deeplab v3+ 源码详解

训练模型：

** 下载好voc数据集，并传入所需的参数即可进行训练。**

参数配置：

"""
训练：
--model deeplabv3plus_mobilenet 
--gpu_id 0 
--year 2012_aug 
--crop_val 
--lr 0.01 
--crop_size 513 
--batch_size 4
--output_stride 16
测试：
--model deeplabv3plus_mobilenet 
--gpu_id 0 --year 2012_aug 
--crop_val 
--lr 0.01 
--crop_size 513 
--batch_size 16 
--output_stride 16 
--ckpt checkpoints/best_deeplabv3plus_mobilenet_voc_os16.pth 
--test_only 
--save_val_results
"""

1.数据预处理部分

    deeplab v3+默认使用voc数据集和cityspace数据集，图片预处理部分仅仅读取图片和对应的标签，同时对图片进行随机翻转、随机裁剪等常见图片预处理方式。

def get_dataset(opts):
    """ Dataset And Augmentation
    """
    if opts.dataset == 'voc':
        train_transform = et.ExtCompose([
            #et.ExtResize(size=opts.crop_size),
            et.ExtRandomScale((0.5, 2.0)),
            et.ExtRandomCrop(size=(opts.crop_size, opts.crop_size), pad_if_needed=True),
            et.ExtRandomHorizontalFlip(),   # 以给定的概率水平翻转给定的图像
            et.ExtToTensor(),
            et.ExtNormalize(mean=[0.485, 0.456, 0.406],
                            std=[0.229, 0.224, 0.225]),
        ])
        if opts.crop_val:
            val_transform = et.ExtCompose([
                et.ExtResize(opts.crop_size),
                et.ExtCenterCrop(opts.crop_size),
                et.ExtToTensor(),
                et.ExtNormalize(mean=[0.485, 0.456, 0.406],
                                std=[0.229, 0.224, 0.225]),
            ])
        else:
            val_transform = et.ExtCompose([
                et.ExtToTensor(),
                et.ExtNormalize(mean=[0.485, 0.456, 0.406],
                                std=[0.229, 0.224, 0.225]),
            ])
        # 读取数据
        train_dst = VOCSegmentation(root=opts.data_root, year=opts.year,
                                    image_set='train', download=opts.download, transform=train_transform)
        val_dst = VOCSegmentation(root=opts.data_root, year=opts.year,
                                  image_set='val', download=False, transform=val_transform)
    if opts.dataset == 'cityscapes':
        train_transform = et.ExtCompose([
            #et.ExtResize( 512 ),
            et.ExtRandomCrop(size=(opts.crop_size, opts.crop_size)),
            et.ExtColorJitter( brightness=0.5, contrast=0.5, saturation=0.5 ),
            et.ExtRandomHorizontalFlip(),
            et.ExtToTensor(),
            et.ExtNormalize(mean=[0.485, 0.456, 0.406],
                            std=[0.229, 0.224, 0.225]),
        ])
        val_transform = et.ExtCompose([
            #et.ExtResize( 512 ),
            et.ExtToTensor(),
            et.ExtNormalize(mean=[0.485, 0.456, 0.406],
                            std=[0.229, 0.224, 0.225]),
        ])
        train_dst = Cityscapes(root=opts.data_root,
                               split='train', transform=train_transform)
        val_dst = Cityscapes(root=opts.data_root,
                             split='val', transform=val_transform)
    return train_dst, val_dst

2.网络结构：

Ecoder部分

    使用resnet作为网络的ecoder部分，resnet作为图像分类模型，将图像下采样了32倍，特征图信息损失比较大，尤其是目标分割而言，无法再下采样32倍的特征图中恢复细节信息，因此，resnet的最后三层，将根据需要的特征图的大小，将下采样换为空洞卷积。输出layer4经过ASPP层的结果（下采样16倍或者8倍），同时也输出layer1的结果。
   ** ASPP层：**

** ** 如图所示，ASPP由不同空洞率的空洞卷积组成。以实现不同感受野的特征信息融合。此外，还有一个细节是空洞卷积的padding=空洞率，以保证输入输出特征图大小不改变。

class ASPPConv(nn.Sequential):
    def __init__(self, in_channels, out_channels, dilation):
        modules = [
            nn.Conv2d(in_channels, out_channels, 3, padding=dilation, dilation=dilation, bias=False),
            nn.BatchNorm2d(out_channels),
            nn.ReLU(inplace=True)
        ]
        super(ASPPConv, self).__init__(*modules)
class ASPPPooling(nn.Sequential):
    def __init__(self, in_channels, out_channels):
        super(ASPPPooling, self).__init__(
            nn.AdaptiveAvgPool2d(1),
            nn.Conv2d(in_channels, out_channels, 1, bias=False),
            nn.BatchNorm2d(out_channels),
            nn.ReLU(inplace=True))
    def forward(self, x):
        size = x.shape[-2:]
        x = super(ASPPPooling, self).forward(x)
        return F.interpolate(x, size=size, mode='bilinear', align_corners=False)
class ASPP(nn.Module):
    def __init__(self, in_channels, atrous_rates):
        super(ASPP, self).__init__()
        out_channels = 256
        modules = []
        modules.append(nn.Sequential(
            nn.Conv2d(in_channels, out_channels, 1, bias=False),
            nn.BatchNorm2d(out_channels),
            nn.ReLU(inplace=True)))
        rate1, rate2, rate3 = tuple(atrous_rates)
        modules.append(ASPPConv(in_channels, out_channels, rate1))
        modules.append(ASPPConv(in_channels, out_channels, rate2))
        modules.append(ASPPConv(in_channels, out_channels, rate3))
        modules.append(ASPPPooling(in_channels, out_channels))
        self.convs = nn.ModuleList(modules)
        self.project = nn.Sequential(
            nn.Conv2d(5 * out_channels, out_channels, 1, bias=False),
            nn.BatchNorm2d(out_channels),
            nn.ReLU(inplace=True),
            nn.Dropout(0.1),)
    def forward(self, x):
        res = []
        for conv in self.convs:
            #print(conv(x).shape)
            res.append(conv(x))
        res = torch.cat(res, dim=1)
        return self.project(res)

Decoder部分

    Layer4的输出首先会经过ASPP模块，然后经过1*1的卷积调整通道数至256，然后上采样至layer1输出结果的大小，将layer1的输出结果的通道数经过1*1的卷积调整至48，将这两个结果进行拼接。经过3*3的卷积后，再经过1*1的卷积对输出进行预测。

代码如下：

class DeepLabHeadV3Plus(nn.Module):
    def __init__(self, in_channels, low_level_channels, num_classes, aspp_dilate=[12, 24, 36]):
        super(DeepLabHeadV3Plus, self).__init__()
        self.project = nn.Sequential( 
            nn.Conv2d(low_level_channels, 48, 1, bias=False),
            nn.BatchNorm2d(48),
            nn.ReLU(inplace=True),
        )
        self.aspp = ASPP(in_channels, aspp_dilate)
        self.classifier = nn.Sequential(
            nn.Conv2d(304, 256, 3, padding=1, bias=False),
            nn.BatchNorm2d(256),
            nn.ReLU(inplace=True),
            nn.Conv2d(256, num_classes, 1)
        )
        self._init_weight()
    def forward(self, feature):
        #print(feature.shape)
        low_level_feature = self.project( feature['low_level'] )#return_layers = {'layer4': 'out', 'layer1': 'low_level'}
        #print(low_level_feature.shape)
        output_feature = self.aspp(feature['out'])
        #print(output_feature.shape)
        output_feature = F.interpolate(output_feature, size=low_level_feature.shape[2:], mode='bilinear', align_corners=False)
        #print(output_feature.shape)
        return self.classifier( torch.cat( [ low_level_feature, output_feature ], dim=1 ) )
    
    def _init_weight(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight)
            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

    dcoder的输出的特征图还是下采样4倍的结果，最终的输出需要继续进行双线性插值，调整到原特征图大小，最终代码如下：

class _SimpleSegmentationModel(nn.Module):
    def __init__(self, backbone, classifier):
        super(_SimpleSegmentationModel, self).__init__()
        self.backbone = backbone
        self.classifier = classifier
        
    def forward(self, x):
        input_shape = x.shape[-2:]
        features = self.backbone(x)
        x = self.classifier(features)
        x = F.interpolate(x, size=input_shape, mode='bilinear', align_corners=False)
        return x