PyTorch TensorBoard SummaryWriter超参数记录后HPARAMS指标为空求助

核心修复方案 #

1. 原生PyTorch SummaryWriter 调用规范 #

问题大概率出在add_hparams的参数传递格式上，必须严格将超参数字典和指标字典分开传递，不能混淆参数顺序或合并字典。

正确示例代码：

from torch.utils.tensorboard import SummaryWriter

writer = SummaryWriter("runs/hparam_fix")

# 多组超参数与对应测试准确率
hparam_groups = [
    {"lr": 0.001, "batch_size": 32},
    {"lr": 0.01, "batch_size": 64},
    {"lr": 0.1, "batch_size": 128}
]
test_accs = [0.85, 0.88, 0.79]

for hparams, acc in zip(hparam_groups, test_accs):
    # 关键：hparam_dict传超参数，metric_dict传指标，二者独立
    writer.add_hparams(
        hparam_dict=hparams,
        metric_dict={"test_acc": acc}
    )

writer.close()

2. PyTorch Lightning 适配方案 #

不要手动调用原生add_hparams，改用Lightning内置的日志接口，确保超参数和指标属于同一run：

import pytorch_lightning as pl
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, TensorDataset

class SimpleClassifier(pl.LightningModule):
    def __init__(self, lr=0.001, batch_size=32):
        super().__init__()
        self.save_hyperparameters()  # 自动将超参数同步到logger
        self.model = nn.Linear(10, 2)
        self.accuracy = pl.metrics.Accuracy(task="multiclass", num_classes=2)

    def test_step(self, batch, batch_idx):
        x, y = batch
        logits = self(x)
        acc = self.accuracy(logits, y)
        self.log("test_acc", acc)  # 指标自动关联当前超参数run

    def configure_optimizers(self):
        return optim.Adam(self.parameters(), lr=self.hparams.lr)

# 测试不同超参数组合
for lr, batch_size in [(0.001,32), (0.01,64), (0.1,128)]:
    dataset = TensorDataset(torch.randn(100,10), torch.randint(0,2,(100,)))
    test_loader = DataLoader(dataset, batch_size=batch_size)
    
    model = SimpleClassifier(lr=lr, batch_size=batch_size)
    trainer = pl.Trainer(
        max_epochs=1,
        logger=pl.loggers.TensorBoardLogger("lightning_runs"),
        enable_checkpointing=False
    )
    trainer.test(model, dataloaders=test_loader)

3. 版本兼容性修复 #

当前tensorboard 2.17.0与torch 2.4.0存在兼容性问题，降级tensorboard到2.16.0即可：

pip install tensorboard==2.16.0

4. 清理日志缓存 #

旧日志文件可能导致加载异常，删除现有日志目录后重新运行代码：

rm -rf runs/ lightning_runs/
# 重新运行代码后启动TensorBoard
tensorboard --logdir=runs/

验证步骤 #

1. 运行修复后的代码生成新日志
2. 启动TensorBoard后等待HPARAMS标签页完全加载（约1-3秒）
3. 确认表格同时显示超参数列和test_acc指标列

内容的提问来源于stack exchange，提问作者Theta