MultiDatasetTrainer#

class MultiDatasetTrainer(cfg)[source]#

Bases: Trainer

Trainer that draws mixed batches from several datasets.

Identical to Trainer except build_train_loader() swaps the standard loader for MultiDatasetDataloader, which samples across the configured datasets (honoring per-dataset ratios and mix_prob) and defines the epoch length. Registered as MultiDatasetTrainer – select via train = dict(type="MultiDatasetTrainer").

build_train_loader()[source]#: Build a multi-dataset train loader and expose its epoch length.