padding=1,dilation=1,output_padding=1)  # x2
        self.bn1 = nn.BatchNorm2d(512)
        # (512, 14, 14)
        self.deconv2 = nn.ConvTranspose2d(512,256,kernel_size=3,stride=2,padding=1,dilation=1,output_padding=1)  # x2
        self.bn2 = nn.BatchNorm2d(256)
        # (256, 28, 28)
        self.deconv3 = nn.ConvTranspose2d(256,128,kernel_size=3,stride=2,padding=1,dilation=1,output_padding=1)  # x2
        self.bn3 = nn.BatchNorm2d(128)
        # (128, 56, 56)
        self.deconv4 = nn.ConvTranspose2d(128,64,kernel_size=3,stride=2,padding=1,dilation=1,output_padding=1)   # x2
        self.bn4 = nn.BatchNorm2d(64)
        # (64, 112, 112)
        self.deconv5 = nn.ConvTranspose2d(64,32,kernel_size=3,stride=2,padding=1,dilation=1,output_padding=1)    # x2
        self.bn5 = nn.BatchNorm2d(32)
        # (32, 224, 224)
        self.classifier = nn.Conv2d(32, num_classes, kernel_size=1)
        # (num_classes, 224, 224)
        

    def forward(self, input):
        x = input
        for i in range(len(self.premodel)):
            x = self.premodel[i](x)
            if i == 16:
                x3 = x  # maxpooling3的feature map (1/8)
            if i == 23:
                x4 = x  # maxpooling4的feature map (1/16)
            if i == 30:
                x5 = x  # maxpooling5的feature map (1/32)

        # 五层转置卷积，每层size放大2倍，与VGG16刚好相反。两个skip-connect
        score = self.relu(self.deconv1(x5))   # out_size = 2*in_size (1/16)
        score = self.bn1(score + x4)

        score = self.relu(self.deconv2(score)) # out_size = 2*in_size (1/8)  
        score = self.bn2(score + x3)

        score = self.bn3(self.relu(self.deconv3(score)))  # out_size = 2*in_size (1/4)
        score = self.bn4(self.relu(self.deconv4(score)))  # out_size = 2*in_size (1/2)
        score = self.bn5(self.relu(self.deconv5(score)))  # out_size = 2*in_size (1)

        score = self.classifier(score)                    # size不变，使输出的channel等于类别数

        return score

if __name__ == "__main__":
    model = FCN8s()
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    model = model.to(device)
    print(model)

• FCN的网络代码实现上，在网上查的都有所差异，不过总体都是卷积+转置卷积+跳链接的结构。实际上只要实现特征提取（提取抽象特征）——转置卷积（恢复原图大小）——给每一个像素分类的过程就够了。
• 本次实验采用vgg16的五层卷积层作为特征提取网络，然后接五个转置卷积（2x）恢复到原图大小，然后再接一个卷积层把feature map的通道调整为类别个数（21）。最后再softmax分类就行了。

2.3 训练

train.py

import torch
import torch.nn as nn
from torch.utils.data import DataLoader,Dataset
from voc_seg_data import VOC_SEG
from fcn_net import FCN8s
import os
import numpy as np
 

# 计算混淆矩阵
def _fast_hist(label_true, label_pred, n_class):
    mask = (label_true >= 0) & (label_true < n_class)
    hist = np.bincount(
        n_class * label_true[mask].astype(int) +
        label_pred[mask], minlength=n_class ** 2).reshape(n_class, n_class)
    return hist


# 根据混淆矩阵计算Acc和mIou
def label_accuracy_score(label_trues, label_preds, n_class):
    """Returns accuracy score evaluation result.
      - overall accuracy
      - mean accuracy
      - mean IU
    """
    hist = np.zeros((n_class, n_class))
    for lt, lp in zip(label_trues, label_preds):
        hist += _fast_hist(lt.flatten(), lp.flatten(), n_class)
    acc = np.diag(hist).sum() / hist.sum()
    with np.errstate(divide='ignore', invalid='ignore'):
        acc_cls = np.diag(hist) / hist.sum(axis=1)
    acc_cls = np.nanmean(acc_cls)
    with np.errstate(divide='ignore', invalid='ignore'):
        iu = np.diag(hist) / (
            hist.sum(axis=1) + hist.sum(axis=0) - np.diag(hist)
        )
    mean_iu = np.nanmean(iu)
    freq = hist.sum(axis=1) / hist.sum()
    return acc, acc_cls, mean_iu


def main():
    # 1. load dataset
    root = "./VOCdevkit/VOC2012"
    batch_size = 32
    height = 224
    width = 224
    voc_train = VOC_SEG(root, width, height, train=True)
    voc_test = VOC_SEG(root, width, height, train=False)
    train_dataloader = DataLoader(voc_train,batch_size=batch_size,shuffle=True)
    val_dataloader = DataLoader(voc_test,batch_size=batch_size,shuffle=True)
    
    # 2. load model
    num_class = 21
    model = FCN8s(num_classes=num_class)
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    model = model.to(device)
    
    # 3. prepare super parameters
    criterion = nn.CrossEntropyLoss() 
    optimizer = torch.optim.SGD(model.parameters(), lr=1e-3, momentum=0.7)
    epoch = 50
 
    # 4. train
    val_acc_list = []
    out_dir = "./checkpoints/"
    if not os.path.exists(out_dir):
        os.makedirs(out_dir)
    for epoch in range(0, epoch):
        print('\nEpoch: %d' % (epoch + 1))
        model.train()
        sum_loss = 0.0
        for batch_idx, (images, labels) in enumerate(train_dataloader):
            length = len(train_dataloader)
            images, labels = images.to(device), labels.to(device)
            optimizer.zero_grad()
            outputs = model(images) # torch.size([batch_size, num_class, width, height])
            loss = criterion(outputs, labels)
            loss.backward()
            optimizer.step()
        
            sum_loss += loss.item()
            predicted = torch.argmax(outputs.data, 1)
            
            label_pred = predicted.data.cpu().numpy()
            label_true = labels.data.cpu().numpy()
            acc, acc_cls, mean_iu = label_accuracy_score(label_true,label_pred,num_class)
            
            print('[epoch:%d, iter:%d] Loss: %.03f | Acc: %.3f%% | Acc_cls: %.03f%% |Mean_iu: %.3f' 
                % (epoch + 1, (batch_idx + 1 + epoch * length), sum_loss / (batch_idx + 1), 
                100. *acc, 100.*acc_cls, mean_iu))
            
        #get the ac with testdataset in each epoch
        print('Waiting Val...')
        mean_iu_epoch = 0.0
        mean_acc = 0.0
        mean_acc_cls = 0.0
        with torch.no_grad():
            for batch_idx, (images, labels) in enumerate(val_dataloader):
                model.eval()
                images, labels = images.to(device), labels.to(device)
                outputs = model(images)
                predicted = torch.argmax(outputs.data, 1)


                label_pred = predicted.data.cpu().numpy()
                label_true = labels.data.cpu().numpy()
                acc, acc_cls, mean_iu = label_accuracy_score(label_true,label_pred,num_class)

                # total += labels.size(0)
                # iou = torch.sum((predicted == labels.data), (1,2)) / float(width*height)
                # iou = torch.sum(iou)
                # correct += iou
                mean_iu_epoch += mean_iu
                mean_acc += acc
                mean_acc_cls += acc_cls
            
            print('Acc_epoch: %.3f%% | Acc_cls_epoch: %.03f%% |Mean_iu_epoch: %.3f' 
                % ((100. *mean_acc / len(val_dataloader)), (100.*mean_acc_cls/len(val_dataloader)), mean_iu_epoch/len(val_dataloader)) )
            
            val_acc_list.append(mean_iu_epoch/len(val_dataloader))
 
 
        torch.save(model.state_dict(), out_dir+"last.pt")
        if mean_iu_epoch/len(val_dataloader) == max(val_acc_list):
            torch.save(model.state_dict(), out_dir+"best.pt")
            print("save epoch {} model".format(epoch))
 
if __name__ == "__main__":
    main()

 整体训练流程没问题，读者可以根据需要更改其模型评价标准和相关代码。在本次训练中，主要使用Acc作为评价指标，其实就是分类正确的像素个数除以全部像素个数。最终训练结果如下：

0.8

训练集的Acc来到了 0.8, 验证集的Acc来到了0.77。由于有一些函数是复制过来的，如_hist等，所以其他指标暂时不参考。