MNSIT手写数字识别 入门
MNIST数据集手写数字识别
数据集介绍
MNIST 包括6万张28x28的训练样本，1万张测试样本，很多教程都会对它”下手”几乎成为一个 “典范”，可以说它就是计算机视觉里面的Hello World。
导入必备的包
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets,transforms
import matplotlib.pyplot as plt
import numpy as np
设置超参数
BATCH_SIZE = 512
EPOCHS = 10
DEVICE = torch.device("cuda")
lr = 0.01
加载数据
train_loader = torch.utils.data.DataLoader(
datasets.MNIST(
root='data',
train=True,
download=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,),(0.3801,))
])),
batch_size = BATCH_SIZE,
shuffle=True)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST(
root='data',
train=False,
download=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,),(0.3801,))
])),
batch_size = BATCH_SIZE,
shuffle=True)
定义模型
首先使用MLP多层感知机尝试一下
class MLP(nn.Module):
def __init__(self):
super().__init__()
self.fc1 = nn.Linear(784,256)
self.fc2 = nn.Linear(256,10)
def forward(self,x):
in_size = x.size(0)
out = x.view(in_size,-1)
out = self.fc1(out)# 输入28*28=784 输出256
out = F.relu(out)
out = self.fc2(out)# 输入256 输出10
out = F.log_softmax(out,dim=1)
return out
model_mlp = MLP().to(DEVICE)
optimizer_mlp = optim.SGD(model_mlp.parameters(),lr=lr)
训练
def train(model,device,train_loader,optimizer,epoch):
model.train()
for batch_idx, (data,target) in enumerate(train_loader):
data,target = data.to(device),target.to(device)
optimizer.zero_grad()
output = model(data)
loss = F.nll_loss(output,target)
loss.backward()
optimizer.step()
if(batch_idx+1)%30 == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data), len(train_loader.dataset),
100. * batch_idx / len(train_loader), loss.item()))
def test(model,device,test_loader):
model.eval()
test_loss = 0
correct = 0
with torch.no_grad():
for data,target in test_loader:
data, target = data.to(device), target.to(device)
output = model(data)
test_loss += F.nll_loss(output,target,reduction='sum').item()
pred = output.max(1,keepdim=True)[1]
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(test_loader.dataset)
print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
test_loss, correct, len(test_loader.dataset),
100. * correct / len(test_loader.dataset)))
for epoch in range(1, EPOCHS + 1):
train(model_mlp, DEVICE, train_loader, optimizer_mlp, epoch)
test(model_mlp, DEVICE, test_loader)
Train Epoch: 1 [14848/60000 (25%)] Loss: 1.943828
Train Epoch: 1 [30208/60000 (50%)] Loss: 1.579372
Train Epoch: 1 [45568/60000 (75%)] Loss: 1.273092
Test set: Average loss: 1.0416, Accuracy: 8151/10000 (82%)
Train Epoch: 2 [14848/60000 (25%)] Loss: 0.892413
Train Epoch: 2 [30208/60000 (50%)] Loss: 0.777179
Train Epoch: 2 [45568/60000 (75%)] Loss: 0.703713
Test set: Average loss: 0.6323, Accuracy: 8635/10000 (86%)
Train Epoch: 3 [14848/60000 (25%)] Loss: 0.581855
Train Epoch: 3 [30208/60000 (50%)] Loss: 0.585811
Train Epoch: 3 [45568/60000 (75%)] Loss: 0.575292
Test set: Average loss: 0.4979, Accuracy: 8789/10000 (88%)
Train Epoch: 4 [14848/60000 (25%)] Loss: 0.524915
Train Epoch: 4 [30208/60000 (50%)] Loss: 0.506676
Train Epoch: 4 [45568/60000 (75%)] Loss: 0.489618
Test set: Average loss: 0.4336, Accuracy: 8860/10000 (89%)
Train Epoch: 5 [14848/60000 (25%)] Loss: 0.502646
Train Epoch: 5 [30208/60000 (50%)] Loss: 0.453554
Train Epoch: 5 [45568/60000 (75%)] Loss: 0.422082
Test set: Average loss: 0.3956, Accuracy: 8945/10000 (89%)
Train Epoch: 6 [14848/60000 (25%)] Loss: 0.428002
Train Epoch: 6 [30208/60000 (50%)] Loss: 0.442678
Train Epoch: 6 [45568/60000 (75%)] Loss: 0.420096
Test set: Average loss: 0.3706, Accuracy: 8983/10000 (90%)
Train Epoch: 7 [14848/60000 (25%)] Loss: 0.350920
Train Epoch: 7 [30208/60000 (50%)] Loss: 0.367541
Train Epoch: 7 [45568/60000 (75%)] Loss: 0.376175
Test set: Average loss: 0.3527, Accuracy: 9011/10000 (90%)
Train Epoch: 8 [14848/60000 (25%)] Loss: 0.371939
Train Epoch: 8 [30208/60000 (50%)] Loss: 0.364408
Train Epoch: 8 [45568/60000 (75%)] Loss: 0.398144
Test set: Average loss: 0.3379, Accuracy: 9041/10000 (90%)
Train Epoch: 9 [14848/60000 (25%)] Loss: 0.381710
Train Epoch: 9 [30208/60000 (50%)] Loss: 0.350960
Train Epoch: 9 [45568/60000 (75%)] Loss: 0.364110
Test set: Average loss: 0.3270, Accuracy: 9071/10000 (91%)
Train Epoch: 10 [14848/60000 (25%)] Loss: 0.261994
Train Epoch: 10 [30208/60000 (50%)] Loss: 0.329554
Train Epoch: 10 [45568/60000 (75%)] Loss: 0.388473
Test set: Average loss: 0.3165, Accuracy: 9103/10000 (91%)
我们可以看到准确率还有提高的空间，下面用CNN尝试
定义CNN卷积神经网络
class ConvNet(nn.Module):
def __init__(self):
super().__init__()
# batch1*28*28
# 下面的卷积层Conv2d的第一个参数指输入通道数，第二个参数指输出通道数，第三个参数指卷积核的大小
self.conv1 = nn.Conv2d(1,10,5)
self.conv2 = nn.Conv2d(10,20,3)
self.fc1 = nn.Linear(20*10*10,500)# 输入2000，输出500
self.fc2 = nn.Linear(500,10)
def forward(self,x):
in_size = x.size(0)
out = self.conv1(x) # 输入为batch_size*1*28*28的张量 经过5*5卷积 输出batc_size*10*24*24
out = F.relu(out)
out = F.max_pool2d(out,2,2) #2*2池化层 输出batc_size*10*12*12
out = self.conv2(out) # 3*3卷积 输出batc_size*20*10*10
out = F.relu(out)
out = out.view(in_size,-1)# in_size为batch_size 展成batc_size*2000
out = self.fc1(out)# 线性层 2000→500
out = F.relu(out)
out = self.fc2(out)# 线性层 500→10
out = F.log_softmax(out,dim=1) #log softmax
return out
model_cnn = ConvNet().to(DEVICE)
optimizer_cnn = optim.Adam(model_cnn.parameters())
for epoch in range(1,EPOCHS+1):
train(model_cnn,DEVICE,train_loader,optimizer_cnn,epoch)
test(model_cnn,DEVICE,test_loader)
Train Epoch: 1 [14848/60000 (25%)] Loss: 0.386750
Train Epoch: 1 [30208/60000 (50%)] Loss: 0.218018
Train Epoch: 1 [45568/60000 (75%)] Loss: 0.137457
Test set: Average loss: 0.1020, Accuracy: 9667/10000 (97%)
Train Epoch: 2 [14848/60000 (25%)] Loss: 0.099728
Train Epoch: 2 [30208/60000 (50%)] Loss: 0.056601
Train Epoch: 2 [45568/60000 (75%)] Loss: 0.094554
Test set: Average loss: 0.0563, Accuracy: 9819/10000 (98%)
Train Epoch: 3 [14848/60000 (25%)] Loss: 0.085307
Train Epoch: 3 [30208/60000 (50%)] Loss: 0.038580
Train Epoch: 3 [45568/60000 (75%)] Loss: 0.059188
Test set: Average loss: 0.0443, Accuracy: 9851/10000 (99%)
Train Epoch: 4 [14848/60000 (25%)] Loss: 0.036482
Train Epoch: 4 [30208/60000 (50%)] Loss: 0.036975
Train Epoch: 4 [45568/60000 (75%)] Loss: 0.036004
Test set: Average loss: 0.0460, Accuracy: 9850/10000 (98%)
Train Epoch: 5 [14848/60000 (25%)] Loss: 0.022025
Train Epoch: 5 [30208/60000 (50%)] Loss: 0.052678
Train Epoch: 5 [45568/60000 (75%)] Loss: 0.022728
Test set: Average loss: 0.0395, Accuracy: 9886/10000 (99%)
Train Epoch: 6 [14848/60000 (25%)] Loss: 0.032698
Train Epoch: 6 [30208/60000 (50%)] Loss: 0.053429
Train Epoch: 6 [45568/60000 (75%)] Loss: 0.034514
Test set: Average loss: 0.0349, Accuracy: 9880/10000 (99%)
Train Epoch: 7 [14848/60000 (25%)] Loss: 0.007919
Train Epoch: 7 [30208/60000 (50%)] Loss: 0.011484
Train Epoch: 7 [45568/60000 (75%)] Loss: 0.044704
Test set: Average loss: 0.0278, Accuracy: 9918/10000 (99%)
Train Epoch: 8 [14848/60000 (25%)] Loss: 0.018610
Train Epoch: 8 [30208/60000 (50%)] Loss: 0.028454
Train Epoch: 8 [45568/60000 (75%)] Loss: 0.012064
Test set: Average loss: 0.0300, Accuracy: 9909/10000 (99%)
Train Epoch: 9 [14848/60000 (25%)] Loss: 0.016430
Train Epoch: 9 [30208/60000 (50%)] Loss: 0.009306
Train Epoch: 9 [45568/60000 (75%)] Loss: 0.025945
Test set: Average loss: 0.0345, Accuracy: 9896/10000 (99%)
Train Epoch: 10 [14848/60000 (25%)] Loss: 0.032782
Train Epoch: 10 [30208/60000 (50%)] Loss: 0.006292
Train Epoch: 10 [45568/60000 (75%)] Loss: 0.009470
Test set: Average loss: 0.0304, Accuracy: 9903/10000 (99%)
准确率达到了99%
尽管如此，我们知道MNIST是一个很简单的数据集，由于它的局限性只能作为研究用途，对实际应用带来的价值非常有限。如果你的模型连MNIST都搞不定，那么你的模型没有任何的价值即使你的模型搞定了MNIST，你的模型也可能没有任何的价值 但我只会这两个
浙公网安备 33010602011771号