0.环境说明

python3.8.5+pytorch

1. 模型结构可视化

1.1 netron

step1：在虚拟环境中安装netron

pip install netron

step2: 在虚拟环境中打开netron

step3:浏览器中输入地址:http://localhost:8080/

step4:选择保存的模型xxx.pt

1.2 使用tensorboard

step1:安装tensorboard，最简单的方式就是直接安装一个tensorflow

pip install tensorflow==1.15.0 -i https://mirrors.aliyun.com/pypi/simple

step2：代码中设置

from torch.utils.tensorboard import SummaryWriter

'''设置在模型构建后'''
 writer = SummaryWriter(log_dir='./output/log')
 writer.add_graph(model, torch.empty(10, 4)) #注意这里要结合你具体的训练样本，10是batch_szie可任意，4是训练样本的特征长度需要和训练样本一致

 '''设置在反向传播过程中，记录loss和acc'''
 # 可视化输出
 writer.add_scalar('loss', _loss, train_step)
 writer.add_scalar('acc', _acc, train_step)
 train_step += 1

'''train损失和test损失共同打印在一张图上,add_scalars注意s'''
writer.add_scalars('epoch_loss',{'train':train_loss,'test':test_loss},epoch)

step3：

1. 进入cmd命令行；
2. 切换当前磁盘到events文件所在的磁盘；
3. 确保events文件所在的路径没有中文字符串；
4. 输入命令:tensorboard --logdir C:\Users\...\output\log
   
5. 浏览器中输入http://localhost:6006/#images
   

2. 训练过程可视化

2.1 tensorboard

上文已经提及，只需要在训练过程中add即可。

 '''设置在反向传播过程中，记录loss和acc'''
 # 可视化输出
 writer.add_scalar('loss', _loss, train_step)
 writer.add_scalar('acc', _acc, train_step)
 train_step += 1

2.2 普通代码

if batch_idx % 100 == 0:
            print(f"Train Epoch:{epoch} [{batch_idx*len(data)}/{len(train_loader.dataset)} ({100.*batch_idx/len(train_loader):.0f}%)]\tloss:{loss.item():.6f}")

效果图：

3. 特征提取可视化

需要tensorboard配合hook，直接上代码。
model: LeNet
data: MNIST

import enum
import sys
import torch
from torch import nn
from torchvision import datasets,transforms
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from myutils.metrics import Acc_Score  #自己写的一个计算准确度的类，继承Module

class LeNet_BN(nn.Module):
    def __init__(self,in_chanel) -> None:
        super(LeNet_BN,self).__init__()
        self.feature_hook_img = {}

        self.features = nn.Sequential(
        nn.Conv2d(in_chanel, 6, kernel_size=5), nn.BatchNorm2d(6), nn.Sigmoid(),
        nn.AvgPool2d(kernel_size=2, stride=2),
        nn.Conv2d(6, 16, kernel_size=5), nn.BatchNorm2d(16), nn.Sigmoid())
        self.classifi = nn.Sequential(
        nn.AvgPool2d(kernel_size=2, stride=2), nn.Flatten(),
        nn.Linear(256, 120), nn.BatchNorm1d(120), nn.Sigmoid(),
        nn.Linear(120, 84), nn.BatchNorm1d(84), nn.Sigmoid(),
        nn.Linear(84, 10))   


    def forward(self,X):
        X = self.features(X)# 特征提取与分类需要分开
        X = self.classifi(X)
        return X

    def add_hooks(self):#可视化钩子
        def create_hook_fn(idx):
            def hook_fn(model,input,output):
                self.feature_hook_img[idx]=output.cpu()
            return hook_fn

        for _idx,_layer in enumerate(self.features):
            _layer.register_forward_hook(create_hook_fn(_idx))

    def add_image_summary(self,writer,step,prefix=None):
        if len(self.feature_hook_img)==0:
            return
        if prefix is None:
            prefix='layer'
        else:
            prefix = f"{prefix}_layer"
        for _k in self.feature_hook_img:# 包含原始图像
            _v = self.feature_hook_img[_k][0:1,...]# 只获取第一张图像
            _v = torch.permute(_v,(1,0,2,3))#(1,c,h,w)->(c,1,h,w)# 交换通道，展示每个维度的提取的图像特征
            writer.add_images(f"{prefix}_{_k}",_v,step)


if __name__=='__main__':
    # 加载数据
    # device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
    tsf = transforms.Compose([transforms.ToTensor()])
    train_data= datasets.MNIST(root='dataset\mnist_train',train=True,transform=tsf,download=True)
    train_data_loader = DataLoader(train_data,batch_size=32,shuffle=True)
    test_data = datasets.MNIST(root='dataset\mnist_test',train=False,transform=tsf,download=True)
    test_data_loader = DataLoader(test_data,batch_size=32,shuffle=32)

    model = LeNet_BN(1)
    model.add_hooks()
    lr=1e-2
    epochs = 10
    loss_f = torch.nn.CrossEntropyLoss()
    acc_f = Acc_Score()
    opt = torch.optim.SGD(model.parameters(),lr)
    writer = SummaryWriter(log_dir='./output/log')
    writer.add_graph(model,torch.empty(10,1,28,28))
    
    for epoch in range(epochs):
        for idx,data in enumerate(train_data_loader):
            X,y = data
            y = y.to(torch.long)
            # 前向传播
            y_pred = model(X)
            train_loss = loss_f(y_pred,y)
            train_acc = acc_f(y_pred,y)
            # 反向传播
            opt.zero_grad()
            train_loss.backward()
            opt.step()
            if (idx+1)%100==0:
                print(f"epoch:{epoch} |{(idx+1)*32}/{len(train_data)}({100.*(idx+1)*32/len(train_data):.2f}%)|\tloss:{train_loss.item():.3f}\tacc:{train_acc.item():.2f}")
                model.add_image_summary(writer,epoch,'train')# 添加本次训练


        test_loss=0
        test_acc=0
        test_numbers = len(test_data)/32
        for data in test_data_loader:
            model.eval()
            X,y = data
            y=y.to(torch.long)
            # print(y)
            # y = y.to(torch.long)
            y_pred = model(X)
            test_loss += loss_f(y_pred,y).item()
            test_acc += acc_f(y_pred,y).item()
        test_loss = test_loss/test_numbers
        test_acc = test_acc/test_numbers
        print('test res:')
        print(f"epoch:{epoch} \tloss:{test_loss:.3f}\tacc:{test_acc:.2f}")
        print('-'*80)
        writer.add_scalars('epoch_loss',{'train':train_loss.item(),'test':test_loss},epoch)
        writer.add_scalars('epoch_acc',{'train':train_acc.item(),'test':test_acc},epoch)
        
    writer.close()# 关闭