-
Notifications
You must be signed in to change notification settings - Fork 47
/
Copy pathtest_treeutils.py
150 lines (119 loc) · 4.66 KB
/
test_treeutils.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
import numpy as np
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import RandomForestClassifier
import adapt._tree_utils as ut
np.random.seed(0)
# Generate training source data
ns = 200
ns_perclass = ns // 2
mean_1 = (1, 1)
var_1 = np.diag([1, 1])
mean_2 = (3, 3)
var_2 = np.diag([2, 2])
Xs = np.r_[np.random.multivariate_normal(mean_1, var_1, size=ns_perclass),
np.random.multivariate_normal(mean_2, var_2, size=ns_perclass)]
ys = np.zeros(ns)
ys[ns_perclass:] = 1
# Generate training target data
nt = 50
# imbalanced
nt_0 = nt // 10
mean_1 = (6, 3)
var_1 = np.diag([4, 1])
mean_2 = (5, 5)
var_2 = np.diag([1, 3])
Xt = np.r_[np.random.multivariate_normal(mean_1, var_1, size=nt_0),
np.random.multivariate_normal(mean_2, var_2, size=nt - nt_0)]
yt = np.zeros(nt)
yt[nt_0:] = 1
# Generate testing target data
nt_test = 1000
nt_test_perclass = nt_test // 2
Xt_test = np.r_[np.random.multivariate_normal(mean_1, var_1, size=nt_test_perclass),
np.random.multivariate_normal(mean_2, var_2, size=nt_test_perclass)]
yt_test = np.zeros(nt_test)
yt_test[nt_test_perclass:] = 1
# Source classifier
RF_SIZE = 10
classes_test = [0,1]
node_test = 5
node_test2 = 4
feats_test = np.array([0,1])
values_test = np.array([5,10])
clf_source_dt = DecisionTreeClassifier(max_depth=None)
clf_source_rf = RandomForestClassifier(n_estimators=RF_SIZE)
clf_source_dt.fit(Xs, ys)
clf_source_rf.fit(Xs, ys)
Nkmin = sum(yt == 0 )
root_source_values = clf_source_dt.tree_.value[0].reshape(-1)
props_s = root_source_values
props_s = props_s / sum(props_s)
props_t = np.zeros(props_s.size)
for k in range(props_s.size):
props_t[k] = np.sum(yt == k) / yt.size
coeffs = np.divide(props_t, props_s)
def test_depth():
ut.depth_tree(clf_source_dt)
ut.depth_rf(clf_source_rf)
ut.depth(clf_source_dt,node_test)
ut.depth_array(clf_source_dt,np.arange(clf_source_dt.tree_.node_count))
def test_rules():
ut.sub_nodes(clf_source_dt.tree_,node_test)
parent,direction = ut.find_parent_vtree(clf_source_dt.tree_, node_test)
parent,direction = ut.find_parent(clf_source_dt, node_test)
p,t,b = ut.extract_rule_vtree(clf_source_dt.tree_,node_test)
p,t,b = ut.extract_rule(clf_source_dt,node_test)
p2,t2,b2 = ut.extract_rule(clf_source_dt,node_test2)
rule = p,t,b
rule2 = p2,t2,b2
split_0 = p[0],t[0]
ut.isinrule(rule, split_0)
ut.isdisj_feat(p[0],t[0],p[1],t[1])
ut.isdisj(rule,rule2)
try:
n_feat = clf_source_dt.n_features_
except:
n_feat = clf_source_dt.n_features_in_
ut.bounds_rule(rule,n_feat)
leaves,rules = ut.extract_leaves_rules(clf_source_dt)
ut.add_to_parents(clf_source_dt, node_test, values_test)
def test_splits():
leaves,rules = ut.extract_leaves_rules(clf_source_dt)
p,t,b = ut.extract_rule(clf_source_dt,node_test)
p2,t2,b2 = ut.extract_rule(clf_source_dt,node_test2)
rule = p,t,b
rule2 = p2,t2,b2
ut.coherent_new_split(p[1],t[1],rule2)
ut.liste_non_coherent_splits(clf_source_dt,rule)
all_splits = np.zeros(clf_source_dt.tree_.node_count - leaves.size,dtype=[("phi",'<i8'),("th",'<f8')])
coh_splits = ut.all_coherent_splits(rule,all_splits)
s = coh_splits.size
ut.filter_feature(all_splits,feats_test)
ut.new_random_split(np.ones(s)/s,coh_splits)
def test_error():
e = ut.error(clf_source_dt.tree_,node_test)
le = ut.leaf_error(clf_source_dt.tree_,node_test)
return e,le
def test_distribution():
ut.get_children_distributions(clf_source_dt,node_test)
ut.get_node_distribution(clf_source_dt,node_test)
ut.compute_class_distribution(classes_test,ys)
phi = clf_source_dt.tree_.feature[0]
threshold = clf_source_dt.tree_.threshold[0]
ut.compute_Q_children_target(Xs,ys,phi,threshold,classes_test)
def test_pruning_risk():
ut.compute_LLR_estimates_homog(clf_source_dt,Nkmin=Nkmin)
ut.contain_leaf_to_not_prune(clf_source_dt,Nkmin=Nkmin,coeffs=coeffs)
def test_divergence_computation():
phi = clf_source_dt.tree_.feature[0]
threshold = clf_source_dt.tree_.threshold[0]
Q_source_parent = ut.get_node_distribution(clf_source_dt,node_test)
Q_source_l, Q_source_r = ut.get_children_distributions(clf_source_dt,node_test)
Q_target_l, Q_target_r = ut.compute_Q_children_target(Xt,yt,phi,threshold,classes_test)
ut.H(Q_source_parent)
ut.GINI(Q_source_parent)
ut.IG(Q_source_parent,[Q_target_l, Q_target_r])
ut.DG(Q_source_l,Q_source_r,Q_target_l,Q_target_r)
ut.JSD(Q_target_l, Q_source_l)
ut.KL_divergence(Q_source_l,Q_target_l)
ut.threshold_selection(Q_source_parent,Q_source_l,Q_source_r,Xt,yt,phi,classes_test)