-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
145 lines (112 loc) · 6.77 KB
/
main.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
import argparse
import pandas as pd
from datasets import Dataset
from use_setfit import do_setfit
from setup_utils import (str2bool,
sample_df_convert_ds,
fix_liar,
write_eval_jsons,
ds_for_orig_liar,
ds_for_general,
fix_amzn)
from modelling import (train_once,
train_head_every_step,
train_every_step,
setfit_once,
setfit_every_step,
linear_probe)
from lagonn import LaGoNN
parser = argparse.ArgumentParser(description='configs LaGoNN')
parser.add_argument('--WRITE', action="store", type=str2bool, dest='write', default='True')
parser.add_argument('--TRANSFORMER_CLF', action="store", dest='transformer_clf', default='roberta-base')
parser.add_argument('--ST_MODEL', action="store", dest='st_model', default="sentence-transformers/paraphrase-mpnet-base-v2")
parser.add_argument('--SEED', action="store", type=int, dest='seed', default=0)
parser.add_argument('--TASK', action="store", dest='task', default='insincere-questions')
parser.add_argument('--MODE', action="store", dest='mode', default='LAGONN_CHEAP')
parser.add_argument('--LAGONN_CONFIG', action="store", dest='lagonnconfig', default='LABEL')
parser.add_argument('--NUM_NEIGHBORS', action="store", type=int, dest='num_neighbors', default=1)
parser.add_argument('--DISTANCE_PRECISION', action="store", dest='dist_precision', default='None')
content_mod = ['amazon_counterfactual_en', 'hate_speech_offensive', 'insincere-questions', 'toxic_conversations']
general = ['student-question-categories', 'emotion', '20_newsgroups', 'imdb', 'sst5']
amz_multi_ling = ['amazon_reviews_multi_ja','amazon_reviews_multi_zh', 'amazon_reviews_multi_de',
'amazon_reviews_multi_fr', 'amazon_reviews_multi_es', 'amazon_reviews_multi_en']
if __name__ == '__main__':
args = parser.parse_args()
if args.task in content_mod + ['liar']:
balances = ['extreme', 'imbalanced', 'moderate', 'balanced']
elif args.task in general + ['orig_liar']:
balances = ['balanced']
elif args.task in amz_multi_ling:
balances = ['balanced']
assert args.mode in ['LAGONN_CHEAP', 'LAGONN', 'LAGONN_LITE', 'LAGONN_EXP', 'PROBE', 'KNN', 'LOG_REG',
'SETFIT', 'SETFIT_LITE', 'ROBERTA_FREEZE', 'ROBERTA_FULL']
if args.task in ['liar']:
input_test_ds = Dataset.from_pandas(fix_liar(args, 'test'))
input_val_ds = Dataset.from_pandas(fix_liar(args, 'val'))
train_df = fix_liar(args, 'train')
elif args.task in ['orig_liar']:
train_df, input_test_ds, input_val_ds = ds_for_orig_liar(args)
elif args.task in content_mod:
input_test_ds = Dataset.from_pandas(pd.read_csv('dataframes_with_val/{}_test.csv'.format(args.task)).dropna()[['text', 'labels', 'label_text']])
input_val_ds = Dataset.from_pandas(pd.read_csv('dataframes_with_val/{}_val.csv'.format(args.task)).dropna()[['text', 'labels', 'label_text']])
train_df = pd.read_csv('dataframes_with_val/{}_train.csv'.format(args.task)).dropna()[['text', 'labels', 'label_text']]
elif args.task in general:
train_df, input_test_ds, input_val_ds = ds_for_general(args.task)
elif args.task in amz_multi_ling:
train_df, input_test_ds, input_val_ds = fix_amzn(args)
for balance in balances:
for step in range(1, 11):
input_train_ds = sample_df_convert_ds(train_df, balance, step, args)
if args.mode in ['PROBE']:
eval_dict = linear_probe(input_train_ds, input_test_ds, args)
elif args.mode in ['KNN', 'LOG_REG']:
if step == 1:
sbert_trainer = do_setfit(args, input_train_ds, input_val_ds)
eval_dict = setfit_once(input_train_ds, input_test_ds, sbert_trainer, step, args)
elif args.mode in ['SETFIT']:
eval_dict = setfit_every_step(input_train_ds, input_val_ds, input_test_ds, args)
elif args.mode in ['SETFIT_LITE']:
if step < 5:
eval_dict, sbert_trainer = setfit_every_step(input_train_ds, input_val_ds, input_test_ds, args)
else:
eval_dict = setfit_once(input_train_ds, input_test_ds, sbert_trainer, step, args)
elif args.mode in ['ROBERTA_FREEZE']:
if step == 1:
transformer_trainer, eval_dict = train_once(input_train_ds, input_val_ds, input_test_ds, args, balance)
elif step > 1:
eval_dict = train_head_every_step(input_train_ds, input_val_ds, input_test_ds, args, transformer_trainer, balance)
elif args.mode in ['ROBERTA_FULL']:
eval_dict = train_every_step(input_train_ds, input_val_ds, input_test_ds, args, balance)
elif args.mode in ['LAGONN_CHEAP', 'LAGONN_EXP']:
lgn = LaGoNN(input_train_ds, input_val_ds, input_test_ds, args, step)
eval_dict = lgn.predict()
elif args.mode in ['LAGONN']:
if step == 1:
lgn = LaGoNN(input_train_ds, input_val_ds, input_test_ds, args, step)
eval_dict, sbert_trainer = lgn.predict()
else:
lgn = LaGoNN(input_train_ds, input_val_ds, input_test_ds, args, step, sbert_trainer)
eval_dict = lgn.predict()
elif args.mode in ['LAGONN_LITE']:
if step < 5:
lgn = LaGoNN(input_train_ds, input_val_ds, input_test_ds, args, step)
eval_dict, sbert_trainer = lgn.predict()
else:
lgn = LaGoNN(input_train_ds, input_val_ds, input_test_ds, args, step, sbert_trainer)
eval_dict = lgn.predict()
train_avg_pre = eval_dict['train_avg_pre']
train_f1 = eval_dict['train_f1']
test_avg_pre = eval_dict['test_avg_pre']
test_f1 = eval_dict['test_f1']
print('TRAINING RESULTS')
print('AP: {}'.format(train_avg_pre))
print('F1: {}'.format(train_f1))
print()
print('TESTING RESULTS')
print('AP: {}'.format(test_avg_pre))
print('F1: {}'.format(test_f1))
print()
if args.write:
write_eval_jsons(eval_dict, args, step, balance)
print('Seed {} done for step {} of {} done! \n'.format(args.seed, step, balance+' '+args.task))
print("Job's done!")