SqueezeNet(2016)笔记及代码

SqueezeNet(2016)笔记及代码

1. 模块化卷积(File Module)

• file module包括两部分：squeeze层 + expand层
• squeeze层：使用1x1卷积对特征进行压缩，输出通道数小于输入
• expand层：使用类似inception结构，使用两路1x1和3x3卷积分别卷积并concat

2. 使用average pooling代替全连接

• 参数大大减少，防止过拟合
• 满足多尺度输入要求

3. 模型结构

4. 模型评估

• 在与AlexNet相当精确的情况下模型大小从240下降到4.8
• 同过深度压缩技术实现标题中0.5大小（标题党）
  

5. pytorch 代码

• https://github.com/gsp-27/pytorch_Squeezenet/blob/master/model.py

import torch
import torch.nn as nn
from torch.autograd import Variable
import torch.functional as F
import numpy as np
import torch.optim as optim
import math

class fire(nn.Module):
    def __init__(self, inplanes, squeeze_planes, expand_planes):
        super(fire, self).__init__()
        self.conv1 = nn.Conv2d(inplanes, squeeze_planes, kernel_size=1, stride=1)
        self.bn1 = nn.BatchNorm2d(squeeze_planes)
        self.relu1 = nn.ReLU(inplace=True)
        self.conv2 = nn.Conv2d(squeeze_planes, expand_planes, kernel_size=1, stride=1)
        self.bn2 = nn.BatchNorm2d(expand_planes)
        self.conv3 = nn.Conv2d(squeeze_planes, expand_planes, kernel_size=3, stride=1, padding=1)
        self.bn3 = nn.BatchNorm2d(expand_planes)
        self.relu2 = nn.ReLU(inplace=True)

        # using MSR initilization
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.in_channels
                m.weight.data.normal_(0, math.sqrt(2./n))

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu1(x)
        out1 = self.conv2(x)
        out1 = self.bn2(out1)
        out2 = self.conv3(x)
        out2 = self.bn3(out2)
        out = torch.cat([out1, out2], 1)
        out = self.relu2(out)
        return out


class SqueezeNet(nn.Module):
    def __init__(self):
        super(SqueezeNet, self).__init__()
        self.conv1 = nn.Conv2d(3, 96, kernel_size=3, stride=1, padding=1) # 32
        self.bn1 = nn.BatchNorm2d(96)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool1 = nn.MaxPool2d(kernel_size=2, stride=2) # 16
        self.fire2 = fire(96, 16, 64)
        self.fire3 = fire(128, 16, 64)
        self.fire4 = fire(128, 32, 128)
        self.maxpool2 = nn.MaxPool2d(kernel_size=2, stride=2) # 8
        self.fire5 = fire(256, 32, 128)
        self.fire6 = fire(256, 48, 192)
        self.fire7 = fire(384, 48, 192)
        self.fire8 = fire(384, 64, 256)
        self.maxpool3 = nn.MaxPool2d(kernel_size=2, stride=2) # 4
        self.fire9 = fire(512, 64, 256)
        self.conv2 = nn.Conv2d(512, 10, kernel_size=1, stride=1)
        self.avg_pool = nn.AvgPool2d(kernel_size=4, stride=4)
        self.softmax = nn.LogSoftmax(dim=1)
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.in_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
            elif isinstance(m, nn.BatchNorm2d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()


    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool1(x)
        x = self.fire2(x)
        x = self.fire3(x)
        x = self.fire4(x)
        x = self.maxpool2(x)
        x = self.fire5(x)
        x = self.fire6(x)
        x = self.fire7(x)
        x = self.fire8(x)
        x = self.maxpool3(x)
        x = self.fire9(x)
        x = self.conv2(x)
        x = self.avg_pool(x)
        x = self.softmax(x)
        return x

def fire_layer(inp, s, e):
    f = fire(inp, s, e)
    return f

def squeezenet(pretrained=False):
    net = SqueezeNet()
    # inp = Variable(torch.randn(64,3,32,32))
    # out = net.forward(inp)
    # print(out.size())
    return net