-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval_experiments.py
171 lines (146 loc) · 7.49 KB
/
eval_experiments.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
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
import os
import sys
import random
import math
import numpy as np
from sklearn.metrics import f1_score
from scoring.dnn import get_model
import scoring
import scoring.globalcost_scoring
import scoring.groupfair_scoring
from utils import downsample
if __name__ == "__main__":
folder_out = "exp-eval-results"
f_in = sys.argv[1]
scoring_desc = sys.argv[2]
cf_desc = sys.argv[3]
use_log_reg = sys.argv[4] == "True"
random_removal = True if sys.argv[5] == "True" else False
use_scikit_learn = False
if scoring_desc == "groupfaircf":
if cf_desc == "mem":
cf_method = scoring.groupfair_scoring.MemoryGroupFairCfScoring
elif cf_desc == "wachter":
cf_method = scoring.groupfair_scoring.WachterGroupFairCfScoring
use_scikit_learn = True
elif cf_desc == "proto":
cf_method = scoring.groupfair_scoring.ProtoGroupFairCfScoring
use_scikit_learn = True
else:
raise ValueError("Unknown CF method!")
elif scoring_desc == "globalrecourse":
if cf_desc == "mem":
cf_method = scoring.globalcost_scoring.MemoryCostResourceScoring
elif cf_desc == "wachter":
cf_method = scoring.globalcost_scoring.WachterCostResourceScoring
use_scikit_learn = True
elif cf_desc == "proto":
cf_method = scoring.globalcost_scoring.ProtoCostResourceScoring
use_scikit_learn = True
else:
raise ValueError("Unknown CF method!")
n_scoring_itr = 20
n_fit_itr = 2
random_removal_n_itr = 5
use_train_eval = False
print(folder_out, f_in, n_scoring_itr, n_fit_itr, cf_desc, scoring_desc, f"random_revmoal={random_removal}", f"logreg={use_log_reg}")
data = np.load(os.path.join(f_in), allow_pickle=True)
if len(data["X_train_results"].shape) > 1:
X_train_results = data["X_train_results"].astype('float32')
y_train_results = data["y_train_results"].astype('int32')
y_train_sensitive_results = data["y_train_sensitive_results"]
else:
X_train_results = [a.astype('float32') for a in data["X_train_results"].tolist()]
y_train_results = [a.astype('int32') for a in data["y_train_results"].tolist()]
y_train_sensitive_results = data["y_train_sensitive_results"].tolist()
t = data["X_test_results"]
if len(data["X_test_results"].shape) > 1:
X_test_results = data["X_test_results"].astype('float32')
y_test_results = data["y_test_results"].astype('int32')
y_test_sensitive_results = data["y_test_sensitive_results"]
else:
X_test_results = [a.astype('float32') for a in data["X_test_results"].tolist()]
y_test_results = [a.astype('int32') for a in data["y_test_results"].tolist()]
y_test_sensitive_results = data["y_test_sensitive_results"].tolist()
infl_scores_results = data["infl_scores_results"].tolist()
results = []
for X_train, y_train, y_train_sensitive, X_test, y_test, y_test_sensitive, infl_scores in \
zip(X_train_results, y_train_results, y_train_sensitive_results, X_test_results,
y_test_results, y_test_sensitive_results, infl_scores_results):
if not isinstance(infl_scores, np.ndarray):
infl_scores = np.array(infl_scores)
# Compute initial predictive accuracy!
pred_scores = []
for _ in range(n_fit_itr):
clf = get_model((X_train.shape[1],), use_scikit_learn=use_scikit_learn,
logreg=use_log_reg)
clf.fit(X_train, y_train)
y_test_pred = clf.predict(X_test)
pred_score = f1_score(y_test, y_test_pred)
pred_scores.append(pred_score)
orig_pred_score = np.mean(pred_scores)
print(f"Test data: {X_test.shape}")
print(f"F1-score: {np.mean(pred_scores), np.var(pred_scores)}")
# Compute initial score using all training sampes
X_eval, y_eval_sensitive = X_test, y_test_sensitive
if use_train_eval is True:
X_eval, y_eval_sensitive = X_train, y_train_sensitive
s = cf_method(get_model((X_test.shape[1],), use_scikit_learn=use_scikit_learn,
logreg=use_log_reg), X_eval, y_eval_sensitive)
s_orig = s.compute_score(X_train, y_train, n_itr=n_scoring_itr)
# Evaluation
idx = np.argsort(infl_scores)
infl_scores_sorted = infl_scores[idx]
max_n_samples_drop = np.sum(infl_scores_sorted > 0)
for k in [0.01, 0.02, 0.03, 0.04, .05, .1, .15, .2, .25, .3]:
n_sampels_drop = min([math.ceil(len(idx) * k), max_n_samples_drop]) # Convert to percentage but do not remove points with a negative impact on the value function
if random_removal is False:
# Remove most influential samples
idx_ = idx[: len(idx) - n_sampels_drop]
X_train_ = X_train[idx_, :]
y_train_ = y_train[idx_]
# Recompute predictive accuracy and fairness scores
pred_scores = []
for _ in range(n_fit_itr):
clf = get_model((X_train_.shape[1],), use_scikit_learn=use_scikit_learn,
logreg=use_log_reg)
clf.fit(X_train_, y_train_)
y_test_pred = clf.predict(X_test)
pred_score = f1_score(y_test, y_test_pred)
pred_scores.append(pred_score)
new_pred_score = np.mean(pred_scores)
s_new = s.compute_score(X_train_, y_train_, n_itr=n_scoring_itr)
print(f"{k} ({n_sampels_drop}): Value: {s_orig} -> {s_new} " +
f"F1-Score: {orig_pred_score} -> {new_pred_score}")
results.append((k, (s_orig, s_new), (orig_pred_score, new_pred_score)))
else:
# Remove random samples and recompute metrics -- repeat this multiple times!
pred_scores = []
scores_new = []
for _ in range(random_removal_n_itr):
idx_ = random.sample(list(idx), k=len(idx) - n_sampels_drop)
X_train_ = X_train[idx_, :]
y_train_ = y_train[idx_]
pred_scores = []
for _ in range(n_fit_itr):
clf = get_model((X_train_.shape[1],), use_scikit_learn=use_scikit_learn)
clf.fit(X_train_, y_train_)
y_test_pred = clf.predict(X_test)
pred_score = f1_score(y_test, y_test_pred)
pred_scores.append(pred_score)
s_new = s.compute_score(X_train_, y_train_, n_itr=n_scoring_itr)
scores_new.append(s_new)
# Aggregate results
new_pred_score = np.mean(pred_scores)
s_new = np.mean(scores_new)
print(s_new, np.var(scores_new))
print(f"{k} ({n_sampels_drop}): Value: {s_orig} -> {s_new} " +
f"F1-Score: {orig_pred_score} -> {new_pred_score}")
results.append((k, (s_orig, s_new), (orig_pred_score, new_pred_score)))
print(results)
f_ext = ""
if f_in.endswith("accuracy.npz"):
f_ext = "_accuracy"
np.savez(os.path.join(folder_out, f"{os.path.basename(f_in)}_{cf_desc}_{scoring_desc}_logreg={use_log_reg}_randomremoval={random_removal}{f_ext}"),
results=np.array(results, dtype=object),
infl_scores_results=np.array(infl_scores_results, dtype=object))