1. 创建训练脚本创建训练脚本 mnist.py，内容如下：
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
# 定义超参数
BATCH_SIZE = 512
EPOCHS = 20
LEARNING_RATE = 1e-3
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# 数据加载和预处理，在 data 目录下载 MNIST 数据集
train_loader = torch.utils.data.DataLoader(
datasets.MNIST('data', train=True, download=True,
transform=transforms.Compose([
transforms.RandomRotation(10), # 随机旋转，提高模型的泛化能力
transforms.RandomAffine(0, shear=10, scale=(0.8, 1.2)), # 仿射变换，提高模型的泛化能力
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=BATCH_SIZE, shuffle=True)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST('data', train=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=BATCH_SIZE, shuffle=False)
# 定义模型的结构，这里是一个简单的卷积神经网络
class ConvNet(nn.Module):
def __init__(self):
super(ConvNet, self).__init__()
self.conv1 = nn.Conv2d(1, 32, 3)
self.conv2 = nn.Conv2d(32, 64, 3)
self.dropout1 = nn.Dropout(0.25)
self.fc1 = nn.Linear(64 * 5 * 5, 128)
self.dropout2 = nn.Dropout(0.5)
self.fc2 = nn.Linear(128, 10)
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.max_pool2d(x, 2)
x = F.relu(self.conv2(x))
x = F.max_pool2d(x, 2)
x = self.dropout1(x)
x = x.view(-1, 64 * 5 * 5)
x = F.relu(self.fc1(x))
x = self.dropout2(x)
x = self.fc2(x)
return F.log_softmax(x, dim=1)
# 初始化模型和优化器
model = ConvNet().to(DEVICE)
optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=5, gamma=0.5) # 学习率衰减
# 训练函数
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(f"Train Epoch: {epoch} [{batch_idx * len(data)}/{len(train_loader.dataset)} "
f"({100. * batch_idx / len(train_loader):.0f}%)]\tLoss: {loss.item():.6f}")
# 测试函数
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.argmax(dim=1, keepdim=True)
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(test_loader.dataset)
accuracy = 100. * correct / len(test_loader.dataset)
print(f"\nTest set: Average loss: {test_loss:.4f}, Accuracy: {correct}/{len(test_loader.dataset)} ({accuracy:.2f}%)\n")
# 训练和测试循环
for epoch in range(1, EPOCHS + 1):
train(model, DEVICE, train_loader, optimizer, epoch)
test(model, DEVICE, test_loader)
scheduler.step() # 更新学习率
# 保存模型
torch.save(model.state_dict(), "mnist_cnn.pth")
print("Model saved to mnist_cnn.pth")
2. 启动训练任务使用指定的卡号的 GPU 进行训练
1
2
3
4
5
docker run --gpus '"device=5"' \
-v "$PWD":/workspace \
-it --rm \
pytorch/pytorch:2.2.2-cuda12.1-cudnn8-runtime \
python mnist.py
可以看到训练过程输出如下日志：
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
Train Epoch: 1 [14848/60000 (25%)] Loss: 0.628275
Train Epoch: 1 [30208/60000 (50%)] Loss: 0.372206
Train Epoch: 1 [45568/60000 (75%)] Loss: 0.319735
Test set: Average loss: 0.0831, Accuracy: 9726/10000 (97.26%)
Train Epoch: 2 [14848/60000 (25%)] Loss: 0.165549
Train Epoch: 2 [30208/60000 (50%)] Loss: 0.184633
Train Epoch: 2 [45568/60000 (75%)] Loss: 0.112471
Test set: Average loss: 0.0579, Accuracy: 9805/10000 (98.05%)
Train Epoch: 3 [14848/60000 (25%)] Loss: 0.137946
Train Epoch: 3 [30208/60000 (50%)] Loss: 0.154524
Train Epoch: 3 [45568/60000 (75%)] Loss: 0.159434
...
Test set: Average loss: 0.0206, Accuracy: 9926/10000 (99.26%)
Train Epoch: 18 [14848/60000 (25%)] Loss: 0.071504
Train Epoch: 18 [30208/60000 (50%)] Loss: 0.050344
Train Epoch: 18 [45568/60000 (75%)] Loss: 0.070272
Test set: Average loss: 0.0198, Accuracy: 9929/10000 (99.29%)
Train Epoch: 19 [14848/60000 (25%)] Loss: 0.039598
Train Epoch: 19 [30208/60000 (50%)] Loss: 0.067286
Train Epoch: 19 [45568/60000 (75%)] Loss: 0.048905
Test set: Average loss: 0.0191, Accuracy: 9929/10000 (99.29%)
Train Epoch: 20 [14848/60000 (25%)] Loss: 0.060975
Train Epoch: 20 [30208/60000 (50%)] Loss: 0.087753
Train Epoch: 20 [45568/60000 (75%)] Loss: 0.067398
Test set: Average loss: 0.0189, Accuracy: 9936/10000 (99.36%)
可以看到，loss 值从最开始的 0.628275 经过若干 epoch 之后降到了 0.07 左右。这意味着模型的误差在不断减小，准确率在不断提高。
3. 创建测试脚本为了验证模型在真实数据上的表现，我们可以使用一个测试脚本 test.py，内容如下：
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
import sys
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision import transforms
from PIL import Image
# 定义与训练时相同的网络结构
class ConvNet(nn.Module):
def __init__(self):
super(ConvNet, self).__init__()
self.conv1 = nn.Conv2d(1, 32, 3)
self.conv2 = nn.Conv2d(32, 64, 3)
self.dropout1 = nn.Dropout(0.25)
self.fc1 = nn.Linear(64 * 5 * 5, 128)
self.dropout2 = nn.Dropout(0.5)
self.fc2 = nn.Linear(128, 10)
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.max_pool2d(x, 2)
x = F.relu(self.conv2(x))
x = F.max_pool2d(x, 2)
x = self.dropout1(x)
x = x.view(-1, 64 * 5 * 5)
x = F.relu(self.fc1(x))
x = self.dropout2(x)
x = self.fc2(x)
return F.log_softmax(x, dim=1)
# 加载模型
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = ConvNet().to(DEVICE)
model.load_state_dict(torch.load("mnist_cnn.pth", map_location=DEVICE))
model.eval()
# 图片预处理
def preprocess_image(image_path):
transform = transforms.Compose([
transforms.Grayscale(num_output_channels=1), # 确保是单通道（灰度图）
transforms.Resize((28, 28)), # 调整大小为28x28
transforms.ToTensor(), # 转为Tensor
transforms.Normalize((0.1307,), (0.3081,)) # 标准化
])
image = Image.open(image_path).convert('RGB') # 确保图片为RGB模式，避免报错
return transform(image).unsqueeze(0) # 添加批次维度
# 预测函数
def predict(image_path):
try:
image_tensor = preprocess_image(image_path).to(DEVICE)
with torch.no_grad():
output = model(image_tensor)
pred = output.argmax(dim=1, keepdim=True)
print(f"Predicted Digit: {pred.item()}")
except Exception as e:
print(f"Error processing image '{image_path}': {e}")
# 脚本入口
if __name__ == "__main__":
if len(sys.argv) != 2:
print("Usage: python test.py
sys.exit(1)
image_path = sys.argv[1]
predict(image_path)
EOF
在使用时，只需要指定图片文件的路径即可：
1
python test.py 4.jpg
4. 启动测试任务下载图片，准备测试数据这里在网上找了一些手写数字图片，用于测试模型的准确率。
1
wget https://img.redocn.com/sheji/20211030/shouxieshuzi0yishuzi_11822055.jpg -O 0.jpg
1
wget https://img-blog.csdn.net/20170413123610276?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvQ2FsbE1lR29EZW5n/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast -O 3.jpg
1
wget https://img.redocn.com/sheji/20211030/shouxieshuzi4yishuzi_11822063.jpg.400.jpg -O 4.jpg
1
wget https://img.redocn.com/sheji/20211030/shouxieshuzi6yishuzi_11822067.jpg.400.jpg -O 6.jpg
1
wget https://code.lardcave.net/2015/12/06/1/67CDB2E1-C19C-4CE7-94A3-32AE4B98F4C8@local -O 7.jpg
测试模型1
2
3
4
5
6
7
for i in 0 3 4 6 7 ; do
docker run --gpus '"device=5"' \
-v "$PWD":/workspace \
-it --rm \
pytorch/pytorch:2.2.2-cuda12.1-cudnn8-runtime \
python test.py $i.jpg
done
1
2
3
4
5
Predicted Digit: 0
Predicted Digit: 3
Predicted Digit: 4
Predicted Digit: 6
Predicted Digit: 0
5. 计算准确率图片文件实际标签预测标签是否正确0.jpg00是3.jpg33是4.jpg44是6.jpg66是7.jpg70否准确率 = 正确预测数 / 总预测数 = 4 / 5 = 80%