-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_bert_multi.delta.py
163 lines (137 loc) · 5.4 KB
/
train_bert_multi.delta.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
import sys
from simpletransformers.classification import (
ClassificationModel)
import pandas as pd
import logging
import math
from sklearn.model_selection import KFold
from sklearn.metrics import confusion_matrix, accuracy_score, f1_score, roc_auc_score
import os
import statistics
import string
import json
import math
from pathlib import Path
# Input
train = sys.argv[1]
valid = sys.argv[2]
test_path = Path(sys.argv[3])
output_dir = sys.argv[4]
num_labels = int(sys.argv[5])
model_args = dict()
model_args["reprocess_input_data"] = True
model_args["num_train_epochs"] = 10
model_args["evaluate_during_training"] = True
model_args["overwrite_output_dir"] = True
model_args["fp16"] = False
model_args["use_early_stopping"] = True
model_args["early_stopping_delta"] = 0.01
model_args["early_stopping_metric"] = 'mcc'
model_args["early_stopping_metric_minimize"] = False
model_args["early_stopping_patience"] = 3
model_args["evaluate_during_training_steps"] = 1000
model_args["save_eval_checkpoints"] = False
model_args["use_cuda"] = True
if len(sys.argv) > 6:
model_dir = sys.argv[6]
else:
model_dir = False
def f1_score_macro(y_true, y_pred):
return f1_score(y_true, y_pred, average='macro')
def f1_score_micro(y_true, y_pred):
return f1_score(y_true, y_pred, average='micro')
# Set loggers
logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger('transformers')
transformers_logger.setLevel(logging.WARNING)
# Get Data
train_data = pd.read_csv(train, sep='\t', header=None)
train_df = pd.DataFrame(train_data)
train_df.dropna(subset=[0], inplace=True)
train_df = train_df.sample(frac=1).reset_index(drop=True)
valid_data = pd.read_csv(valid, sep='\t', header=None)
valid_df = pd.DataFrame(valid_data)
valid_df.dropna(subset=[0], inplace=True)
valid_df = valid_df.sample(frac=1).reset_index(drop=True)
test_data_en = pd.read_csv(test_path / "test_en.tsv", sep='\t', header=None)
test_df_en = pd.DataFrame(test_data_en)
test_df_en.dropna(subset=[0], inplace=True)
test_df_en = test_df_en.sample(frac=1).reset_index(drop=True)
test_data_ar = pd.read_csv(test_path / "test_ar.tsv", sep='\t', header=None)
test_df_ar = pd.DataFrame(test_data_ar)
test_df_ar.dropna(subset=[0], inplace=True)
test_df_ar = test_df_ar.sample(frac=1).reset_index(drop=True)
accs_en = []
f1s_en = []
accs_ar = []
f1s_ar = []
for i in range(0,10):
# Prepare cross validation
# Setup
cur_output = '{}/{}/'.format(output_dir, i)
if not os.path.exists(cur_output):
os.mkdir(cur_output)
model_args["output_dir"] = '{}/output'.format(cur_output)
model_args["cache_dir"] = '{}/cache/'.format(cur_output)
model_args["tensorboard_dir"] = '{}/runs/'.format(cur_output)
model_args["best_model_dir"] = '{}/output'.format(cur_output)
if model_dir:
model_pretrained = ClassificationModel(
"bert",
model_dir,
use_cuda=True,
)
model_downstream = ClassificationModel(
"bert",
"bert-base-multilingual-cased",
args=model_args,
num_labels=num_labels
)
model_downstream.model.bert = model_pretrained.model.bert
model = model_downstream
else:
model = ClassificationModel(
"bert",
"bert-base-multilingual-cased",
args=model_args,
num_labels=num_labels
)
# Train model
model.train_model(train_df, eval_df=valid_df)
# Validate model
result, model_outputs, wrong_predictions = model.eval_model(test_df_en,
cm=confusion_matrix,
acc=accuracy_score,
f1_macro=f1_score_macro,
f1_micro=f1_score_micro)
print(result, flush=True)
accs_en.append(result['acc'])
f1s_en.append(result['f1_macro'])
result, model_outputs, wrong_predictions = model.eval_model(test_df_ar,
cm=confusion_matrix,
acc=accuracy_score,
f1_macro=f1_score_macro,
f1_micro=f1_score_micro)
print(result, flush=True)
accs_ar.append(result['acc'])
f1s_ar.append(result['f1_macro'])
def avg(lst):
return sum(lst) / len(lst)
acc_stdev_en = statistics.stdev(accs_en)
acc_avg_en = avg(accs_en)
acc_se_en = acc_stdev_en/math.sqrt(10)
f1_stdev_en = statistics.stdev(f1s_en)
f1_avg_en = avg(f1s_en)
f1_se_en = f1_stdev_en/math.sqrt(10)
print("Results for: EN")
print('Acc... Mean: {}\tStandard Deviaton: {}\tStandard Error: {}'.format(acc_avg_en, acc_stdev_en, acc_se_en))
print('F1 Macro... Mean: {}\tStandard Deviaton: {}\tStandard Error: {}'.format(f1_avg_en, f1_stdev_en, f1_se_en))
acc_stdev_ar = statistics.stdev(accs_ar)
acc_avg_ar = avg(accs_ar)
acc_se_ar = acc_stdev_ar/math.sqrt(10)
f1_stdev_ar = statistics.stdev(f1s_ar)
f1_avg_ar = avg(f1s_ar)
f1_se_ar = f1_stdev_ar/math.sqrt(10)
print("Results for: AR")
print('Acc... Mean: {}\tStandard Deviaton: {}\tStandard Error: {}'.format(acc_avg_ar, acc_stdev_ar, acc_se_ar))
print('F1 Macro... Mean: {}\tStandard Deviaton: {}\tStandard Error: {}'.format(f1_avg_ar, f1_stdev_ar, f1_se_ar))