文件结构：

data是放数据集，img是放模型训练完成之后测试的，module放模型，model是模型，test是测试img中的，train是训练模型

模型

# -*- coding: utf-8 -*-from torch import nnclass Model(nn.Module):def __init__(self):super(Model, self).__init__()self.model = nn.Sequential(nn.Conv2d(in_channels=3, out_channels=32, kernel_size=5, padding=2), # 卷积操作nn.MaxPool2d(kernel_size=2), # 最大池化nn.Conv2d(in_channels=32, out_channels=32, kernel_size=5, padding=2),nn.MaxPool2d(kernel_size=2),nn.Conv2d(32, 64, 5, padding=2),nn.MaxPool2d(kernel_size=2),nn.Flatten(), # 展开nn.Linear(1024, 64), # 线形变换nn.Linear(64, 10) # 线形变换)def forward(self, input):output = self.model(input)return output

加载数据集

该数据集Pytorch可以直接下载，直接download=True，若是现在太慢，可以在迅雷下载

import torchvision
train_dataset = torchvision.datasets.CIFAR10(root='./data',train=True,transform=torchvision.transforms.ToTensor(),download=True)
test_dataset = torchvision.datasets.CIFAR10(root='./data',train=False,transform=torchvision.transforms.ToTensor(),download=True)

训练

其中可以调用gpu进行训练，若是没有gpu则会选用cpu训练，训练速度lr =0.001，用的SGD做优化器，平方差计算loss，训练次数为500次，并且如果验证集的准确率大于0.68

# -*- coding: utf-8 -*-from torch.utils.data import DataLoader
from model import Model
import torch
import torchvision
import torch.nn as nndevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')train_dataset = torchvision.datasets.CIFAR10(root='./data',train=True,transform=torchvision.transforms.ToTensor(),download=True)
test_dataset = torchvision.datasets.CIFAR10(root='./data',train=False,transform=torchvision.transforms.ToTensor(),download=True)train_len = len(train_dataset)
test_len = len(test_dataset)
print(f"训练数据量有{train_len}个")
print(f"测试数据量有{test_len}个")# 加载数据
train_loader = DataLoader(train_dataset, batch_size=64)
test_loader = DataLoader(test_dataset, batch_size=64)# 加载网络
model = Model().to(device)
# 损失函数
losser = nn.CrossEntropyLoss()
losser.to(device)
# 优化器
speed = 0.001
optimzer = torch.optim.SGD(model.parameters(), lr=speed)# 训练轮数
total_train_step = 500# 记录训练次数和测试次数model.train()
for i in range(1,total_train_step+1):print(f"------第{i}轮训练开始------")train_loss = 0for data in train_loader:img, target = dataimg, target = img.to(device), target.to(device)output = model(img)loss = losser(output, target)# 模型优化optimzer.zero_grad()loss.backward()optimzer.step()train_loss += lossprint(f"\t\t训练Loss：{loss}")# 测试数据model.eval()total_test_Loss = 0total_accuracy = 0with torch.no_grad():print(f"\t测试开始")for data in test_loader:img, target = dataimg, target = img.to(device), target.to(device)output = model(img)loss = losser(output, target)total_test_Loss += lossaccuracy = (output.argmax(1) == target).sum()total_accuracy += accuracytotal_accuracy = int(total_accuracy)print(f"\t\t测试总损失：{total_test_Loss}")print(f"\t\t测试正确率：{round((total_accuracy/test_len), 4)*100}%")if (total_accuracy/test_len) > 0.68:torch.save(model.state_dict(), f"./module/CIFAR_{i}_{round((total_accuracy/test_len), 2)}.pth")print(f"------第{i}轮模型已保存------")

测试模型

注意：png图片要转化成RGB图片，CIFAR总类顺序是'airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck'。并且在测试是不需要梯度。

# -*- coding: utf-8 -*-import torch
from model import Model
from PIL import Image
import torchvisionmodel = Model()
model.load_state_dict(torch.load('./module/CIFAR_313_0.68.pth'))kind = ['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']img = Image.open('./img/img.png')transf = torchvision.transforms.Compose([torchvision.transforms.Resize((32,32)),torchvision.transforms.ToTensor()])
img = transf(img)
img = torch.reshape(img, (1,3,32,32))
model.eval()
with torch.no_grad():output = model(img)print(output.argmax(1))

总结

这个模型训练最高的准确率在68.9%左右，这也是这个模型能训练最高的准确率。如果你发现后面的准确率在降低，但是训练集loss也在减少，说明模型过拟合了，就可以停止。

如果想要再度提高准确率，你可以选择增加数据集的数据量，或者改变模型（用残差等）