实验环境
win10 + anaconda + jupyter notebook
Pytorch1.1.0
Python3.7
gpu环境（可选）
MNIST数据集介绍
MNIST 包括6万张28x28的训练样本，1万张测试样本，可以说是CV里的“Hello Word”。本文使用的CNN网络将MNIST数据的识别率提高到了99%。下面我们就开始进行实战。
导入包
|
1
2
3
4
5
6
|
import torchimport torch.nn as nnimport torch.nn.functional as Fimport torch.optim as optimfrom torchvision import datasets, transformstorch.__version__ |
定义超参数
|
1
2
3
|
BATCH_SIZE=512EPOCHS=20DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu") |
数据集
我们直接使用PyTorch中自带的dataset，并使用DataLoader对训练数据和测试数据分别进行读取。如果下载过数据集这里download可选择False
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
|
train_loader = torch.utils.data.DataLoader( datasets.MNIST('data', train=True, download=True, transform=transforms.Compose([ 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=True) |
定义网络
该网络包括两个卷积层和两个线性层，最后输出10个维度，即代表0-9十个数字。
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
|
class ConvNet(nn.Module): def __init__(self): super().__init__() self.conv1=nn.Conv2d(1,10,5) # input:(1,28,28) output:(10,24,24) self.conv2=nn.Conv2d(10,20,3) # input:(10,12,12) output:(20,10,10) self.fc1 = nn.Linear(20*10*10,500) self.fc2 = nn.Linear(500,10) def forward(self,x): in_size = x.size(0) out = self.conv1(x) out = F.relu(out) out = F.max_pool2d(out, 2, 2) out = self.conv2(out) out = F.relu(out) out = out.view(in_size,-1) out = self.fc1(out) out = F.relu(out) out = self.fc2(out) out = F.log_softmax(out,dim=1) return out |
实例化网络
|
1
2
|
model = ConvNet().to(DEVICE) # 将网络移动到gpu上optimizer = optim.Adam(model.parameters()) # 使用Adam优化器 |
定义训练函数
|
1
2
3
4
5
6
7
8
9
10
11
12
13
|
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())) |
定义测试函数
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
|
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() # 将一批的损失相加 pred = output.max(1, keepdim=True)[1] # 找到概率最大的下标 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))) |
开始训练
|
1
2
3
|
for epoch in range(1, EPOCHS + 1): train(model, DEVICE, train_loader, optimizer, epoch) test(model, DEVICE, test_loader) |
实验结果
|
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
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
|
Train Epoch: 1 [14848/60000 (25%)] Loss: 0.375058Train Epoch: 1 [30208/60000 (50%)] Loss: 0.255248Train Epoch: 1 [45568/60000 (75%)] Loss: 0.128060Test set: Average loss: 0.0992, Accuracy: 9690/10000 (97%)Train Epoch: 2 [14848/60000 (25%)] Loss: 0.093066Train Epoch: 2 [30208/60000 (50%)] Loss: 0.087888Train Epoch: 2 [45568/60000 (75%)] Loss: 0.068078Test set: Average loss: 0.0599, Accuracy: 9816/10000 (98%)Train Epoch: 3 [14848/60000 (25%)] Loss: 0.043926Train Epoch: 3 [30208/60000 (50%)] Loss: 0.037321Train Epoch: 3 [45568/60000 (75%)] Loss: 0.068404Test set: Average loss: 0.0416, Accuracy: 9859/10000 (99%)Train Epoch: 4 [14848/60000 (25%)] Loss: 0.031654Train Epoch: 4 [30208/60000 (50%)] Loss: 0.041341Train Epoch: 4 [45568/60000 (75%)] Loss: 0.036493Test set: Average loss: 0.0361, Accuracy: 9873/10000 (99%)Train Epoch: 5 [14848/60000 (25%)] Loss: 0.027688Train Epoch: 5 [30208/60000 (50%)] Loss: 0.019488Train Epoch: 5 [45568/60000 (75%)] Loss: 0.018023Test set: Average loss: 0.0344, Accuracy: 9875/10000 (99%)Train Epoch: 6 [14848/60000 (25%)] Loss: 0.024212Train Epoch: 6 [30208/60000 (50%)] Loss: 0.018689Train Epoch: 6 [45568/60000 (75%)] Loss: 0.040412Test set: Average loss: 0.0350, Accuracy: 9879/10000 (99%)Train Epoch: 7 [14848/60000 (25%)] Loss: 0.030426Train Epoch: 7 [30208/60000 (50%)] Loss: 0.026939Train Epoch: 7 [45568/60000 (75%)] Loss: 0.010722Test set: Average loss: 0.0287, Accuracy: 9892/10000 (99%)Train Epoch: 8 [14848/60000 (25%)] Loss: 0.021109Train Epoch: 8 [30208/60000 (50%)] Loss: 0.034845Train Epoch: 8 [45568/60000 (75%)] Loss: 0.011223Test set: Average loss: 0.0299, Accuracy: 9904/10000 (99%)Train Epoch: 9 [14848/60000 (25%)] Loss: 0.011391Train Epoch: 9 [30208/60000 (50%)] Loss: 0.008091Train Epoch: 9 [45568/60000 (75%)] Loss: 0.039870Test set: Average loss: 0.0341, Accuracy: 9890/10000 (99%)Train Epoch: 10 [14848/60000 (25%)] Loss: 0.026813Train Epoch: 10 [30208/60000 (50%)] Loss: 0.011159Train Epoch: 10 [45568/60000 (75%)] Loss: 0.024884Test set: Average loss: 0.0286, Accuracy: 9901/10000 (99%)Train Epoch: 11 [14848/60000 (25%)] Loss: 0.006420Train Epoch: 11 [30208/60000 (50%)] Loss: 0.003641Train Epoch: 11 [45568/60000 (75%)] Loss: 0.003402Test set: Average loss: 0.0377, Accuracy: 9894/10000 (99%)Train Epoch: 12 [14848/60000 (25%)] Loss: 0.006866Train Epoch: 12 [30208/60000 (50%)] Loss: 0.012617Train Epoch: 12 [45568/60000 (75%)] Loss: 0.008548Test set: Average loss: 0.0311, Accuracy: 9908/10000 (99%)Train Epoch: 13 [14848/60000 (25%)] Loss: 0.010539Train Epoch: 13 [30208/60000 (50%)] Loss: 0.002952Train Epoch: 13 [45568/60000 (75%)] Loss: 0.002313Test set: Average loss: 0.0293, Accuracy: 9905/10000 (99%)Train Epoch: 14 [14848/60000 (25%)] Loss: 0.002100Train Epoch: 14 [30208/60000 (50%)] Loss: 0.000779Train Epoch: 14 [45568/60000 (75%)] Loss: 0.005952Test set: Average loss: 0.0335, Accuracy: 9897/10000 (99%)Train Epoch: 15 [14848/60000 (25%)] Loss: 0.006053Train Epoch: 15 [30208/60000 (50%)] Loss: 0.002559Train Epoch: 15 [45568/60000 (75%)] Loss: 0.002555Test set: Average loss: 0.0357, Accuracy: 9894/10000 (99%)Train Epoch: 16 [14848/60000 (25%)] Loss: 0.000895Train Epoch: 16 [30208/60000 (50%)] Loss: 0.004923Train Epoch: 16 [45568/60000 (75%)] Loss: 0.002339Test set: Average loss: 0.0400, Accuracy: 9893/10000 (99%)Train Epoch: 17 [14848/60000 (25%)] Loss: 0.004136Train Epoch: 17 [30208/60000 (50%)] Loss: 0.000927Train Epoch: 17 [45568/60000 (75%)] Loss: 0.002084Test set: Average loss: 0.0353, Accuracy: 9895/10000 (99%)Train Epoch: 18 [14848/60000 (25%)] Loss: 0.004508Train Epoch: 18 [30208/60000 (50%)] Loss: 0.001272Train Epoch: 18 [45568/60000 (75%)] Loss: 0.000543Test set: Average loss: 0.0380, Accuracy: 9894/10000 (99%)Train Epoch: 19 [14848/60000 (25%)] Loss: 0.001699Train Epoch: 19 [30208/60000 (50%)] Loss: 0.000661Train Epoch: 19 [45568/60000 (75%)] Loss: 0.000275Test set: Average loss: 0.0339, Accuracy: 9905/10000 (99%)Train Epoch: 20 [14848/60000 (25%)] Loss: 0.000441Train Epoch: 20 [30208/60000 (50%)] Loss: 0.000695Train Epoch: 20 [45568/60000 (75%)] Loss: 0.000467Test set: Average loss: 0.0396, Accuracy: 9894/10000 (99%) |
总结
一个实际项目的工作流程：找到数据集，对数据做预处理，定义我们的模型，调整超参数，测试训练，再通过训练结果对超参数进行调整或者对模型进行调整。
以上这篇使用PyTorch实现MNIST手写体识别代码就是小编分享给大家的全部内容了，希望能给大家一个参考，也希望大家多多支持服务器之家。
原文链接：https://blog.csdn.net/abcgkj/article/details/100884143
