$devvkit learn --librarie mlflow-guide

MLflow Guide

[mllifecycle][experiments][tracking][mlops]

AI / LLM Tools

Install

pip install mlflow
# or: uv add mlflow
# Start UI:
mlflow ui
# Open: http://localhost:5000

MLflow tracks ML experiments: log parameters, metrics, artifacts (models, plots, data), and source code for every run. The UI lets you compare runs side-by-side and find the best hyperparameters. Use `mlflow.start_run()` as a context manager.

Beyond tracking, MLflow includes a Model Registry (version and stage models), a Deployment module (serve models as REST APIs), and an Evaluation module (benchmark LLMs and classical models). Models can be any framework: PyTorch, TensorFlow, sklearn, XGBoost, or custom.

MLflow supports autologging: `mlflow.autolog()` automatically logs params, metrics, and models for sklearn, PyTorch Lightning, XGBoost, and more. The MLflow UI is extensible with custom plugins. For production, use MLflow with a PostgreSQL backend and S3-compatible storage.

Tracking

Start UI· Launch tracking dashboard.

mlflow ui                         # Default port 5000
mlflow ui --port 8080
mlflow ui --backend-store-uri sqlite:///mlflow.db
# Open http://localhost:5000

Log experiment· Track parameters and metrics.

import mlflow

with mlflow.start_run():
    # Log parameters
    mlflow.log_param('learning_rate', 0.01)
    mlflow.log_param('batch_size', 32)
    
    # Log metrics
    for epoch in range(10):
        accuracy = 0.8 + epoch * 0.02
        mlflow.log_metric('accuracy', accuracy, step=epoch)
    
    # Log artifacts
    mlflow.log_artifact('confusion_matrix.png')
    mlflow.log_artifact('model.pkl')
    
    # Log model
    mlflow.sklearn.log_model(model, 'model')

MLflow projects· Packaging ML code.

# MLproject file:
# name: my_project
# conda_env: conda.yaml
# entry_points:
#   main:
#     parameters:
#       data: {type: str, default: ./data}
#     command: "python train.py {data}"

# Run project:
mlflow run . -P data=./dataset

# Remote execution:
mlflow run https://github.com/user/repo -P param=value

Autologging

Autologging· Automatic tracking.

import mlflow

mlflow.autolog()  # Enable autolog for all supported frameworks

# Or specific framework:
mlflow.sklearn.autolog()
mlflow.pytorch.autolog()
mlflow.xgboost.autolog()

# Now any model training is auto-logged:
from sklearn.ensemble import RandomForestClassifier
model = RandomForestClassifier().fit(X_train, y_train)
# All params, metrics, and model are logged automatically

Model Registry

Model Registry· Version and stage models.

# Register a model:
mlflow.register_model(
    'runs:/<RUN_ID>/model',
    'MyModel'
)

# CLI:
mlflow models -h  # List registered models

# Load from registry:
import mlflow.pyfunc
model = mlflow.pyfunc.load_model(
    model_uri='models:/MyModel/Production'
)

# Transition stage via UI or API:
from mlflow.tracking import MlflowClient
client = MlflowClient()
client.transition_model_version_stage(
    name='MyModel',
    version=3,
    stage='Production'
)

Serving

Serve model as API· REST API for model.

# Serve a registered model:
mlflow models serve -m models:/MyModel/Production -p 5001

# Test:
curl -X POST http://localhost:5001/invocations
  -H 'Content-Type: application/json'
  -d '{"inputs": {"feature1": 1.0, "feature2": 2.0}}'

# Serve from run:
mlflow models serve -m runs:/<RUN_ID>/model --port 5001

Evaluation

LLM evaluation· Benchmark LLMs.

import mlflow
from mlflow.metrics import rouge, flesch_kincaid

with mlflow.start_run():
    eval_data = {
        'inputs': ['What is MLflow?', 'Explain RAG'],
        'ground_truth': [
            'MLflow is an ML platform',
            'RAG is retrieval-augmented generation'
        ]
    }
    
    results = mlflow.evaluate(
        data=eval_data,
        model='openai:/gpt-4o-mini',
        model_type='text-summarization',
        extra_metrics=[rouge(), flesch_kincaid()]
    )
    print(results.metrics)