Pytorch 使用CNN圖像分類的實(shí)現(xiàn)

更新時(shí)間：2020年06月16日 09:56:21 作者：NULL

這篇文章主要介紹了Pytorch 使用CNN圖像分類的實(shí)現(xiàn)，文中通過(guò)示例代碼介紹的非常詳細(xì)，對(duì)大家的學(xué)習(xí)或者工作具有一定的參考學(xué)習(xí)價(jià)值，需要的朋友們下面隨著小編來(lái)一起學(xué)習(xí)學(xué)習(xí)吧

需求

在4*4的圖片中，比較外圍黑色像素點(diǎn)和內(nèi)圈黑色像素點(diǎn)個(gè)數(shù)的大小將圖片分類

如上圖圖片外圍黑色像素點(diǎn)5個(gè)大于內(nèi)圈黑色像素點(diǎn)1個(gè)分為0類反之1類

想法

通過(guò)numpy、PIL構(gòu)造4*4的圖像數(shù)據(jù)集
構(gòu)造自己的數(shù)據(jù)集類
讀取數(shù)據(jù)集對(duì)數(shù)據(jù)集選取減少偏斜
cnn設(shè)計(jì)因?yàn)樘卣魃?，直?*1卷積層
或者在4*4外圍添加padding成6*6,設(shè)計(jì)2*2的卷積核得出3*3再接上全連接層

代碼

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

構(gòu)造數(shù)據(jù)集

import csv
import collections
import os
import shutil

def buildDataset(root,dataType,dataSize):
  """構(gòu)造數(shù)據(jù)集
  構(gòu)造的圖片存到root/{dataType}Data
  圖片地址和標(biāo)簽的csv文件存到 root/{dataType}DataInfo.csv
  Args:
    root:str
      項(xiàng)目目錄
    dataType:str
      'train'或者‘test'
    dataNum:int
      數(shù)據(jù)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):
    # 創(chuàng)建0，1 數(shù)組
    imageArray=np.random.randint(0,2,(4,4))
    # 計(jì)算0，1數(shù)量得到標(biāo)簽
    allBlackNum = collections.Counter(imageArray.flatten())[0]
    innerBlackNum = collections.Counter(imageArray[1:3,1:3].flatten())[0]
    label = 0 if (allBlackNum-innerBlackNum)>innerBlackNum else 1
    # 將圖片保存
    path = f'{dataPath}/{i}.jpg'
    dataInfo.append([path,label])
    im = Image.fromarray(np.uint8(imageArray*255))
    im = im.convert('1') 
    im.save(path)
  # 將圖片地址和標(biāo)簽存入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'

構(gòu)造訓(xùn)練數(shù)據(jù)集

buildDataset(root,'train',20000)

構(gòu)造測(cè)試數(shù)據(jù)集

buildDataset(root,'test',10000)

讀取數(shù)據(jù)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 = []
      # 讀取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())

處理數(shù)據(jù)集使得數(shù)據(jù)集不偏斜

import itertools

def chooseData(dataset,scale):
  # 將類別為1的排序到前面
  dataset.imgs.sort(key=lambda x:x[1],reverse=True)
  # 獲取類別1的數(shù)目 ，取scale倍的數(shù)組，得數(shù)據(jù)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

# 超參數(shù)
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)

定義模型

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層常數(shù)1,設(shè)定卷積核為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

定義損失函數(shù)

# 交叉熵?fù)p失函數(shù)
loss = nn.CrossEntropyLoss()
loss2 = nn.CrossEntropyLoss()

定義優(yōu)化算法

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

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

訓(xùn)練模型

# 計(jì)算準(zhǔn)確率
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()
      # 計(jì)算訓(xùn)練準(zhǔn)確度和loss
      trainLossSum += l.item()
      trainAccSum += (yHat.argmax(dim=1) == y).sum().item()
      n += y.shape[0]
    # 評(píng)估測(cè)試準(zhǔn)確度
    testAcc = evaluateAccuracy(testIter, net)
    print('epoch {:d}, loss {:.4f}, train acc {:.3f}, test acc {:.3f}'.format(epoch + 1, trainLossSum / n, trainAccSum / n, testAcc))

Net模型訓(xùn)練

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模型訓(xùn)練

# batchSize = 50 
# lr = 0.1
# numEpochs = 15 下得出的結(jié)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

測(cè)試

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)
# 比較結(jié)果
print(f'Net測(cè)試結(jié)果{predictions.eq(target)}')
print(f'Net2測(cè)試結(jié)果{predictions2.eq(target)}')
Net測(cè)試結(jié)果tensor([ True, True, False, True, True, True, True])
Net2測(cè)試結(jié)果tensor([False, True, False, True, True, False, True])

到此這篇關(guān)于Pytorch 使用CNN圖像分類的實(shí)現(xiàn)的文章就介紹到這了,更多相關(guān)Pytorch CNN圖像分類內(nèi)容請(qǐng)搜索腳本之家以前的文章或繼續(xù)瀏覽下面的相關(guān)文章希望大家以后多多支持腳本之家！

您可能感興趣的文章: