What it does LabML is a Python toolkit that pipes training metrics, hardware stats, and experiment metadata to a web UI you can check from a phone or laptop. You add two lines of instrumentation to your PyTorch loop, run a local Mongo-backed server, and get live plots plus git-commit tracking without signing up for anything.

The interesting bit The mobile angle is genuine, not marketing garnish. The server is self-hosted, the client is a browser, and the whole thing is deliberately lighter than the usual experiment-platform cargo cult. It also doubles as a one-command hardware monitor via labml monitor using psutil and py3nvml.

Key highlights

• Two-line integration: wrap your loop in experiment.record() and tracker.save()
• Self-hosted server (labml app-server) with MongoDB as the only external dependency
• Distributed training support with manual UUID sync across nodes
• Custom visualization API with Colab-ready notebooks for stocks, poker CFR, etc.
• Stylized terminal logs and TensorBoard-compatible analytics views

Caveats

• MongoDB is mandatory for the server; no SQLite fallback is mentioned
• The .labml.yaml config and server URL setup are manual steps, not zero-config
• Hardware monitoring requires extra dependencies (psutil, py3nvml) installed separately

Verdict Good fit for solo researchers or small teams who want experiment tracking without SaaS lock-in or wallet drain. Skip it if you need multi-user auth, automatic hyperparameter search, or managed cloud hosting.