forked from phenixace/Comp90042-Final_Project
-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
196 lines (164 loc) · 8.92 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
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
import argparse
import torch
import transformers
from tqdm import tqdm
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
from utils import FixedScheduler, WarmupLinearScheduler, save
from metrics import calc_accuracy_score, calc_f1_score
from dataset import MyDataset, Collator
parser = argparse.ArgumentParser(description='Model parameters')
parser.add_argument('--random_seed', type=int, default=1022, help='Choose random_seed')
parser.add_argument('--model', default='bart',help='Sevral models are available: Bart | Bert | T5')
parser.add_argument('--model_path', default='none', help="use local model weights if specified")
parser.add_argument('--save_path', default='./models/roberta-base', help="where to save the model")
parser.add_argument('--total_steps', type=int, default=5000, help='Set training steps')
parser.add_argument('--eval_steps', type=int, default=500, help='Set evaluation steps')
parser.add_argument('--batch_size', type=int, default=4, help='Set batch size')
parser.add_argument('--lr', type=float, default=2e-5, help='Set learning rate')
parser.add_argument('--optim', default='Adam', help='Choose optimizer')
parser.add_argument('--warmup_steps', type=int, default=1000, help='WarmUp Steps')
parser.add_argument('--lrscheduler', action='store_true', help='Apply LRScheduler')
parser.add_argument('--mode', default='train',help='train, test, or inference')
parser.add_argument('--device', default='cuda:0',help='Device')
if __name__ == '__main__':
args = parser.parse_args()
torch.manual_seed(args.random_seed)
# load models
if args.model == 'bart':
if args.model_path == "none":
model = transformers.BartForSequenceClassification.from_pretrained('facebook/bart-base', num_labels = 2)
tokenizer = transformers.BartTokenizer.from_pretrained('facebook/bart-base')
else:
model = transformers.BartForSequenceClassification.from_pretrained(args.model_path, num_labels = 2)
tokenizer = transformers.BartTokenizer.from_pretrained(args.model_path)
model.to(args.device)
elif args.model == 'roberta':
if args.model_path == "none":
model = transformers.RobertaForSequenceClassification.from_pretrained('roberta-base',num_labels = 2)
tokenizer = transformers.RobertaTokenizer.from_pretrained('roberta-base', num_labels = 2)
else:
model = transformers.RobertaForSequenceClassification.from_pretrained(args.model_path, num_labels = 2)
tokenizer = transformers.RobertaTokenizer.from_pretrained(args.model_path, num_labels = 2)
model.to(args.device)
elif args.model == 'bert':
if args.model_path == "none":
model = transformers.BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels = 2)
tokenizer = transformers.BertTokenizer.from_pretrained('bert-base-uncased', num_labels = 2)
else:
model = transformers.BertForSequenceClassification.from_pretrained(args.model_path, num_labels = 2)
tokenizer = transformers.BertTokenizer.from_pretrained(args.model_path, num_labels = 2)
model.to(args.device)
else:
raise RuntimeError("[Error] Model not in the scope!")
# load data
train_set = MyDataset('train')
dev_set = MyDataset('dev')
test_set = MyDataset('test')
collator = Collator(tokenizer, max_length=512)
# optimizer
if args.optim == 'adam':
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
else:
optimizer = torch.optim.SGD(model.parameters(), lr=args.lr)
# scheduler
if args.lrscheduler:
scheduler = WarmupLinearScheduler(optimizer, warmup_steps=args.warmup_steps, scheduler_steps=args.total_steps, min_ratio=0., fixed_lr=args.fixed_lr)
else:
scheduler = FixedScheduler(optimizer)
# training
if args.mode == 'train':
train_sampler = RandomSampler(train_set)
train_loader = DataLoader(train_set, batch_size=args.batch_size, sampler=train_sampler, num_workers=0, collate_fn=collator)
step, epoch = 0, 0
loss, curr_loss = 0.0, 0.0
best_f1 = 0
model.train()
model.zero_grad()
while step < args.total_steps:
epoch += 1
with tqdm(train_loader) as t:
t.set_description("Epochs "+ str(epoch))
for i, batch in enumerate(t):
step += 1
(texts, labels) = batch
# print(texts, labels)
train_loss = model(**texts.to(args.device), labels=labels.to(args.device)).loss
t.set_postfix(loss=train_loss.item())
train_loss.backward()
optimizer.step()
scheduler.step()
model.zero_grad()
curr_loss += train_loss.item()
if step % args.eval_steps == 0:
model.eval()
# evaluation
dev_sampler = SequentialSampler(dev_set)
dev_loader = DataLoader(dev_set, batch_size=args.batch_size, sampler=dev_sampler, num_workers=0, collate_fn=collator)
all_pred_labels = []
all_ture_labels = []
with tqdm(dev_loader) as t:
t.set_description("Dev")
for i, batch in enumerate(t):
(texts, labels) = batch
all_ture_labels += labels
output = model(**texts.to(args.device), labels=labels.to(args.device)).logits
all_pred_labels += [item.argmax().item() for item in output]
# metrics here
acc = calc_accuracy_score(all_pred_labels, all_ture_labels)
f1, precision, recall = calc_f1_score(all_pred_labels, all_ture_labels)
print("Step {}: Acc = {:2f}; F1 = {:2f}, P = {:2f}; R = {:2f}\n".format(step, acc, f1, precision, recall))
# save best
if f1 >= best_f1:
save(model, tokenizer, args.save_path, args.model+'-step-'+str(step))
best_f1 = f1
model.train()
# testing
elif args.mode == 'test':
model.eval()
test_sampler = SequentialSampler(test_set)
test_loader = DataLoader(test_set, batch_size=args.batch_size, sampler=test_sampler, num_workers=0, collate_fn=collator)
labels = []
with tqdm(test_loader) as t:
t.set_description('Test')
for i, batch in enumerate(t):
text = batch
output = model(**text.to(args.device)).logits
labels += [item.argmax().item() for item in output]
with open('./project-data/test_pred.label.csv', 'w+') as f:
f.write('Id,Predicted\n')
for i in range(len(labels)):
f.write(str(i) + ',' + str(labels[i]) + '\n')
# inference
elif args.mode == 'inference':
model.eval()
dev_sampler = SequentialSampler(dev_set)
dev_loader = DataLoader(dev_set, batch_size=args.batch_size, sampler=dev_sampler, num_workers=0, collate_fn=collator)
index = 0
with tqdm(dev_loader) as t:
t.set_description("Dev")
for i, batch in enumerate(t):
(texts, labels) = batch
output = model(**texts.to(args.device), labels=labels.to(args.device)).logits
pred_labels = [item.argmax().item() for item in output]
with open('results/inference.log.txt', 'a+', encoding='utf-8') as f:
for i in range(0, len(pred_labels)):
if pred_labels[i] != labels[i]:
f.write('Text:' + dev_set[index + i]['text'] + '\n')
f.write('Raw Text:' + dev_set[index + i]['raw_text'] + '\n')
f.write('True:'+str(labels[i].data)+'\tPred:'+str(pred_labels[i])+'\n')
index += len(pred_labels)
elif args.mode == 'process':
model.eval()
index = 0
import pandas as pd
covid_data = pd.read_csv('./project-data/covid19_tweets.csv')
with open('./results/covid19_tweets_processed.csv', 'w+', encoding='utf-8') as f:
f.write('label\n')
for i in range(len(covid_data)):
text = covid_data['text'][i]
text = tokenizer(text, return_tensors='pt', padding='max_length', max_length=512, truncation=True)
output = model(**text.to(args.device)).logits
pred_labels = [item.argmax().item() for item in output]
f.write(str(pred_labels[0]) + '\n')
else:
raise RuntimeError("[Error] Mode not in the scope!")