MNIST手写数字识别项目因为数据量小、识别任务简单而成为图像识别入门的第一课,MNIST手写数字识别项目有如下特点:
- 识别难度低,即使把图片展开为一维数据,且只使用全连接层也能获得超过98%的识别准确度;
- 计算量小,不需要GPU加速也可以快速训练完成;
- 数据易得,教程易得。
本文中的数据来源于Kaggle网站,并非Pytorch自带的数据集(个人认为使用Pytorch封装好的MNIST数据集不利于入门)。接下来我们展示不同思路下的MNIST识别结果。
1. 将MNIST视为回归问题和分类问题的差别
一般来说,我们认为MNIST是一个图像分类问题,而非回归问题,因为图像的特征并没有随着0-9数字增大而显现出的对应的特征值变化关系,即图像中的像素值与数字没有表现出相关关系。
全连接回归模式(使用MSELoss)准确率:
全连接回归网络
排行榜上清一色的0.9以上,这个成绩真的拿不出手啊。
全连接分类模式(使用CrossEntropy)准确率:
全连接分类模式——两次迭代
全连接分类模式——两百次迭代
使用分类模式,仅需两次迭代就能达到0.93的准确率,并且200次迭代之后,准确率达到了0.97。
2. 卷积神经网络
将图片一维展开尚且能达到0.97的准确率,如果让图片保持原状,使用卷积神经网络来识别能够得到什么结果呢。
卷积神经网络准确率:
卷积神经网络准确率
项目包含两个py文件,mnist_models.py和main.py,下面是mnist_models.py中的源代码:
import torch as t
def to_image(data):
data = data.view(-1,1,28,28)
return data
class fc_net(t.nn.Module):
'''
全连接网络
'''
def __init__(self):
super(fc_net,self).__init__()
self.fc1 = t.nn.Sequential(t.nn.Linear(784,200),t.nn.ReLU())
self.fc2 = t.nn.Sequential(t.nn.Linear(200,100),t.nn.ReLU())
self.fc3 = t.nn.Sequential(t.nn.Linear(100,20),t.nn.ReLU())
self.fc4 = t.nn.Linear(20,10)
def forward(self,x):
x = self.fc1(x)
x = self.fc2(x)
x = self.fc3(x)
x = self.fc4(x)
return x
class conv_net(t.nn.Module):
'''
卷积网络,需先将数据转为2维图片形式
'''
def __init__(self):
super(conv_net,self).__init__()
self.conv1 = t.nn.Sequential(
t.nn.Conv2d(1,10,5,1,1),
t.nn.MaxPool2d(2),
t.nn.ReLU(),
t.nn.BatchNorm2d(10)
)
self.conv2 = t.nn.Sequential(
t.nn.Conv2d(10,20,5,1,1),
t.nn.MaxPool2d(2),
t.nn.ReLU(),
t.nn.BatchNorm2d(20) # num_features为通道数
)
self.fc1 = t.nn.Sequential(
t.nn.Linear(500,60),
t.nn.Dropout(0.5),
t.nn.ReLU()
)
self.fc2 = t.nn.Sequential(
t.nn.Linear(60,20),
t.nn.Dropout(0.5),
t.nn.ReLU()
)
self.fc3 = t.nn.Linear(20,10)
def forward(self, x):
x = self.conv1(x)
x = self.conv2(x)
x = x.view(-1,500)
x = self.fc1(x)
x = self.fc2(x)
x = self.fc3(x)
return x
class AlexNet(t.nn.Module):
'''
类似AlexNet的神经网络,因为电脑配置及MNIST数据集图片尺寸问题,将Kernel_size和stride都改小了
'''
def __init__(self, num_classes=1000):
super(AlexNet, self).__init__()
self.features = t.nn.Sequential(
t.nn.Conv2d(1, 64, kernel_size=5, stride=1, padding=2),
t.nn.ReLU(inplace=True),
t.nn.MaxPool2d(kernel_size=3, stride=1),
t.nn.Conv2d(64, 192, kernel_size=3, padding=2),
t.nn.ReLU(inplace=True),
t.nn.MaxPool2d(kernel_size=3, stride=2),
t.nn.Conv2d(192, 384, kernel_size=3, padding=1),
t.nn.ReLU(inplace=True),
t.nn.Conv2d(384, 256, kernel_size=3, padding=1),
t.nn.ReLU(inplace=True),
t.nn.Conv2d(256, 256, kernel_size=3, padding=1),
t.nn.ReLU(inplace=True),
t.nn.MaxPool2d(kernel_size=3, stride=2),
)
self.classifier = t.nn.Sequential(
t.nn.Dropout(),
t.nn.Linear(256 * 6 * 6, 4096),
t.nn.ReLU(inplace=True),
t.nn.Dropout(),
t.nn.Linear(4096, 4096),
t.nn.ReLU(inplace=True),
t.nn.Linear(4096, num_classes),
)
def forward(self, x):
x = self.features(x)
print(x.shape)
x = x.view(x.size(0), 256 * 6 * 6)
print(x.shape)
x = self.classifier(x)
return x
下面是main.py中的源代码
import os
import pandas as pd
import torch.nn.functional as F
from torchvision import models
import matplotlib.pyplot as plt
import torch as t
from tqdm import tqdm
from torch.autograd import Variable
from mnist_models import conv_net,to_image,fc_net,AlexNet
import signal
# 设置模型参数
TYPE = 'cla'
METHOD = 'conv'
EPOCHS = 400
BATCH_SIZE = 500
LR = 0.001
# 读取数据
train = pd.read_csv('./data/train.csv')
data = train.drop('label',axis=1)
test = pd.read_csv('./data/test.csv')
test_data = t.from_numpy(test.values).float()
data = data.values
# 标签与自变量处理
y = train['label'].values
y = t.from_numpy(y).long()
data = t.from_numpy(data).float()
data,y = Variable(data),Variable(y)
# 初始化模型
if METHOD == 'conv':
data = to_image(data) # 将数据转为二维
test_data = to_image(test_data)
net = conv_net()
elif METHOD == 'fc':
net = fc_net()
elif METHOD == 'res':
# 使用resnet18进行迁移学习,微调参数,如果冻结参数,将resnet作为特征选择器的话,训练速度更快。
# 因为resnet参数过多,不建议使用CPU运算,使用Xeon E5620一个EPOCH要训练三个小时
data = to_image(data)
test_data = to_image(test_data)
net = models.resnet18(pretrained=True)
# 固定参数
for p in net.parameters():
p.requires_grad = False
# 因为MNIST图片是单通道,并且尺寸较小,所以需要对resnet进行一些细节修改
net.conv1 = t.nn.Conv2d(1, 64, kernel_size=3, stride=1, padding=3,
bias=False)
net.maxpool = t.nn.MaxPool2d(kernel_size=2, stride=1, padding=1)
net.avgpool = t.nn.AvgPool2d(5, stride=1)
num_ftrs = net.fc.in_features
net.fc = t.nn.Linear(num_ftrs,10)
elif METHOD == 'alex':
data = to_image(data)
test_data = to_image(test_data)
net = AlexNet()
else:
raise Exception("Wrong Method!")
# 如果模型文件存在则尝试加载模型参数
if os.path.exists('H:/learning_notes/MNIST/%s.pth' % METHOD):
try:
net.load_state_dict(t.load('H:/learning_notes/MNIST/%s.pth' % METHOD))
except Exception as e:
print(e)
print("Parameters Error")
# 定义模型代价函数
if TYPE == 'reg':
criterion = t.nn.MSELoss()
elif TYPE == 'cla':
criterion = t.nn.CrossEntropyLoss()
else:
raise Exception("Wrong Type!")
# 定义优化器
if METHOD == 'res':
# 如果是用的resnet,则只训练最后的全连接层的参数
optim = t.optim.Adam(net.fc.parameters(),lr = 0.001,weight_decay=0.0)
else:
optim = t.optim.Adam(net.parameters(),lr=0.001,weight_decay=0.0)
# plt.ion() # 用于绘制动态图
# losses = []
# 用于捕捉KeyboardInterrupt错误,效果比try except好得多
# 可以人为终止训练,并将训练得到的参数保存下来,实现断点训练
def exit(signum, frame):
print("Model Saved")
t.save(net.state_dict(), 'H:/learning_notes/MNIST/%s.pth' % METHOD)
raise KeyboardInterrupt
signal.signal(signal.SIGINT, exit)
signal.signal(signal.SIGTERM, exit)
# 开始训练
for epoch in tqdm(range(EPOCHS)):
index = 0
if epoch % 100 == 0:
for param_group in optim.param_groups:
LR = LR * 0.9
param_group['lr'] = LR
for i in tqdm(range(int(len(data)/BATCH_SIZE)),total=int(len(data)/BATCH_SIZE)):
batch_x = data[index:index + BATCH_SIZE]
batch_y = y[index:index + BATCH_SIZE]
prediction = net.forward(batch_x)
loss = criterion(prediction, batch_y)
optim.zero_grad()
loss.backward()
optim.step()
index += BATCH_SIZE # 进入下一个batch
# if loss <= 0.3:
# losses.append(loss)
# plt.plot(losses)
# plt.pause(0.001)
print(loss)
t.save(net.state_dict(),'H:/learning_notes/MNIST/%s.pth' % METHOD)
# plt.ioff()
submission = pd.read_csv("./data/sample_submission.csv")
print('=======Predicting========')
# 切换成验证模式,验证模式下DROPOUT将不起作用
net.eval()
test_data = Variable(test_data)
result = t.Tensor()
index = 0
# 分段进行预测,节省内存
for i in tqdm(range(int(test_data.shape[0]/BATCH_SIZE)),total=int(test_data.shape[0]/BATCH_SIZE)):
label_prediction = net(test_data[index:index+BATCH_SIZE])
index += BATCH_SIZE
result = t.cat((result,label_prediction),0)
# 结果处理
if TYPE == 'cla':
_,submission['Label'] = t.max(result.data,1) # t.max返回一个元祖,第一个元素是最大元素值,第二个元素是最大元素位置
elif TYPE == 'reg':
submission['Label'] = submission['Label'].astype('int')
submission['Label'] = submission['Label'].apply(lambda x:9 if x>= 10 else x)
submission.to_csv("submission.csv",index=False)
网友评论