笔记3：torch训练测试VGG网络

（1）利用Netron查看网络实际情况

上图链接
python生成上图代码如下，其中GETVGGnet是搭建VGG网络的程序GETVGGnet.py，VGGnet是该程序中的搭建网络类。netron是需要pip安装的可视化库，注意do_constant_folding=False可以防止Netron中不显示Batchnorm2D层，禁用参数隐藏。

import torch
from torch.autograd import Variable
from GetVGGnet import VGGnet
import netron

net = VGGnet()
x = Variable(torch.FloatTensor(1,3,28,28))
y = net(x)
print(y.data.shape)
onnx_path = "./save_model/VGGnet.onnx"
torch.onnx.export(net, x, onnx_path,do_constant_folding=False)
print(net)
netron.start(onnx_path)

（2）VGG训练测试全过程

此次训练在CPU上进行，迭代次epoch = 10，迭代内轮次batch=300，训练集10000张，测试集2000张。
train loss和train corre分别代表损失和正确率，横轴是不同迭代下每一个伦次的loss&corre累加，一个迭代进行33个轮次，每个迭代最后一个伦次数据不足被网络舍弃，10个迭代总共320次。test loss和test corre是每个一个迭代下所有伦次的正确率平均值。根据图可以看出，训练和测试结果都较好。

训练的损失和正确率在波动，但总体趋势较好。

数据集大小可以在此处修改：

代码：cifar10_handle和GetVGGnet在上几篇文章有说明

#!/usr/bin/env python
# -*- coding:utf-8 -*-
"""
@author: 楠楠星球
@time: 2024/5/10 10:15 
@file: VGGTrain.py-->test
@project: pythonProject
@# ------------------------------------------(one)--------------------------------------
@# ------------------------------------------(two)--------------------------------------
"""
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from GetVGGnet import VGGnet
from cifar10_handle import train_dataset,test_dataset
import matplotlib.pyplot as plt

epoch = 10  #迭代次数
learn_rate = 0.01 #初始学习率

net = VGGnet().to(device='cpu') #模型实例化
loss_fun = nn.CrossEntropyLoss() #调用损失函数
train_data_loder = DataLoader(dataset=train_dataset,
                              batch_size=300,  #每一次迭代的调用的波次
                              shuffle=True,    #这个波次是否打乱数据集
                              num_workers=4,   # 线程数
                              drop_last=True)  # 最后一个波次数据不足是否舍去

test_data_loder = DataLoader(dataset=test_dataset,
                             batch_size=300,
                             shuffle=False,
                             num_workers=4,
                             drop_last=True)

# optimizer = torch.optim.Adam(net.parameters(), lr=learn_rate)
optimizer = torch.optim.SGD(net.parameters(), lr=learn_rate, momentum=0.5) #优化器

# scheduler = torch.optim.lr_scheduler.StepLR(optijumizer, step_size=5, gamma=0.9) #step_size=1表示每迭代一次更新一下学习率
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.7) #学习率调整器


def train(epoch_num,train_net):
	# ------------------------------------------()--------------------------------------
	loss_base = []
	corre_base = []
	
	test_loss_base = []
	test_corre_base =[]
	for epoch in range(epoch_num):
		# ------------------------------------------(TRAIN)--------------------------------------
		train_net.train()
		for i, data in enumerate(train_data_loder):
			input_tensor, label = data
			input_tensor = input_tensor.to(device='cpu')
			label = label.to(device='cpu')
			
			output_tensor = train_net(input_tensor)
			loss = loss_fun(output_tensor, label)
			
			optimizer.zero_grad()
			loss.backward()
			optimizer.step()
			
			_, pred = torch.max(output_tensor.data, dim=1)
			correct = pred.eq(label.data).cpu().sum()
			
			print(f"训练中：第{epoch + 1}次迭代的小迭代{i}的损失率为：{1.00 * loss.item()},正确率为：{100.00 * correct / 300}")
			loss_base.append(loss.item())
			corre_base.append(100.00 * correct.item() / 300)
	
		scheduler.step()
		
		# ------------------------------------------(TEST)--------------------------------------
		sum_test_loss = 0
		sum_test_corre = 0
		train_net.eval()
		for i, test_data in enumerate(test_data_loder):
			input_tensor, label = test_data
			input_tensor = input_tensor.to(device='cpu')
			label = label.to(device='cpu')
			
			output_tensor = train_net(input_tensor)
			loss = loss_fun(output_tensor, label)
			_, pred = torch.max(output_tensor.data, dim=1)
			correct = pred.eq(label.data).cpu().sum()

			sum_test_loss += loss.item()
			sum_test_corre += correct.item()
			
		
		test_loss = sum_test_loss * 1.0 / len(test_data_loder)
		test_corre = sum_test_corre * 100.0 / len(test_data_loder) / 300
		test_loss_base.append(test_loss)
		test_corre_base.append(test_corre)
		print(f"测试中：当前迭代的测试集损失为：{test_loss},正确率为：{test_corre}")
	return loss_base,corre_base,test_loss_base,test_corre_base
	# ------------------------------------------()--------------------------------------

if __name__ == '__main__':
	[train_loss,train_corre,test_loss,test_corr] = train(epoch,net)
	fig, axes = plt.subplots(2, 2)
	
	axes[0, 0].plot(list(range(1, len(train_loss)+1 )), train_loss,color ='r')
	axes[0, 0].set_title('train loss')
	
	axes[0, 1].plot(list(range(1, len(train_corre) + 1)), train_corre, color ='r')
	axes[0, 1].set_title('train corre')
	
	axes[1, 0].plot(list(range(1, len(test_loss) + 1)), test_loss,color ='r')
	axes[1, 0].set_title('test loss')

	axes[1, 1].plot(list(range(1, len(test_corr) + 1)), test_corr,color ='r')
	axes[1, 1].set_title('test corre')
	plt.show()
	
	# torch.save(net.state_dict(), './save_model/example1.pt')