Experiment Tracking — MLflow, W&B, Neptune

Notebook 1: model_v1.ipynb → Accuracy 92%
Notebook 2: model_v2.ipynb → Accuracy 93%
Notebook 3: model_final.ipynb → Accuracy 93.2%

Aber: Welche Hyperparameter in v3? Welche Daten? Welcher Code?
→ Unreproducible, unmaintainable

Run ID: exp-2026-03-21-001
- Hyperparameters: lr=0.001, batch_size=32
- Dataset: training_data_v2.csv (hash: abc123)
- Code version: commit abc123def456
- Metrics: accuracy=0.932, loss=0.15
- Artifacts: model.pkl (100 MB), plots/

→ Vollständig reproducible!

import mlflow

mlflow.set_experiment("fraud-detection")

with mlflow.start_run():
    # Log Hyperparameters
    mlflow.log_param("learning_rate", 0.001)
    mlflow.log_param("batch_size", 32)
    
    # Training...
    model = train()
    
    # Log Metrics
    accuracy = evaluate(model)
    mlflow.log_metric("accuracy", accuracy)
    
    # Log Artifacts
    mlflow.log_artifact("model.pkl")
    mlflow.log_artifact("plots/")

Run 1: lr=0.001, accuracy=0.92
Run 2: lr=0.01, accuracy=0.89
Run 3: lr=0.0001, accuracy=0.91

MLflow UI: http://localhost:5000
→ Visualisiere welcher Run am besten war

import wandb

wandb.init(project="fraud-detection", name="run-v2")

wandb.log({
    "accuracy": 0.932,
    "loss": 0.15,
    "learning_rate": 0.001
})

wandb.log_model(path="model.pkl", name="fraud-model-v1")

wandb.finish()