文章目录
- [DAY 46 TensorBoard 使用介绍](#DAY 46 TensorBoard 使用介绍)
DAY 46 TensorBoard 使用介绍
学习目标
- 理解 TensorBoard 的作用与数据流转方式
- 掌握 SummaryWriter 的核心用法（标量、图像、直方图、计算图）
- 通过 CIFAR-10 的 MLP / CNN 实战，生成可视化日志
一、TensorBoard 概览
TensorBoard 是深度学习训练过程的可视化面板，可用于：
- 观察 loss / acc 曲线，判断收敛或过拟合
- 查看模型结构图，快速确认网络连接
- 记录样本图像、参数分布，辅助排查训练异常
工作原理：训练时把指标、图像、直方图等写入日志文件（*.tfevents），TensorBoard 读取该目录并在网页展示。
二、准备环境与数据
python
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, Subset
from torch.utils.tensorboard import SummaryWriter
import torchvision
import torchvision.transforms as transforms
from torchvision.utils import make_grid
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('Using device:', device)Using device: cuda
python
# 为了演示更快，这里截取少量样本；想要完整训练可去掉 Subset
def get_loaders(batch_size=128, limit_train=5000, limit_test=1000):
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616)),
])
train_set = torchvision.datasets.CIFAR10(root='data', train=True, download=True, transform=transform)
test_set = torchvision.datasets.CIFAR10(root='data', train=False, download=True, transform=transform)
if limit_train:
train_set = Subset(train_set, range(limit_train))
if limit_test:
test_set = Subset(test_set, range(limit_test))
train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True, num_workers=2, pin_memory=True)
test_loader = DataLoader(test_set, batch_size=batch_size, shuffle=False, num_workers=2, pin_memory=True)
print(f'Train samples: {len(train_set)}, Test samples: {len(test_set)}')
return train_loader, test_loader
train_loader, test_loader = get_loaders()
images, labels = next(iter(train_loader))
print('Sample batch shape:', images.shape)Train samples: 5000, Test samples: 1000
Sample batch shape: torch.Size([128, 3, 32, 32])三、创建 SummaryWriter 与基础可视化
python
# 创建 writer，日志会自动追加编号避免覆盖
writer = SummaryWriter(log_dir='runs/day46_intro')
# 记录一组训练图像
img_grid = make_grid(images[:16], nrow=8, normalize=True, scale_each=True)
writer.add_image('TrainSamples', img_grid, global_step=0)
writer.flush()
print('Logged sample images to runs/day46_intro')Logged sample images to runs/day46_intro记录模型结构（Graph）
python
class SimpleMLP(nn.Module):
def __init__(self):
super().__init__()
self.model = nn.Sequential(
nn.Flatten(),
nn.Linear(3 * 32 * 32, 256),
nn.ReLU(),
nn.Linear(256, 10)
)
def forward(self, x):
return self.model(x)
mlp = SimpleMLP().to(device)
dummy_input = images[:1].to(device)
writer.add_graph(mlp, dummy_input)
writer.flush()
print('Logged MLP graph')Logged MLP graph四、MLP 训练 + TensorBoard 日志
python
def train_mlp(epochs=2, log_dir='runs/day46_mlp'):
model = SimpleMLP().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=1e-3)
writer = SummaryWriter(log_dir=log_dir)
writer.add_graph(model, images[:1].to(device))
global_step = 0
for epoch in range(epochs):
model.train()
running_loss, correct, total = 0.0, 0, 0
for batch_idx, (inputs, targets) in enumerate(train_loader):
inputs, targets = inputs.to(device), targets.to(device)
optimizer.zero_grad()
outputs = model(inputs)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
running_loss += loss.item() * inputs.size(0)
_, predicted = outputs.max(1)
total += targets.size(0)
correct += predicted.eq(targets).sum().item()
if batch_idx % 50 == 0:
writer.add_scalar('Loss/train', loss.item(), global_step)
writer.add_scalar('Acc/train', correct / total, global_step)
global_step += 1
epoch_loss = running_loss / total
epoch_acc = correct / total
writer.add_scalar('Epoch/Loss', epoch_loss, epoch)
writer.add_scalar('Epoch/Acc', epoch_acc, epoch)
for name, param in model.named_parameters():
if 'weight' in name:
writer.add_histogram(name, param, epoch)
print(f'Epoch {epoch+1}: loss={epoch_loss:.4f}, acc={epoch_acc:.4f}')
writer.close()
return model
mlp_model = train_mlp()Epoch 1: loss=2.0121, acc=0.3234
Epoch 2: loss=1.6385, acc=0.4348五、CNN 训练 + TensorBoard 日志
python
class SimpleCNN(nn.Module):
def __init__(self):
super().__init__()
self.features = nn.Sequential(
nn.Conv2d(3, 32, kernel_size=3, padding=1),
nn.ReLU(),
nn.MaxPool2d(2),
nn.Conv2d(32, 64, kernel_size=3, padding=1),
nn.ReLU(),
nn.MaxPool2d(2),
)
self.classifier = nn.Sequential(
nn.Flatten(),
nn.Linear(64 * 8 * 8, 128),
nn.ReLU(),
nn.Linear(128, 10)
)
def forward(self, x):
x = self.features(x)
return self.classifier(x)
def train_cnn(epochs=2, log_dir='runs/day46_cnn'):
model = SimpleCNN().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=1e-3)
writer = SummaryWriter(log_dir=log_dir)
writer.add_graph(model, images[:1].to(device))
global_step = 0
for epoch in range(epochs):
model.train()
running_loss, correct, total = 0.0, 0, 0
for batch_idx, (inputs, targets) in enumerate(train_loader):
inputs, targets = inputs.to(device), targets.to(device)
optimizer.zero_grad()
outputs = model(inputs)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
running_loss += loss.item() * inputs.size(0)
_, predicted = outputs.max(1)
total += targets.size(0)
correct += predicted.eq(targets).sum().item()
if batch_idx % 50 == 0:
writer.add_scalar('Loss/train', loss.item(), global_step)
writer.add_scalar('Acc/train', correct / total, global_step)
global_step += 1
epoch_loss = running_loss / total
epoch_acc = correct / total
writer.add_scalar('Epoch/Loss', epoch_loss, epoch)
writer.add_scalar('Epoch/Acc', epoch_acc, epoch)
writer.add_histogram('features.conv1.weight', model.features[0].weight, epoch)
writer.add_histogram('features.conv2.weight', model.features[3].weight, epoch)
print(f'Epoch {epoch+1}: loss={epoch_loss:.4f}, acc={epoch_acc:.4f}')
writer.close()
return model
cnn_model = train_cnn()Epoch 1: loss=1.9511, acc=0.2924
Epoch 2: loss=1.5537, acc=0.4464六、启动 TensorBoard
训练完成后在项目根目录执行：
python
# tensorboard --logdir runs
# 浏览器打开 http://localhost:6006七、常见问题与建议
- 直方图记录频率不宜过高，可按 epoch 记录减少日志体积
- 图像可用于检查数据增强是否符合预期
- 若曲线剧烈抖动，优先检查学习率、数据预处理和 batch size