, Community Led Stage
Training loops shouldn't need boilerplate just to see a loss curve. As part of my GSoC 2026 work with FOSSASIA on Visdom, I built VisdomLogger — a native logging API that streams metrics straight to Visdom's live dashboard. It shipped in Visdom v0.3.0 and is already usable today.
VisdomLogger's design keeps Visdom's philosophy intact it receives values through a simple log(name, value) call and never instruments training itself. That boundary is what makes true autologging possible on top of it — and in this talk I'll show the real thing a Keras callback that hooks into model.fit() and streams loss, accuracy, and any tracked metric to Visdom automatically, with zero manual logging calls in the training loop.
In this lightning talk, I'll run a live Keras training script with no logging boilerplate and watch the dashboard populate in real time, then briefly show the same pattern working for PyTorch. I'll close with what's next extending true autologging to PyTorch Lightning, XGBoost, and scikit-learn on a shared backend.
Jayant Parashar is an open source contributor and ML engineer. He's a GSoC 2026 contributor at FOSSASIA, where he's building native autologging for Visdom — extending it across PyTorch, Keras, PyTorch Lightning, XGBoost, and scikit-learn so training metrics stream to the dashboard without manual logging code.