-
Notifications
You must be signed in to change notification settings - Fork 10
/
score.py
194 lines (156 loc) · 7.92 KB
/
score.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
import pandas as pd
import numpy as np
from sklearn.metrics import log_loss
from models.data_pipeline import data_pipeline_v2, Dataset, CorefAnnotator
def load_data_version(data_dir=None,
exp_dir=None,
tst_data_version=None,
sanitize_labels=None,
persist=True,
coref_extractor=None,
proref_extractor=None
):
train = {
'input': Dataset().transform('{}/gap-test.tsv'.format(data_dir),
label_corrections='{}/gap_corrections/my_corrections_tst.csv'.format(data_dir))
}
val = {
'input': Dataset().transform('{}/gap-validation.tsv'.format(data_dir),
label_corrections='{}/gap_corrections/my_corrections_val.csv'.format(data_dir))
}
test = {
'input': Dataset().transform('{}/gap-development.tsv'.format(data_dir),
label_corrections='{}/gap_corrections/my_corrections_dev.csv'.format(data_dir))
}
neither = {
'input': Dataset().transform('{}/gpr-neither.tsv'.format(data_dir),
)
}
test_stage2 = {
'input': Dataset().transform('{}/{}.tsv'.format(data_dir, tst_data_version),
label_corrections='{}/gap_corrections/my_corrections_tst_stage2.csv'.format(data_dir),
shift_by_one=False)
}
dpl_trn = data_pipeline_v2(exp_dir,
mode='train',
annotate_mentions=True,
annotate_coref_mentions=True,
pretrained_proref=True,
sanitize_labels=sanitize_labels,
persist=persist,
coref_extractor=coref_extractor,
proref_extractor=proref_extractor
)
dpl_val = data_pipeline_v2(exp_dir,
mode='val',
annotate_mentions=True,
annotate_coref_mentions=True,
pretrained_proref=True,
sanitize_labels=sanitize_labels,
persist=persist,
coref_extractor=coref_extractor,
proref_extractor=proref_extractor
)
dpl_tst = data_pipeline_v2(exp_dir,
mode='test',
annotate_mentions=True,
annotate_coref_mentions=True,
pretrained_proref=True,
sanitize_labels=sanitize_labels,
persist=persist,
coref_extractor=coref_extractor,
proref_extractor=proref_extractor
)
dpl_neither = data_pipeline_v2(exp_dir,
mode='neither',
annotate_mentions=True,
annotate_coref_mentions=True,
pretrained_proref=True,
sanitize_labels=sanitize_labels,
persist=persist,
coref_extractor=coref_extractor,
proref_extractor=proref_extractor
)
dpl_tst_2 = data_pipeline_v2(exp_dir,
mode='test_stage2',
annotate_mentions=True,
annotate_coref_mentions=True,
pretrained_proref=True,
sanitize_labels=sanitize_labels,
persist=persist,
coref_extractor=coref_extractor,
proref_extractor=proref_extractor
)
X_trn = dpl_trn.gather_step.transform(train)['X']
X_val = dpl_val.gather_step.transform(val)['X']
X_tst = dpl_tst.gather_step.transform(test)['X']
X_neither = dpl_neither.gather_step.transform(neither)['X']
X_tst_2 = dpl_tst_2.gather_step.transform(test_stage2)['X']
train = pd.concat([X_trn, X_val, X_tst, X_neither, X_neither.head(3)]).reset_index(drop=True)
print(train.shape)
return X_trn, X_val, X_tst, train, X_tst_2
def get_score(probs, data):
y_true = data['label']
return round(log_loss(y_true, probs[:len(y_true), :])*100, 3)
def get_val_scores(predictions,
lms,
seeds,
train_san,
train_unsan,
tst_san,
tst_unsan,
X_trn,
X_val):
# CV ensemble score
probs_all = pd.concat([pd.read_csv(file) for file in predictions], axis=1).values.reshape(-1, len(lms), len(seeds), 3).transpose(1, 2, 0, 3)
index = []
oof_all_rows = []
oof_tst_rows = []
for j, lm in enumerate(lms):
for i, seed in enumerate(seeds):
index.append('{} {}'.format(lm, seed))
probs = probs_all[j][i]
oof_all_rows.append((get_score(probs, train_san), get_score(probs, train_unsan)))
probs = probs_all[j][i, len(X_trn)+len(X_val):len(X_trn)+len(X_val)+len(tst_san), :]
oof_tst_rows.append((get_score(probs, tst_san), get_score(probs, tst_unsan)))
index.append('{} {}'.format(lm, 'mean cv'))
oof_all_rows.append(np.mean(oof_all_rows[-5:], axis=0))
oof_tst_rows.append(np.mean(oof_tst_rows[-5:], axis=0))
index.append('{} {}'.format(lm, 'seed-based ensemble'))
probs = probs_all[j].mean(axis=0)
oof_all_rows.append((get_score(probs, train_san), get_score(probs, train_unsan)))
probs = probs_all[j].mean(axis=0)[len(X_trn)+len(X_val):len(X_trn)+len(X_val)+len(tst_san), :]
oof_tst_rows.append((get_score(probs, tst_san), get_score(probs, tst_unsan)))
index.append('lm-based ensemble')
cv_probs = probs = probs_all.mean(axis=0).mean(axis=0)
oof_all_rows.append((get_score(probs, train_san), get_score(probs, train_unsan)))
probs = probs_all.mean(axis=0).mean(axis=0)[len(X_trn)+len(X_val):len(X_trn)+len(X_val)+len(tst_san), :]
oof_tst_rows.append((get_score(probs, tst_san), get_score(probs, tst_unsan)))
cols = pd.MultiIndex.from_product([['oof_all', 'oof_tst'], ['sanitized', 'unsanitized']])
return pd.DataFrame(np.hstack((oof_all_rows, oof_tst_rows)), index=index, columns=cols), cv_probs, probs
def get_tst_scores(predictions,
lms,
seeds,
n_folds,
tst_san,
tst_unsan):
probs_raw = pd.concat([pd.read_csv(file) for file in predictions], axis=1).values.reshape(-1, len(lms), len(seeds), n_folds, 3).transpose(1, 2, 3, 0, 4)
index = []
tst_rows = []
for i, lm in enumerate(lms):
for j, seed in enumerate(seeds):
for k in range(n_folds):
index.append('{} {} {}'.format(lm, seed, k))
probs = probs_raw[i][j][k]
tst_rows.append((get_score(probs, tst_san), get_score(probs, tst_unsan)))
index.append('{} {} {}'.format(lm, seed, 'fold-based ensemble'))
probs = probs_raw[i][j].mean(axis=0)
tst_rows.append((get_score(probs, tst_san), get_score(probs, tst_unsan)))
index.append('{} {}'.format(lm, 'seed-based ensemble'))
probs = probs_raw[i].mean(axis=0).mean(axis=0)
tst_rows.append((get_score(probs, tst_san), get_score(probs, tst_unsan)))
index.append('lm-based ensemble')
probs = probs_raw.mean(axis=0).mean(axis=0).mean(axis=0)
tst_rows.append((get_score(probs, tst_san), get_score(probs, tst_unsan)))
cols = pd.MultiIndex.from_product([['tst'], ['sanitized', 'unsanitized']])
return pd.DataFrame(tst_rows, index=index, columns=cols)