Aspara

The loss must go on. Aspara tracks every step of the descent, all the way to convergence.

Why Aspara?

• Fast by design: LTTB-based metric downsampling keeps dashboards responsive
• Built for scale: manage hundreds runs without friction
• Flexible UI: Web dashboard and TUI dashboard from the same data

Try the Demo

Want to see Aspara in action without installing? Try the live demo.

https://prednext-aspara.hf.space/

The demo lets you explore the experiment results dashboard with sample data.

Requirements

• Python 3.10+

Installation

# Install with all features
pip install aspara[all]

# Or install components separately
pip install aspara              # Client only
pip install aspara[dashboard]   # Dashboard only
pip install aspara[tracker]     # Tracker only

Quick Start

1. Log your experiments (just 3 lines!)

import aspara

aspara.init(project="my_project", config={"lr": 0.01, "batch_size": 32})

for epoch in range(100):
    loss, accuracy = train_one_epoch()
    aspara.log({"train/loss": loss, "train/accuracy": accuracy}, step=epoch)

aspara.finish()

2. Visualize results

aspara dashboard

Open http://localhost:3141 to compare runs, explore metrics, and share insights.

3. Or use the Terminal UI

pip install aspara[tui]
aspara tui

Navigate projects, runs, and metrics with Vim-style keybindings. Perfect for SSH sessions and terminal workflows.

Documentation

• Getting Started
• Dashboard Guide
• User Guide
• API Reference

Development

See DEVELOPMENT.md for development setup and guidelines.

Quick setup:

pnpm install && pnpm build  # Build frontend assets
uv sync --dev               # Install Python dependencies