Example of how to use multi-class metrics (and do they work in in feature_importance)?

[ebagjones]

I've tried using Precision(average='micro') as an accuracy metric and then using key_metrix_idx=n to select this for feature_importance, but I'm getting "ValueError: Classification metrics can't handle a mix of continuous-multioutput and multiclass targets"

I believe I'm probably wrong with Precision being used this way for a multi-class classification problem, but honestly I'm a little stuck on what I'm supposed to pass for 'inp' and 'targ' for the multiclass metrics like precision_multi. I tried following:
https://github.com/timeseriesAI/tsai/blob/main/tutorial_nbs/01a_MultiClass_MultiLabel_TSClassification.ipynb but this throws errors:
learn = ts_learner(dls, InceptionTime, metrics=accuracy_multi)
gives TypeError: __init__() got an unexpected keyword argument 'custom_head'
and when trying to fit:
ValueError: Target size (torch.Size([384])) must be the same as input size (torch.Size([2304]))

On a then related note, how would you recommend one proceeds on trying to increase precision or recall at the expense of accuracy? I have an unbalanced classification problem and I'm finding that feature importance sometimes indicates accuracy gets better if ALL inputs are removed, presumably because always classifying as the dominant class actually gives the highest net accuracy. Thanks in advance.