DeepLabV3+ 基本原理及Pytorch版注解

1.原理图

在原理图中，整个执行过程如下：
（1）一张图片A，送进改造过后的主流深度卷积网络B（DCNN,加入了一个空洞卷积Atrous Conv）提取特征，得到高级语义特征C和低级语义特征G。

（2）高级语义特征C进入到空洞金字塔池化模块ASPP，分别与四个空洞卷积层和一个池化层进行卷积和池化，得到五个特征图，然后连接成五层D。D再通地一个1*1的卷积进行运算后得到E; E再经过上采样得到F。

（3）通过在深度卷积网络层找到一个与F分辨率相同的低级语义特征图G；经过1*1卷积进行降通道数使之与F所占通道比重一样，更有利于模型学习

（4）合并成H，然后再通过一个3*3细化卷积进行细化；后通过双线性上采样4倍，得到预测结果。

具体细节，请参考下面pytorch版本的源码。

2.pytorch版本的源码

注意的地方：

A.用到的sync_batchnorm.batchnorm比归一化包，pytorch版本的可以在https://github.com/acgtyrant/Synchronized-BatchNorm-PyTorch下载

B.在python2.7中，引用的队列是大写import Queue

import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.model_zoo as model_zoo
from sync_batchnorm.batchnorm import SynchronizedBatchNorm2d
 
BatchNorm2d = SynchronizedBatchNorm2d
import cv2
import matplotlib.pyplot as plt
 
class Bottleneck(nn.Module):#'resnet网络的基本框架’
    expansion = 4
    def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None):
        super(Bottleneck, self).__init__()
        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
        self.bn1 = BatchNorm2d(planes)
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride,dilation=dilation, padding=dilation, bias=False)
        self.bn2 = BatchNorm2d(planes)
        self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False)
        self.bn3 = BatchNorm2d(planes * 4)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride
        self.dilation = dilation
 
    def forward(self, x):
        residual = x
        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)
        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)
        out = self.conv3(out)
        out = self.bn3(out)
        if self.downsample is not None:
            residual = self.downsample(x)
        out += residual
        out = self.relu(out)
        return out
 
 
class ResNet(nn.Module):#renet网络的构成部分
    def __init__(self, nInputChannels, block, layers, os=16, pretrained=False):
        self.inplanes = 64
        super(ResNet, self).__init__()
        if os == 16:
            strides = [1, 2, 2, 1]
            dilations = [1, 1, 1, 2]
            blocks = [1, 2, 4]
        elif os == 8:
            strides = [1, 2, 1, 1]
            dilations = [1, 1, 2, 2]
            blocks = [1, 2, 1]
        else:
            raise NotImplementedError
        # Modules
        self.conv1 = nn.Conv2d(nInputChannels, 64, kernel_size=7, stride=2, padding=3,bias=False)
        self.bn1 = BatchNorm2d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
 
        self.layer1 = self._make_layer(block, 64, layers[0], stride=strides[0], dilation=dilations[0])
        self.layer2 = self._make_layer(block, 128, layers[1], stride=strides[1], dilation=dilations[1])
        self.layer3 = self._make_layer(block, 256, layers[2], stride=strides[2], dilation=dilations[2])
        self.layer4 = self._make_MG_unit(block, 512, blocks=blocks, stride=strides[3], dilation=dilations[3])
        self._init_weight()
 
 
        if pretrained:
            self._load_pretrained_model()
 
    def _make_layer(self, block, planes, blocks, stride=1, dilation=1):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False),
                BatchNorm2d(planes * block.expansion),
            )
 
        layers = []
        layers.append(block(self.inplanes, planes, stride, dilation, downsample))
        self.inplanes = planes * block.expansion
        for i in range(1, blocks):
            layers.append(block(self.inplanes, planes))
 
        return nn.Sequential(*layers)
 
 
 
    def _make_MG_unit(self, block, planes, blocks=[1, 2, 4], stride=1, dilation=1):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.inplanes, planes * block.expansion,
                          kernel_size=1, stride=stride, bias=False),
                BatchNorm2d(planes * block.expansion),
            )
 
        layers = []
        layers.append(block(self.inplanes, planes, stride, dilation=blocks[0]*dilation, downsample=downsample))
        self.inplanes = planes * block.expansion
        for i in range(1, len(blocks)):
            layers.append(block(self.inplanes, planes, stride=1, dilation=blocks[i]*dilation))
 
        return nn.Sequential(*layers)
 
 
    def forward(self, input):
        x = self.conv1(input)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)
        x = self.layer1(x)
        low_level_feat = x
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        return x, low_level_feat
 
    def _init_weight(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
            elif isinstance(m, BatchNorm2d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()
 
 
 
    def _load_pretrained_model(self):
        pretrain_dict = model_zoo.load_url('https://download.pytorch.org/models/resnet101-5d3b4d8f.pth')
        model_dict = {}
        state_dict = self.state_dict()
        for k, v in pretrain_dict.items():
            if k in state_dict:
                model_dict[k] = v
        state_dict.update(model_dict)
        self.load_state_dict(state_dict)
 
 
def ResNet101(nInputChannels=3, os=16, pretrained=False):
    model = ResNet(nInputChannels, Bottleneck, [3, 4, 23, 3], os, pretrained=pretrained)
    return model
 
 
 
class ASPP_module(nn.Module):#ASpp模块的组成
    def __init__(self, inplanes, planes, dilation):
        super(ASPP_module, self).__init__()
        if dilation == 1:
            kernel_size = 1
            padding = 0
        else:
            kernel_size = 3
            padding = dilation
        self.atrous_convolution = nn.Conv2d(inplanes, planes, kernel_size=kernel_size,
                                            stride=1, padding=padding, dilation=dilation, bias=False)
        self.bn = BatchNorm2d(planes)
        self.relu = nn.ReLU()
        self._init_weight()
 
    def forward(self, x):
        x = self.atrous_convolution(x)
        x = self.bn(x)
        return self.relu(x)
 
    def _init_weight(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
            elif isinstance(m, BatchNorm2d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()
 
#正式开始deeplabv3+的结构组成
class DeepLabv3_plus(nn.Module):
    def __init__(self, nInputChannels=3, n_classes=21, os=16, pretrained=False, freeze_bn=False, _print=True):
        if _print:
            print("Constructing DeepLabv3+ model...")
            print("Backbone: Resnet-101")
            print("Number of classes: {}".format(n_classes))
            print("Output stride: {}".format(os))
            print("Number of Input Channels: {}".format(nInputChannels))
        super(DeepLabv3_plus, self).__init__()
 
        # Atrous Conv  首先获得从resnet101中提取的features map
        self.resnet_features = ResNet101(nInputChannels, os, pretrained=pretrained)
 
        # ASPP,挑选参数
        if os == 16:
            dilations = [1, 6, 12, 18]
        elif os == 8:
            dilations = [1, 12, 24, 36]
        else:
            raise NotImplementedError
#四个不同带洞卷积的设置，获取不同感受野
        self.aspp1 = ASPP_module(2048, 256, dilation=dilations[0])
        self.aspp2 = ASPP_module(2048, 256, dilation=dilations[1])
        self.aspp3 = ASPP_module(2048, 256, dilation=dilations[2])
        self.aspp4 = ASPP_module(2048, 256, dilation=dilations[3])
        self.relu = nn.ReLU()
 
#全局平均池化层的设置
        self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
                                             nn.Conv2d(2048, 256, 1, stride=1, bias=False),
                                             BatchNorm2d(256),
                                             nn.ReLU())
 
 
 
        self.conv1 = nn.Conv2d(1280, 256, 1, bias=False)
        self.bn1 = BatchNorm2d(256)
 
        # adopt [1x1, 48] for channel reduction.
        self.conv2 = nn.Conv2d(256, 48, 1, bias=False)
        self.bn2 = BatchNorm2d(48)
#结构图中的解码部分的最后一个3*3的卷积块
        self.last_conv = nn.Sequential(nn.Conv2d(304, 256, kernel_size=3, stride=1, padding=1, bias=False),
                                       BatchNorm2d(256),
                                       nn.ReLU(),
                                       nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1, bias=False),
                                       BatchNorm2d(256),
                                       nn.ReLU(),
                                       nn.Conv2d(256, n_classes, kernel_size=1, stride=1))
 
        if freeze_bn:
            self._freeze_bn()
 
#前向传播
 
    def forward(self, input):
        x, low_level_features = self.resnet_features(input)
        x1 = self.aspp1(x)
        x2 = self.aspp2(x)
        x3 = self.aspp3(x)
        x4 = self.aspp4(x)
        x5 = self.global_avg_pool(x)
        x5 = F.upsample(x5, size=x4.size()[2:], mode='bilinear', align_corners=True)
        #把四个ASPP模块以及全局池化层拼接起来
        x = torch.cat((x1, x2, x3, x4, x5), dim=1)
        #上采样
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = F.upsample(x, size=(int(math.ceil(input.size()[-2]/4)),
                                int(math.ceil(input.size()[-1]/4))), mode='bilinear', align_corners=True)
 
        low_level_features = self.conv2(low_level_features)
        low_level_features = self.bn2(low_level_features)
        low_level_features = self.relu(low_level_features)
 
 
        #拼接低层次的特征，然后再通过插值获取原图大小的结果
        x = torch.cat((x, low_level_features), dim=1)
        x = self.last_conv(x)
        #实现插值和上采样
        x = F.interpolate(x, size=input.size()[2:], mode='bilinear', align_corners=True)
        return x
 
    def _freeze_bn(self):
        for m in self.modules():
            if isinstance(m, BatchNorm2d):
                m.eval()
 
 
 
    def _init_weight(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
            elif isinstance(m, BatchNorm2d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()
 
 
 
def get_1x_lr_params(model):
    """
    This generator returns all the parameters of the net except for
    the last classification layer. Note that for each batchnorm layer,
    requires_grad is set to False in deeplab_resnet.py, therefore this function does not return
    any batchnorm parameter
    """
    b = [model.resnet_features]
    for i in range(len(b)):
        for k in b[i].parameters():
            if k.requires_grad:
                yield k
 
 
 
def get_10x_lr_params(model):
    """
    This generator returns all the parameters for the last layer of the net,
    which does the classification of pixel into classes
    """
    b = [model.aspp1, model.aspp2, model.aspp3, model.aspp4, model.conv1, model.conv2, model.last_conv]
    for j in range(len(b)):
        for k in b[j].parameters():
            if k.requires_grad:
                yield k
 
 
 
 
 
if __name__ == "__main__":
    model = DeepLabv3_plus(nInputChannels=3, n_classes=21, os=16, pretrained=True, _print=True)
    model.eval()
    image = torch.randn(1, 3, 512, 512)
    
    with torch.no_grad():
        output = model.forward(image)
    print(output.size())