featureclass
Feature engineering library that helps you keep track of feature dependencies, documentation and schema
This library helps define a featureclass.
featureclass is inspired by dataclass, and is meant to provide alternative way to define features engineering classes.
I have noticed that the below code is pretty common when doing feature engineering:
from statistics import variance
from math import sqrt
class MyFeatures:
def calc_all(self, datapoint):
out = {}
out['var'] = self.calc_var(datapoint),
out['stdev'] = self.calc_std(out['var'])
return out
def calc_var(self, data) -> float:
return variance(data)
def calc_stdev(self, var) -> float:
return sqrt(var)
Some things were missing for me from this type of implementation:
- Implicit dependencies between features
- No simple schema
- No documentation for features
- Duplicate declaration of the same feature - once as a function and one as a dict key
This is why I created this library.
I turned the above code into this:
float: """Calc stdev""" return sqrt(self.var) print(feature_names(MyFeatures)) # ('var', 'stdev') print(feature_annotations(MyFeatures)) # {'var': float, 'stdev': float} print(asDict(MyFeatures([1,2,3,4,5]))) # {'var': 2.5, 'stdev': 1.5811388300841898} print(asDataclass(MyFeatures([1,2,3,4,5]))) # MyFeatures(stdev=1.5811388300841898, var=2.5) ">
from featureclass import feature, featureclass, feature_names, feature_annotations, asDict, asDataclass from statistics import variance from math import sqrt @featureclass class MyFeatures: def __init__(self, datapoint): self.datapoint = datapoint @feature() def var(self) -> float: """Calc variance""" return variance(self.datapoint) @feature() def stdev(self) -> float: """Calc stdev""" return sqrt(self.var) print(feature_names(MyFeatures)) # ('var', 'stdev') print(feature_annotations(MyFeatures)) # {'var': float, 'stdev': float} print(asDict(MyFeatures([1,2,3,4,5]))) # {'var': 2.5, 'stdev': 1.5811388300841898} print(asDataclass(MyFeatures([1,2,3,4,5]))) # MyFeatures(stdev=1.5811388300841898, var=2.5)
The feature decorator is using cached_property to cache the feature calculation,
making sure that each feature is calculated once per datapoint
36 Nov 2, 2022
7 Dec 15, 2021
15 Nov 19, 2022
2 Mar 1, 2022
5 Feb 12, 2022
38 Feb 20, 2022
3 Dec 2, 2022
6 Aug 7, 2022
3 Jan 4, 2022
152 Dec 28, 2022
1.7k Jan 03, 2023
403 Jan 02, 2023
2 Sep 12, 2022
2 Jul 20, 2022
10 Jun 05, 2022
14 Aug 18, 2022
13 May 24, 2022
11 Nov 30, 2022
31 Sep 15, 2022
4 Dec 05, 2021
77 Nov 04, 2022
22k Jan 06, 2023
1 Feb 09, 2022
3.8k Dec 31, 2022
1 Feb 08, 2022
8 Dec 04, 2021
9 Nov 04, 2022
14 May 29, 2022