1.引用

   torchvision提供了一些常用的数据集、模型、转换函数等

import torchvision
from torchvision import transforms
import numpy as np
import matplotlib.pyplot as plt

2.内置图片数据集加载

  torch的内置图片数据集均在datasets模块下，包含Catletch、CelebA、CIFAR、Cityscapes、COCO、Fashion-MNIST、ImageNet、MNIST等。
  MNIST数据集是0-9手写数字数据集。

  train=True表示是训练数据
  torchvision.transforms包含了转换函数
  这里用到了ToTensor类，该类的主要作用有以下3点：
  ①将输入转换成张量
  ②读取图片的格式规范为(channel,heigth,width)
  ③将图片像素的取值范围归一化0-1

train_ds=torchvision.datasets.MNIST('data/',train=True,transform=transforms.ToTensor(),download=True)
test_ds=torchvision.datasets.MNIST('data/',train=False,transform=transforms.ToTensor(),download=True)

3.处理为batch类型

  DataLoader有以下4个目的：
  ①使用shuffle参数对数据集做乱序的操作(随机打乱)
  ②将数据采样为小批次，可用batch_size参数指定批次大小(小批次)
  ③可以充分利用多个子进程加速数据预处理(多线程)
  ④可通过collate_fn参数传递批次数据中的处理函数，实现对批次数据进行转换处理(转换处理)

train_dl=torch.utils.data.DataLoader(train_ds,batch_size=64,shuffle=True)
test_dl=torch.utils.data.DataLoader(test_ds,batch_size=64)

  上述两行代码创建了DataLoader类型的train_dl和test_dl
  DataLoader是可迭代对象，next方法返回一个批次的图像imgs和对应一个批次的标签labels

4.设置运行设备

  机器学习或者深度学习需要选择程序运行的设备是CPU还是GPU，GPU就是通常所说的需要有显卡。

device='cuda' if torch.cuda.is_available() else 'cpu'
print('use {} device'.format(device))

5.查看数据

imgs,labels=next(iter(train_dl))
print(imgs.shape)
print(labels.shape)
结果：
torch.Size([64, 1, 28, 28])
torch.Size([64])

6.绘图查看数据图片

  imgs[:10]查看前10条数据
  np.squeeze从数组的形状中删除维度为 1 的维度。
  np.unsqueeze从数组的形状中添加维度为 1 的维度。
  注：只有数组长度在该维度上为 1，那么该维度才可以被删除。如果不是1，那么删除的话会报错
  报错信息：cannot select an axis to squeeze out which has size not equal to one

(1)不显示图片标签

plt.figure(figsize=(10,1))
for i,img in enumerate(imgs[:10]):npimg=img.numpy()npimg=np.squeeze(npimg)#形状由(1,28,28)转换为(28,28)plt.subplot(1,10,i+1)plt.imshow(npimg)      #在子图中绘制单张图片plt.axis('off')        #关闭显示子图坐标    
print(labels[:10])
plt.show()

(2)打印图片标签

classes = ('0', '1', '2', '3', '4', '5', '6', '7', '8', '9')
class_label_str=''img_label_list=list(zip(imgs,labels))
for i,(img,label) in enumerate(img_label_list):nimg=np.array(img)nimg=np.squeeze(nimg)plt.subplot(8,8,i+1)plt.title(str(label.item()))plt.imshow(nimg)plt.axis('off')'''按照图片显示格式打印所有标签：i!=0实现按行打印的同时第一行前面无空行，按每行8列打印'''if i!=0 and i%8==0:class_label_str +='\n'class_label_str += classes[label.item()]+'\t'
print(class_label_str)
plt.show()

(3)图片显示标签

  我这里以'airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck'这些类别示例，作用于手写字体图像分类时，要更改成0-9

classes = ('airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')img_label_list=list(zip(imgs,labels))
for i,(img,label) in enumerate(img_label_list):nimg=img.transpose(0, 2)nimg=nimg.numpy()plt.subplot(5,5,i+1)plt.title(classes[label.item()])plt.imshow(nimg)plt.axis('off')
plt.show()

7.定义卷积函数

  定义卷积函数才是算法模型的真正开始，卷积层一般是必不可少的，是机器学习和深度学习的灵魂与基石所在。

class net(nn.Module):def __init__(self):super().__init__()self.conv1=nn.Conv2d(1,6,5)self.conv2=nn.Conv2d(6,16,5)self.linear1 = nn.Linear(16*4*4,20)self.linear2 = nn.Linear(20,10)def forward(self,input):x=torch.max_pool2d(torch.relu(self.conv1(input)),2)x=torch.max_pool2d(torch.relu(self.conv2(x)),2)x=x.view(x.size(0),-1)x=torch.relu(self.linear1(x))x=self.linear2(x)return x

  算法模型流程示意：

8.卷积实例化、损失函数、优化器

model=net().to(device)
loss_f=nn.CrossEntropyLoss()
opti=optim.Adam(model.parameters(), lr=0.005)

9.训练和测试损失、正确率

(1)训练

def train(dataloader,model,loss_f,opt):model.train()                  #模型为训练模式num_batches=len(dataloader)    #总批次数size=len(dataloader.dataset)   #样本总数（所有的批次里所有的数据点）loss_zhi=0                     #所有批次的损失之和correct=0                      #预测正确的样本总数for x,y in dataloader:x,y=x.to(device),y.to(device)pred=model(x)loss=loss_f(pred,y)        '''梯度清零、反向传播、梯度更新是专属'''opt.zero_grad()loss.backward()opt.step()with torch.no_grad():loss_zhi+=loss.item()correct+=(pred.argmax(1)==y).type(torch.float).sum().item()loss_zhi/=num_batches  #loss_zhi是所有批次的损失之和，所以计算全部样本的平均损失需要除以总批次数correct/=size          #correct是预测正确的样本总数，若计算每个批次总体正确率，需除以样本总数量return loss_zhi,correct

  注：当前代码里的pred.argmax(1)会返会类似Tensor([4,6,...,0])的Tensor，而y也是类似形状的tensor，因此二者可以用==比较

(2)测试

def test(dataloader, model):model.eval()                   #模型为测试模式num_batches=len(dataloader)    #总批次数size=len(dataloader.dataset)   #样本总数（所有的批次里所有的数据点）loss_zhi=0                     #所有批次的损失之和correct=0                      #预测正确的样本总数for x, y in dataloader:x,y=x.to(device),y.to(device)pred = model(x)loss = loss_f(pred, y)with torch.no_grad():loss_zhi += loss.item()correct += (pred.argmax(1) == y).type(torch.float).sum().item()loss_zhi /= num_batchescorrect /= sizereturn loss_zhi, correct

(3)循环

epochs = 200
train_loss = []
train_acc = []
test_loss = []
test_acc = []
for epoch in range(epochs):epoch_train_loss,epoch_train_acc=train(train_dl,model,loss_f,opt)epoch_test_loss,epoch_test_acc=test(test_dl,model)train_loss.append(epoch_train_loss)train_acc.append(epoch_train_acc)test_loss.append(epoch_test_loss)test_acc.append(epoch_test_acc)tishi='epoch:{},train_loss:{:.4f},train_acc:{:.2f}%,test_loss:{:.4f},test_acc:{:.2f}%'print(tishi.format(epoch,train_loss[-1],train_acc[-1]*100,test_loss[-1],test_acc[-1]*100))
print('batch over!')

(4)损失和正确率曲线

plt.figure(figsize=(10,4))
plt.subplot(121)
#打印损失
plt.plot(range(1,epochs+1),train_loss,label='train_loss')
plt.plot(range(1,epochs+1),test_loss,label='test_loss')
plt.title('train+test:loss')
plt.xlabel('epoch')
plt.legend(loc='upper right')
plt.subplot(122)
#打印正确率
plt.plot(range(1,epochs+1),train_acc,label='train_acc')
plt.plot(range(1,epochs+1),test_acc,label='test_acc')
plt.title('train+test:acc')
plt.xlabel('epoch')
plt.legend(loc='lower right')
plt.show()
plt.savefig('D:/loss+acc.png')

(5)输出数据到表格

table={'train_loss':train_loss,'train_acc':train_acc,'test_loss':test_loss,'test_acc':test_acc}
data_shuju=pd.DataFrame(table,index=list(range(1,epochs+1)))
data_shuju.to_excel('D:/loss+acc.xlsx')

10.额外添加

(1)添加dropout减少过拟合

  卷积后添加Dropout层较少使用，效果也不是很明显，这是因为相邻元素之间有相关性，随机地丢弃卷积输出特征像素点，抑制过拟合的效果有限。
  Dropout的第一个参数是输入的tensor，第二个参数p代表的是丢弃的神经元的比例，默认为0.5。

①未添加dropout层

class net(nn.Module):def __init__(self):super().__init__()self.conv1=nn.Conv2d(1,6,5)self.conv2=nn.Conv2d(6,16,5)self.linear1 = nn.Linear(16*4*4,20)self.linear2 = nn.Linear(20,10)def forward(self,input):x=torch.max_pool2d(torch.relu(self.conv1(input)),2)x=torch.max_pool2d(torch.relu(self.conv2(x)),2)x=x.view(x.size(0),-1)x=torch.relu(self.linear1(x))x=self.linear2(x)return x

  dropout前loss+acc图像：

②添加dropout层

class net(nn.Module):def __init__(self):super().__init__()self.conv1=nn.Conv2d(1,6,5)self.conv2=nn.Conv2d(6,16,5)self.linear1 = nn.Linear(16*4*4,20)self.linear2 = nn.Linear(20,10)def forward(self,input):x=torch.max_pool2d(torch.relu(self.conv1(input)),2)x=torch.max_pool2d(torch.relu(self.conv2(x)),2)x=x.view(x.size(0),-1)x=torch.dropout(x, p=0.5, train=self.training)x=torch.relu(self.linear1(x))x=torch.dropout(x, p=0.5, train=self.training)x=self.linear2(x)return x

  dropout后loss+acc图像：

(2)循环同时输出时间

import timeepochs=1
train_loss=[]
train_acc=[]
test_loss=[]
test_acc=[]
epoch_time=[]
start=time.time()
for epoch in range(epochs):epoch_train_loss,epoch_train_acc=train(train_dl,model,loss_f,opt)epoch_test_loss, epoch_test_acc = test(test_dl, model)train_loss.append(epoch_train_loss)train_acc.append(epoch_train_acc)test_loss.append(epoch_test_loss)test_acc.append(epoch_test_acc)epoch_time.append(time.time()-start)tishi='epoch:{},train_loss:{:.4f},train_acc:{:.2f}%,test_loss:{:.4f},test_acc:{:.2f}%,time:{:.2f}'print(tishi.format(epoch,train_loss[-1],train_acc[-1]*100,test_loss[-1],test_acc[-1]*100,epoch_time[-1]))
print('epoch over!')table={'train_loss':train_loss,'train_acc':train_acc,'test_loss':test_loss,'test_acc':test_acc,'epoch_time':epoch_time}
data_shuju=pd.DataFrame(table,index=list(range(1,epochs+1)))
data_shuju.to_excel('loss+acc.xlsx')
print('save over!')

(3)每个类别分类正确率输出

①输出到控制台

  注：需要在网络循环loss+acc之后使用，若是提前使用正确率只有个位数

class_correct = list(0 for i in range(10))  #每个类别预测正确的数量
class_total   = list(0 for i in range(10))  #每个类别的总数量
with torch.no_grad():# 从测试数据中取出数据for x, y in test_dl: x, y= x.to(device), y.to(device)outputs = model(x)_, predicted = torch.max(outputs, 1)# 预测正确的返回True，预测错误的返回False;squeeze将数据转换为一维数据c = (predicted == y).squeeze()for i in range(10):label = y[i]  # 提取标签class_correct[label] += c[i].item()  # 预测正确个数class_total[label] += 1  # 总数
for i in range(10):print('{}的准确率:{:.2f}%'.format(classes[i], 100 * class_correct[i] / class_total[i]))
结果：
0的准确率:98.03%
1的准确率:100.00%
2的准确率:100.00%
3的准确率:98.08%
4的准确率:97.74%
5的准确率:97.62%
6的准确率:98.44%
7的准确率:99.39%
8的准确率:98.60%
9的准确率:97.60%

②输出到表格

class_test_dic={}
for i in range(10):print('{:.10s}的准确率:{:.2f}%'.format(classes[i], 100 * class_correct[i] / class_total[i]))class_test_dic['{:.12s}'.format(classes[i])]=[100 * class_correct[i] / class_total[i],'{:.2f}%'.format(100 * class_correct[i] / class_total[i])]
class_dic = pd.DataFrame(class_test_dic,index=list(range(2)))
class_dic.to_excel('model5s_class_test_dic.xlsx')
print('save over!')

(4)模型保存/加载

torch.save(model,"K:\\classifier3.pt")    #保存完整模型
load_model = torch.load("K:\\classifier3.pt")

  测试图片：

path='./MNIST_data.pth'
test_model = net()
test_model.load_state_dict(torch.load(path))test_image = Image.open(file)  # 加载要测试的图片test_transform = torchvision.transforms.Compose([torchvision.transforms.Resize((28, 28)),torchvision.transforms.Grayscale(),  # 训练的是灰色图片需要加上，不然通道数不对torchvision.transforms.ToTensor(),torchvision.transforms.Normalize((0.1307,), (0.3081,))
])test_image = test_transform(test_image)
test_image = test_image.unsqueeze(0)  # 添加批次维度output = test_model(test_image)  # 输入图片到模型中进行推理
_, predicted = torch.max(output, 1)  # 获取预测结果
self.label_result.setText(str(predicted.item()))

(5)保存后的网络模型可视化

  浏览器输入链接：netron

  点击按钮，打开保存的.pt文件，就可以显示网络机构
  例：

(6)训练过程可视化

def visualize(train_loss,val_loss,val_acc):train_loss = np.array(train_loss)val_loss = np.array(val_loss)val_acc = np.array(val_acc)plt.grid(True)plt.xlabel("epoch")plt.ylabel("value")plt.title("train_loss and valid_acc")plt.plot(np.arange(len(val_acc)),val_acc, label=r"valid_acc",c="g")plt.plot(np.arange(len(train_loss)),train_loss,label=r"train_loss",c="r")plt.legend()plt.savefig("K:\\a.png")visualize(train_loss_list,valid_loss_list,valid_accuracy_list)

(7)显示彩色图片

  问题1：TypeError: Invalid shape (3, 224, 224) for image data

①灰色图片

img_label_list=list(zip(img,label))
for i,(img,label) in enumerate(img_label_list):nimg=np.array(img)nimg=np.squeeze(nimg)plt.subplot(8,8,i+1)plt.title(str(label.item()))plt.imshow(nimg)plt.axis('off')
plt.show()

②彩色图片

img_label_list=list(zip(img,label))
for i,(img,label) in enumerate(img_label_list):nimg=img.transpose(0, 2)  彩色图像是3*n*n需要先将3移到最后！nimg=nimg.numpy()plt.subplot(5,5,i+1)plt.title(str(label.item()))plt.imshow(nimg)plt.axis('off')
plt.show()

  问题2：Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
  matplotlib.pyplot.imshow()函数在处理灰度图像时，自动将其值做归一化处理而在处理彩色图像时则不会，而是将浮点值变换至[0，1]，整数值变换到[0, 255]范围

(8)每次卷积后特征图显示

  需要先修改transform：

transform=transforms.Compose([transforms.ToTensor(),   #转换成张量transforms.Normalize(mean=[0.485,0.456,0.406],std=[0.229,0.224,0.225])])

#保存数据集一张图像
imgs,labels=next(iter(train_dl))
img_label_list=list(zip(imgs,labels))
for i,(img,label) in enumerate(img_label_list):nimg=img.transpose(0, 2)nimg=nimg.numpy()plt.imshow(nimg)plt.axis('off')plt.savefig('1.png')if i==0:break# 传入图片
from PIL import Imageimage=Image.open(str(r'1.png')).convert('RGB')
image=transform(image)
print('\n输入图尺寸:{}'.format(image.shape))image=image.unsqueeze(0)
image=image.to(device)# 计算卷积个数
model_weights=[] #卷积层参数
conv_layers=[]   #卷积层本身
model_children=list(model.children())
counter=0        #卷积层个数
for i in range(len(model_children)):if type(model_children[i])==nn.Conv2d:counter+=1model_weights.append(model_children[i].weight)conv_layers.append(model_children[i])elif type(model_children[i])==nn.Sequential:for j in range(len(model_children[i])):for child in model_children[i][j].children():if type(child)==nn.Conv2d:counter+=1model_weights.append(child.weight)conv_layer.append(child)outputs=[]
names=[]
for layer in conv_layers[0:]:image=layer(image)outputs.append(image)names.append(str(layer))
print('特征图尺寸:{}'.format(outputs[1].shape))#具体绘制特征图
processed=[]
for feature_map in outputs:feature_map=feature_map.squeeze(0)gray_scale=torch.sum(feature_map,0)gray_scale=gray_scale/feature_map.shape[0]processed.append(gray_scale.data.cpu().numpy())
fig=plt.figure()for i in range(len(processed)):a=fig.add_subplot(5,4,i+1)imgplot=plt.imshow(processed[i])a.axis('off')a.set_title(names[i].split('(')[0],fontsize=10)
plt.savefig('feature_map.png',bbox_inches='tight')
print('over!')

(9)运行过程中忽视警告

import warnings
warnings.filterwarnings("ignore")