trackr

A lightweight, local-first ML experiment tracker — a minimal alternative to W&B/MLflow. Works fully offline, stores everything in a single SQLite file, and ships with a small web UI for comparing runs. Framework-agnostic: not tied to any particular project or ML library.

Install

pip install -e .

Quickstart

import trackr

run = trackr.init(project="arvyo", name="cnn-baseline", config={"lr": 3e-4, "epochs": 20})
for step in range(100):
    run.log({"loss": 0.5, "val_acc": 0.81}, step=step)
run.log_artifact("confusion_matrix.png")   # copies file into run folder
run.finish(status="completed")

run also works as a context manager — finish() is called automatically, with status="failed" if an exception propagates out of the with block. If the process dies without calling finish(), the run stays "running" in the DB; see trackr doctor below.

CLI

trackr list [--project NAME]        # table of runs with final metric values
trackr compare RUN1 RUN2 ...        # side-by-side config diff + final metrics
trackr rm RUN1 [RUN2 ...] [-y]      # delete run(s) + their artifacts (prompts unless -y/--yes)
trackr ui [--host H] [--port P]     # launch the web UI (default 127.0.0.1:8000)
trackr doctor [--stale-minutes N]   # mark runs with no heartbeat in N min (default 10) as "crashed"

Storage

Everything lives under ~/.trackr (override with the TRACKR_DIR env var):

• trackr.db — SQLite database with runs, metrics, and artifacts tables
• artifacts/<run_id>/ — files copied in via run.log_artifact()

Web UI

trackr ui starts a local FastAPI server: a runs table (filterable by project), and a per-run detail page with metric line charts, a config viewer, and artifact previews (images render inline). No auth, no multi-user support — it's meant to run on localhost against your own DB.

Demo

python examples/fake_training.py   # simulates 3 runs with noisy loss curves
trackr list
trackr ui

Tests

pytest

Non-goals for v1

Remote sync, teams, sweeps, and GPU monitoring are out of scope — see the TODO stubs in trackr/core.py.