peal.entrypoints.train_predictor¶
Train a predictor (student or test model) from a PredictorConfig YAML.
Drives ModelTrainer in peal/training/trainers.py, which builds the
architecture named in the config, fits it on the dataset of the referenced
data config and tracks accuracy and group accuracies on the validation split.
Invocation:
python train_predictor.py --config <predictor.yaml> \
[--<any PredictorConfig field> value] [--wandb_project NAME]
Every PredictorConfig field is exposed as an optional --flag that
overrides the YAML value, e.g. --continue_training True to resume from
model_path or --is_loaded True to start from its saved weights.
Unknown flags are ignored. A <PEAL_BASE> prefix in model_path is
resolved to the project directory. A W&B run named
predictor_<dataset_class> is opened when the package and an API key are
available.
Writes the run directory model_path with config.yaml and the trained
model pickled as model.cpl.
Functions
Load the |