MultiConfigTask

[mafrahm]

There are many tasks that might require to access columns/histograms from multiple campaigns simultaneously.

• Machine Learning
• creating datacards
• rebinning
• plotting (e.g. combining 2022preEE and 2022postEE)

There is already the MLModelTrainingMixin [1] that allows us to train DNNs with multiple campaigns. However, this mixin cannot be reused for other types of MultiConfigTasks.
Instead we should build one central MultiConfigTask base task.

The MLModelTrainingMixin currently resolves Calibrators, Selector and Producers (CSPs) for each year individually. Therefore, we cannot rely on the already defined mixins for CSPs but have to reimplement the resolving of each of those.
This could be simplified by removing the resolving for each year individually, since users can simply write their CSPs such that they are year-independent. Such a Producer might look like this:

@producer
def producer_run2(self: Producer, events: ak.Array, **kwargs) -> ak.Array:
    year = self.config_inst.campaign.x.year

    if year == 2016:
        events = self[producer_2016](events, **kwargs)
    elif year == 2017:
        events = self[producer_2017](events, **kwargs)
    elif year == 2018:
        events = self[producer_2018](events, **kwargs)


def producer_run2_init(self: Producer) -> None:
    year = self.config_inst.campaign.x.year

    if year == 2016:
        self.uses.add(producer_2016)
        self.produces.add(producer_2016)
    elif year == 2017:
        self.uses.add(producer_2017)
        self.produces.add(producer_2017)
    elif year == 2018:
        self.uses.add(producer_2018)
        self.produces.add(producer_2018)

The main changes that are required in our current configs/tasks is that our mixins for CSPs + ML need to inherit from the AnalysisTask instead of the ConfigTask and that the default CSPs need to be defined in the analysis_inst instead of the config_inst

@riga

[1]

columnflow/columnflow/tasks/framework/mixins.py

Line 1014 in e786a35

class MLModelTrainingMixin(MLModelMixinBase):