Packora-ckpt

Packora checkpoints

Project links

• Project page
• Try Packora (live demo)
• Paper
• Code
• Dataset manifests
• Hugging Face collection

This repository contains Packora-M and Packora-L checkpoints for crystal structure prediction. Both files retain complete PyTorch Lightning state for inference or full-state stage-2 training resume. Public filenames intentionally do not encode the training epoch.

The nested layout is part of the loading interface: keep each checkpoint under its model's checkpoints/ directory and retain the sibling .hydra/config.yaml. Set the common data root before loading a model so its standardizing scaler can read csd_clari/dataset_stats.json:

export PACKORA_DATA_ROOT=/path/to/Packora-data

Python API

from pathlib import Path

from src.prediction.api import PackoraPredictor


model_root = Path("/path/to/Packora-ckpt")
predictor = PackoraPredictor(
    model_manifest_path=model_root / "model_manifest.json",
    z_prior_path=model_root / "z_distribution.json",
)
result = predictor.predict(
    {
        "model": "packora-m",
        "components": [{"smiles": "N#Cc1ccc(cc1)C#N", "ratio": 1}],
        "z": 1,
    },
    seed=42,
)
print(result.summary)

The manifest configures both released checkpoints from repository-relative entries. Pass checkpoint_paths to override either entry explicitly; direct checkpoint paths take precedence and can also be used without a manifest.

The licensed CSD Python API is required for training-equivalent SMILES featurization. Omitting z uses the included sanitized empirical z_distribution.json prior.

Resume training

Use ckpt_path, not init_ckpt_path, to restore model, EMA, optimizer, scheduler, loop, epoch, and global-step state. Set trainer.max_epochs to a total greater than the epoch stored in the selected checkpoint:

python src/finetune.py \
  data=csd_clari_finetune \
  ckpt_path=/path/to/Packora-ckpt/packora-m/checkpoints/packora-m.ckpt \
  paths.data_dir="$PACKORA_DATA_ROOT" \
  trainer.max_epochs=<new-total-epoch-limit>

New checkpoint bookkeeping and outputs are rooted in the new Hydra run. Verify downloaded artifacts with sha256sum -c SHA256SUMS.