-
Notifications
You must be signed in to change notification settings - Fork 0
/
cv.py
202 lines (169 loc) · 6.73 KB
/
cv.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
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
# the below functions are taken from https://github.com/VGligorijevic/deepNF
import numpy as np
from sklearn import svm
from sklearn.multiclass import OneVsRestClassifier
from sklearn.metrics import accuracy_score, f1_score
from sklearn.model_selection import ShuffleSplit, KFold
from sklearn.metrics.pairwise import rbf_kernel, linear_kernel
from sklearn.utils import resample
#import pickle5 as pickle
def kernel_func(X, Y=None, param=0):
if param != 0:
K = rbf_kernel(X, Y, gamma=param)
else:
K = linear_kernel(X, Y)
return K
def real_AUPR(label, score):
"""Computing real AUPR . By Vlad and Meet"""
label = label.flatten()
score = score.flatten()
order = np.argsort(score)[::-1]
label = label[order]
P = np.count_nonzero(label)
# N = len(label) - P
TP = np.cumsum(label, dtype=float)
PP = np.arange(1, len(label)+1, dtype=float) # python
x = np.divide(TP, P) # recall
y = np.divide(TP, PP) # precision
pr = np.trapz(y, x)
f = np.divide(2*x*y, (x + y))
idx = np.where((x + y) != 0)[0]
if len(idx) != 0:
f = np.max(f[idx])
else:
f = 0.0
return pr, f
def ml_split(y):
"""Split annotations"""
kf = KFold(n_splits=5, shuffle=True)
splits = []
for t_idx, v_idx in kf.split(y):
splits.append((t_idx, v_idx))
return splits
def evaluate_performance(y_test, y_score, y_pred):
"""Evaluate performance"""
n_classes = y_test.shape[1]
perf = dict()
# Compute macro-averaged AUPR
perf["M-aupr"] = 0.0
n = 0
for i in range(n_classes):
perf[i], _ = real_AUPR(y_test[:, i], y_score[:, i])
if sum(y_test[:, i]) > 0:
n += 1
perf["M-aupr"] += perf[i]
perf["M-aupr"] /= n
# Compute micro-averaged AUPR
perf["m-aupr"], _ = real_AUPR(y_test, y_score)
# Computes accuracy
perf['acc'] = accuracy_score(y_test, y_pred)
# Computes F1-score
alpha = 3
y_new_pred = np.zeros_like(y_pred)
for i in range(y_pred.shape[0]):
top_alpha = np.argsort(y_score[i, :])[-alpha:]
y_new_pred[i, top_alpha] = np.array(alpha*[1])
perf["F1"] = f1_score(y_test, y_new_pred, average='micro')
return perf
def cross_validation(X, y, n_trials=5, ker='rbf'):
"""Perform model selection via 5-fold cross validation"""
# filter samples with no annotations
del_rid = np.where(y.sum(axis=1) == 0)[0]
y = np.delete(y, del_rid, axis=0)
X = np.delete(X, del_rid, axis=0)
# range of hyperparameters
C_range = 10.**np.arange(-1, 3)
if ker == 'rbf':
gamma_range = 10.**np.arange(-3, 1)
elif ker == 'lin':
gamma_range = [0]
else:
print ("### Wrong kernel.")
# pre-generating kernels
print ("### Pregenerating kernels...")
K_rbf = {}
for gamma in gamma_range:
K_rbf[gamma] = kernel_func(X, param=gamma)
print ("### Done.")
# performance measures
pr_micro = []
pr_macro = []
fmax = []
acc = []
# shuffle and split training and test sets
trials = ShuffleSplit(n_splits=n_trials, test_size=0.2, random_state=None)
ss = trials.split(X)
trial_splits = []
for train_idx, test_idx in ss:
trial_splits.append((train_idx, test_idx))
it = 0
for jj in range(0, n_trials):
train_idx = trial_splits[jj][0]
test_idx = trial_splits[jj][1]
it += 1
y_train = y[train_idx]
y_test = y[test_idx]
print ("### [Trial %d] Perfom cross validation...." % (it))
print ("Train samples=%d; #Test samples=%d" % (y_train.shape[0], y_test.shape[0]))
# setup for neasted cross-validation
splits = ml_split(y_train)
# parameter fitting
C_opt = None
gamma_opt = None
max_aupr = 0
for C in C_range:
for gamma in gamma_range:
# Multi-label classification
cv_results = []
for train, valid in splits:
clf = OneVsRestClassifier(svm.SVC(C=C, kernel='precomputed',
random_state=123,
probability=True), n_jobs=-1)
K_train = K_rbf[gamma][train_idx[train], :][:, train_idx[train]]
K_valid = K_rbf[gamma][train_idx[valid], :][:, train_idx[train]]
y_train_t = y_train[train]
y_train_v = y_train[valid]
y_score_valid = np.zeros(y_train_v.shape, dtype=float)
y_pred_valid = np.zeros_like(y_train_v)
idx = np.where(y_train_t.sum(axis=0) > 0)[0]
clf.fit(K_train, y_train_t[:, idx])
y_score_valid[:, idx] = clf.predict_proba(K_valid)
y_pred_valid[:, idx] = clf.predict(K_valid)
perf_cv = evaluate_performance(y_train_v,
y_score_valid,
y_pred_valid)
cv_results.append(perf_cv['m-aupr'])
cv_aupr = np.median(cv_results)
print ("### gamma = %0.3f, C = %0.3f, AUPR = %0.3f" % (gamma, C, cv_aupr))
if cv_aupr > max_aupr:
C_opt = C
gamma_opt = gamma
max_aupr = cv_aupr
print ("### Optimal parameters: ")
print ("C_opt = %0.3f, gamma_opt = %0.3f" % (C_opt, gamma_opt))
print ("### Train dataset: AUPR = %0.3f" % (max_aupr))
print
print ("### Using full training data...")
clf = OneVsRestClassifier(svm.SVC(C=C_opt, kernel='precomputed',
random_state=123,
probability=True), n_jobs=-1)
y_score = np.zeros(y_test.shape, dtype=float)
y_pred = np.zeros_like(y_test)
idx = np.where(y_train.sum(axis=0) > 0)[0]
clf.fit(K_rbf[gamma_opt][train_idx, :][:, train_idx], y_train[:, idx])
y_score[:, idx] = clf.predict_proba(K_rbf[gamma_opt][test_idx, :][:, train_idx])
y_pred[:, idx] = clf.predict(K_rbf[gamma_opt][test_idx, :][:, train_idx])
perf_trial = evaluate_performance(y_test, y_score, y_pred)
pr_micro.append(perf_trial['m-aupr'])
pr_macro.append(perf_trial['M-aupr'])
fmax.append(perf_trial['F1'])
acc.append(perf_trial['acc'])
print ("### Test dataset: AUPR['micro'] = %0.3f, AUPR['macro'] = %0.3f, F1 = %0.3f, Acc = %0.3f" % (perf_trial['m-aupr'], perf_trial['M-aupr'], perf_trial['F1'], perf_trial['acc']))
print
print
perf = dict()
perf['pr_micro'] = pr_micro
perf['pr_macro'] = pr_macro
perf['fmax'] = fmax
perf['acc'] = acc
return perf