原论文:SQUEEZENET: ALEXNET-LEVEL ACCURACY WITH 50X FEWER PARAMETERS AND <0.5MB MODEL SIZE
摘要
对CNN的研究主要集中在提高准确性上,这篇文章则不同,关注CNN更小的架构。较小的架构具有如下优点:
- 分布式训练更加高效(服务器间需要较少的交流)
- 新模型导入到客户端时开销更小(使得频繁的更新可行)
- FPGA的部署更加可行(FPGA内存受限)
设计策略
- 将3x3卷积核替换为1x1卷积核;因为1x1卷积核参数比3x3卷积核参数少9倍。
- 减少输入通道的数量;一个3x3卷积层上的参数总数为(输入通道数)x(卷积核数)x 3 x 3。
- 延迟下采样;可以在网络早期具有较大的特征映射,学到的特征更加丰富,因此可以提高分类精度。
设计了一个Fire模块来满足设计策略
- fire模块中自由使用1x1卷积核是策略1的应用;
- 压缩部分将s1x1设为小于(e1x1 + e3x3)是策略2的应用;
整个网络中使用了相对较晚的池化层,这是策略3的应用。
pytorch实现以及各阶段特征映射变化
import torch
import numpy as np
import torch.nn as nn
from torch.nn import init
x = np.random.rand(96, 3, 224, 224)
x = torch.tensor(x, dtype=torch.float32)
class 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))
return torch.cat([
self.expand1x1_activation(self.expand1x1(x)),
self.expand3x3_activation(self.expand3x3(x))
], 1)
class SqueezeNet(nn.Module):
def __init__(self, version=1.0, num_classes=1000):
super(SqueezeNet, self).__init__()
if version not in [1.0, 1.1]:
raise ValueError("Unsupported SqueezeNet version {version}:"
"1.0 or 1.1 expected".format(version=version))
self.num_classes = num_classes
if version == 1.0:
self.features = nn.Sequential(
nn.Conv2d(3, 96, kernel_size=7, stride=2), # (96, 96, 109, 109)
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), # (96, 96, 54, 54)
Fire(96, 16, 64, 64), # (96, 128, 54, 54)
Fire(128, 16, 64, 64), # (96, 128, 54, 54)
Fire(128, 32, 128, 128), # (96, 256, 54, 54)
nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True), # (96, 256, 27, 27)
Fire(256, 32, 128, 128), # (96, 256, 27, 27)
Fire(256, 48, 192, 192), # (96, 384, 27, 27)
Fire(384, 48, 192, 192), # (96, 384, 27, 27)
Fire(384, 64, 256, 256), # (96, 512, 27, 27)
nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True), # (96, 512, 13, 13)
Fire(512, 64, 256, 256), # (96, 512, 13, 13)
)
else:
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),
)
# Final convolution is initialized differently form 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)
x.view(x.size(0), self.num_classes)
return x
model = SqueezeNet()
x = model(x)
print(x.shape)