接辅导、项目定制

数据导入及处理部分本次数据导入没有使用torchvision自带的数据集需要将原始数据进行处理包括数据导入查看数据分类情况定义transforms进行数据类型转换等操作。

划分数据集划定训练集测试集后再使用torch.utils.data中的DataLoader()分别加载上一步处理好的训练及测试数据查看批处理维度.

模型构建部分resnet-50

设置超参数在这之前需要定义损失函数学习率动态学习率以及根据学习率定义优化器例如SGD随机梯度下降用来在训练中更新参数最小化损失函数。

定义训练函数函数的传入的参数有四个分别是设置好的DataLoader(),定义好的模型损失函数优化器。

函数内部初始化损失准确率为0接着开始循环使用DataLoader()获取一个批次的数据对这个批次的数据带入模型得到预测值然后使用损失函数计算得到损失值。

接下来就是进行反向传播以及使用优化器优化参数梯度清零放在反向传播之前或者是使用优化器优化之后都是可以的一般是默认放在反向传播之前。

定义测试函数函数传入的参数相比训练函数少了优化器只需传入设置好的DataLoader(),定义好的模型损失函数。

此外除了处理批次数据时无需再设置梯度清零、返向传播以及优化器优化参数其余部分均和训练函数保持一致。

训练过程定义训练次数有几次就使用整个数据集进行几次训练初始化四个空list分别存储每次训练及测试的准确率及损失。

使用model.train()开启训练模式调用训练函数得到准确率及损失。

使用model.eval()将模型设置为评估模式调用测试函数得到准确率及损失。

接着就是将得到的训练及测试的准确率及损失存储到相应list中并合并打印出来得到每一次整体训练后的准确率及损失。

‘model.pth’)

model.load_state_dict(torch.load(‘model.pth’))

需要改进优化的地方在保证整体流程没有问题的情况下继续细化细节研究比如一些函数的原理及作用如何提升训练集准确率等问题。

import

warningswarnings.filterwarnings(ignore)

忽略警告信息plt.rcParams[font.sans-serif]

[SimHei]

plt.rcParams[axes.unicode_minus]

False

transforms.Compose([transforms.Resize([224,224]),

将输入图片resize成统一尺寸transforms.RandomHorizontalFlip(),

将PIL

装换为tensor,并归一化到[0,1]之间transforms.Normalize(

标准化处理

mean[0.485,0.456,0.406]与std[0.229,0.224,0.225]

[0.229,0.224,0.225])

transforms.Compose([transforms.Resize([224,224]),transforms.ToTensor(),transforms.Normalize(mean

[0.229,0.224,0.225])

datasets.ImageFolder(./data/bird_photos/,transform

train_transforms)

./data/bird_photos/StandardTransform

Transform:

antialiasTrue)RandomHorizontalFlip(p0.5)ToTensor()Normalize(mean[0.485,

0.456,

0.225]))total_data.class_to_idx{Bananaquit:

0,Black

train_sizetrain_dataset,test_dataset

torch.utils.data.random_split(total_data,[train_size,test_size])

train_dataset,test_dataset(torch.utils.data.dataset.Subset

0x23b545994b0,torch.utils.data.dataset.Subset

0x23b54599300)#

torch.utils.data.DataLoader(train_dataset,batch_size

batch_size,shuffle

torch.utils.data.DataLoader(test_dataset,batch_size

batch_size,shuffle

nn.BatchNorm2d(filters1)self.conv2

nn.Conv2d(filters1,

nn.BatchNorm2d(filters2)self.conv3

nn.Conv2d(filters2,

nn.BatchNorm2d(filters3)self.shortcut

nn.Sequential()if

nn.Sequential(nn.Conv2d(in_channels,

filters3,

stridestrides),nn.BatchNorm2d(filters3))def

forward(self,

F.relu(self.bn1(self.conv1(x)))x

F.relu(self.bn2(self.conv2(x)))x

shortcutx

num_classes1000):super(ResNet50,

nn.Conv2d(3,

strides)layers.append(ConvBlock(in_channels,

filters,

blocks):layers.append(ConvBlock(in_channels,

filters,

F.relu(self.bn1(self.conv1(x)))x

self.maxpool(x)x

model.to(device)ResNet50((pad):

ZeroPad2d((3,

track_running_statsTrue)(maxpool):

stride2,

track_running_statsTrue)(conv2):

Conv2d(64,

track_running_statsTrue)(conv3):

Conv2d(64,

track_running_statsTrue)(shortcut):

Sequential((0):

track_running_statsTrue)(conv2):

Conv2d(64,

track_running_statsTrue)(conv3):

Conv2d(64,

track_running_statsTrue)(shortcut):

Sequential())(2):

track_running_statsTrue)(conv2):

Conv2d(64,

track_running_statsTrue)(conv3):

Conv2d(64,

track_running_statsTrue)(shortcut):

Sequential((0):

track_running_statsTrue)(conv2):

Conv2d(128,

track_running_statsTrue)(conv3):

Conv2d(128,

track_running_statsTrue)(shortcut):

Sequential((0):

track_running_statsTrue)(conv2):

Conv2d(128,

track_running_statsTrue)(conv3):

Conv2d(128,

track_running_statsTrue)(shortcut):

Sequential())(2):

track_running_statsTrue)(conv2):

Conv2d(128,

track_running_statsTrue)(conv3):

Conv2d(128,

track_running_statsTrue)(shortcut):

Sequential())(3):

track_running_statsTrue)(conv2):

Conv2d(128,

track_running_statsTrue)(conv3):

Conv2d(128,

track_running_statsTrue)(shortcut):

Sequential((0):

track_running_statsTrue)(conv2):

Conv2d(256,

track_running_statsTrue)(conv3):

Conv2d(256,

track_running_statsTrue)(shortcut):

Sequential((0):

track_running_statsTrue)(conv2):

Conv2d(256,

track_running_statsTrue)(conv3):

Conv2d(256,

track_running_statsTrue)(shortcut):

Sequential())(2):

track_running_statsTrue)(conv2):

Conv2d(256,

track_running_statsTrue)(conv3):

Conv2d(256,

track_running_statsTrue)(shortcut):

Sequential())(3):

track_running_statsTrue)(conv2):

Conv2d(256,

track_running_statsTrue)(conv3):

Conv2d(256,

track_running_statsTrue)(shortcut):

Sequential())(4):

track_running_statsTrue)(conv2):

Conv2d(256,

track_running_statsTrue)(conv3):

Conv2d(256,

track_running_statsTrue)(shortcut):

Sequential())(5):

track_running_statsTrue)(conv2):

Conv2d(256,

track_running_statsTrue)(conv3):

Conv2d(256,

track_running_statsTrue)(shortcut):

Sequential((0):

track_running_statsTrue)(conv2):

Conv2d(512,

track_running_statsTrue)(conv3):

Conv2d(512,

track_running_statsTrue)(shortcut):

Sequential((0):

track_running_statsTrue)(conv2):

Conv2d(512,

track_running_statsTrue)(conv3):

Conv2d(512,

track_running_statsTrue)(shortcut):

Sequential())(2):

track_running_statsTrue)(conv2):

Conv2d(512,

track_running_statsTrue)(conv3):

Conv2d(512,

track_running_statsTrue)(shortcut):

AdaptiveAvgPool2d(output_size(1,

1))(fc):

adjust_learning_rate(optimizer,epoch,start_lr):

每两个epoch

torch.optim.Adam(model.parameters(),lrlearn_rate)#

调用官方接口示例

nn.CrossEntropyLoss()learn_rate

1e-4

epoch:(0.92**(epoch//2))optimizer

torch.optim.Adam(model.parameters(),lr

learn_rate)

torch.optim.lr_scheduler.LambdaLR(optimizer,lr_lambdalambda1)

选定调整方法6.

train(dataloader,model,loss_fn,optimizer):size

训练集大小num_batches

反向传播optimizer.zero_grad()loss.backward()optimizer.step()#

记录acc与losstrain_acc

(pred.argmax(1)y).type(torch.float).sum().item()train_loss

loss.item()train_acc

test(dataloader,model,loss_fn):size

len(dataloader.dataset)num_batches

len(dataloader)test_acc,test_loss

0,0with

(pred.argmax(1)y).type(torch.float).sum().item()test_loss

loss.item()test_acc

adjust_learning_rate(optimizer,epoch,learn_rate)model.train()epoch_train_acc,epoch_train_loss

train(train_dl,model,loss_fn,optimizer)scheduler.step()

更新学习率——调用官方动态学习率时使用model.eval()epoch_test_acc,epoch_test_loss

保存最佳模型到

copy.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)#

获取当前学习率lr

optimizer.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))print(Done)Epoch:

1,Train_acc:35.4%,Train_loss:3.431,Test_acc:23.9%,Test_loss:3.033,Lr:1.00E-04

Epoch:

2,Train_acc:69.2%,Train_loss:0.899,Test_acc:26.5%,Test_loss:3.017,Lr:9.20E-05

Epoch:

3,Train_acc:74.1%,Train_loss:0.652,Test_acc:59.3%,Test_loss:1.278,Lr:9.20E-05

Epoch:

4,Train_acc:78.3%,Train_loss:0.587,Test_acc:67.3%,Test_loss:1.353,Lr:8.46E-05

Epoch:

5,Train_acc:82.5%,Train_loss:0.521,Test_acc:75.2%,Test_loss:0.829,Lr:8.46E-05

Epoch:

6,Train_acc:86.3%,Train_loss:0.433,Test_acc:66.4%,Test_loss:1.308,Lr:7.79E-05

Epoch:

7,Train_acc:90.3%,Train_loss:0.340,Test_acc:67.3%,Test_loss:1.600,Lr:7.79E-05

Epoch:

8,Train_acc:89.2%,Train_loss:0.374,Test_acc:71.7%,Test_loss:1.014,Lr:7.16E-05

Epoch:

9,Train_acc:88.7%,Train_loss:0.305,Test_acc:77.0%,Test_loss:0.841,Lr:7.16E-05

Epoch:10,Train_acc:90.7%,Train_loss:0.309,Test_acc:79.6%,Test_loss:1.094,Lr:6.59E-05

Epoch:11,Train_acc:91.8%,Train_loss:0.318,Test_acc:72.6%,Test_loss:0.976,Lr:6.59E-05

Epoch:12,Train_acc:93.6%,Train_loss:0.243,Test_acc:73.5%,Test_loss:1.209,Lr:6.06E-05

Epoch:13,Train_acc:94.7%,Train_loss:0.159,Test_acc:71.7%,Test_loss:0.947,Lr:6.06E-05

Epoch:14,Train_acc:98.0%,Train_loss:0.076,Test_acc:80.5%,Test_loss:0.707,Lr:5.58E-05

Epoch:15,Train_acc:97.8%,Train_loss:0.083,Test_acc:79.6%,Test_loss:0.923,Lr:5.58E-05

Epoch:16,Train_acc:98.2%,Train_loss:0.059,Test_acc:82.3%,Test_loss:0.650,Lr:5.13E-05

Epoch:17,Train_acc:98.5%,Train_loss:0.072,Test_acc:76.1%,Test_loss:0.828,Lr:5.13E-05

Epoch:18,Train_acc:98.0%,Train_loss:0.175,Test_acc:78.8%,Test_loss:0.834,Lr:4.72E-05

Epoch:19,Train_acc:94.2%,Train_loss:0.173,Test_acc:61.9%,Test_loss:2.606,Lr:4.72E-05

Epoch:20,Train_acc:96.2%,Train_loss:0.123,Test_acc:77.9%,Test_loss:0.959,Lr:4.34E-05

Epoch:21,Train_acc:96.5%,Train_loss:0.166,Test_acc:76.1%,Test_loss:1.266,Lr:4.34E-05

Epoch:22,Train_acc:96.9%,Train_loss:0.196,Test_acc:85.0%,Test_loss:0.698,Lr:4.00E-05

Epoch:23,Train_acc:98.7%,Train_loss:0.082,Test_acc:82.3%,Test_loss:0.626,Lr:4.00E-05

Epoch:24,Train_acc:96.0%,Train_loss:0.136,Test_acc:81.4%,Test_loss:0.805,Lr:3.68E-05

Epoch:25,Train_acc:98.5%,Train_loss:0.101,Test_acc:83.2%,Test_loss:0.576,Lr:3.68E-05

Epoch:26,Train_acc:98.5%,Train_loss:0.062,Test_acc:80.5%,Test_loss:0.597,Lr:3.38E-05

Epoch:27,Train_acc:99.6%,Train_loss:0.039,Test_acc:83.2%,Test_loss:0.574,Lr:3.38E-05

Epoch:28,Train_acc:99.3%,Train_loss:0.080,Test_acc:85.0%,Test_loss:0.758,Lr:3.11E-05

Epoch:29,Train_acc:99.6%,Train_loss:0.059,Test_acc:84.1%,Test_loss:0.608,Lr:3.11E-05

Epoch:30,Train_acc:98.9%,Train_loss:0.054,Test_acc:82.3%,Test_loss:0.753,Lr:2.86E-05

Epoch:31,Train_acc:98.9%,Train_loss:0.035,Test_acc:83.2%,Test_loss:0.617,Lr:2.86E-05

Epoch:32,Train_acc:98.7%,Train_loss:0.046,Test_acc:78.8%,Test_loss:0.847,Lr:2.63E-05

Epoch:33,Train_acc:98.9%,Train_loss:0.028,Test_acc:82.3%,Test_loss:0.746,Lr:2.63E-05

Epoch:34,Train_acc:99.6%,Train_loss:0.032,Test_acc:79.6%,Test_loss:0.629,Lr:2.42E-05

Epoch:35,Train_acc:99.3%,Train_loss:0.027,Test_acc:82.3%,Test_loss:0.597,Lr:2.42E-05

Epoch:36,Train_acc:99.6%,Train_loss:0.029,Test_acc:87.6%,Test_loss:0.488,Lr:2.23E-05

Epoch:37,Train_acc:99.6%,Train_loss:0.026,Test_acc:87.6%,Test_loss:0.552,Lr:2.23E-05

Epoch:38,Train_acc:99.1%,Train_loss:0.029,Test_acc:79.6%,Test_loss:0.572,Lr:2.05E-05

Epoch:39,Train_acc:99.8%,Train_loss:0.107,Test_acc:84.1%,Test_loss:0.704,Lr:2.05E-05

Epoch:40,Train_acc:99.1%,Train_loss:0.094,Test_acc:76.1%,Test_loss:0.772,Lr:1.89E-05

Done9.

range(epochs)plt.figure(figsize

plt.plot(epochs_range,train_acc,label

Training

plt.plot(epochs_range,test_acc,label

Test

plt.plot(epochs_range,train_loss,label

Test

plt.plot(epochs_range,test_loss,label

Test

torch.save(model.state_dict(),./模型参数/J1_resnet50_model_state_dict.pth)

仅保存状态字典#

定义官方vgg16模型用来加载参数best_model.load_state_dict(torch.load(./模型参数/J1_resnet50_model_state_dict.pth))

加载状态字典到模型All

list(total_data.class_to_idx)def

predict_one_image(image_path,model,transform,classes):test_img

Image.open(image_path).convert(RGB)#

plt.imshow(test_img)

test_img.to(device).unsqueeze(0)model.eval()output

classes[pred]print(f预测结果是:{pred_class})#

predict_one_image(image_path./data/bird_photos/Bananaquit/007.jpg,model

model,transform

预测结果是:Bananaquitclasses[Bananaquit,

Black