// the find
polyaxon/traceml
Engine for AI/ML/Data tracking, visualization, explainability, drift detection, and dashboards for Polyaxon.
TraceML is the tracking and visualization layer extracted from Polyaxon, the ML platform. It handles experiment logging, artifact tracking, and DataFrame profiling, and works either standalone (offline mode) or wired into a Polyaxon backend. Aimed at ML engineers who want experiment tracking without committing to a full MLflow/W&B setup.
The offline mode is genuinely useful — you get local experiment tracking with zero server dependency, which MLflow requires by default. Framework coverage is broad and consistent: Keras, PyTorch, Lightning, HuggingFace, fastai, XGBoost, LightGBM, scikit-learn all get first-class callbacks, not afterthoughts. The DataFrameSummary module gives you substantially more than pandas describe() — per-column type inference, outlier flagging, top correlations — useful for quick EDA without pulling in a full profiling library. The serialization layer uses a custom .plx binary format with typed event schemas, so you're not just dumping JSON blobs.
The README is littered with [WIP] sections — column-pair analysis, visualizations, catalog/versioning — and those items have been sitting incomplete long enough to raise questions about whether they ever ship. Standalone value is limited: the offline tracking writes local files but there's no built-in viewer; you need the full Polyaxon platform to actually see dashboards, which makes this feel like a library that escaped its parent product rather than a standalone tool. The vendored matplotlylib code (a fork of an old mpl-to-plotly converter) is unmaintained upstream and likely to break against current matplotlib. The drift detection advertised in the description doesn't appear to exist in the codebase at all.