|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
import numpy as np |
|
6
|
|
|
|
|
7
|
|
|
from sklearn.linear_model import BayesianRidge |
|
8
|
|
|
from sklearn.gaussian_process import GaussianProcessRegressor |
|
9
|
|
|
from sklearn.gaussian_process.kernels import Matern, WhiteKernel, RBF |
|
10
|
|
|
from sklearn.ensemble import ExtraTreesRegressor as _ExtraTreesRegressor_ |
|
11
|
|
|
from sklearn.ensemble import RandomForestRegressor as _RandomForestRegressor_ |
|
12
|
|
|
from sklearn.ensemble import ( |
|
13
|
|
|
GradientBoostingRegressor as _GradientBoostingRegressor_, |
|
14
|
|
|
) |
|
15
|
|
|
|
|
16
|
|
|
|
|
17
|
|
|
class EnsembleRegressor: |
|
18
|
|
|
def __init__(self, estimators, min_std=0.001): |
|
19
|
|
|
self.estimators = estimators |
|
20
|
|
|
self.min_std = min_std |
|
21
|
|
|
|
|
22
|
|
|
def fit(self, X, y): |
|
23
|
|
|
for estimator in self.estimators: |
|
24
|
|
|
estimator.fit(X, np.ravel(y)) |
|
25
|
|
|
|
|
26
|
|
|
def predict(self, X, return_std=False): |
|
27
|
|
|
predictions = [] |
|
28
|
|
|
for estimator in self.estimators: |
|
29
|
|
|
predictions.append(estimator.predict(X).reshape(-1, 1)) |
|
30
|
|
|
|
|
31
|
|
|
predictions = np.array(predictions) |
|
32
|
|
|
mean = predictions.mean(axis=0) |
|
33
|
|
|
std = predictions.std(axis=0) |
|
34
|
|
|
|
|
35
|
|
|
std[std < self.min_std] = self.min_std |
|
36
|
|
|
|
|
37
|
|
|
if return_std: |
|
38
|
|
|
|
|
39
|
|
|
return mean, std |
|
40
|
|
|
return mean |
|
41
|
|
|
|
|
42
|
|
|
|
|
43
|
|
|
def _return_std(X, trees, predictions, min_variance): |
|
44
|
|
|
""" |
|
45
|
|
|
used from: |
|
46
|
|
|
https://github.com/scikit-optimize/scikit-optimize/blob/master/skopt/learning/forest.py |
|
47
|
|
|
""" |
|
48
|
|
|
std = np.zeros(len(X)) |
|
49
|
|
|
trees = list(trees) |
|
50
|
|
|
|
|
51
|
|
|
for tree in trees: |
|
52
|
|
|
if isinstance(tree, np.ndarray): |
|
53
|
|
|
tree = tree[0] |
|
54
|
|
|
|
|
55
|
|
|
var_tree = tree.tree_.impurity[tree.apply(X)] |
|
56
|
|
|
var_tree[var_tree < min_variance] = min_variance |
|
57
|
|
|
mean_tree = tree.predict(X) |
|
58
|
|
|
std += var_tree + mean_tree ** 2 |
|
59
|
|
|
|
|
60
|
|
|
std /= len(trees) |
|
61
|
|
|
std -= predictions ** 2.0 |
|
62
|
|
|
std[std < 0.0] = 0.0 |
|
63
|
|
|
std = std ** 0.5 |
|
64
|
|
|
return std |
|
65
|
|
|
|
|
66
|
|
|
|
|
67
|
|
|
class TreeEnsembleBase: |
|
68
|
|
|
def __init__(self, min_variance=0.001, **kwargs): |
|
69
|
|
|
self.min_variance = min_variance |
|
70
|
|
|
super().__init__(**kwargs) |
|
71
|
|
|
|
|
72
|
|
|
def fit(self, X, y): |
|
73
|
|
|
super().fit(X, np.ravel(y)) |
|
74
|
|
|
|
|
75
|
|
|
def predict(self, X, return_std=False): |
|
76
|
|
|
mean = super().predict(X) |
|
77
|
|
|
|
|
78
|
|
|
if return_std: |
|
79
|
|
|
std = _return_std(X, self.estimators_, mean, self.min_variance) |
|
80
|
|
|
|
|
81
|
|
|
return mean, std |
|
82
|
|
|
return mean |
|
83
|
|
|
|
|
84
|
|
|
|
|
85
|
|
|
class RandomForestRegressor(TreeEnsembleBase, _RandomForestRegressor_): |
|
86
|
|
|
def __init__(self, min_variance=0.001, **kwargs): |
|
87
|
|
|
super().__init__(**kwargs) |
|
88
|
|
|
|
|
89
|
|
|
|
|
90
|
|
|
class ExtraTreesRegressor(TreeEnsembleBase, _ExtraTreesRegressor_): |
|
91
|
|
|
def __init__(self, min_variance=0.001, **kwargs): |
|
92
|
|
|
super().__init__(**kwargs) |
|
93
|
|
|
|
|
94
|
|
|
|
|
95
|
|
|
class GradientBoostingRegressor(TreeEnsembleBase, _GradientBoostingRegressor_): |
|
96
|
|
|
def __init__(self, min_variance=0.001, **kwargs): |
|
97
|
|
|
super().__init__(**kwargs) |
|
98
|
|
|
|
|
99
|
|
|
|
|
100
|
|
|
class GPR: |
|
101
|
|
|
def __init__(self): |
|
102
|
|
|
length_scale_param = 1 |
|
103
|
|
|
length_scale_bounds_param = (1e-05, 100000.0) |
|
104
|
|
|
nu_param = 0.5 |
|
105
|
|
|
matern = Matern( |
|
106
|
|
|
# length_scale=length_scale_param, |
|
107
|
|
|
# length_scale_bounds=length_scale_bounds_param, |
|
108
|
|
|
nu=nu_param, |
|
109
|
|
|
) |
|
110
|
|
|
|
|
111
|
|
|
self.gpr = GaussianProcessRegressor( |
|
112
|
|
|
kernel=matern + WhiteKernel(), n_restarts_optimizer=0 |
|
113
|
|
|
) |
|
114
|
|
|
|
|
115
|
|
|
def fit(self, X, y): |
|
116
|
|
|
self.gpr.fit(X, y) |
|
117
|
|
|
|
|
118
|
|
|
def predict(self, X, return_std=False): |
|
119
|
|
|
return self.gpr.predict(X, return_std=return_std) |
|
120
|
|
|
|
|
121
|
|
|
|
|
122
|
|
|
class GPR_linear: |
|
123
|
|
|
def __init__(self): |
|
124
|
|
|
self.gpr = BayesianRidge() |
|
125
|
|
|
|
|
126
|
|
|
def fit(self, X, y): |
|
127
|
|
|
self.gpr.fit(X, y) |
|
128
|
|
|
|
|
129
|
|
|
def predict(self, X, return_std=False): |
|
130
|
|
|
return self.gpr.predict(X, return_std=return_std) |
|
131
|
|
|
|
SimonBlanke /
Gradient-Free-Optimizers
