第P7周：马铃薯病害识别（VGG-16复现）

「声明」文为「365天深度学习训练营」内部文章一、 前期准备1. 设置GPU如果设备上支持GPU就使用GPU,否则使用CPUimporttorchimporttorch.nnasnnimporttorchvision.transformsastransformsimporttorchvisionfromtorchvisionimporttransforms,datasetsimportos,PIL,pathlib,warnings warnings.filterwarnings(ignore)#忽略警告信息devicetorch.device(cudaiftorch.cuda.is_available()elsecpu)device2. 导入数据importos,PIL,random,pathlib data_dir./PotatoPlants/data_dirpathlib.Path(data_dir)data_pathslist(data_dir.glob(*))classeNames[str(path).split(\\)[1]forpathindata_paths]classeNames# 关于transforms.Compose的更多介绍可以参考https://blog.csdn.net/qq_38251616/article/details/124878863train_transformstransforms.Compose([transforms.Resize([224,224]),# 将输入图片resize成统一尺寸# transforms.RandomHorizontalFlip(), # 随机水平翻转transforms.ToTensor(),# 将PIL Image或numpy.ndarray转换为tensor并归一化到[0,1]之间transforms.Normalize(# 标准化处理--转换为标准正太分布高斯分布使模型更容易收敛mean[0.485,0.456,0.406],std[0.229,0.224,0.225])# 其中 mean[0.485,0.456,0.406]与std[0.229,0.224,0.225] 从数据集中随机抽样计算得到的。])test_transformtransforms.Compose([transforms.Resize([224,224]),# 将输入图片resize成统一尺寸transforms.ToTensor(),# 将PIL Image或numpy.ndarray转换为tensor并归一化到[0,1]之间transforms.Normalize(# 标准化处理--转换为标准正太分布高斯分布使模型更容易收敛mean[0.485,0.456,0.406],std[0.229,0.224,0.225])# 其中 mean[0.485,0.456,0.406]与std[0.229,0.224,0.225] 从数据集中随机抽样计算得到的。])total_datadatasets.ImageFolder(./PotatoPlants/,transformtrain_transforms)total_datatotal_data.class_to_idx3. 划分数据集train_sizeint(0.8*len(total_data))test_sizelen(total_data)-train_size train_dataset,test_datasettorch.utils.data.random_split(total_data,[train_size,test_size])train_dataset,test_datasetbatch_size32train_dltorch.utils.data.DataLoader(train_dataset,batch_sizebatch_size,shuffleTrue,num_workers1)test_dltorch.utils.data.DataLoader(test_dataset,batch_sizebatch_size,shuffleTrue,num_workers1)forX,yintest_dl:print(Shape of X [N, C, H, W]: ,X.shape)print(Shape of y: ,y.shape,y.dtype)break二、手动搭建VGG-16模型1. 搭建模型importtorch.nn.functionalasFclassvgg16(nn.Module):def__init__(self):super(vgg16,self).__init__()# 卷积块1self.block1nn.Sequential(nn.Conv2d(3,64,kernel_size(3,3),stride(1,1),padding(1,1)),nn.ReLU(),nn.Conv2d(64,64,kernel_size(3,3),stride(1,1),padding(1,1)),nn.ReLU(),nn.MaxPool2d(kernel_size(2,2),stride(2,2)))# 卷积块2self.block2nn.Sequential(nn.Conv2d(64,128,kernel_size(3,3),stride(1,1),padding(1,1)),nn.ReLU(),nn.Conv2d(128,128,kernel_size(3,3),stride(1,1),padding(1,1)),nn.ReLU(),nn.MaxPool2d(kernel_size(2,2),stride(2,2)))# 卷积块3self.block3nn.Sequential(nn.Conv2d(128,256,kernel_size(3,3),stride(1,1),padding(1,1)),nn.ReLU(),nn.Conv2d(256,256,kernel_size(3,3),stride(1,1),padding(1,1)),nn.ReLU(),nn.Conv2d(256,256,kernel_size(3,3),stride(1,1),padding(1,1)),nn.ReLU(),nn.MaxPool2d(kernel_size(2,2),stride(2,2)))# 卷积块4self.block4nn.Sequential(nn.Conv2d(256,512,kernel_size(3,3),stride(1,1),padding(1,1)),nn.ReLU(),nn.Conv2d(512,512,kernel_size(3,3),stride(1,1),padding(1,1)),nn.ReLU(),nn.Conv2d(512,512,kernel_size(3,3),stride(1,1),padding(1,1)),nn.ReLU(),nn.MaxPool2d(kernel_size(2,2),stride(2,2)))# 卷积块5self.block5nn.Sequential(nn.Conv2d(512,512,kernel_size(3,3),stride(1,1),padding(1,1)),nn.ReLU(),nn.Conv2d(512,512,kernel_size(3,3),stride(1,1),padding(1,1)),nn.ReLU(),nn.Conv2d(512,512,kernel_size(3,3),stride(1,1),padding(1,1)),nn.ReLU(),nn.MaxPool2d(kernel_size(2,2),stride(2,2)))# 全连接网络层用于分类self.classifiernn.Sequential(nn.Linear(in_features512*7*7,out_features4096),nn.ReLU(),nn.Linear(in_features4096,out_features4096),nn.ReLU(),nn.Linear(in_features4096,out_features3))defforward(self,x):xself.block1(x)xself.block2(x)xself.block3(x)xself.block4(x)xself.block5(x)xtorch.flatten(x,start_dim1)xself.classifier(x)returnx devicecudaiftorch.cuda.is_available()elsecpuprint(Using {} device.format(device))modelvgg16().to(device)model2. 查看模型详情# 统计模型参数量以及其他指标importtorchsummaryassummary summary.summary(model,(3,224,224))三、 训练模型1. 编写训练函数# 训练循环deftrain(dataloader,model,loss_fn,optimizer):sizelen(dataloader.dataset)# 训练集的大小num_batcheslen(dataloader)# 批次数目, (size/batch_size向上取整)train_loss,train_acc0,0# 初始化训练损失和正确率forX,yindataloader:# 获取图片及其标签X,yX.to(device),y.to(device)# 计算预测误差predmodel(X)# 网络输出lossloss_fn(pred,y)# 计算网络输出和真实值之间的差距targets为真实值计算二者差值即为损失# 反向传播optimizer.zero_grad()# grad属性归零loss.backward()# 反向传播optimizer.step()# 每一步自动更新# 记录acc与losstrain_acc(pred.argmax(1)y).type(torch.float).sum().item()train_lossloss.item()train_acc/size train_loss/num_batchesreturntrain_acc,train_loss3. 编写测试函数测试函数和训练函数大致相同但是由于不进行梯度下降对网络权重进行更新所以不需要传入优化器deftest(dataloader,model,loss_fn):sizelen(dataloader.dataset)# 测试集的大小num_batcheslen(dataloader)# 批次数目, (size/batch_size向上取整)test_loss,test_acc0,0# 当不进行训练时停止梯度更新节省计算内存消耗withtorch.no_grad():forimgs,targetindataloader:imgs,targetimgs.to(device),target.to(device)# 计算losstarget_predmodel(imgs)lossloss_fn(target_pred,target)test_lossloss.item()test_acc(target_pred.argmax(1)target).type(torch.float).sum().item()test_acc/size test_loss/num_batchesreturntest_acc,test_loss4. 正式训练importcopy optimizertorch.optim.Adam(model.parameters(),lr1e-4)loss_fnnn.CrossEntropyLoss()# 创建损失函数epochs40train_loss[]train_acc[]test_loss[]test_acc[]best_acc0# 设置一个最佳准确率作为最佳模型的判别指标forepochinrange(epochs):model.train()epoch_train_acc,epoch_train_losstrain(train_dl,model,loss_fn,optimizer)model.eval()epoch_test_acc,epoch_test_losstest(test_dl,model,loss_fn)# 保存最佳模型到 best_modelifepoch_test_accbest_acc:best_accepoch_test_acc best_modelcopy.deepcopy(model)train_acc.append(epoch_train_acc)train_loss.append(epoch_train_loss)test_acc.append(epoch_test_acc)test_loss.append(epoch_test_loss)# 获取当前的学习率lroptimizer.state_dict()[param_groups][0][lr]template(Epoch:{:2d}, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%, Test_loss:{:.3f}, Lr:{:.2E})print(template.format(epoch1,epoch_train_acc*100,epoch_train_loss,epoch_test_acc*100,epoch_test_loss,lr))# 保存最佳模型到文件中PATH./best_model.pth# 保存的参数文件名torch.save(model.state_dict(),PATH)print(Done)四、 结果可视化1. Loss与Accuracy图importmatplotlib.pyplotasplt#隐藏警告importwarnings warnings.filterwarnings(ignore)#忽略警告信息plt.rcParams[font.sans-serif][SimHei]# 用来正常显示中文标签plt.rcParams[axes.unicode_minus]False# 用来正常显示负号plt.rcParams[figure.dpi]100#分辨率fromdatetimeimportdatetime current_timedatetime.now()# 获取当前时间epochs_rangerange(epochs)plt.figure(figsize(12,3))plt.subplot(1,2,1)plt.plot(epochs_range,train_acc,labelTraining Accuracy)plt.plot(epochs_range,test_acc,labelTest Accuracy)plt.legend(loclower right)plt.title(Training and Validation Accuracy)plt.xlabel(current_time)# 打卡请带上时间戳否则代码截图无效plt.subplot(1,2,2)plt.plot(epochs_range,train_loss,labelTraining Loss)plt.plot(epochs_range,test_loss,labelTest Loss)plt.legend(locupper right)plt.title(Training and Validation Loss)plt.show()2. 指定图片进行预测fromPILimportImage classeslist(total_data.class_to_idx)defpredict_one_image(image_path,model,transform,classes):test_imgImage.open(image_path).convert(RGB)plt.imshow(test_img)# 展示预测的图片test_imgtransform(test_img)imgtest_img.to(device).unsqueeze(0)model.eval()outputmodel(img)_,predtorch.max(output,1)pred_classclasses[pred]print(f预测结果是{pred_class})# 预测训练集中的某张照片predict_one_image(image_path./PotatoPlants/Early_blight/1.JPG,modelmodel,transformtrain_transforms,classesclasses)3. 模型评估best_model.eval()epoch_test_acc,epoch_test_losstest(test_dl,best_model,loss_fn)epoch_test_acc,epoch_test_loss五、感悟1.本周跑VGG的一个最大的问题是速度极慢不知道是因为我这个早年的游戏本的原因还是因为VGG参数多显存占用高导致的尝试将batch_size降低后可以有所改善。2.更换优化器可以观察得到SGD对学习率更为敏感。