CODEWITHSUNDEEP
pythonmachine-learningscikit-learn
ncgoodbody
ncgoodbody

Reputation: 243

Why is my sklearn custom transformer not saving an attribute when used in a ColumnTransformer?

I am working with the California Housing data set in scikit-learn. I want to engineer two binary features: "within 10 km of San Francisco" and "within 10 km of Los Angeles". I created a custom transformer that works fine on its own, but throws a TypeError when I put it into a ColumnTransformer. Here's the code:

from math import radians
from sklearn.base import BaseEstimator, TransformerMixin
from sklearn.compose import ColumnTransformer
from sklearn.metrics.pairwise import haversine_distances
from sklearn.datasets import fetch_california_housing
import numpy as np
import pandas as pd

# Import data into DataFrame
data = fetch_california_housing()
X = pd.DataFrame(data['data'], columns=data['feature_names'])
y = data['target']

# Custom transformer for 'Latitude' and 'Longitude' cols
class NearCity(BaseEstimator, TransformerMixin):
    def __init__(self, distance=10):
        self.la = (34.05, -118.24)
        self.sf = (37.77, -122.41)
        self.dis = distance

    def calc_dist(self, coords_1, coords_2):
        coords_1 = [radians(_) for _ in coords_1]
        coords_2 = [radians(_) for _ in coords_2]
        result = haversine_distances([coords_1, coords_2])[0,-1]
        return result * 6_371

    def fit(self, X, y=None):
        return self

    def transform(self, X):
        dist_to_sf = np.apply_along_axis(self.calc_dist, 1, X, coords_2=self.sf)
        dist_to_sf = (dist_to_sf < self.dis).astype(int)

        dist_to_la = np.apply_along_axis(self.calc_dist, 1, X, coords_2=self.la)
        dist_to_la = (dist_to_la < self.dis).astype(int)

        X_trans = np.column_stack((X, dist_to_sf, dist_to_la))
        return X_trans

ct = ColumnTransformer([('near_city', NearCity(), ['Latitude', 'Longitude'])],
                       remainder='passthrough')

ct.fit_transform(X)
#> /Users/.../anaconda3/envs/data3/lib/python3.7/site-packages/sklearn/base.py:197: FutureWarning: From version 0.24, get_params will raise an AttributeError if a parameter cannot be retrieved as an instance attribute. Previously it would return None.
#>   FutureWarning)
#> Traceback (most recent call last):
#> <ipython-input-13-603f6cd4afd3> in transform(self, X)
#>      17     def transform(self, X):
#>      18         dist_to_sf = np.apply_along_axis(self.calc_dist, 1, X, coords_2=self.sf)
#> ---> 19         dist_to_sf = (dist_to_sf < self.dis).astype(int)
#>      20 
#>      21         dist_to_la = np.apply_along_axis(self.calc_dist, 1, X, coords_2=self.la)
#> TypeError: '<' not supported between instances of 'float' and 'NoneType'

Created on 2020-04-23 by the reprexpy package

The issue is that the self.dis attribute doesn't persist. If I instantiate the transformer by itself, no problem: self.dis = distance = 10. But in the ColumnTransformer, it ends up as NoneType. Oddly enough, if I just hard-code in self.dis = 10, it works.

What do folks think is going on?

Session info --------------------------------------------------------------------
Platform: Darwin-18.7.0-x86_64-i386-64bit (64-bit)
Python: 3.7
Date: 2020-04-23
Packages ------------------------------------------------------------------------
numpy==1.18.1
pandas==1.0.1
reprexpy==0.3.0
scikit-learn==0.22.1

Upvotes: 3

Views: 1282

Answers (1)

ncgoodbody
ncgoodbody

Reputation: 243

Turns out the problem lies in sklearn.base. 

deep_items = value.get_params().items()

The get_params() function looks at the init arguments to figure out what the class parameters are and then assumes that they're the same as the internal variable names.

So I can solve this problem by changing my init method to: 

def __init__(self, distance=10):
    self.la = (34.05, -118.24)
    self.sf = (37.77, -122.41)
    self.distance = distance # <-- give same name

Many thanks to a colleague of mine who figured this out!

Upvotes: 6

Related Questions