some sweep and ap features are computed by relying on the selection of a representative sweep / ap. The use should be able to define how this is chosen.

The dream would be to somehow merge SweepFeature and SweepSetFeature. This would enable the following:

• Features could have the same name, i.e. Num_AP could exist on the sweep and sweepset level. Depending on if the input is a sweep or sweepset Num_AP either gets computed for only one sweep or the compute method gets broadcasted across the entire sweepset, i.e. applied to every sweep.
• only one feature class would need to be defined for every feature containing _compute, _select and _aggregate methods.
• Then calling value on a sweep feature only calls _compute, while calling it on a sweepset also calls _select and _aggregate afterwards

from abc import ABC, abstractmethod

class Feature(ABC):
    def __init__(self, data, **kwargs):
        self.data_init(data)
        self.feature_init()
        # depending on isinstance(data, Sweep) or isinstance(data, SweepSet), 
        # methods get called on data directly or broadcasted.
    
    @abstractmethod
    def _compute(self):
     # compute can be defined however and operate on both sweepsets and sweeps
         if isinstance(data, Sweep):
            # compute feature for sweep
         elif isinstance(data, SweepSet)
             # could operate on sweepset directly or just broadcast compute function to sweeps
        return value

    @abstractmethod
    def _select(self, fts):
        return fts

    @abstractmethod
    def _aggregate(self, fts):
        return value

Issues with this approach that need to be solved:

• features are currently stored as objects in sweep.features and they + their methods/attrs can be accessed easily. If only one feature object is instantiated for all sweeps of a set this needs to be handled differently (only storing values / diagnostics potentially)
• broadcasting is not trivial, since not all functions can be equally applied across all sweeps individually, i.e. _select or _data_init.
• initialising self for every sweep does not work.
• Should a sweep be handled as a sweepset with only one index? I think this is not ideal.

Ideas and drafts for this proposal are welcome. This is work in progress.

Feasability is questionable at this point.

idea sketch for new API

it would be nice if each feature was a self contained class somehow. This would have the advantage that all feature related metadata could be initialised and stored in the object. However, will potentially lead to a lot of overhead and instantiations of object. Possible solution would be to instantiate each object only once and then passing it to SweepSetFeatureExtractor which can call __call__ to compute the feature. feature would take the current job of get_[sweep/sweepset]_[feature_name]. The diagnostic functionality could also be contained in the feature.

from abc import ABC, abstractmethod

class EphysFeature(ABC):
    def __init__(self):
        pass

    @abstractmethod
    def __call__(self, sweep):
        pass

    @abstractmethod
    def plot(self, sweep):
        pass


class SweepNumAPs(EphysFeature):
    def __init__(self):
        super().__init__()

    def __call__(self, sweep):
        pass

    def plot(self, sweep):
        pass

# instantiated once 
{"num_ap": SweepNumAPs()}

Hey,
two points for the Sweep_Burstiness feature:

• it often does not identify bursts. But is there a more detailed documentation what it should do?
• running my pipeline it worked for some voltage traces but then I got the following error:

  "[...]/ephyspy/features/utils.py", line 122, in get_sweep_burst_metrics
  burst_metrics = sweep._process_bursts()
  "[...]/ephyspy/allen_sdk/ephys_extractor.py", line 368, in _process_bursts
    bursts = ft.detect_bursts(
 "[...]/ephyspy/allen_sdk/ephys_features.py", line 1048, in detect_bursts
    isi_types[pauses] = "pauselike"
ValueError: could not convert string to float: 'pauselike'

No idea why what pauselike does....

Maybe you can have a look?
But in the meanwhile, I will skip this feature as it seems to be not very robust in my case.

mark features that are computed mostly via AllenSDK.
probably best to add it to descriptions of features.

• Check how to deal with clipped features and loc vs iloc!
• Add diagnostic plotting functions
  • spike ft diagnostics
  • sweep ft diagnostics
  • sweepset ft diagnostics
• add convenience function to plot diagnostic summary for sweep / sweepset
• Check carefully if features make sense for a few examples and hardcode tests for those (fts should be in a given range)
• Crosscheck with old pipeline
• (? make compound features that automatically satisfy dependencies! (or try to calculate them in the get_feature function))
• Add more documentation
• Fix remainder of failing tests!
• add github actions (include pyright, black, isort)
• add other stimulus modalities (ramp, short square) and change current features to long square
• #5
• #6
• #8

add suitable software license

This should only be tackled after closing #4 . Should be part of v0.0.2 release.

Integrate the additional features and diagnostic functionality better with the AllenSDK package.

A lot of modifications were made to the ephys_extractor.py and ephys_features.py from the AllenSDK. These modifications should be moved / integrated with features.py and added to the extractor via the add_sweep_feature function.

pytest raises a multitude of warnings. Reduce them if possible.