tracking

# 综合跟踪指南 使用 MLflow 进行实验跟踪的完整指南。 ## 目录 - 记录参数 - 记录指标 - 记录产出物 - 记录模型 - 自动记录 - 运行和实验 - 搜索和比较 ## 记录参数 ### 基础参数记录 python import mlflow with mlflow.start_run(): # 单个参数 mlflow.log_param("learning_rate", 0.001) mlflow.log_param("batch_size", 32) mlflow.log_param("optimizer", "Adam") # 一次记录多个参数 mlflow.log_params({ "epochs": 50, "dropout": 0.2, "weight_decay": 1e-4, "momentum": 0.9 }) ### 结构化参数 python # 嵌套配置 config = { "model": { "architecture": "ResNet50", "pretrained": True, "num_classes": 10 }, "training": { "lr": 0.001, "batch_size": 32, "epochs": 50 }, "data": { "dataset": "ImageNet", "augmentation": True } } with mlflow.start_run(): # 记录为扁平化参数 for section, params in config.items(): for key, value in params.items(): mlflow.log_param(f"{section}.{key}", value) # 或者将整个配置记录为产出物 mlflow.log_dict(config, "config.json") ### 参数最佳实践 python with mlflow.start_run(): # ✅ 推荐：记录所有超参数 mlflow.log_params({ "learning_rate": 0.001, "batch_size": 32, "optimizer": "Adam", "scheduler": "CosineAnnealing", "weight_decay": 1e-4 }) # ✅ 推荐：记录数据信息 mlflow.log_params({ "dataset": "ImageNet", "train_samples": len(train_dataset), "val_samples": len(val_dataset), "num_classes": 1000 }) # ✅ 推荐：记录环境信息 mlflow.log_params({ "framework": "PyTorch 2.0", "cuda_version": torch.version.cuda, "gpu": torch.cuda.get_device_name(0) }) ## 记录指标 ### 时间序列指标 python with mlflow.start_run(): for epoch in range(num_epochs): # 训练 train_loss, train_acc = train_epoch() # 验证 val_loss, val_acc = validate() # 按步长记录指标 mlflow.log_metric("train_loss", train_loss, step=epoch) mlflow.log_metric("train_accuracy", train_acc, step=epoch) mlflow.log_metric("val_loss", val_loss, step=epoch) mlflow.log_metric("val_accura