-
Notifications
You must be signed in to change notification settings - Fork 2
/
origin_bert.py
166 lines (134 loc) · 5.53 KB
/
origin_bert.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
from seqeval.metrics import classification_report, f1_score, precision_score, accuracy_score, recall_score
from transformers import Trainer, BertForTokenClassification, TrainingArguments
from CC.loaders.utils.vocab import Vocab
import json
from typing import Any
import torch
from torch.utils.data import Dataset
from transformers.trainer_utils import EvalPrediction, set_seed
from transformers import BertTokenizer
from torch import nn
import os
from tqdm import tqdm
from torch.utils.tensorboard import SummaryWriter
import argparse
parser = argparse.ArgumentParser(description="bert arguments")
parser.add_argument("dataset", type=str)
parser.add_argument("scale", type=int)
global_args = parser.parse_args()
class SimpleDataset(Dataset):
def __init__(self, path, tokenizer: BertTokenizer, padding_length=150, tokens2ids: Any = None):
self.data = []
with open(path, "r", encoding="utf-8") as f:
for line in tqdm(f, desc=f"load {path}"):
data = json.loads(line.strip())
text, labels = data["text"][:padding_length -
2], data["label"][:padding_length-2]
ids = tokenizer(''.join(
text), max_length=padding_length, padding="max_length", truncation=True, return_tensors="pt")
for key in ids:
ids[key] = ids[key][0]
labels_ids = tokens2ids(labels)
labels = torch.tensor(
[-100]+labels_ids+[-100]*(padding_length-1-len(labels_ids)), dtype=torch.long)
ids["labels"] = labels
self.data.append(ids)
def __getitem__(self, index):
return self.data[index]
def __len__(self):
return len(self.data)
report_path = f"data_record/{global_args.dataset}_bert_150padding_{global_args.scale}/"
args = {
"train_file": f"data/few_shot/{global_args.dataset}/train_{global_args.scale}_split.json",
"eval_file": f"data/few_shot/{global_args.dataset}/dev_split.json",
"tag_file": f"data/few_shot/{global_args.dataset}/labels.txt",
}
tokenizer = BertTokenizer.from_pretrained("model/chinese_wwm_ext")
label_vocab = Vocab().from_files([args["tag_file"]])
train_set = SimpleDataset(
args["train_file"], tokenizer, tokens2ids=label_vocab.token2id)
eval_set = SimpleDataset(
args["eval_file"], tokenizer, tokens2ids=label_vocab.token2id)
train_arguments = TrainingArguments(
output_dir=f"./runs/{global_args.dataset}/{global_args.scale}",
do_eval=True,
do_train=True,
evaluation_strategy="epoch",
learning_rate=2e-5,
per_device_train_batch_size=2,
per_device_eval_batch_size=8,
num_train_epochs=30,
warmup_steps=190,
logging_strategy="epoch",
save_strategy="epoch",
no_cuda=False,
)
def generate_csv(path, data):
os.makedirs(path, exist_ok=True)
with open(os.path.join(path, "eval.csv"), "w", encoding="utf-8") as f:
f.write(f"f1,precision,recall,accuracy\n")
for line in data:
f.write(
f'{line["f1"]},{line["precision"]},{line["recall"]},{line["accuracy"]}\n')
f.flush()
def get_item(obj):
if isinstance(obj, dict):
for k in obj:
obj[k] = get_item(obj[k])
else:
if getattr(obj, "item"):
return obj.item()
return obj
def generate_report(path, epoch, data):
data = get_item(data)
path = os.path.join(path, "reports")
os.makedirs(path, exist_ok=True)
with open(os.path.join(path, f"{epoch}_epoch.json"), "w", encoding="utf-8") as f:
json.dump(data, f, ensure_ascii=False, indent=2)
data_collections = []
def compute_metrics(predictions: EvalPrediction):
global label_vocab, data_collections, report_path
epoch = len(data_collections)+1
labels = predictions.label_ids
preds = predictions.predictions.argmax(-1)
preds_list = []
trues_list = []
for label, pred in zip(labels, preds):
label = torch.tensor(label)
pred = torch.tensor(pred)
mask = label.gt(-1)
label = label[mask].tolist()
pred = pred[mask].tolist()
assert len(label) == len(pred)
preds_list.append(list(i.replace("M-", "I-")
for i in label_vocab.id2token(pred)))
trues_list.append(list(i.replace("M-", "I-")
for i in label_vocab.id2token(label)))
result = {
"accuracy": accuracy_score(trues_list, preds_list),
"recall": recall_score(trues_list, preds_list),
"f1": f1_score(trues_list, preds_list),
"precision": precision_score(trues_list, preds_list),
# "reports": classification_report(trues_list, preds_list, output_dict=True)
}
data_collections.append(result)
generate_report(report_path, epoch, classification_report(
trues_list, preds_list, output_dict=True))
return result
def model_init():
set_seed(2021)
model = BertForTokenClassification.from_pretrained("model/chinese_wwm_ext", num_labels=len(label_vocab),
id2label=dict(
zip(range(len(label_vocab)), label_vocab.idx2item)),
label2id=label_vocab.item2idx)
model = nn.DataParallel(model, device_ids=[0, 1, 2, 3])
return model
trainer = Trainer(
model_init=model_init,
args=train_arguments,
train_dataset=train_set,
eval_dataset=eval_set,
compute_metrics=compute_metrics
)
trainer.train()
generate_csv(report_path, data_collections)