Summary:
- The aim of this fork is to simplify installation and ensure usage out-of-the-box
- I forked Python wrapper, added Win and Linux binaries, and made some changes in code
- So now you just need to install this package and you'll get working binaries with it
- Tested on: Ubuntu 14.04 x64, Win 7 x64
- Implementation (binaries) is slow (single thread)
- License for rgf_python: Apache License v2.0
- License for RGF: GNU GPL v3
- RGF maintainer page
- RGF page
- There are also multi-thread implementation FastRGF, but I didn't find Python wrapper for it and didn't try it
Installation:
git clone https://github.com/vecxoz/rgf_python
cd rgf_python
python setup.py install --user
Usage:
from sklearn.datasets import load_iris, load_boston
from sklearn.model_selection import cross_val_score
from rgf.rgf import RGFClassifier, RGFRegressor
# Classification
iris = load_iris()
X = iris.data
y = iris.target
model = RGFClassifier()
print(cross_val_score(model, X, y, cv = 5))
# array([ 0.96666667, 0.96666667, 0.93333333, 0.9 , 1. ])
# Regression
boston = load_boston()
X = boston.data
y = boston.target
model = RGFRegressor()
print(cross_val_score(model, X, y, cv = 5))
# [ 0.7286153 0.79284581 0.7961001 0.47978064 0.1185657 ]Hyperparameter tuning:
- Ditails on hyperparameter tuning
- max_leaf: Appropriate values are data-dependent and vary from 1000 to 10000.
- test_interval: For efficiency, it must be either multiple or divisor of 100.
- algorithm: 'RGF', 'RGF_Opt' or 'RGF_Sib'
- loss: "LS", "Log" or "Expo".
- reg_depth: Must be no smaller than 1. Meant for being used with algorithm='RGF_Opt' or 'RGF_Sib'.
- l2: Either 1, 0.1, or 0.01 often produces good results though with exponential loss (loss='Expo') and logistic loss (loss='Log') some data requires smaller values such as 1e-10 or 1e-20
- sl2: By default equal to l2. On some data, l2/100 works well