-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathCV_pipeline.py
More file actions
190 lines (180 loc) · 8.15 KB
/
CV_pipeline.py
File metadata and controls
190 lines (180 loc) · 8.15 KB
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
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
import numpy as np
import pandas as pd
import torch
from sklearn.model_selection import KFold
from sklearn.metrics import pairwise
from tqdm import tqdm
from losses import ce_risk
from ece_bin import ECE_bin_fit
from ece_kde import ECE_kde_fit
from KRR_estimators import ECE_krr_fit, ECE_kkrr_fit
def ECE_bin_CV(
Ps, Ys, bin_range=np.arange(5, 95, 5), k_folds_val=5, k_folds_test=1, disable_pb=False, seed=0
):
# use 20% as test set and do CV on the other 80%
if k_folds_test < 2:
n_all = Ps.shape[0]
np.random.seed(seed)
shuffle_ids = np.random.choice(n_all, n_all, replace=False)
Ps = Ps[shuffle_ids]
Ys = Ys[shuffle_ids]
split = int(n_all*0.8)
Ps_train_val = Ps[:split]
Ys_train_val = Ys[:split]
Ps_test = Ps[split:]
Ys_test = Ys[split:]
#test_ind = np.random.binomial(n=1, p=0.2, size=n_all)
#Ps_train_val = Ps[test_ind==0]
#Ys_train_val = Ys[test_ind==0]
#Ps_test = Ps[test_ind==1]
#Ys_test = Ys[test_ind==1]
kf = KFold(n_splits=k_folds_val)
val_results = []
fold = 0
for train_ind, val_ind in tqdm(kf.split(Ps_train_val), total=k_folds_val, disable=disable_pb):
Ps_train = Ps_train_val[train_ind]
Ys_train = Ys_train_val[train_ind]
Ps_val = Ps_train_val[val_ind]
Ys_val = Ys_train_val[val_ind]
for n_bins in bin_range:
cal_model = ECE_bin_fit(Ps_train, Ys_train, n_bins)
cal_preds_val = cal_model(Ps_val)
val_risk = ce_risk(Ps_val, Ys_val, cal_preds_val)
avg_pred_val = np.diag(cal_preds_val).mean()
cal_preds_test = cal_model(Ps_test)
test_risk = ce_risk(Ps_test, Ys_test, cal_preds_test)
avg_pred_test = np.diag(cal_preds_test).mean()
val_results += [{
'n_bins': n_bins, 'fold': fold, 'val_risk': val_risk, 'test_risk': test_risk,
'avg_pred_val': avg_pred_val, 'avg_pred_test': avg_pred_test
}]
fold += 1
pd_results = pd.DataFrame(val_results).groupby(['n_bins']).mean()
optim = pd_results['val_risk'].argmin()
ece_estimate = pd_results.iloc[optim]['avg_pred_test'].item()
return ece_estimate, val_results
def CWCE_bin_CV(Ps, Ys, **kwargs):
n_classes = Ps.shape[1]
one_hot_Ys = np.zeros((Ys.shape[0], n_classes))
one_hot_Ys[np.arange(Ys.shape[0]), Ys] = 1
CWCE_est = 0
for k in range(n_classes):
ce_est, _ = ECE_bin_CV(Ps[:, k], one_hot_Ys[:, k], **kwargs)
CWCE_est += ce_est
return CWCE_est
def ECE_krr_CV(
Ps, Ys, reg_range=10.**np.arange(-1, -18, -2), k_folds_val=5, k_folds_test=1,
disable_pb=False, kernel_X=lambda x,y: pairwise.rbf_kernel(x, y, gamma=1), seed=0,
use_kkrr=False
):
if len(Ps.shape) > 1 and Ps.shape[1] > 1:
is_binary = False
Y_onehot = np.zeros((Ys.size, Ys.max() + 1))
Y_onehot[np.arange(Ys.size), Ys] = 1
else:
is_binary = True
ECE_model_fit = ECE_kkrr_fit if use_kkrr else ECE_krr_fit
# use 20% as test data and do CV with the other 80%
if k_folds_test < 2:
n_all = Ps.shape[0]
np.random.seed(seed)
shuffle_ids = np.random.choice(n_all, n_all, replace=False)
Ps = Ps[shuffle_ids]
Ys = Ys[shuffle_ids]
split = int(n_all*0.8)
Ps_train_val = Ps[:split]
Ys_train_val = Ys[:split]
Ps_test = Ps[split:]
Ys_test = Ys[split:]
if is_binary:
K_XX = kernel_X(Ps.reshape(-1, 1), Ps.reshape(-1, 1))
K_YY = np.outer(Ps-Ys, Ps-Ys)
else:
K_XX = kernel_X(Ps, Ps)
K_YY = (Ps-Y_onehot) @ (Ps-Y_onehot).T
K_XX_train_val_train_val = K_XX[:split, :][:, :split]
K_YY_train_val_train_val = K_YY[:split, :][:, :split]
K_XX_train_val_test = K_XX[:split, :][:, split:]
#test_ind = np.random.binomial(n=1, p=0.2, size=n_all)
#Ps_train_val = Ps[test_ind==0]
#Ys_train_val = Ys[test_ind==0]
#Ps_test = Ps[test_ind==1]
#Ys_test = Ys[test_ind==1]
# K_XX_train_val_train_val = K_XX[test_ind==0, :][:, test_ind==0]
# K_YY_train_val_train_val = K_YY[test_ind==0, :][:, test_ind==0]
# K_XX_train_val_test = K_XX[test_ind==0, :][:, test_ind==1]
kf = KFold(n_splits=k_folds_val)
val_results = []
fold = 0
for train_ind, val_ind in tqdm(kf.split(Ps_train_val), total=k_folds_val, disable=disable_pb):
Ps_train = Ps_train_val[train_ind]
Ys_train = Ys_train_val[train_ind]
Ps_val = Ps_train_val[val_ind]
Ys_val = Ys_train_val[val_ind]
K_XX_train_train = K_XX_train_val_train_val[train_ind, :][:, train_ind]
K_YY_train_train = K_YY_train_val_train_val[train_ind, :][:, train_ind]
K_XX_train_val = K_XX_train_val_train_val[train_ind, :][:, val_ind]
K_XX_train_test = K_XX_train_val_test[train_ind, :]
cal_model = ECE_model_fit(K_XX_train_train, K_YY_train_train)
for reg_const in reg_range:
cal_preds_val = cal_model(K_XX_train_val, reg_const)
nan_val = np.isnan(cal_preds_val).mean()
cal_preds_test = cal_model(K_XX_train_test, reg_const)
val_risk = ce_risk(Ps_val, Ys_val, cal_preds_val)
test_risk = ce_risk(Ps_test, Ys_test, cal_preds_test)
avg_pred_val = np.diag(cal_preds_val).mean()
avg_pred_test = np.diag(cal_preds_test).mean()
val_results += [{
'reg_const': reg_const, 'fold': fold, 'val_risk': val_risk, 'test_risk': test_risk,
'avg_pred_val': avg_pred_val, 'avg_pred_test': avg_pred_test, 'nan_val': nan_val,
}]
fold += 1
pd_results = pd.DataFrame(val_results).groupby(['reg_const']).mean()
optim = pd_results['val_risk'].argmin()
ece_estimate = pd_results.iloc[optim]['avg_pred_test']
return ece_estimate, val_results
def ECE_kde_CV(
Ps, Ys, bw_range=torch.cat((torch.logspace(start=-5, end=-1, steps=15), torch.linspace(0.2, 1, steps=5))), seed=0, disable_pb=False, k_folds_val=5, k_folds_test=1
):
if len(Ps.shape) > 1 and Ps.shape[1] > 1:
is_binary = False
else:
is_binary = True
# use 20% as test data and do CV with the other 80%
if k_folds_test < 2:
n_all = Ps.shape[0]
np.random.seed(seed)
shuffle_ids = np.random.choice(n_all, n_all, replace=False)
Ps = Ps[shuffle_ids]
Ys = Ys[shuffle_ids]
split = int(n_all*0.8)
Ps_train_val = Ps[:split]
Ys_train_val = Ys[:split]
Ps_test = Ps[split:]
Ys_test = Ys[split:]
kf = KFold(n_splits=k_folds_val)
val_results = []
fold = 0
for train_ind, val_ind in tqdm(kf.split(Ps_train_val), total=k_folds_val, disable=disable_pb):
Ps_train = Ps_train_val[train_ind]
Ys_train = Ys_train_val[train_ind]
Ps_val = Ps_train_val[val_ind]
Ys_val = Ys_train_val[val_ind]
for bw in bw_range:
cal_model = ECE_kde_fit(Ps_train, Ys_train, bw=bw)
cal_preds_val = cal_model(Ps_val)
nan_val = np.isnan(cal_preds_val).mean()
val_risk = ce_risk(Ps_val, Ys_val, cal_preds_val, ignore_nan=True)
avg_pred_val = np.nanmean(np.diag(cal_preds_val))
cal_preds_test = cal_model(Ps_test)
test_risk = ce_risk(Ps_test, Ys_test, cal_preds_test, ignore_nan=True)
avg_pred_test = np.nanmean(np.diag(cal_preds_test))
val_results += [{
'bandwidth': bw, 'fold': fold, 'val_risk': val_risk, 'test_risk': test_risk,
'avg_pred_val': avg_pred_val, 'avg_pred_test': avg_pred_test, 'nan_val': nan_val,
}]
fold += 1
pd_results = pd.DataFrame(val_results).groupby(['bandwidth']).mean()
optim = pd_results['val_risk'].argmin()
ece_estimate = pd_results.iloc[optim]['avg_pred_test'].item()
return ece_estimate, val_results