-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy path__main__.py
167 lines (140 loc) · 6.8 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
import os
import random
import time
import numpy as np
import torch
from common.evaluators.bert_evaluator import BertEvaluator
from common.trainers.bert_trainer import BertTrainer
from datasets.bert_processors.aapd_processor import AAPDProcessor
from datasets.bert_processors.agnews_processor import AGNewsProcessor
from datasets.bert_processors.imdb_processor import IMDBProcessor
from datasets.bert_processors.reuters_processor import ReutersProcessor
from datasets.bert_processors.sogou_processor import SogouProcessor
from datasets.bert_processors.sst_processor import SST2Processor
from datasets.bert_processors.yelp2014_processor import Yelp2014Processor
from datasets.bert_processors.mbti_processor import MBTIProcessor
from models.bert.args import get_args
from models.bert.model import BertForSequenceClassification
from utils.io import PYTORCH_PRETRAINED_BERT_CACHE
from utils.optimization import BertAdam
from utils.tokenization import BertTokenizer
# String templates for logging results
LOG_HEADER = 'Split Dev/Acc. Dev/Hamm. Dev/Jacc. Dev/Prec Dev/Rec Dev/micro-F1 Dev/F1 Dev/Loss'
LOG_TEMPLATE = ' '.join('{:>5s},{:>6.4f},{:>8.4f},{:8.4f},{:8.4f},{:>8.4f},{:8.4f},{:8.4f},{:10.4f}'.split(','))
def evaluate_split(model, processor, args, split='dev'):
evaluator = BertEvaluator(model, processor, args, split)
start_time = time.time()
accuracy, hamming, jaccard, precision, recall, microf1, f1, avg_loss = evaluator.get_scores(silent=True)[0]
print("Inference time", time.time() - start_time)
print('\n' + LOG_HEADER)
print(LOG_TEMPLATE.format(split.upper(), accuracy, hamming, jaccard, precision, recall, f1, microf1, f1, avg_loss))
if __name__ == '__main__':
# Set default configuration in args.py
args = get_args()
if args.local_rank == -1 or not args.cuda:
device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
print('Device:', str(device).upper())
print('Number of GPUs:', n_gpu)
print('Distributed training:', bool(args.local_rank != -1))
print('FP16:', args.fp16)
# Set random seed for reproducibility
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
dataset_map = {
'SST-2': SST2Processor,
'Reuters': ReutersProcessor,
'IMDB': IMDBProcessor,
'AAPD': AAPDProcessor,
'AGNews': AGNewsProcessor,
'Yelp2014': Yelp2014Processor,
'Sogou': SogouProcessor,
'MBTI': MBTIProcessor,
}
if args.gradient_accumulation_steps < 1:
raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
args.gradient_accumulation_steps))
if args.dataset not in dataset_map:
raise ValueError('Unrecognized dataset')
args.batch_size = args.batch_size // args.gradient_accumulation_steps
args.device = device
args.n_gpu = n_gpu
args.num_labels = dataset_map[args.dataset].NUM_CLASSES
args.is_multilabel = dataset_map[args.dataset].IS_MULTILABEL
if not args.trained_model:
save_path = os.path.join(args.save_path, dataset_map[args.dataset].NAME)
os.makedirs(save_path, exist_ok=True)
processor = dataset_map[args.dataset]()
args.is_lowercase = 'uncased' in args.model
args.is_hierarchical = False
tokenizer = BertTokenizer.from_pretrained(args.model, is_lowercase=args.is_lowercase)
train_examples = None
num_train_optimization_steps = None
if not args.trained_model:
train_examples = processor.get_train_examples(args.data_dir)
num_train_optimization_steps = int(
len(train_examples) / args.batch_size / args.gradient_accumulation_steps) * args.epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
cache_dir = args.cache_dir if args.cache_dir else os.path.join(str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(args.local_rank))
model = BertForSequenceClassification.from_pretrained(args.model, cache_dir=cache_dir, num_labels=args.num_labels)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError("Install NVIDIA Apex to use distributed and FP16 training.")
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError("Please install NVIDIA Apex for distributed and FP16 training")
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.lr,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.lr,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
trainer = BertTrainer(model, optimizer, processor, args)
if not args.trained_model:
trainer.train()
model = torch.load(trainer.snapshot_path)
else:
model = BertForSequenceClassification.from_pretrained(args.model, num_labels=args.num_labels)
model_ = torch.load(args.trained_model, map_location=lambda storage, loc: storage)
state={}
for key in model_.state_dict().keys():
new_key = key.replace("module.", "")
state[new_key] = model_.state_dict()[key]
model.load_state_dict(state)
model = model.to(device)
evaluate_split(model, processor, args, split='dev')
evaluate_split(model, processor, args, split='test')