sakura.sakuraAE.sakuraAE.train_hybrid_fastload

sakuraAE.train_hybrid_fastload(split_configs: dict, ticks=50000, hybrid_mode='interleave', prog_loss_weight_mode='epoch_end', make_logs=True, log_prefix='', log_loss_groups=['loss', 'regularization'], save_raw_loss=False, perform_test=False, test_segment=2000, tests: dict | None = None, perform_checkpoint=False, checkpoint_segment=2000, checkpoint_prefix='', checkpoint_save_arch=False, loss_prog_on_test: dict | None = None, resume=False, resume_dict=None, prefetch_strategy='reuse', reuse_factor=8, reuse_shuffle_when_reassign=False)

Implement the multithread dataloader version of hybrid mode training, where model module splits are trained with flexibility.

Parameters:

• prefetch_strategy (Literal['fresh','reuse'], optional) – The strategy for prefetching data in the multithread dataloader, defaults to ‘reuse’ sets of loaded data batches to reduce I/O overhead

• reuse_factor (int, optional) – The number of sets of prefetched data batch is reused, defaults to 8

• reuse_shuffle_when_reassign (bool, optional) – Whether to shuffle a set of data batches when reassigning them for reuse, defaults to False

Note

See also train_hybrid() for omitted parameter details on how to perform training in a configured hybrid mode.

Returns:

None