CurveFit trainer (train model with simple curve fit)

[sunya-ch]

What would you like to be added?

I would like to add a new trainer class in addition to scikit and xgboost called curvefit (based on scipy curve_fit function).
This will be simple trainer that based on only single main feature of the group, for example, bpf_cpu_time_us of BPFOnly group.

On Trainer instance creation, fit_func such as logarithm func, logistic func is needed to be set.
In addition, p0 and main feature can be alternatively set via p0_func and feature_index, index of the target feature in the FeatureGroup, correspondingly.

class CurveFitModel():
    def __init__(self, fit_func, p0_func=None):
        self.fit_func = fit_func
        self.popt = None
        self.pcov = None
        self.feature_index = None
        self.p0_func = p0_func

Signed-off-by: Sunyanan Choochotkaew sunyanan.choochotkaew1@ibm.com

Why is this needed?

The main objective is to expand the range of power model training to existing data set such as SPECPower where only CPU usage metric is available.

[marceloamaral]

@sunya-ch Interesting
Btw, why scikit or xgboost are not enough?

[sunya-ch]

The curvefit allows more flexibility for adding a new simple model function that based on a single independent variable not limited to what provided from the scikit.
I put my PR here: #215.

[rootfs]

The curve fitting algorithms can be implemented in golang for kepler without introducing additional pkg or libraries.