The way we delay transformer calls (look https://gitlab.idiap.ch/bob/bob.pipelines/-/blob/master/bob/pipelines/wrappers.py#L132) makes our pipeline super memory greedy.

I'm running a simple experiment LOCALLY, no dask involved, on bob.bio.base wrapping everything with the CheckpointWrapper; and my experiment blows 32GB of my RAM + my swap without writing one single file from mine experiment.

Do you have any thoughts on this @amohammadi ? Do you think is a good call the DelayedSampleCall?

Thanks

ping @ydayer