Configuration

Configuration files live under uqdd/config/ and control data preparation, model defaults, and sweeps.

Data configuration

• papyrus.json: defines data sources, descriptor choices, and splitting strategy for Papyrus++.

  • Typical keys: activity_type (xc50|kx), descriptor_protein, descriptor_chemical, split_type ( random|scaffold|time), file_ext (pkl|parquet|csv), and flags (e.g., sanitize).

• desc_dim.json: maps descriptor names to vector dimensions (e.g., ecfp2048: 2048, ankh-large: <dim>), used to size model inputs correctly.

Model configuration (defaults)

• pnn.json, ensemble.json, mcdropout.json, evidential.json, eoe.json, emc.json:
  • Set default hyperparameters per model family: architecture (layers, hidden sizes), dropout, learning rate, epochs, batch size, and uncertainty-specific knobs (e.g., number of MC samples, ensemble size).
  • These defaults are read by training scripts and can be overridden via CLI (see Models & Training).

Sweep configurations

• pnn-sweep.json, evidential-sweep.json:
  • Define parameter grids or distributions for hyperparameter searches (often used with Weights & Biases sweeps), e.g., varying learning rates, widths, and regularization.

How configuration is used

• The main parser uqdd/models/model_parser.py reads CLI flags and merges them with the chosen config JSON to produce the final run configuration.
• Data preparation scripts reference papyrus.json and desc_dim.json to validate descriptor dimensions and save outputs consistently.

Customizing experiments

• Start with a base model JSON (e.g., evidential.json), then override via CLI flags for quick experiments:
  • Example: --epochs 100 --batch_size 512 --lr 3e-4
• For consistent runs across teams, keep custom JSONs under uqdd/config/ and refer to them with a flag or by adjusting the default in model_parser.py.

Best practices

• Keep descriptor names consistent across data and model configs.
• Update desc_dim.json when adding new embeddings to avoid shape mismatches.
• Version control any changes to config files and document rationale in commit messages.