MAPS Structure

[mdiazmel]

Currently MAPS structure stores a huge amount of information on the same place. This was the initial wish at the moment of designing this structure, in order to facilitate the reproduction/replication other experiments. However, some inconveniences arise from the original design. E.g. how to manage the write of outputs inside this folder when several instances of the clinicadl are working at the same time (multiple splits parallelized in multiple nodes, one per split).

Management of this information can be inspired from the way how MLflow store and serve the training (meta)information and the models (same for Weight&Biases). Another possibility can include to propose an option for "expert" users, with a fine detail of management for the assets generated by ClinicaDL.