【python深度学习】Day 45 Tensorboard使用介绍

作业：对resnet18在cifar10上采用微调策略下，用tensorboard监控训练过程。

PS:

1. tensorboard和torch版本存在一定的不兼容性，如果报错请新建环境尝试。
2. tensorboard的代码还有有一定的记忆量，实际上深度学习的经典代码都是类似于八股文，看多了就习惯了，难度远远小于考研数学等需要思考的内容
3. 实际上对目前的 AI 而言，你只需要先完成最简单的demo，然后让他给你加上tensorboard需要打印的部分即可。——核心是弄懂tensorboard可以打印什么信息，以及如何看可视化后的结果，把ai当成记忆大师用到的时候通过它来调取对应的代码即可。

一、介绍

     之前在神经网络训练中，为了帮助自己理解，借用了很多的组件，比如训练进度条、可视化的loss下降曲线、权重分布图，运行结束后还可以查看单张图的推理效果。

        tensorboard这个库，集成了以上所有可视化工具

二、代码实战

MLP

import torch
import torch.nn as nn
import torch.optim as optim
import torchvision
from torchvision import datasets, transforms
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
import numpy as np
import matplotlib.pyplot as plt
import os# 设置随机种子以确保结果可复现
torch.manual_seed(42)
np.random.seed(42)# 1. 数据预处理
transform = transforms.Compose([transforms.ToTensor(),                # 转换为张量transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))  # 标准化处理
])# 2. 加载CIFAR-10数据集
train_dataset = datasets.CIFAR10(root='./data',train=True,download=True,transform=transform
)test_dataset = datasets.CIFAR10(root='./data',train=False,transform=transform
)# 3. 创建数据加载器
batch_size = 64
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False)# CIFAR-10的类别名称
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')# 4. 定义MLP模型（适应CIFAR-10的输入尺寸）
class MLP(nn.Module):def __init__(self):super(MLP, self).__init__()self.flatten = nn.Flatten()  # 将3x32x32的图像展平为3072维向量self.layer1 = nn.Linear(3072, 512)  # 第一层：3072个输入，512个神经元self.relu1 = nn.ReLU()self.dropout1 = nn.Dropout(0.2)  # 添加Dropout防止过拟合self.layer2 = nn.Linear(512, 256)  # 第二层：512个输入，256个神经元self.relu2 = nn.ReLU()self.dropout2 = nn.Dropout(0.2)self.layer3 = nn.Linear(256, 10)  # 输出层：10个类别def forward(self, x):# 第一步：将输入图像展平为一维向量x = self.flatten(x)  # 输入尺寸: [batch_size, 3, 32, 32] → [batch_size, 3072]# 第一层全连接 + 激活 + Dropoutx = self.layer1(x)   # 线性变换: [batch_size, 3072] → [batch_size, 512]x = self.relu1(x)    # 应用ReLU激活函数x = self.dropout1(x) # 训练时随机丢弃部分神经元输出# 第二层全连接 + 激活 + Dropoutx = self.layer2(x)   # 线性变换: [batch_size, 512] → [batch_size, 256]x = self.relu2(x)    # 应用ReLU激活函数x = self.dropout2(x) # 训练时随机丢弃部分神经元输出# 第三层（输出层）全连接x = self.layer3(x)   # 线性变换: [batch_size, 256] → [batch_size, 10]return x  # 返回未经过Softmax的logits# 检查GPU是否可用
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")# 初始化模型
model = MLP()
model = model.to(device)  # 将模型移至GPU（如果可用）criterion = nn.CrossEntropyLoss()  # 交叉熵损失函数
optimizer = optim.Adam(model.parameters(), lr=0.001)  # Adam优化器# 创建TensorBoard的SummaryWriter，指定日志保存目录
log_dir = 'runs/cifar10_mlp_experiment'
# 如果目录已存在，添加后缀避免覆盖
if os.path.exists(log_dir):i = 1while os.path.exists(f"{log_dir}_{i}"):i += 1log_dir = f"{log_dir}_{i}"
writer = SummaryWriter(log_dir)# 5. 训练模型（使用TensorBoard记录各种信息）
def train(model, train_loader, test_loader, criterion, optimizer, device, epochs, writer):model.train()  # 设置为训练模式# 记录训练开始时间，用于计算训练速度global_step = 0# 可视化模型结构dataiter = iter(train_loader)images, labels = next(dataiter)images = images.to(device)writer.add_graph(model, images)  # 添加模型图# 可视化原始图像样本img_grid = torchvision.utils.make_grid(images[:8].cpu())writer.add_image('原始训练图像', img_grid)for epoch in range(epochs):running_loss = 0.0correct = 0total = 0for batch_idx, (data, target) in enumerate(train_loader):data, target = data.to(device), target.to(device)  # 移至GPUoptimizer.zero_grad()  # 梯度清零output = model(data)  # 前向传播loss = criterion(output, target)  # 计算损失loss.backward()  # 反向传播optimizer.step()  # 更新参数# 统计准确率和损失running_loss += loss.item()_, predicted = output.max(1)total += target.size(0)correct += predicted.eq(target).sum().item()# 每100个批次记录一次信息到TensorBoardif (batch_idx + 1) % 100 == 0:batch_loss = loss.item()batch_acc = 100. * correct / total# 记录标量数据（损失、准确率）writer.add_scalar('Train/Batch_Loss', batch_loss, global_step)writer.add_scalar('Train/Batch_Accuracy', batch_acc, global_step)# 记录学习率writer.add_scalar('Train/Learning_Rate', optimizer.param_groups[0]['lr'], global_step)# 每500个批次记录一次直方图（权重和梯度）if (batch_idx + 1) % 500 == 0:for name, param in model.named_parameters():writer.add_histogram(f'weights/{name}', param, global_step)if param.grad is not None:writer.add_histogram(f'grads/{name}', param.grad, global_step)print(f'Epoch: {epoch+1}/{epochs} | Batch: {batch_idx+1}/{len(train_loader)} 'f'| 单Batch损失: {batch_loss:.4f} | 累计平均损失: {running_loss/(batch_idx+1):.4f}')global_step += 1# 计算当前epoch的平均训练损失和准确率epoch_train_loss = running_loss / len(train_loader)epoch_train_acc = 100. * correct / total# 记录每个epoch的训练损失和准确率writer.add_scalar('Train/Epoch_Loss', epoch_train_loss, epoch)writer.add_scalar('Train/Epoch_Accuracy', epoch_train_acc, epoch)# 测试阶段model.eval()  # 设置为评估模式test_loss = 0correct_test = 0total_test = 0# 用于存储预测错误的样本wrong_images = []wrong_labels = []wrong_preds = []with torch.no_grad():for data, target in test_loader:data, target = data.to(device), target.to(device)output = model(data)test_loss += criterion(output, target).item()_, predicted = output.max(1)total_test += target.size(0)correct_test += predicted.eq(target).sum().item()# 收集预测错误的样本wrong_mask = (predicted != target).cpu()if wrong_mask.sum() > 0:wrong_batch_images = data[wrong_mask].cpu()wrong_batch_labels = target[wrong_mask].cpu()wrong_batch_preds = predicted[wrong_mask].cpu()wrong_images.extend(wrong_batch_images)wrong_labels.extend(wrong_batch_labels)wrong_preds.extend(wrong_batch_preds)epoch_test_loss = test_loss / len(test_loader)epoch_test_acc = 100. * correct_test / total_test# 记录每个epoch的测试损失和准确率writer.add_scalar('Test/Loss', epoch_test_loss, epoch)writer.add_scalar('Test/Accuracy', epoch_test_acc, epoch)# 计算并记录训练速度（每秒处理的样本数）# 这里简化处理，假设每个epoch的时间相同samples_per_epoch = len(train_loader.dataset)# 实际应用中应该使用time.time()来计算真实时间print(f'Epoch {epoch+1}/{epochs} 完成 | 训练准确率: {epoch_train_acc:.2f}% | 测试准确率: {epoch_test_acc:.2f}%')# 可视化预测错误的样本（只在最后一个epoch进行）if epoch == epochs - 1 and len(wrong_images) > 0:# 最多显示8个错误样本display_count = min(8, len(wrong_images))wrong_img_grid = torchvision.utils.make_grid(wrong_images[:display_count])# 创建错误预测的标签文本wrong_text = []for i in range(display_count):true_label = classes[wrong_labels[i]]pred_label = classes[wrong_preds[i]]wrong_text.append(f'True: {true_label}, Pred: {pred_label}')writer.add_image('错误预测样本', wrong_img_grid)writer.add_text('错误预测标签', '\n'.join(wrong_text), epoch)# 关闭TensorBoard写入器writer.close()return epoch_test_acc  # 返回最终测试准确率# 6. 执行训练和测试
epochs = 20  # 训练轮次
print("开始训练模型...")
print(f"TensorBoard日志保存在: {log_dir}")
print("训练完成后，使用命令 `tensorboard --logdir=runs` 启动TensorBoard查看可视化结果")final_accuracy = train(model, train_loader, test_loader, criterion, optimizer, device, epochs, writer)
print(f"训练完成！最终测试准确率: {final_accuracy:.2f}%")