peal.training.trainers.distill_1ofn_dataset¶
- peal.training.trainers.distill_1ofn_dataset(predictor_distillation, base_path, predictor, predictor_datasets)[source]¶
Relabel
Image2ClassDataset(folder-per-class) data with a predictor.Every sample of the first two datasets is classified and saved as
<base_path>/dataset_<i>/<predicted class>/<idx>.png(skipped when the directory exists), then reloaded as a dataset without confounder information; index 0 is the train split, index 1 the validation split. Samples with more than three channels, or that fail to load or save, are skipped with a printed message.- Parameters:
predictor_distillation (
str,dictorPredictorConfig) – Config of the student.base_path (
str) – Directory the relabelled image folders are written to.predictor (
torch.nn.Module) – Teacher that produces the labels.predictor_datasets (
list) – Two datasets, dataloader mixers or weighted dataloader lists.
- Returns:
The two relabelled datasets.
- Return type: