SqueezeNet

论文地址:https://arxiv.org/abs/1602.07360

和别的轻量级模型一样,模型的设计目标就是在保证精度的情况下尽量减少模型参数.核心是论文提出的一种叫"fire module"的卷积方式.

设计策略

• 主要用1x1卷积核,而不是3x3.
• 减少3x3卷积核作用的channel.
• 推迟下采样的时间.以获取更大尺寸的feature map.这一点是处于精度的考虑.毕竟feature map的resolution越大,信息越丰富.下采样主要通过pool来完成或者卷积的时候控制stride大小.

Fire Module

这个就是网络的核心组件了.

分2部分:

• squeeze convolution layer
• expand layer
  其中squeeze只有1x1filter,expand layer由1x1和3x3filter组成.
  在squeeze层卷积核数记为\(s_{1x1}\)，在expand层，记1x1卷积核数为\(e_{1x1}\)，而3x3卷积核数为\(e_{3x3}\),这三个属于超参数,可调。为了尽量降低3x3的输入通道数，让\(s_{1x1}<e_{1x1}+e_{3x3}\)。
  这两层的设计分别体现了策略1(多用1x1filter)和策略2(减少3x3filter作用channel).

首先,squeeze convolution layer通过控制1x1卷积核数量达到把输入的channel数目降低的目的.这个是降低参数的最关键的一步.
然后,分别用1x1卷积核和3x3卷积核去做卷积.然后得到不同depth的输出,concat起来.([x,x,depth1],[x,x,depth2]-->[x,x,depth1+depth2])

代码实现
torch的官方实现:https://pytorch.org/docs/stable/_modules/torchvision/models/squeezenet.html

import torchimport torch.nn as nnclass Fire(nn.Module): def __init__(self, inplanes, squeeze_planes, expand1x1_planes, expand3x3_planes): super(Fire, self).__init__() self.inplanes = inplanes self.squeeze = nn.Conv2d(inplanes, squeeze_planes, kernel_size=1) self.squeeze_activation = nn.ReLU(inplace=True) self.expand1x1 = nn.Conv2d(squeeze_planes, expand1x1_planes, kernel_size=1) self.expand1x1_activation = nn.ReLU(inplace=True) self.expand3x3 = nn.Conv2d(squeeze_planes, expand3x3_planes, kernel_size=3, padding=1) self.expand3x3_activation = nn.ReLU(inplace=True) def forward(self, x): x = self.squeeze_activation(self.squeeze(x)) print(x.shape) e_1 = self.expand1x1(x) print(e_1.shape) e_3 = self.expand3x3(x) print(e_3.shape) return torch.cat([ self.expand1x1_activation(e_1), self.expand3x3_activation(e_3) ], 1)

很显然地,squeeze convolution layer把channel数量降下来了,所以参数少了很多.
以输入tensor为[n,c,h,w]=[1,96,224,224]举例,假设fire module的squeeze layer的卷积核数量为6,expand layer中1x1卷积核数量为5,3x3卷积核数量为4.
则fire module的参数数量为1x1x96x6 + 1x1x6x5 + 3x3x6x4=822.
普通的3x3卷积,得到depth=9的feature map的话需要3x3x96x9=7776个参数.
所以模型才可以做到很小.

网络结构

基本就是fire module的堆叠.中间穿插了一些maxpool对feature map下采样. 注意一下最后用了dropout以及全局平均池化而不是全连接来完成分类.

最左边的就是类似vgg的堆叠式的结构.中间和右边的参考了resnet的skip-connection.

class SqueezeNet(nn.Module): def __init__(self, version='1_0', num_classes=1000): super(SqueezeNet, self).__init__() self.num_classes = num_classes if version == '1_0': self.features = nn.Sequential( nn.Conv2d(3, 96, kernel_size=7, stride=2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), Fire(96, 16, 64, 64), Fire(128, 16, 64, 64), Fire(128, 32, 128, 128), nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), Fire(256, 32, 128, 128), Fire(256, 48, 192, 192), Fire(384, 48, 192, 192), Fire(384, 64, 256, 256), nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), Fire(512, 64, 256, 256), ) elif version == '1_1': self.features = nn.Sequential( nn.Conv2d(3, 64, kernel_size=3, stride=2), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), Fire(64, 16, 64, 64), Fire(128, 16, 64, 64), nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), Fire(128, 32, 128, 128), Fire(256, 32, 128, 128), nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), Fire(256, 48, 192, 192), Fire(384, 48, 192, 192), Fire(384, 64, 256, 256), Fire(512, 64, 256, 256), ) else: # FIXME: Is this needed? SqueezeNet should only be called from the # FIXME: squeezenet1_x() functions # FIXME: This checking is not done for the other models raise ValueError("Unsupported SqueezeNet version {version}:" "1_0 or 1_1 expected".format(version=version)) # Final convolution is initialized differently from the rest final_conv = nn.Conv2d(512, self.num_classes, kernel_size=1) self.classifier = nn.Sequential( nn.Dropout(p=0.5), final_conv, nn.ReLU(inplace=True), nn.AdaptiveAvgPool2d((1, 1)) ) for m in self.modules(): if isinstance(m, nn.Conv2d): if m is final_conv: init.normal_(m.weight, mean=0.0, std=0.01) else: init.kaiming_uniform_(m.weight) if m.bias is not None: init.constant_(m.bias, 0) def forward(self, x): x = self.features(x) x = self.classifier(x) return torch.flatten(x, 1)

https://pytorch.org/hub/pytorch_vision_squeezenet/
这里有一个用torch中的模型做推理的例子,

.....with torch.no_grad(): output = model(input_batch)# Tensor of shape 1000, with confidence scores over Imagenet's 1000 classesprint(output[0])# The output has unnormalized scores. To get probabilities, you can run a softmax on it.print(torch.nn.functional.softmax(output[0], dim=0))

CNN结构设计的探索

主要从2个方面做实验探讨了不同结构对模型精度和模型大小的影响.

• fire module怎么设计,squeeze layer和expand layer的filter数量怎么设计
• fire module怎么串起来形成一个网络,是简单堆叠还是引入bypass

关于第一点fire module中各种filter占比的实验结果如下图:

这里的sr指的是squeeze layer的卷积核数量/expand layer比例.3x3filter比例指expand layer里3x3filter比例.
具体设计参考论文:

一点思考:
1x1的卷积核关联了某个位置的feature所有channel上的信息.3x3的卷积核关联了多个位置的feature的所有channel的信息.按道理说3x3的越多应该模型精度越好,但实验数据显示并非如此.可能有些是噪音,3x3卷积参考太多周围的feature反而导致精度下降. 这也是深度学习现在被比较多诟病的一点,太黑盒了.只知道能work,为啥work不好解释.不同的数据集可能不同的参数表现会不一样,很难讲哪个最优,调参比较依赖经验.

关于第二点在layer之间采用不同的连接方式,实验结果如下: