glzn

glzn (/²ɡlɪsːn/) is a minimal library to facilitate rapid training and inference of ML / AI for research, developed by the Digital Signal Processing and Image Analysis Group at the Institute of Informatics at the University of Oslo.

The name glzn is a textese disemvowelment of glissen, meaning sparse in Norwegian.

---

Goal and Scope

glzn is packaged in submodules, each designed to facilitate a specific role in running machine learning experiments with PyTorch.

• glzn.data: WDS format data wrapper for handling data of different modalities.
• glzn.log: local logging functionality, can be paired with Aim for more extensive reporting.
• glzn.cfg: config and argparse module to handle experiments with pydantic.
• glzn.proc: training and validation processor for runs, providing optimization in a neatly packaged context manager.

Planned submodules:

• glzn.optim: Commonly used optimizers for large scale image training.
• glzn.aug: Augmentation factories for commonly used setups.
• glzn.parse: Run factories for common supervised / self-supervised vision pipelines.

glzn is designed to stay minimal and efficient for HPC resources, and minimalisim is what drives the development.

---

Installation

pip install git+https://github.com/dsbifi/glzn.git

pip install git+ssh://git@github.com/dsbifi/glzn.git

---

TODOs:

• data submodule for data handling.
  • iTar implementation.
  • Basic grouping support.
  • Stem search and extraction.
  • Improved extension filtering.
  • Low overhead stateful sampling capability.
  • Add-ons (low priority):
    • Add optional encoders.
      • blosc2-openhtj2k.
      • pillow-jxl-plugin.
      • Additional video codecs.
      • Seismic data support.
    • Add encoder based grouping format.
    • Collator factory with support for NamedTuple or dict from Dataloader.
• aug submodule for augmentations.
  • Standard ViT Augmentations.
  • DEIT3 Augmentations.
  • DINO / iBOT Augmentations.
    • DINOv2 / v3 support.
  • MAE Augmentations.
• cfg submodule for config declaration.
  • Pydantic type verification.
  • Presedence logic.
• parse module for modular approach to central config / run parsing.
  • LLRD parsing support.
  • Factories for creating runs for supervised training.
    • IN1k training.
    • IN22k training.
    • COCO Segmentation training.
    • COCO Detection / Instance Seg. training.
  • Factories for creating runs for self-supervised training.
    • DINO (no MIM)
    • iBOT / DINOv2 / DINOv3
    • MAE / MIMR (MIM Refiner)
• log module for rudimentary logging to jsonl and stdout.
  • Basic logging support.
  • Add-ons (medium priority):
    • Add Aim support.
    • W&B support (low priority, locks users into pay-to-use)
• optim module with commonly used optimizers not covered by PyTorch.
  • cAdamW, StableAdamW, cStableAdamW.
  • LAMB, cLAMB.
  • Flags / registry for adaptive selection of gradient clipping (based on optimizer functionality).
  • Add-ons (medium priority):
    • Scion.
• proc submodule for train / validation processing and wrappers.
  • Simple ema wrapper.
  • Simple sched module.
    • Meta style precomputed array based schedulers.
    • wrap module for wrapping scheduled events.
  • step module, tracks relevant training / validation phases.
    • StepState class, for immutables.
    • StepTelemetry class (clock for run start, etc.).
    • StepTracker class for full experiment tracking.
  • Main proc module for context-based batch processing.
    • Gradient clipping support
    • Gradient accumulation support, in conj. with step.
    • AMP support / gradient scaling.
    • Scheduling support through wrap.
    • Simple logging via log + step modules.
    • Context manager implementation.