HRNet网络代码解读：Deep High-Resolution Representation Learning for Human Pose Estimation

论文地址：https://arxiv.org/abs/1902.09212
官方源码：https://github.com/leoxiaobin/deep-high-resolution-net.pytorch

HRNet网络结构

这里引用“太阳花的小绿豆”绘制的一张基于HRNet-32模型的结构图。便于后续理解。

重要的部分写在代码注释里了，阅读的时候注意。

代码详解

forward函数

def get_pose_net(cfg, is_train, **kwargs):
    model = PoseHighResolutionNet(cfg, **kwargs)

    if is_train and cfg['MODEL']['INIT_WEIGHTS']:
        model.init_weights(cfg['MODEL']['PRETRAINED'])

    return model

使用了PoseHighresolutionNet类，让我们进入到这个类看一下。

首先看forward函数：

    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)
        x = self.layer1(x)

所对应的stem net为：

# stem net
        self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=2, padding=1,
                               bias=False)
        self.bn1 = nn.BatchNorm2d(64, momentum=BN_MOMENTUM)
        self.conv2 = nn.Conv2d(64, 64, kernel_size=3, stride=2, padding=1,
                               bias=False)
        self.bn2 = nn.BatchNorm2d(64, momentum=BN_MOMENTUM)
        self.relu = nn.ReLU(inplace=True)
        self.layer1 = self._make_layer(Bottleneck, 64, 4)

这里经过两个卷积bn激活函数的操作，后接一个layer1模块，特征通道数下采样4倍，通道变为256.

其中layer1由_make_layer(Bottleneck, 64, 4)构建。

layer1函数

让我们看下_make_layer的具体操作。

    def _make_layer(self, block, planes, blocks, stride=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
                ),
                nn.BatchNorm2d(planes * block.expansion, momentum=BN_MOMENTUM),
            )

        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample))
        # 通道数由64变为256
        self.inplanes = planes * block.expansion
        # self.inplanes = 64 * 4 = 256
        for i in range(1, blocks):
        	# 重复堆叠三次，不使用downsample，其实这里的downsample操作也并没有进行下采样。
        	# 输入通道数为256，输出通道数也为256
        	# 最后得到特征图的大小为下采样4倍，输出通道256的featuremap
            layers.append(block(self.inplanes, planes))

        return nn.Sequential(*layers)

Bottleneck类中expansion = 4, self.inplanes = 64 != 64 *4 执行downsample操作。注意这里downsample并没有对模型进行下采样，stride= 1，只是沿用了Resnet的名称，叫成了downsample。

Bottleneck的搭建如下代码：

class Bottleneck(nn.Module):
    expansion = 4

    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(Bottleneck, self).__init__()
        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
        self.bn1 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM)
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride,
                               padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM)
        self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=1,
                               bias=False)
        self.bn3 = nn.BatchNorm2d(planes * self.expansion,
                                  momentum=BN_MOMENTUM)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    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

其中Bottlenck的输入为inplanes，输出为4倍的planes。

配置文件

	STAGE2:
      NUM_MODULES: 1
      NUM_BRANCHES: 2
      BLOCK: BASIC
      NUM_BLOCKS:
      - 4
      - 4
      NUM_CHANNELS:
      - 32
      - 64
      FUSE_METHOD: SUM

forward函数

		x_list = []
		# NUM_BRANCHES = 2
        for i in range(self.stage2_cfg['NUM_BRANCHES']):
            if self.transition1[i] is not None:
                x_list.append(self.transition1[i](x))
            else:
                x_list.append(x)
        y_list = self.stage2(x_list)

分为Transition模块和Stage模块，其中Transition模块是为了进行下采样，并联不同下采样倍率的操作，Stage模块则是进行特征融合。由低下采样倍率和高下采样倍率的特征图融合在一起。

所对应的 __ init __ 方法里的代码如下：

		self.stage2_cfg = extra['STAGE2']
        num_channels = self.stage2_cfg['NUM_CHANNELS']
        # num_channels [32, 64]
        block = blocks_dict[self.stage2_cfg['BLOCK']]
        # basic block
        num_channels = [
            num_channels[i] * block.expansion for i in range(len(num_channels))
        ]
        # num_channels [32 * 1, 64 * 1]
        self.transition1 = self._make_transition_layer([256], num_channels)
        self.stage2, pre_stage_channels = self._make_stage(
            self.stage2_cfg, num_channels)

Transition函数

由_make_transision_layer函数定义

    def _make_transition_layer(
            self, num_channels_pre_layer, num_channels_cur_layer):
        # num_channels_pre_layer 之前layer层的channels个数，在stage2之前是256
        # num_channels_cur_layer 现在layer层的channels个数，在stage2为[32, 64]
        num_branches_cur = len(num_channels_cur_layer) # 2
        num_branches_pre = len(num_channels_pre_layer) # 1

        transition_layers = []
        # 对应图片上Transition1上的两层3 * 3卷积
        for i in range(num_branches_cur): # i = 0, 1
            if i < num_branches_pre:
                if num_channels_cur_layer[i] != num_channels_pre_layer[i]:
                # 如果通道数不相等，则通过卷积层改变通道数
                # 如果通道数相等，则无需卷积操作，可以直接使用，接到下面第一个else语句
                    transition_layers.append(
                        nn.Sequential(
                            nn.Conv2d(
                                num_channels_pre_layer[i],
                                num_channels_cur_layer[i],
                                3, 1, 1, bias=False
                            ),
                            nn.BatchNorm2d(num_channels_cur_layer[i]),
                            nn.ReLU(inplace=True)
                        )
                    )
                else:
                # 对应Transition2, 3中不进行卷积操作的分支
                    transition_layers.append(None)
            else:
            # 对应Transition模块上多出的那一个分支，使用stride = 2 再进行下采样
                conv3x3s = []
                for j in range(i+1-num_branches_pre):
                    # 利用num_channels_pre_layer之前shape最小的特征层来生成新的分支
                    inchannels = num_channels_pre_layer[-1]
                    outchannels = num_channels_cur_layer[i] \
                        if j == i-num_branches_pre else inchannels
                    conv3x3s.append(
                        nn.Sequential(
                            nn.Conv2d(
                                inchannels, outchannels, 3, 2, 1, bias=False
                            ),
                            # 这里卷积进行下采样，stride = 2
                            nn.BatchNorm2d(outchannels),
                            nn.ReLU(inplace=True)
                        )
                    )
                transition_layers.append(nn.Sequential(*conv3x3s))

        return nn.ModuleList(transition_layers)

Stage函数

由_make_stage函数定义：

    def _make_stage(self, layer_config, num_inchannels,
                    multi_scale_output=True):
        num_modules = layer_config['NUM_MODULES'] # 1
        num_branches = layer_config['NUM_BRANCHES'] # 2
        num_blocks = layer_config['NUM_BLOCKS'] # [4, 4]
        num_channels = layer_config['NUM_CHANNELS'] # [32, 64]
        block = blocks_dict[layer_config['BLOCK']] # BasicBlock
        fuse_method = layer_config['FUSE_METHOD'] # SUM

        modules = []
        # num_modules 表示一个stage中融合进行几次
        # 最后一次融合是将其他分支的特征融合到最高分辨率的特征图上,只输出最高分辨率的特征图(multi_scale_output = False)
        # 前几次融合是将所有分支的特征融合到每个特征图上，输出所有尺寸特征图(multi_scale_output = True)
        for i in range(num_modules):
            # multi_scale_output is only used last module
            if not multi_scale_output and i == num_modules - 1:
                reset_multi_scale_output = False
            else:
                reset_multi_scale_output = True

            modules.append(
                HighResolutionModule(
                    num_branches,
                    block,
                    num_blocks,
                    num_inchannels,
                    num_channels,
                    fuse_method,
                    reset_multi_scale_output
                )
            )
            num_inchannels = modules[-1].get_num_inchannels()

        return nn.Sequential(*modules), num_inchannels

HighResolutionModule

forward函数

    def forward(self, x):
        if self.num_branches == 1:
        # 如果只有一个分支，则直接将单个分支特征图作为输入送进self.branches
            return [self.branches[0](x[0])]
        # 如果有多个分支，则分别将每个分支特征图作为输入送进self.branches[i]，得到x[i]
        for i in range(self.num_branches):
            x[i] = self.branches[i](x[i])

        x_fuse = []
		# 把不同分支分别进行上采样和下采样然后融合
        for i in range(len(self.fuse_layers)):
            y = x[0] if i == 0 else self.fuse_layers[i][0](x[0])
            for j in range(1, self.num_branches):
                if i == j:
                    y = y + x[j]
                else:
                    y = y + self.fuse_layers[i][j](x[j])
            # 整体部分最后加Relu激活函数
            x_fuse.append(self.relu(y))

        return x_fuse

__ init __ 函数

    def __init__(self, num_branches, blocks, num_blocks, num_inchannels,
                 num_channels, fuse_method, multi_scale_output=True):
        super(HighResolutionModule, self).__init__()
        self._check_branches(
            num_branches, blocks, num_blocks, num_inchannels, num_channels)

        self.num_inchannels = num_inchannels
        self.fuse_method = fuse_method
        self.num_branches = num_branches

        self.multi_scale_output = multi_scale_output

        self.branches = self._make_branches(
            num_branches, blocks, num_blocks, num_channels)
        self.fuse_layers = self._make_fuse_layers()
        self.relu = nn.ReLU(True)

_make_branches函数

    def _make_branches(self, num_branches, block, num_blocks, num_channels):
        branches = []
		# 反复堆叠_make_one_branch，重复num_branches次数
        for i in range(num_branches):
            branches.append(
                self._make_one_branch(i, block, num_blocks, num_channels)
            )

        return nn.ModuleList(branches)

_make_one_branch函数

    def _make_one_branch(self, branch_index, block, num_blocks, num_channels,
                         stride=1):
        downsample = None
        if stride != 1 or \
           self.num_inchannels[branch_index] != num_channels[branch_index] * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(
                    self.num_inchannels[branch_index],
                    num_channels[branch_index] * block.expansion,
                    kernel_size=1, stride=stride, bias=False
                ),
                nn.BatchNorm2d(
                    num_channels[branch_index] * block.expansion,
                    momentum=BN_MOMENTUM
                ),
            )

        layers = []
        # 第一个layer接入downsample，但是这里不会进行下采样，堆叠一次basicblock
        layers.append(
            block(
                self.num_inchannels[branch_index],
                num_channels[branch_index],
                stride,
                downsample
            )
        )
        # 通道数[32, 64]
        self.num_inchannels[branch_index] = \
            num_channels[branch_index] * block.expansion
        # num_blocks 为 [4, 4]，所以循环次数为3，重复堆叠basicblock
        for i in range(1, num_blocks[branch_index]):
            layers.append(
                block(
                    self.num_inchannels[branch_index],
                    num_channels[branch_index]
                )
            )

        return nn.Sequential(*layers)

以上这部分为堆叠BasicBlock四次，对应图中Stage2中左侧的部分。

_make_fuse_layers函数

    def _make_fuse_layers(self):
        if self.num_branches == 1:
            return None

        num_branches = self.num_branches # 2
        num_inchannels = self.num_inchannels # [32, 64]
        fuse_layers = []
        # 把j分支的特征融入到i分支中。
        for i in range(num_branches if self.multi_scale_output else 1):
            fuse_layer = []
            
            for j in range(num_branches):
                if j > i:
                # 如果j分支大于i分支，则说明j下采样倍率更高，需要进行上采样与i分支融合。
                    fuse_layer.append(
                        nn.Sequential(
                            nn.Conv2d(
                                num_inchannels[j],
                                num_inchannels[i],
                                1, 1, 0, bias=False
                            ),
                            nn.BatchNorm2d(num_inchannels[i]),
                            nn.Upsample(scale_factor=2**(j-i), mode='nearest')
                        )
                    )
                elif j == i:
                # j分支等于i分支，不需要进行操作
                    fuse_layer.append(None)
                else:
                # j分支大于i分支，需要进行下采样，这里stride = 2
                # 判断k是否是最后一层，不是最后一层需要加Relu激活函数，最后一层则不需要添加
                    conv3x3s = []
                    for k in range(i-j):
                        if k == i - j - 1:
                            num_outchannels_conv3x3 = num_inchannels[i]
                            conv3x3s.append(
                                nn.Sequential(
                                    nn.Conv2d(
                                        num_inchannels[j],
                                        num_outchannels_conv3x3,
                                        3, 2, 1, bias=False
                                    ),
                                    nn.BatchNorm2d(num_outchannels_conv3x3)
                                )
                            )
                        else:
                            num_outchannels_conv3x3 = num_inchannels[j]
                            conv3x3s.append(
                                nn.Sequential(
                                    nn.Conv2d(
                                        num_inchannels[j],
                                        num_outchannels_conv3x3,
                                        3, 2, 1, bias=False
                                    ),
                                    nn.BatchNorm2d(num_outchannels_conv3x3),
                                    nn.ReLU(True)
                                )
                            )
                    fuse_layer.append(nn.Sequential(*conv3x3s))
            fuse_layers.append(nn.ModuleList(fuse_layer))

        return nn.ModuleList(fuse_layers)

FinalLayer

		self.final_layer = nn.Conv2d(
            in_channels=pre_stage_channels[0],
            out_channels=cfg['MODEL']['NUM_JOINTS'],
            kernel_size=extra['FINAL_CONV_KERNEL'],
            stride=1,
            padding=1 if extra['FINAL_CONV_KERNEL'] == 3 else 0
        )

Conv2d的参数为k = 1, s = 1, p = 0, out_channels = 17对应17个关键点。

后记

其中关键的部分已经再代码中以注释的形式展现，请认真读注释。

另外只介绍了Stage2的部分，Stage3，4堆叠策略同上，就不再赘述了。