项目实战 DeepLabV1,V2,V3 Google三大语义分割算法源码解析
- 2019 年 12 月 9 日
- 筆記
前言
算法和工程是算法工程师不可缺少的两种能力，之前我介绍了DeepLab V1，V2, V3，但总是感觉少了点什么？只有Paper，没有源码那不相当于是纸上谈兵了，所以今天尝试结合论文的源码来进行仔细的分析这三个算法。等我们分析清楚这三个算法之后，有机会再解析一下DeepLabV3。由于最近正在看Pytorch版本的《动手学深度学习》，不妨用Pytorch的源码来进行分析。我分析的源码均来自这个Pytorch工程：https://github.com/kazuto1011/deeplab-pytorch/tree/master/libs/models
DeepLab V1源码分析
DeepLab V1的算法原理可以看我之前的推文，地址是：https://mp.weixin.qq.com/s/rvP8-Y-CRuq4HFzR0qJWcg 。我们今天解析的DeepLab系列网络模型是在ResNet残差模块的基础上配合空洞卷积实现的，对于DeepLab V1, 第一层为普通卷积，stride = 2，紧跟着 stride = 2 的 max-pooling，然后一个普通的 bottleneck ，一个 stride = 2 的 bottleneck，然后 dilation =2、dilation =4 的bottleneck。
from __future__ import absolute_import, print_function import torch import torch.nn as nn import torch.nn.functional as F # 定义DeepLabV1的网络结构 class DeepLabV1(nn.Sequential): """ DeepLab v1: Dilated ResNet + 1x1 Conv Note that this is just a container for loading the pretrained COCO model and not mentioned as "v1" in papers. """ def __init__(self, n_classes, n_blocks): super(DeepLabV1, self).__init__() ch = [64 * 2 ** p for p in range(6)] self.add_module("layer1", _Stem(ch[0])) self.add_module("layer2", _ResLayer(n_blocks[0], ch[0], ch[2], 1, 1)) self.add_module("layer3", _ResLayer(n_blocks[1], ch[2], ch[3], 2, 1)) self.add_module("layer4", _ResLayer(n_blocks[2], ch[3], ch[4], 1, 2)) self.add_module("layer5", _ResLayer(n_blocks[3], ch[4], ch[5], 1, 4)) self.add_module("fc", nn.Conv2d(2048, n_classes, 1)) # 这里是看一下是使用torch的nn模块中BatchNorm还是在encoding文件中定义的BatchNorm try: from encoding.nn import SyncBatchNorm _BATCH_NORM = SyncBatchNorm except: _BATCH_NORM = nn.BatchNorm2d _BOTTLENECK_EXPANSION = 4 # 定义卷积+BN+ReLU的组件 class _ConvBnReLU(nn.Sequential): """ Cascade of 2D convolution, batch norm, and ReLU. """ BATCH_NORM = _BATCH_NORM def __init__( self, in_ch, out_ch, kernel_size, stride, padding, dilation, relu=True ): super(_ConvBnReLU, self).__init__() self.add_module( "conv", nn.Conv2d( in_ch, out_ch, kernel_size, stride, padding, dilation, bias=False ), ) self.add_module("bn", _BATCH_NORM(out_ch, eps=1e-5, momentum=0.999)) if relu: self.add_module("relu", nn.ReLU()) # 定义Bottleneck，先1*1卷积降维，然后使用3*3卷积，最后再1*1卷积升维，然后再shortcut连接。 # 降维到多少是由_BOTTLENECK_EXPANSION参数决定的，这是ResNet的Bottleneck。 class _Bottleneck(nn.Module): """ Bottleneck block of MSRA ResNet. """ def __init__(self, in_ch, out_ch, stride, dilation, downsample): super(_Bottleneck, self).__init__() mid_ch = out_ch // _BOTTLENECK_EXPANSION self.reduce = _ConvBnReLU(in_ch, mid_ch, 1, stride, 0, 1, True) self.conv3x3 = _ConvBnReLU(mid_ch, mid_ch, 3, 1, dilation, dilation, True) self.increase = _ConvBnReLU(mid_ch, out_ch, 1, 1, 0, 1, False) self.shortcut = ( _ConvBnReLU(in_ch, out_ch, 1, stride, 0, 1, False) if downsample else lambda x: x # identity ) def forward(self, x): h = self.reduce(x) h = self.conv3x3(h) h = self.increase(h) h += self.shortcut(x) return F.relu(h) # 定义ResLayer，整个DeepLabv1是用ResLayer堆叠起来的，下采样是在每个ResLayer的第一个 # Bottleneck发生的。 class _ResLayer(nn.Sequential): """ Residual layer with multi grids """ def __init__(self, n_layers, in_ch, out_ch, stride, dilation, multi_grids=None): super(_ResLayer, self).__init__() if multi_grids is None: multi_grids = [1 for _ in range(n_layers)] else: assert n_layers == len(multi_grids) # Downsampling is only in the first block for i in range(n_layers): self.add_module( "block{}".format(i + 1), _Bottleneck( in_ch=(in_ch if i == 0 else out_ch), out_ch=out_ch, stride=(stride if i == 0 else 1), dilation=dilation * multi_grids[i], downsample=(True if i == 0 else False), ), ) # 在进入ResLayer之前，先用7*7的卷积核在原图滑动，增大感受野。padding方式设为same，大小不变。 # Pool层的核大小为3，步长为2，这会导致特征图的分辨率发生变化。 class _Stem(nn.Sequential): """ The 1st conv layer. Note that the max pooling is different from both MSRA and FAIR ResNet. """ def __init__(self, out_ch): super(_Stem, self).__init__() self.add_module("conv1", _ConvBnReLU(3, out_ch, 7, 2, 3, 1)) self.add_module("pool", nn.MaxPool2d(3, 2, 1, ceil_mode=True)) # 相当于Reshape，网络并没有用到 class _Flatten(nn.Module): def forward(self, x): return x.view(x.size(0), -1) # 主函数，输出构建的DeepLab V1模型的结构还有原始图像分辨率和结果图像的分辨率 if __name__ == "__main__": model = DeepLabV1(n_classes=21, n_blocks=[3, 4, 23, 3]) #model.eval() image = torch.randn(1, 3, 513, 513) print(model) print("input:", image.shape) print("output:", model(image).shape)
我们看一下网络的输入和输出特征图尺寸：
input: torch.Size([1, 3, 513, 513]) output: torch.Size([1, 21, 65, 65])
网络结构已经非常清晰了，可以直接运行Python代码打印出网络结构或者按照我的源码注释来理解。注意，训练的时候ground truth要resize到和模型的输出特征图尺寸一样大才可以。
DeepLab V2源码分析
DeepLab V2的论文解读请看我前面发的文章：https://mp.weixin.qq.com/s/ylv3QfOe_BOuVuxQTd_m_g 。简单的说，DeepLab V2就是DeepLab V1的基础上加了一个ASPP模块，这是一个类似于Inception模块的结构，包含不同膨胀系数的空洞卷积，增强模型识别同一物体的多尺度能力。这里仍然只分析源码：为了方便理解把上篇文章中的ASPP模块的示意图放在这里：
在这里插入图片描述
from __future__ import absolute_import, print_function import torch import torch.nn as nn import torch.nn.functional as F # 定义ASPP模块，这是DeepLab V2和V1的主要区别，可以看到其他部分和V1的代码一模一样 class _ASPP(nn.Module): """ Atrous spatial pyramid pooling (ASPP) """ def __init__(self, in_ch, out_ch, rates): super(_ASPP, self).__init__() for i, rate in enumerate(rates): self.add_module( "c{}".format(i), nn.Conv2d(in_ch, out_ch, 3, 1, padding=rate, dilation=rate, bias=True), ) for m in self.children(): nn.init.normal_(m.weight, mean=0, std=0.01) nn.init.constant_(m.bias, 0) def forward(self, x): return sum([stage(x) for stage in self.children()]) class DeepLabV2(nn.Sequential): """ DeepLab v2: Dilated ResNet + ASPP Output stride is fixed at 8 """ def __init__(self, n_classes, n_blocks, atrous_rates): super(DeepLabV2, self).__init__() ch = [64 * 2 ** p for p in range(6)] self.add_module("layer1", _Stem(ch[0])) self.add_module("layer2", _ResLayer(n_blocks[0], ch[0], ch[2], 1, 1)) self.add_module("layer3", _ResLayer(n_blocks[1], ch[2], ch[3], 2, 1)) self.add_module("layer4", _ResLayer(n_blocks[2], ch[3], ch[4], 1, 2)) self.add_module("layer5", _ResLayer(n_blocks[3], ch[4], ch[5], 1, 4)) self.add_module("aspp", _ASPP(ch[5], n_classes, atrous_rates)) def freeze_bn(self): for m in self.modules(): if isinstance(m, _ConvBnReLU.BATCH_NORM): m.eval() try: from encoding.nn import SyncBatchNorm _BATCH_NORM = SyncBatchNorm except: _BATCH_NORM = nn.BatchNorm2d _BOTTLENECK_EXPANSION = 4 class _ConvBnReLU(nn.Sequential): """ Cascade of 2D convolution, batch norm, and ReLU. """ BATCH_NORM = _BATCH_NORM def __init__( self, in_ch, out_ch, kernel_size, stride, padding, dilation, relu=True ): super(_ConvBnReLU, self).__init__() self.add_module( "conv", nn.Conv2d( in_ch, out_ch, kernel_size, stride, padding, dilation, bias=False ), ) self.add_module("bn", _BATCH_NORM(out_ch, eps=1e-5, momentum=0.999)) if relu: self.add_module("relu", nn.ReLU()) class _Bottleneck(nn.Module): """ Bottleneck block of MSRA ResNet. """ def __init__(self, in_ch, out_ch, stride, dilation, downsample): super(_Bottleneck, self).__init__() mid_ch = out_ch // _BOTTLENECK_EXPANSION self.reduce = _ConvBnReLU(in_ch, mid_ch, 1, stride, 0, 1, True) self.conv3x3 = _ConvBnReLU(mid_ch, mid_ch, 3, 1, dilation, dilation, True) self.increase = _ConvBnReLU(mid_ch, out_ch, 1, 1, 0, 1, False) self.shortcut = ( _ConvBnReLU(in_ch, out_ch, 1, stride, 0, 1, False) if downsample else lambda x: x # identity ) def forward(self, x): h = self.reduce(x) h = self.conv3x3(h) h = self.increase(h) h += self.shortcut(x) return F.relu(h) class _ResLayer(nn.Sequential): """ Residual layer with multi grids """ def __init__(self, n_layers, in_ch, out_ch, stride, dilation, multi_grids=None): super(_ResLayer, self).__init__() if multi_grids is None: multi_grids = [1 for _ in range(n_layers)] else: assert n_layers == len(multi_grids) # Downsampling is only in the first block for i in range(n_layers): self.add_module( "block{}".format(i + 1), _Bottleneck( in_ch=(in_ch if i == 0 else out_ch), out_ch=out_ch, stride=(stride if i == 0 else 1), dilation=dilation * multi_grids[i], downsample=(True if i == 0 else False), ), ) class _Stem(nn.Sequential): """ The 1st conv layer. Note that the max pooling is different from both MSRA and FAIR ResNet. """ def __init__(self, out_ch): super(_Stem, self).__init__() self.add_module("conv1", _ConvBnReLU(3, out_ch, 7, 2, 3, 1)) self.add_module("pool", nn.MaxPool2d(3, 2, 1, ceil_mode=True)) if __name__ == "__main__": model = DeepLabV2( n_classes=21, n_blocks=[3, 4, 23, 3], atrous_rates=[6, 12, 18, 24] ) model.eval() image = torch.randn(1, 3, 513, 513) print(model) print("input:", image.shape) print("output:", model(image).shape)
可以看到DeepLab V2的代码除了ASPP模块，其他部分和V1完全一样，所以就没什么好解释的了。但需要注意的一个点是，训练的时候，DeepLabV2的学习率采用了Poly的策略，公式为： ，当时，模型可以取得不普通的分段学习策略MAP值高1.17%的效果。这部分作者也在他的代码中实现了，如下所示：
作者：Uno Whoiam 链接：https://zhuanlan.zhihu.com/p/68531147 来源：知乎 著作权归作者所有。商业转载请联系作者获得授权，非商业转载请注明出处。 from torch.optim.lr_scheduler import _LRScheduler class PolynomialLR(_LRScheduler): def __init__(self, optimizer, step_size, iter_max, power, last_epoch=-1): self.step_size = step_size self.iter_max = iter_max self.power = power super(PolynomialLR, self).__init__(optimizer, last_epoch) def polynomial_decay(self, lr): return lr * (1 - float(self.last_epoch) / self.iter_max) ** self.power def get_lr(self): if ( (self.last_epoch == 0) or (self.last_epoch % self.step_size != 0) or (self.last_epoch > self.iter_max) ): return [group["lr"] for group in self.optimizer.param_groups] return [self.polynomial_decay(lr) for lr in self.base_lrs]
可以看到这个类是直接继承了Pytorch中的学习率调整类_LRScheduler，可以方便的在每个epoch进行学习率调整。
最后网络的输入分辨率和输出分辨率和DeepLab V1一样，具体训练和数据制作请看作者的github工程：https://github.com/kazuto1011/deeplab-pytorch/tree/master/libs/models 。
DeepLab V3源码分析
DeepLab V3论文原理请看我之前发的推文：https://mp.weixin.qq.com/s/D9OX89mklaU4tv74OZMqNg 。这里再简单回归一下DeepLab V3使用的关键Trick。
- 将BN层加到了ASPP模块中。
- 使用了Multi-Grid策略，即在模型后端多加几层不同rate的空洞卷积。
- 具有不同 atrous rates 的 ASPP 能够有效的捕获多尺度信息。不过，论文发现，随着sampling rate的增加，有效filter特征权重(即有效特征区域，而不是补零区域的权重)的数量会变小，极端情况下，当空洞卷积的 rate 和 feature map 的大小一致时，卷积会退化为卷积。为了解决这一问题，并将全局内容信息整合到模型中，则采用图像级特征。即，采用全局平均池化(global average pooling)对模型的 feature map 进行处理，将得到的图像级特征输入到一个 1×1 convolution with 256 filters(加入 batch normalization)中，然后将特征进行双线性上采样(bilinearly upsample)到特定的空间维度。
DeepLab V3的源码如下：
from __future__ import absolute_import, print_function from collections import OrderedDict import torch import torch.nn as nn import torch.nn.functional as F # 全局平均池化，将得到的图像特征输入到一个拥有256个通道的1*1卷积中，最后将特征进行 # 双线性上采样到特定的维度(就是输入到ImagePool之前特征图的维度) class _ImagePool(nn.Module): def __init__(self, in_ch, out_ch): super().__init__() self.pool = nn.AdaptiveAvgPool2d(1) self.conv = _ConvBnReLU(in_ch, out_ch, 1, 1, 0, 1) def forward(self, x): _, _, H, W = x.shape h = self.pool(x) h = self.conv(h) h = F.interpolate(h, size=(H, W), mode="bilinear", align_corners=False) return h # ASPP模块，DeepLabV3改进后的，新增了1*1卷积以及图像全局池化。 class _ASPP(nn.Module): """ Atrous spatial pyramid pooling with image-level feature """ def __init__(self, in_ch, out_ch, rates): super(_ASPP, self).__init__() self.stages = nn.Module() self.stages.add_module("c0", _ConvBnReLU(in_ch, out_ch, 1, 1, 0, 1)) for i, rate in enumerate(rates): self.stages.add_module( "c{}".format(i + 1), _ConvBnReLU(in_ch, out_ch, 3, 1, padding=rate, dilation=rate), ) self.stages.add_module("imagepool", _ImagePool(in_ch, out_ch)) def forward(self, x): return torch.cat([stage(x) for stage in self.stages.children()], dim=1) # 完整的DeepLabV3的结构，使用带空洞卷积的ResNet+multi-grid策略+改进后的ASPP class DeepLabV3(nn.Sequential): """ DeepLab v3: Dilated ResNet with multi-grid + improved ASPP """ def __init__(self, n_classes, n_blocks, atrous_rates, multi_grids, output_stride): super(DeepLabV3, self).__init__() # Stride and dilation if output_stride == 8: s = [1, 2, 1, 1] d = [1, 1, 2, 4] elif output_stride == 16: s = [1, 2, 2, 1] d = [1, 1, 1, 2] ch = [64 * 2 ** p for p in range(6)] self.add_module("layer1", _Stem(ch[0])) self.add_module("layer2", _ResLayer(n_blocks[0], ch[0], ch[2], s[0], d[0])) self.add_module("layer3", _ResLayer(n_blocks[1], ch[2], ch[3], s[1], d[1])) self.add_module("layer4", _ResLayer(n_blocks[2], ch[3], ch[4], s[2], d[2])) self.add_module( "layer5", _ResLayer(n_blocks[3], ch[4], ch[5], s[3], d[3], multi_grids) ) self.add_module("aspp", _ASPP(ch[5], 256, atrous_rates)) # 连接所有分支的最终特征，输入到256个通道的1*1卷积中，并加入BN，再进入最终的1*1卷积， # 得到logits结果。 concat_ch = 256 * (len(atrous_rates) + 2) self.add_module("fc1", _ConvBnReLU(concat_ch, 256, 1, 1, 0, 1)) self.add_module("fc2", nn.Conv2d(256, n_classes, kernel_size=1)) try: from encoding.nn import SyncBatchNorm _BATCH_NORM = SyncBatchNorm except: _BATCH_NORM = nn.BatchNorm2d _BOTTLENECK_EXPANSION = 4 # 和DeepLabV1定义一样 class _ConvBnReLU(nn.Sequential): """ Cascade of 2D convolution, batch norm, and ReLU. """ BATCH_NORM = _BATCH_NORM def __init__( self, in_ch, out_ch, kernel_size, stride, padding, dilation, relu=True ): super(_ConvBnReLU, self).__init__() self.add_module( "conv", nn.Conv2d( in_ch, out_ch, kernel_size, stride, padding, dilation, bias=False ), ) self.add_module("bn", _BATCH_NORM(out_ch, eps=1e-5, momentum=0.999)) if relu: self.add_module("relu", nn.ReLU()) class _Bottleneck(nn.Module): """ Bottleneck block of MSRA ResNet. """ def __init__(self, in_ch, out_ch, stride, dilation, downsample): super(_Bottleneck, self).__init__() mid_ch = out_ch // _BOTTLENECK_EXPANSION self.reduce = _ConvBnReLU(in_ch, mid_ch, 1, stride, 0, 1, True) self.conv3x3 = _ConvBnReLU(mid_ch, mid_ch, 3, 1, dilation, dilation, True) self.increase = _ConvBnReLU(mid_ch, out_ch, 1, 1, 0, 1, False) self.shortcut = ( _ConvBnReLU(in_ch, out_ch, 1, stride, 0, 1, False) if downsample else lambda x: x # identity ) def forward(self, x): h = self.reduce(x) h = self.conv3x3(h) h = self.increase(h) h += self.shortcut(x) return F.relu(h) class _ResLayer(nn.Sequential): """ Residual layer with multi grids """ def __init__(self, n_layers, in_ch, out_ch, stride, dilation, multi_grids=None): super(_ResLayer, self).__init__() if multi_grids is None: multi_grids = [1 for _ in range(n_layers)] else: assert n_layers == len(multi_grids) # Downsampling is only in the first block for i in range(n_layers): self.add_module( "block{}".format(i + 1), _Bottleneck( in_ch=(in_ch if i == 0 else out_ch), out_ch=out_ch, stride=(stride if i == 0 else 1), dilation=dilation * multi_grids[i], downsample=(True if i == 0 else False), ), ) class _Stem(nn.Sequential): """ The 1st conv layer. Note that the max pooling is different from both MSRA and FAIR ResNet. """ def __init__(self, out_ch): super(_Stem, self).__init__() self.add_module("conv1", _ConvBnReLU(3, out_ch, 7, 2, 3, 1)) self.add_module("pool", nn.MaxPool2d(3, 2, 1, ceil_mode=True)) if __name__ == "__main__": model = DeepLabV3( n_classes=21, n_blocks=[3, 4, 23, 3], atrous_rates=[6, 12, 18], multi_grids=[1, 2, 4], output_stride=8, ) model.eval() image = torch.randn(1, 3, 513, 513) print(model) print("input:", image.shape) print("output:", model(image).shape)
和V1，V2的区别在源码里详细注释了。最后DeepLab V3得到输出结果和V1/V2得到输出结果是一致的，训练标签的设置也是一致的。
结论
通过源码解析，应该可以对DeepLab V1,V2,V3的原理和特征图维度变化以及 训练有清楚的认识了，所以暂时就讲到这里了。之后有时间再补上DeepLab V3 Plus的论文理解和源码解析语义分割就算暂时完结了。之后准备做目标检测/分类网络的解析，敬请期待吧。
代码链接
https://github.com/kazuto1011/deeplab-pytorch/tree/master/libs/models
