cifar10 - pytorch - 模型源文件、train、test、use等源代码，kaggle上训练好的模型

# @Time : 2022/8/25 15:02 # @Author :lgl # @e-mail :GuanlinLi_BIT@163.com import matplotlib.pyplot as plt import glob import time import numpy import torch import torchvision from torch.optim import SGD import winsound from res_model import * from model import * from torch.utils.data import DataLoader from torch import nn from torch.utils.tensorboard import SummaryWriter # 把网络模型、数据（输入和标注）、损失函数加上 .cuda 就可以用GPU old_time = time.time() # 创建日志 writer = SummaryWriter("../logs") # 加载数据 train_data = torchvision.datasets.CIFAR10(r"D:\AI_project\20220825_cifar10demo\cifar10", train=True, transform=torchvision.transforms.ToTensor(), download=True) train_dataloader = DataLoader(train_data, batch_size=64, shuffle=True) train_len = len(train_data) # 训练数据集长度 print("训练数据集长度：{}".format(train_len)) # 加载测试数据 test_data = torchvision.datasets.CIFAR10(r"D:\AI_project\20220825_cifar10demo\cifar10", train=False, transform=torchvision.transforms.ToTensor(), download=True) test_dataloader = DataLoader(test_data, batch_size=64, shuffle=True) test_len = len(test_data) # 训练数据集长度 # 加载最新历史模型 path_file_num = glob.glob(r'D:\AI_project\20220825_cifar10demo\model\*.pth') # 匹配给定路径下的所有pattern，并以列表形式返回。 pth_num = len(path_file_num) # 读取路径中有多少.pth文件 if pth_num != 0: history_model = torch.load(r"D:\AI_project\20220825_cifar10demo\model\neural_networks_all_{}.pth".format(pth_num)) print("当前最新历史数据：{}".format("neural_networks_all_{}.pth".format(pth_num))) # 选择训练的模型。 neural_networks 或者 history_model model = res_neural_networks print("残差f网络结构： {}".format(res_neural_networks)) print("网络结构： {}".format(neural_networks)) if torch.cuda.is_available(): model = model.cuda() print("GPU is working") else: print("CPU is working") # 设置损失函数 ce_loss = nn.CrossEntropyLoss() # 交叉熵损失函数，分类问题常用 if torch.cuda.is_available(): ce_loss = ce_loss.cuda() # 设置学习速率 # 设置优化器 # 记录训练次数 total_train_step = 0 # 本次训练轮数 epoch = 10 for i in range(epoch): # 设置学习率 learning_rate_max = 0.05 learning_rate_min = 0.001 learning_rate = learning_rate_max - (learning_rate_max - learning_rate_min) * (i + 1) / epoch print("learning_rate:{}".format(learning_rate)) # 设置优化器 opt = SGD(model.parameters(), lr=learning_rate) # 随机梯度下降 print("\n\n---------第{}轮训练开始---------".format(i + 1)) # loss_history = numpy.load(r"D:\AI_project\20220825_cifar10demo\loss_history\loss_history.npy") total_train_loss = 0.0 min_total_loss = 1400.0 # 最优模型loss值的记录 for data in train_dataloader: imgs, targets = data if torch.cuda.is_available(): imgs = imgs.cuda() targets = targets.cuda() output = model(imgs) # 载入数据 loss = ce_loss(output, targets) # 计算loss # 优化器 opt.zero_grad() # 将之前的梯度清零，准备好这一轮的优化 loss.backward() # 反向传播，求梯度 opt.step() # 执行优化 total_train_loss += loss total_train_step += 1 # 每隔100次显示训练情况 if total_train_step % 100 == 0: print("训练次数：{}, Loss:{}".format(total_train_step, loss.item())) # 本轮训练结束 print("本轮总Loss:{}".format(total_train_loss)) # 由于total_train_loss是张量，所以要转换成numpy类型保存 # new_loss_history = numpy.append(loss_history, total_train_loss.detach().numpy()) # numpy.save(r"D:\AI_project\20220825_cifar10demo\loss_history\loss_history.npy", new_loss_history) # print("len hisloss:{}".format(len(new_loss_history))) writer.add_scalar("loss", total_train_loss, i) if i == 0: min_total_loss = total_train_loss if total_train_loss <= min_total_loss: path_file_num = glob.glob(r'D:\AI_project\20220825_cifar10demo\model\*.pth') pth_num = len(path_file_num) # 读取路径中有多少.pth文件 if i == 0: pth_num += 1 torch.save(model, r"D:\AI_project\20220825_cifar10demo\model\neural_networks_all_{}.pth".format(pth_num)) print("save, loss={}".format(total_train_loss)) min_total_loss = total_train_loss # 新的最优网络出现提示音 ############### winsound.Beep(222, 500) # 主板蜂鸣器 ########### winsound.MessageBeep() # 喇叭 # 跑完一轮，提示音 ######## # winsound.Beep(2222, 500) # 主板蜂鸣器 ########### winsound.MessageBeep() # 喇叭 # 每次隔10轮测试一次准确率 if i % 2 == 0: total_accuracy = 0.0 for data in test_dataloader: test_imgs, test_targets = data if torch.cuda.is_available(): test_imgs = test_imgs.cuda() test_targets = test_targets.cuda() test_output = model(test_imgs) # 载入数据 accuracy = (test_output.argmax(1) == test_targets).sum() # 每轮64个图片里面对了几个 argmax(1) total_accuracy += accuracy print("整体测试集上的准确率：{:.2%}".format(total_accuracy / test_len)) current_time = time.time() writer.close() print("\n\n本次训练共训练{}轮".format(epoch)) print("运行时间为 {} s".format(current_time - old_time)) # loss_history = #numpy.load(r"D:\AI_project\20220825_cifar10demo\loss_history\loss_history.npy") # p#rint("历史loss情况:{}".format(los#s_history)) # plt.plot(range(len(loss_history)), loss_history) # plt.title('history_loss') # plt.show() # 显示 # 在Terminal里面输入：tensorboard --logdir D:\AI_project\20220825_cifar10demo\logs 可以打开网页，查看lost