Time efficiency

• use preprocessing: mml pp ...

• optimise number of workers: num_workers=12 - use neither to many nor too few

• use caching: sampling.enable_caching=true

• reduce validation checks: trainer.check_vals_every_n_epochs=10 - but keep in mind this may prohibit early stopping or reduce learning rate on plateau

• early stopping: callbacks=early

• pruning (not supported yet see issue #19)

Memory efficiency

• remove intermediates: remove.parameters=true - remove any NEW intermediate of the specified type AFTER an experiment is done

• store_parameters (only for train mode): mml train mode.store_parameters=false - does not store parameters at all