Skip to content

Migrate from other tools

You may have existing experiment logs from TensorBoard, MLflow, or Weights & Biases. Matyan provides a built-in TensorBoard converter in the client; for other tools you can use custom scripts or the public API.

TensorBoard

TensorBoard conversion lives in matyan-client. It reads TensorBoard event logs and produces a Matyan backup archive (same format as backend backups). You can then restore that archive with the client or backend (see Backups and restore).

CLI:

matyan-client convert tensorboard <input_dir> <output_path> [--experiment NAME] [--compress] [--workers N]
  • input_dir — Directory containing TensorBoard run directories (each with event files).
  • output_path — Directory where the backup will be created (a timestamped subdirectory is added). Use --compress to produce a .tar.gz archive instead.
  • --experiment — Optional; assign all converted runs to this experiment name.
  • --compress — Produce a single .tar.gz archive.
  • --workers — Number of parallel workers (default: CPU count).

After conversion, restore the backup with matyan-client restore-reingest (client) or matyan-backend restore (backend), as described in Backups and restore.

For image and audio events, the converter uses tbparse; TensorFlow is optional (install with uv sync --extra convert in matyan-client if you need image/audio support).

MLflow / Weights & Biases

There is no built-in matyan-client convert mlflow or matyan-client convert wandb yet. If you need this, please open an issue on GitHub (check for an existing one first) so we can track demand and prioritize.

Other options

  • Re-track with Matyan — Run your training (or a short replay) and log to Matyan using the same SDK patterns. Best when you can re-run or have a small dataset.
  • Future converters — Community or project-specific converters may be added later; check the repository for updates.