Skip to content

Tracelet

Tracelet Logo

Intelligent experiment tracking for PyTorch and PyTorch Lightning
Automagic hyperparameter detection and multi-backend logging

Release Build status Python PyPI License Commit activity

What is Tracelet?

Tracelet is a powerful Python library that automatically captures and logs your machine learning experiments without requiring code modifications. Simply add one line to start tracking, and Tracelet will:

  • 🔮 Automagic instrumentation - Zero-config hyperparameter detection and logging
  • 🔍 Automatically capture TensorBoard metrics, PyTorch Lightning logs, and system metrics
  • 🔄 Route to multiple backends simultaneously (MLflow, ClearML, W&B, AIM)
  • 📊 Track everything - scalars, histograms, images, audio, text, and artifacts
  • ⚡ Zero code changes required for existing TensorBoard workflows
  • 🎯 Plugin architecture for extensible functionality

Quick Start

Installation

pip install tracelet

uv add tracelet

conda install -c conda-forge tracelet

Demo

See Tracelet in action! The video above shows how easy it is to get started with automatic experiment tracking.

Basic Usage

import tracelet
import torch
from torch.utils.tensorboard import SummaryWriter

# 1. Start tracking (one line!)
tracelet.start_logging(
    exp_name="my_experiment",
    project="my_project",
    backend="mlflow"  # or "clearml", "wandb", "aim"
)

# 2. Use TensorBoard as normal - metrics automatically captured
writer = SummaryWriter()
for epoch in range(100):
    loss = train_one_epoch()  # Your existing training code
    writer.add_scalar('Loss/train', loss, epoch)
    # ✨ Metrics automatically sent to MLflow!

# 3. Stop tracking
tracelet.stop_logging()

That's it!

Your existing TensorBoard code now logs to MLflow, ClearML, W&B, or AIM with zero changes!

Key Features

🔌 Multi-Backend Support

Choose from 4 popular experiment tracking backends:

  • MLflow - Open source ML lifecycle management
  • ClearML - Enterprise-grade MLOps platform
  • Weights & Biases - Collaborative ML platform
  • AIM - Open source experiment tracking

🎯 Automatic Instrumentation

Tracelet automatically captures:

  • TensorBoard metrics - Scalars, histograms, images, audio, text
  • PyTorch Lightning - Training/validation metrics, hyperparameters
  • System metrics - CPU, memory, GPU usage
  • Git information - Repository state, commit hash, branch
  • Environment - Python version, package versions, hardware info

📊 Rich Data Types

Log and visualize various data types:

  • Scalars - Loss curves, accuracy, learning rates
  • Histograms - Weight distributions, gradients
  • Images - Sample predictions, confusion matrices
  • Audio - Speech samples, music generation
  • Text - Training summaries, generated text
  • Artifacts - Models, datasets, configuration files

⚡ Performance Optimized

  • Thread-safe orchestrator for concurrent logging
  • Batched operations to minimize overhead
  • Smart buffering for high-throughput scenarios
  • Configurable routing for different metric types

Architecture Overview

graph TB
    A[Your PyTorch Code] --> B[TensorBoard SummaryWriter]
    A --> C[PyTorch Lightning Trainer]
    A --> D[Direct Tracelet API]

    B --> E[Tracelet Orchestrator]
    C --> E
    D --> E

    E --> F[Plugin System]

    F --> G[MLflow Backend]
    F --> H[ClearML Backend]
    F --> I[W&B Backend]
    F --> J[AIM Backend]

    G --> K[MLflow Server]
    H --> L[ClearML Platform]
    I --> M[W&B Platform]
    J --> N[AIM Repository]

Why Tracelet?

Before Tracelet 😤

# Different APIs for each backend
import mlflow
import wandb
from clearml import Task

# Separate logging calls
mlflow.log_metric("loss", loss)
wandb.log({"loss": loss})
Task.current_task().logger.report_scalar("loss", loss)

# Manual setup for each backend
mlflow.start_run()
wandb.init(project="my-project")
task = Task.init(project_name="my-project")

With Tracelet 🎉

# One API, any backend
import tracelet

tracelet.start_logging(backend="mlflow")  # or any backend
writer.add_scalar("loss", loss)  # Works everywhere!

What's Next?

Get up and running in under 5 minutes with your first experiment

Detailed installation instructions for all backends and environments

Complete API documentation with examples and type hints

Real-world examples and Jupyter notebooks to learn from

Community & Support

  • 📚 Documentation - Comprehensive guides and API docs
  • 🐛 Issues - Bug reports and feature requests
  • 💬 Discussions - Questions and community support
  • 📧 Email - Direct contact with maintainers

License

Tracelet is released under the MIT License.

Ready to supercharge your ML experiments?
Get Started View on GitHub