labml

Why we included this project

Long training runs are easier to live with when you can check on them from anywhere. labml wraps a PyTorch, TensorFlow, or Keras training loop with a couple of lines of code and streams live metrics, including loss, accuracy, and hardware usage like GPU and CPU load, to a web dashboard you can open from a phone or laptop. The same package records experiment metadata such as the git commit, configuration, and hyperparameters, so you keep a lightweight record of what actually produced a given run. Distributed training works through a shared experiment ID across machines, which makes it practical for multi-GPU setups. Teams that want a self-hosted alternative to hosted experiment trackers, or just a quick way to watch hardware on any machine with a single command, will find the setup straightforward and the MIT license easy to work with.