What it does TraceML is the standalone tracking engine for Polyaxon’s MLOps platform. It lets you log hyperparameters, metrics, model artifacts, and visualizations (matplotlib, Plotly, Altair, Bokeh) from Python scripts or notebooks. It also includes a DataFrameSummary tool that extends pandas’ describe() with column-type detection, missing-value stats, and distribution details.

The interesting bit The offline mode is the quietly useful feature: set is_offline=True or export POLYAXON_OFFLINE=true and you get full experiment tracking without any API server running. That’s unusual for a platform-tied tool — most competitors hard-require a hosted backend.

Key highlights

• Callback integrations for Keras, PyTorch, TensorFlow, FastAI, PyTorch Lightning, and HuggingFace
• Direct logging of matplotlib, Plotly, Altair, and Bokeh figures as artifacts
• DataFrameSummary with per-column type inference, missing-value percentages, and extended distribution stats
• Offline-first: works without Polyaxon server; sync later if needed
• Part of the broader Polyaxon ecosystem (533 stars) but installable standalone via pip install traceml

Caveats

• The “Local sandbox” feature is marked [WIP] and “Coming soon” — not available yet
• Some README examples show placeholder arguments like ... in log_artifact_ref, suggesting API surface may still be shifting

Verdict Worth a look if you want structured experiment logging without vendor lock-in to a hosted service. Skip it if you need a fully self-contained, batteries-included ML platform today — this is a component, not a complete replacement.