Core Concepts

Project

The top-level grouping for experiments. Tracker(project="demo") puts all runs under the demo project. The dashboard lists experiments per project.

Run

A single experiment run, created by Tracker. Each run has:

• run_id: a unique identifier (e.g. run_a1b2c3), usable with resume_from
• state: running → finished | crashed | archived
• created_at / heartbeat_at: creation and last heartbeat time
• owner_id: owning user (isolation in multi-user mode)

Run state is managed by heartbeats: the SDK sends one every 30s; the server marks a run crashed after 60s without one.

Step

The training iteration number — the x-axis of your metrics. Pass it explicitly with t.log({...}, step=step), or let it auto-increment.

Metric

Scalars logged with log(). TensorBoard-style key/context naming is supported (e.g. train/loss, val/accuracy), and the UI groups them automatically.

Summary

Incremental statistics per run per metric: last, min, max, best, best_step. best respects the direction declared in metric_directions. The table view reads summaries directly — no need to fetch full series.

Sweep

A hyperparameter search group. In practice, compare hyperparameter combinations with the Parallel chart in the Explore workspace.

Config

Tracker(config={...}) logs hyperparameters. It supports:

• Explicit dicts
• Automatic capture of argparse / OmegaConf / Hydra parameters
• A Config tab in the run view, and config-diff comparison in the Compare page

Data flow

Python SDK [log()] → RingBuffer → FlushThread → PyO3/HTTP
    → Rust Ingest [Channel → Writer → SQLite/PostgreSQL/File]
    → Axum API → Web UI