void land space 主页文章标签分类相册
LeNet
| DL

简介

  • LeNet-5模型是Yann LeCun教授于1998年在论文Gradient-based learning applied to document recognition中提出。它是第一个成功应用于手写数字识别问题并产生实际商业价值的卷积神经网络

网络结构

LeNet-5网络结构

  • 2个卷积层 (C1, C3), 2个下采样(池化) 层 (S2 and S4), 2个全连接层 (C5,F6), 随后是输出层
层 说明 输出数据维度
input 输入层，图片 batchx3x32x32
conv1 卷积层
in=3
out=16
kernel=5x5
padding=0
stride=1		 batchx16x28x28
relu 激活函数
pool1 最大下采样
kernel=2x2
stride=2
padding=0		 batchx16x14x14
conv2 卷积层
in=16
out=32
kernel=5x5
padding=0
stride=1		 batchx32x10*10
relu 激活函数
pool2 最大下采样
kernel=2x2
stride=2
padding=0		 batchx32x5x5
fc1 全连接层
in=32x5x5
out=120		 batchx120
fc2 全连接层
in=120
out=84		 batchx84
fc3 全连接层
in=84
out=10		 batchx10

卷积层计算公式：N=W−F+2PS+1N=\frac{W-F+2P}{S}+1
WW：输入图片大小
FF：filter大小
S：步长
P：Padding

代码

pytorch

import torch.nn as nn
import torch.nn.functional as F
class LeNet(nn.Module):
  def __init__(self):
    super(LeNet, self).__init__()
    self.conv1 = nn.Conv2d(3,16,5) # in_channel out_channel kernel_size stride=1
    self.pool1 = nn.MaxPool2d(2,2) # kernel_size stride padding=0
    self.conv2 = nn.Conv2d(16,32,5)
    self.pool2 = nn.MaxPool2d(2,2)
    self.fc1 = nn.Linear(32*5*5,120)
    self.fc2 = nn.Linear(120,84)
    self.fc3 = nn.Linear(84,10)
  def forward(self,x):
    x = F.relu(self.conv1(x))
    x = self.pool1(x)
    x = F.relu(self.conv2(x))
    x = self.pool2(x)
    x = x.view(-1, 32*5*5)
    x = F.relu(self.fc1(x))
    x = F.relu(self.fc2(x))
    x = self.fc3(x)
    return x
  • tensorflow

padding: VALID(N=W−F+1SN=\frac{W-F+1}{S}，向上取整)、SAME(N=WSN=\frac{W}{S}，向上取整)

from tensorflow.keras import Model
from tensorflow.keras.layers import Dense, Flatten, Conv2D, MaxPool2D
class LeNet(Model):
    def __init__(self):
        super(LeNet, self).__init__()
        self.conv1 = Conv2D(16, 5, activation='relu')
        self.pool1 = MaxPool2D()
        self.conv2 = Conv2D(32,5, activation='relu')
        self.pool2 = MaxPool2D()
        self.flatten = Flatten()
        self.d1 = Dense(120, activation='relu')
        self.d2 = Dense(84, activation='relu')
        self.d3 = Dense(10, activation='softmax')
    def call(self, x, **kwargs):
        x = self.pool1(self.conv1(x))     # input[batch, 32, 32, 3] output[batch, 14, 14, 16]
        x = self.pool2(self.conv2(x))     # input[batch, 14, 14, 16] output[batch, 5, 5, 32]
        x = self.flatten(x)               # output [batch, 5*5*32]
        x = self.d1(x)                    # output [batch, 120]
        x = self.d2(x)                    # output [batch, 84]
        return self.d3(x)                 # output [batch, 10]
algorithm