Skip to content

About

UFL models

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

UFL model registry

UFL_MODELS is the authoritative database for CAESAR model registrations. It stores registered checkpoints, their active and retired identities, URLs and owners. The registration client lives in CAESAR; the checks the workflow runs live here in test/check_registration.py.

Register or retire a model

Use CAESAR's register_model.py, after gh auth login:

python3 register_model.py add --name my_model --display-name "My model" \
  --description "Training data and intended use" \
  --checkpoint /path/to/my_model.pt --min-dims 3 --max-dims 5
python3 register_model.py remove my_model

Each command opens a [model-registry] pull request from your fork. add commits the checkpoint and its catalog entry; remove deletes the checkpoint file and moves its entry to retired_models as registration history. The workflow (.github/workflows/model-registry.yml) checks the pull request with test/check_registration.py from this repository, never with code from the pull request, and merges it when every check passes; otherwise it comments the reason. It checks that:

  • exactly one model is registered or retired and nothing else changed;
  • the owner is the pull request author's login and numeric GitHub ID, and the registration number is the next unused one; names, files, hashes and numbers are unique among active and retired models;
  • the checkpoint is a real CAESAR model: a PyTorch state_dict of finite float32 tensors with the default model's parameter names and shapes, within 5% of its size. It is read as data by a restricted unpickler and never executed;
  • the URL names the pull request commit that adds the checkpoint, the SHA-256 matches, and the README model list matches the catalog.

Only E53klasky (GitHub ID 173562962) may retire models. Model owners and other administrators cannot remove them. The stable numeric ID also works after an account rename. Only an administrator may change the default using remove NAME --default-model REPLACEMENT. Removal deletes the checkpoint from the current branch and preserves identities; numbers, names, hashes and filenames remain reserved. Git history retains older checkpoint versions; removal does not shrink the full repository history. Contributors need no write access. Direct write access to the catalog, checks or workflow bypasses this service; keep it restricted to administrators. The workflow merges with merge commits, so keep merge commits allowed.

The compatibility register_model.py here forwards to an adjacent CAESAR checkout. See CAESAR's registration guide for repository settings, testing the workflow, and synchronizing CAESAR's release snapshot.

Database checks

python3 ../CAESAR/register_model.py check --database .
python3 -m unittest discover -s test -v
Registration Model Owner Status Checkpoint Dimensions Description
1 CAESAR v1 (caesar_v1) E53klasky Active caesar_v.pt 3D–5D Foundation model, original CAESAR v1 release; trained for scientific data with 3D–5D inputs.
2 CAESAR v2 (caesar_v2) E53klasky Active (default) model_bs64_ep100k.pt 3D–5D Foundation model, optimized CAESAR v2 release; trained for scientific data with 3D–5D inputs.
3 eelsM1 (eelsM1) E53klasky Active eelsM1.pt 3D–5D Domain fine-tune of the CAESAR foundation model for EELS; trained and tested on one EELS dataset, with 3D–5D inputs.
4 microscopy (microscopy) E53klasky Active microscopy.pt 3D–5D Domain-specific microscopy model; trained and tested across four microscopy datasets, with 3D–5D inputs.
5 Xcompact3d latent 16x16 (Xcompact3d_latent_16_16) E53klasky Active Xcompact3d_latent_16_16.pt 3D–5D Xcompact3d checkpoint; document training data and intended use
6 rahul_isabel_tc_v1 (workflow test) (rahul_isabel_tc_v1) arsenious Active rahul_isabel_tc_v1.pt 3D–5D Workflow test, not for production. caesar_v2 fine-tuned 3k steps on Hurricane Isabel TCf48 frames 0-79; held-out NRMSE 0.00623.
7 RE3200 BS8 100K (re3200_bs8_ep100k) E53klasky Removed model_bs8_ep100k.pt 3D–5D CAESAR VAE3D model trained on RE3200 with batch size 8 for 100K iterations

About

UFL models

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages