Pytorch 使用CNN图像分类的实现


Posted in Python onJune 16, 2020

需求

在4*4的图片中,比较外围黑色像素点和内圈黑色像素点个数的大小将图片分类

Pytorch 使用CNN图像分类的实现

如上图图片外围黑色像素点5个大于内圈黑色像素点1个分为0类反之1类

想法

  • 通过numpy、PIL构造4*4的图像数据集
  • 构造自己的数据集类
  • 读取数据集对数据集选取减少偏斜
  • cnn设计因为特征少,直接1*1卷积层
  • 或者在4*4外围添加padding成6*6,设计2*2的卷积核得出3*3再接上全连接层

代码

import torch
import torchvision
import torchvision.transforms as transforms
import numpy as np
from PIL import Image

构造数据集

import csv
import collections
import os
import shutil

def buildDataset(root,dataType,dataSize):
  """构造数据集
  构造的图片存到root/{dataType}Data
  图片地址和标签的csv文件存到 root/{dataType}DataInfo.csv
  Args:
    root:str
      项目目录
    dataType:str
      'train'或者‘test'
    dataNum:int
      数据大小
  Returns:
  """
  dataInfo = []
  dataPath = f'{root}/{dataType}Data'
  if not os.path.exists(dataPath):
    os.makedirs(dataPath)
  else:
    shutil.rmtree(dataPath)
    os.mkdir(dataPath)
    
  for i in range(dataSize):
    # 创建0,1 数组
    imageArray=np.random.randint(0,2,(4,4))
    # 计算0,1数量得到标签
    allBlackNum = collections.Counter(imageArray.flatten())[0]
    innerBlackNum = collections.Counter(imageArray[1:3,1:3].flatten())[0]
    label = 0 if (allBlackNum-innerBlackNum)>innerBlackNum else 1
    # 将图片保存
    path = f'{dataPath}/{i}.jpg'
    dataInfo.append([path,label])
    im = Image.fromarray(np.uint8(imageArray*255))
    im = im.convert('1') 
    im.save(path)
  # 将图片地址和标签存入csv文件
  filePath = f'{root}/{dataType}DataInfo.csv'
  with open(filePath, 'w') as f:
    writer = csv.writer(f)
    writer.writerows(dataInfo)
root=r'/Users/null/Documents/PythonProject/Classifier'

构造训练数据集

buildDataset(root,'train',20000)

构造测试数据集

buildDataset(root,'test',10000)

读取数据集

class MyDataset(torch.utils.data.Dataset):

  def __init__(self, root, datacsv, transform=None):
    super(MyDataset, self).__init__()
    with open(f'{root}/{datacsv}', 'r') as f:
      imgs = []
      # 读取csv信息到imgs列表
      for path,label in map(lambda line:line.rstrip().split(','),f):
        imgs.append((path, int(label)))
    self.imgs = imgs
    self.transform = transform if transform is not None else lambda x:x
    
  def __getitem__(self, index):
    path, label = self.imgs[index]
    img = self.transform(Image.open(path).convert('1'))
    return img, label

  def __len__(self):
    return len(self.imgs)
trainData=MyDataset(root = root,datacsv='trainDataInfo.csv', transform=transforms.ToTensor())
testData=MyDataset(root = root,datacsv='testDataInfo.csv', transform=transforms.ToTensor())

处理数据集使得数据集不偏斜

import itertools

def chooseData(dataset,scale):
  # 将类别为1的排序到前面
  dataset.imgs.sort(key=lambda x:x[1],reverse=True)
  # 获取类别1的数目 ,取scale倍的数组,得数据不那么偏斜
  trueNum =collections.Counter(itertools.chain.from_iterable(dataset.imgs))[1]
  end = min(trueNum*scale,len(dataset))
  dataset.imgs=dataset.imgs[:end]
scale = 4
chooseData(trainData,scale)
chooseData(testData,scale)
len(trainData),len(testData)
(2250, 1122)
import torch.utils.data as Data

# 超参数
batchSize = 50
lr = 0.1
numEpochs = 20

trainIter = Data.DataLoader(dataset=trainData, batch_size=batchSize, shuffle=True)
testIter = Data.DataLoader(dataset=testData, batch_size=batchSize)

定义模型

from torch import nn
from torch.autograd import Variable
from torch.nn import Module,Linear,Sequential,Conv2d,ReLU,ConstantPad2d
import torch.nn.functional as F
class Net(Module):  
  def __init__(self):
    super(Net, self).__init__()

    self.cnnLayers = Sequential(
      # padding添加1层常数1,设定卷积核为2*2
      ConstantPad2d(1, 1),
      Conv2d(1, 1, kernel_size=2, stride=2,bias=True)
    )
    self.linearLayers = Sequential(
      Linear(9, 2)
    )

  def forward(self, x):
    x = self.cnnLayers(x)
    x = x.view(x.shape[0], -1)
    x = self.linearLayers(x)
    return x
class Net2(Module):  
  def __init__(self):
    super(Net2, self).__init__()

    self.cnnLayers = Sequential(
      Conv2d(1, 1, kernel_size=1, stride=1,bias=True)
    )
    self.linearLayers = Sequential(
      ReLU(),
      Linear(16, 2)
    )

  def forward(self, x):
    x = self.cnnLayers(x)
    x = x.view(x.shape[0], -1)
    x = self.linearLayers(x)
    return x

定义损失函数

# 交叉熵损失函数
loss = nn.CrossEntropyLoss()
loss2 = nn.CrossEntropyLoss()

定义优化算法

net = Net()
optimizer = torch.optim.SGD(net.parameters(),lr = lr)
net2 = Net2()
optimizer2 = torch.optim.SGD(net2.parameters(),lr = lr)

训练模型

# 计算准确率
def evaluateAccuracy(dataIter, net):
  accSum, n = 0.0, 0
  with torch.no_grad():
    for X, y in dataIter:
      accSum += (net(X).argmax(dim=1) == y).float().sum().item()
      n += y.shape[0]
  return accSum / n
def train(net, trainIter, testIter, loss, numEpochs, batchSize,
       optimizer):
  for epoch in range(numEpochs):
    trainLossSum, trainAccSum, n = 0.0, 0.0, 0
    for X,y in trainIter:
      yHat = net(X)
      l = loss(yHat,y).sum()
      optimizer.zero_grad()
      l.backward()
      optimizer.step()
      # 计算训练准确度和loss
      trainLossSum += l.item()
      trainAccSum += (yHat.argmax(dim=1) == y).sum().item()
      n += y.shape[0]
    # 评估测试准确度
    testAcc = evaluateAccuracy(testIter, net)
    print('epoch {:d}, loss {:.4f}, train acc {:.3f}, test acc {:.3f}'.format(epoch + 1, trainLossSum / n, trainAccSum / n, testAcc))

Net模型训练

train(net, trainIter, testIter, loss, numEpochs, batchSize,optimizer)
epoch 1, loss 0.0128, train acc 0.667, test acc 0.667
epoch 2, loss 0.0118, train acc 0.683, test acc 0.760
epoch 3, loss 0.0104, train acc 0.742, test acc 0.807
epoch 4, loss 0.0093, train acc 0.769, test acc 0.772
epoch 5, loss 0.0085, train acc 0.797, test acc 0.745
epoch 6, loss 0.0084, train acc 0.798, test acc 0.807
epoch 7, loss 0.0082, train acc 0.804, test acc 0.816
epoch 8, loss 0.0078, train acc 0.816, test acc 0.812
epoch 9, loss 0.0077, train acc 0.818, test acc 0.817
epoch 10, loss 0.0074, train acc 0.824, test acc 0.826
epoch 11, loss 0.0072, train acc 0.836, test acc 0.819
epoch 12, loss 0.0075, train acc 0.823, test acc 0.829
epoch 13, loss 0.0071, train acc 0.839, test acc 0.797
epoch 14, loss 0.0067, train acc 0.849, test acc 0.824
epoch 15, loss 0.0069, train acc 0.848, test acc 0.843
epoch 16, loss 0.0064, train acc 0.864, test acc 0.851
epoch 17, loss 0.0062, train acc 0.867, test acc 0.780
epoch 18, loss 0.0060, train acc 0.871, test acc 0.864
epoch 19, loss 0.0057, train acc 0.881, test acc 0.890
epoch 20, loss 0.0055, train acc 0.885, test acc 0.897

Net2模型训练

# batchSize = 50 
# lr = 0.1
# numEpochs = 15 下得出的结果
train(net2, trainIter, testIter, loss2, numEpochs, batchSize,optimizer2)

epoch 1, loss 0.0119, train acc 0.638, test acc 0.676
epoch 2, loss 0.0079, train acc 0.823, test acc 0.986
epoch 3, loss 0.0046, train acc 0.987, test acc 0.977
epoch 4, loss 0.0030, train acc 0.983, test acc 0.973
epoch 5, loss 0.0023, train acc 0.981, test acc 0.976
epoch 6, loss 0.0019, train acc 0.980, test acc 0.988
epoch 7, loss 0.0016, train acc 0.984, test acc 0.984
epoch 8, loss 0.0014, train acc 0.985, test acc 0.986
epoch 9, loss 0.0013, train acc 0.987, test acc 0.992
epoch 10, loss 0.0011, train acc 0.989, test acc 0.993
epoch 11, loss 0.0010, train acc 0.989, test acc 0.996
epoch 12, loss 0.0010, train acc 0.992, test acc 0.994
epoch 13, loss 0.0009, train acc 0.993, test acc 0.994
epoch 14, loss 0.0008, train acc 0.995, test acc 0.996
epoch 15, loss 0.0008, train acc 0.994, test acc 0.998

测试

test = torch.Tensor([[[[0,0,0,0],[0,1,1,0],[0,1,1,0],[0,0,0,0]]],
         [[[1,1,1,1],[1,0,0,1],[1,0,0,1],[1,1,1,1]]],
         [[[0,1,0,1],[1,0,0,1],[1,0,0,1],[0,0,0,1]]],
         [[[0,1,1,1],[1,0,0,1],[1,0,0,1],[0,0,0,1]]],
         [[[0,0,1,1],[1,0,0,1],[1,0,0,1],[1,0,1,0]]],
         [[[0,0,1,0],[0,1,0,1],[0,0,1,1],[1,0,1,0]]],
         [[[1,1,1,0],[1,0,0,1],[1,0,1,1],[1,0,1,1]]]
         ])

target=torch.Tensor([0,1,0,1,1,0,1])
test
tensor([[[[0., 0., 0., 0.],
     [0., 1., 1., 0.],
     [0., 1., 1., 0.],
     [0., 0., 0., 0.]]],

​

    [[[1., 1., 1., 1.],
     [1., 0., 0., 1.],
     [1., 0., 0., 1.],
     [1., 1., 1., 1.]]],

​

    [[[0., 1., 0., 1.],
     [1., 0., 0., 1.],
     [1., 0., 0., 1.],
     [0., 0., 0., 1.]]],

​

    [[[0., 1., 1., 1.],
     [1., 0., 0., 1.],
     [1., 0., 0., 1.],
     [0., 0., 0., 1.]]],

​

    [[[0., 0., 1., 1.],
     [1., 0., 0., 1.],
     [1., 0., 0., 1.],
     [1., 0., 1., 0.]]],

​

    [[[0., 0., 1., 0.],
     [0., 1., 0., 1.],
     [0., 0., 1., 1.],
     [1., 0., 1., 0.]]],

​

    [[[1., 1., 1., 0.],
     [1., 0., 0., 1.],
     [1., 0., 1., 1.],
     [1., 0., 1., 1.]]]])



with torch.no_grad():
  output = net(test)
  output2 = net2(test)
predictions =output.argmax(dim=1)
predictions2 =output2.argmax(dim=1)
# 比较结果
print(f'Net测试结果{predictions.eq(target)}')
print(f'Net2测试结果{predictions2.eq(target)}')
Net测试结果tensor([ True, True, False, True, True, True, True])
Net2测试结果tensor([False, True, False, True, True, False, True])

到此这篇关于Pytorch 使用CNN图像分类的实现的文章就介绍到这了,更多相关Pytorch CNN图像分类内容请搜索三水点靠木以前的文章或继续浏览下面的相关文章希望大家以后多多支持三水点靠木!

Python 相关文章推荐
Python的print用法示例
Feb 11 Python
python通过pil模块获得图片exif信息的方法
Mar 16 Python
对比Python中__getattr__和 __getattribute__获取属性的用法
Jun 21 Python
python自动化脚本安装指定版本python环境详解
Sep 14 Python
修改 CentOS 6.x 上默认Python的方法
Sep 06 Python
Python利用全连接神经网络求解MNIST问题详解
Jan 14 Python
关于pytorch中全连接神经网络搭建两种模式详解
Jan 14 Python
解决Python命令行下退格,删除,方向键乱码(亲测有效)
Jan 16 Python
PyCharm中如何直接使用Anaconda已安装的库
May 28 Python
Python requests及aiohttp速度对比代码实例
Jul 16 Python
基于python实现复制文件并重命名
Sep 16 Python
利用Python实现学生信息管理系统的完整实例
Dec 30 Python
利用python中的matplotlib打印混淆矩阵实例
Jun 16 #Python
Python SMTP配置参数并发送邮件
Jun 16 #Python
基于matplotlib中ion()和ioff()的使用详解
Jun 16 #Python
Python数据相关系数矩阵和热力图轻松实现教程
Jun 16 #Python
matplotlib.pyplot.matshow 矩阵可视化实例
Jun 16 #Python
使用python matploblib库绘制准确率,损失率折线图
Jun 16 #Python
为什么称python为胶水语言
Jun 16 #Python
You might like
基于jquery的设置页面文本框 只能输入数字的实现代码
2011/04/19 Javascript
JQuery中判断一个元素下面是否有内容或者有某个标签的判断代码
2012/02/02 Javascript
jquery应该如何来设置改变按钮input的onclick事件
2012/12/10 Javascript
javascript实现文字图片上下滚动的具体实例
2013/06/28 Javascript
正负小数点后两位浮点数实现原理及代码
2013/09/06 Javascript
JS小功能(操作Table--动态添加删除表格及数据)实现代码
2013/11/28 Javascript
一个不错的仿携程自定义数据下拉选择select
2014/09/01 Javascript
AngularJS基于provider实现全局变量的读取和赋值方法
2017/06/28 Javascript
js实现图片上传预览原理分析
2017/07/13 Javascript
vue项目中用cdn优化的方法
2018/01/03 Javascript
js和jQuery以及easyui实现对下拉框的指定赋值方法
2018/01/23 jQuery
AngularJS中ng-options实现下拉列表的数据绑定方法
2018/08/13 Javascript
利用vue重构有赞商城的思路以及总结整理
2019/02/21 Javascript
如何利用vue+vue-router+elementUI实现简易通讯录
2019/05/13 Javascript
微信小程序 弹窗输入组件的实现解析
2019/08/12 Javascript
[55:44]OG vs NAVI 2019国际邀请赛小组赛 BO2 第一场 8.15
2019/08/17 DOTA
Python爬取qq music中的音乐url及批量下载
2017/03/23 Python
python 遍历字符串(含汉字)实例详解
2017/04/04 Python
浅析python递归函数和河内塔问题
2017/04/18 Python
Python简单实现自动删除目录下空文件夹的方法
2017/08/29 Python
学python安装的软件总结
2019/10/12 Python
html5自动播放mov格式视频的实例代码
2020/01/14 HTML / CSS
Nasty Gal英国:美国女性服饰销售网站
2021/03/02 全球购物
Burt’s Bees英国官网:世界领先的天然个人护理品牌
2020/08/17 全球购物
校本教研工作制度
2014/01/22 职场文书
搞笑婚礼主持词
2014/03/13 职场文书
会计核算科岗位职责
2014/03/19 职场文书
计算机应用专业毕业生求职信
2014/06/03 职场文书
电子工程求职信
2014/07/17 职场文书
销售员态度差检讨书
2014/10/26 职场文书
个人简历自我评价怎么写
2015/03/10 职场文书
举起手来观后感
2015/06/09 职场文书
城南旧事观后感
2015/06/11 职场文书
为什么在foreach循环中JAVA集合不能添加或删除元素
2021/06/11 Java/Android
python 对图片进行简单的处理
2021/06/23 Python
MySQL中rank() over、dense_rank() over、row_number() over用法介绍
2022/03/23 MySQL