Ensemble Training

Train model ensembles across split/fold directory structure.

Command

admet model ensemble -c <config_path> --max-parallel <N>

Directory Structure

Expects data organized as:

assets/dataset/splits/
├── split_0/
│   ├── fold_0/
│   │   ├── train.csv
│   │   └── val.csv
│   ├── fold_1/
│   └── ...
├── split_1/
└── ...

Example

# Train 5 splits x 5 folds = 25 models
admet model ensemble -c configs/3-production/ensemble_chemprop.yaml --max-parallel 4

Config Requirements

ensemble:
  enabled: true
  n_models: 5
  aggregation: mean  # or median

data:
  data_dir: assets/dataset/splits/  # Parent directory

Resource Management

• --max-parallel 4 - Limit concurrent training jobs
• Uses Ray for parallelization
• Each fold trains independently

Output

• Per-fold models saved to split_N/fold_M directories
• Ensemble predictions with uncertainty (mean +/- std)