脚本专栏

Pytorch 使用CNN图像分类的实现
脚本专栏 2024/12/26 佚名
3 2

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])
标签：
Pytorch,CNN图像分类
免责声明：本站文章均来自网站采集或用户投稿，网站不提供任何软件下载或自行开发的软件！ 如有用户或公司发现本站内容信息存在侵权行为，请邮件告知！ 858582#qq.com
白云城资源网 Copyright www.dyhadc.com

评论“Pytorch 使用CNN图像分类的实现”

暂无“Pytorch 使用CNN图像分类的实现”评论...
www.dyhadc.com 白云城资源网
129,905影音资源
244,626技术资源
111,817软件资源
578,645站长资源

P70系列延期，华为新旗舰将在下月发布

3月20日消息，近期博主@数码闲聊站 透露，原定三月份发布的华为新旗舰P70系列延期发布，预计4月份上市。

而博主@定焦数码 爆料，华为的P70系列在定位上已经超过了Mate60，成为了重要的旗舰系列之一。它肩负着重返影像领域顶尖的使命。那么这次P70会带来哪些令人惊艳的创新呢？

根据目前爆料的消息来看，华为P70系列将推出三个版本，其中P70和P70 Pro采用了三角形的摄像头模组设计，而P70 Art则采用了与上一代P60 Art相似的不规则形状设计。这样的外观是否好看见仁见智，但辨识度绝对拉满。

更新日志

友情链接

杰晶网络 DDR爱好者之家 南强小屋 黑松山资源网 白云城资源网 更多链接
白云城资源网 Copyright www.dyhadc.com