-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
198 lines (177 loc) · 8.49 KB
/
train.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
197
198
from transformers import AutoAdapterModel, AutoTokenizer, AutoConfig
from datasets import load_dataset
from torch.utils.data import Dataset
import torch
from torch import nn
import torch.nn.functional as F
from torch.utils.data.distributed import DistributedSampler
from tqdm import tqdm
import os
import copy
import argparse
import numpy as np
import logging
import json
from utils.trainer import get, train
from utils.config import read_config, get_tgt_dataset, get_src_dataset, set_seed
from utils.modeling import BertForTokenClassification
adapter_dir = "adapter/"
def get_dataloaders(cfg, tokenizer, is_src):
if cfg.local_rank not in [-1, 0]:
torch.distributed.barrier()
if is_src:
data = get_src_dataset(cfg)
cfg.data = data
else:
data = get_tgt_dataset(cfg)
cfg.data = data
dataset = data.load(tokenizer)
if cfg.local_rank == 0:
print(dataset)
torch.distributed.barrier()
if cfg.local_rank == -1:
train_sampler = None
else:
train_sampler = DistributedSampler(dataset['training'], shuffle=True, seed=cfg.TRAIN.SEED)
batch_size = cfg.TRAIN.SRC_BATCH_SIZE if is_src else cfg.TRAIN.BATCH_SIZE
train_dataloader = torch.utils.data.DataLoader(dataset['training'], sampler=train_sampler, batch_size=batch_size // cfg.world_size)
dev_dataloader = torch.utils.data.DataLoader(dataset['development'], batch_size=batch_size // cfg.world_size)
return train_dataloader, dev_dataloader, data
def train_single(cfg, model, tokenizer):
# load data
train_dataloader, dev_dataloader, data = get_dataloaders(cfg, tokenizer, False)
# add adapter
adapter_name = cfg.DATA.TGT_DATASET + "_ner_" + cfg.MODEL.BACKBONE
head_name = cfg.DATA.TGT_DATASET + "_ner_" + cfg.MODEL.BACKBONE + "_head"
if cfg.ADAPTER.TRAIN != "None" and os.path.exists(cfg.ADAPTER.TRAIN):
if cfg.ADAPTER.ENABLE:
with open(os.path.join(cfg.ADAPTER.TRAIN, "adapter_config.json"), "r") as f:
config = json.load(f)
adapter_name = config["name"]
model.load_adapter(cfg.ADAPTER.TRAIN)
else:
model = BertForTokenClassification.from_pretrained(cfg.ADAPTER.TRAIN)
else:
if cfg.ADAPTER.ENABLE:
model.add_adapter(adapter_name)
else:
model = BertForTokenClassification.from_pretrained(cfg.MODEL.PATH, num_labels=len(data.labels), id2label=data.id2label)
if cfg.ADAPTER.ENABLE:
model.add_tagging_head(head_name, num_labels=len(data.labels), id2label=data.id2label)
model.train_adapter([adapter_name])
else:
model.train()
model.to(cfg.device)
# train
cfg.TRAIN.EPOCHS = cfg.TRAIN.TGT_EPOCHS
if cfg.TGT_LOSS.NAME == "CE_MS":
model, best_f1, valid_f1s = train(cfg, model, tokenizer, train_dataloader, dev_dataloader, adapter_name, head_name, use_ms=True)
elif cfg.TGT_LOSS.NAME == "CrossEntropy":
model, best_f1, valid_f1s = train(cfg, model, tokenizer, train_dataloader, dev_dataloader, adapter_name, head_name)
else:
raise NotImplementedError()
return model, adapter_name, head_name, best_f1, valid_f1s
def train_two_stage(cfg, model, tokenizer):
# load data
train_dataloader, dev_dataloader, data = get_dataloaders(cfg, tokenizer, True)
# add adapter
adapter_name = cfg.DATA.SRC_DATASET + "_ner_" + cfg.MODEL.BACKBONE
head_name = cfg.DATA.SRC_DATASET + "_ner_" + cfg.MODEL.BACKBONE + "_head"
if cfg.ADAPTER.ENABLE:
model.add_adapter(adapter_name)
model.add_tagging_head(head_name, num_labels=len(data.labels), id2label=data.id2label)
model.train_adapter([adapter_name])
else:
model = BertForTokenClassification.from_pretrained(cfg.MODEL.PATH, num_labels=len(data.labels), id2label=data.id2label)
model.train()
model.to(cfg.device)
# train 4 src
cfg.TRAIN.EPOCHS = cfg.TRAIN.SRC_EPOCHS
if cfg.SRC_LOSS.NAME == "CE_MS":
model, _, _ = train(cfg, model, tokenizer, train_dataloader, dev_dataloader, adapter_name, head_name, use_ms=True, pretrain=True)
elif cfg.TGT_LOSS.NAME == "CrossEntropy":
model, _, _ = train(cfg, model, tokenizer, train_dataloader, dev_dataloader, adapter_name, head_name, pretrain=True)
else:
raise NotImplemented
if cfg.local_rank in [-1, 0]:
if cfg.ADAPTER.ENABLE:
os.makedirs(os.path.join(os.path.dirname(cfg.OUTPUT.ADAPTER_SAVE_DIR), adapter_name + "_inter"), exist_ok=True)
os.makedirs(os.path.join(os.path.dirname(cfg.OUTPUT.HEAD_SAVE_DIR), head_name + "_inter"), exist_ok=True)
model.save_adapter(os.path.join(os.path.dirname(cfg.OUTPUT.ADAPTER_SAVE_DIR), adapter_name + "_inter"), adapter_name)
model.save_head(os.path.join(os.path.dirname(cfg.OUTPUT.HEAD_SAVE_DIR), head_name + "_inter"), head_name)
else:
model.save_pretrained(os.path.join(os.path.dirname(cfg.OUTPUT.ADAPTER_SAVE_DIR), "inter"))
cfg.logger.info("Best model for the 1st stage saved")
# prepare for target
set_seed(cfg.TRAIN.SEED)
train_dataloader, dev_dataloader, data = get_dataloaders(cfg, tokenizer, False)
if cfg.ADAPTER.ENABLE:
model = AutoAdapterModel.from_pretrained(cfg.MODEL.PATH)
model.load_adapter(os.path.join(os.path.dirname(cfg.OUTPUT.ADAPTER_SAVE_DIR), adapter_name + "_inter"))
head_name = cfg.DATA.TGT_DATASET + "_ner_" + cfg.MODEL.BACKBONE + "_head"
model.add_tagging_head(head_name, num_labels=len(data.labels), id2label=data.id2label, overwrite_ok=True)
model.train_adapter([adapter_name])
else:
model = BertForTokenClassification.from_pretrained(os.path.join(os.path.dirname(cfg.OUTPUT.ADAPTER_SAVE_DIR), "inter"))
model.train()
model.to(cfg.device)
# train 4 tgt
cfg.TRAIN.EPOCHS = cfg.TRAIN.TGT_EPOCHS
if cfg.TGT_LOSS.NAME == "CE_MS":
model, best_f1, valid_f1s = train(cfg, model, tokenizer, train_dataloader, dev_dataloader, adapter_name, head_name, use_ms=True)
elif cfg.TGT_LOSS.NAME == "CrossEntropy":
model, best_f1, valid_f1s = train(cfg, model, tokenizer, train_dataloader, dev_dataloader, adapter_name, head_name)
else:
raise NotImplementedError()
return model, adapter_name, head_name, best_f1, valid_f1s
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--cfg_file", type=str, default="configs/sample.yaml")
args = parser.parse_args()
cfg = read_config(args.cfg_file)
if "LOCAL_RANK" in os.environ:
local_rank = int(os.environ["LOCAL_RANK"])
world_size = int(os.environ["WORLD_SIZE"])
else:
local_rank = -1
world_size = 1
cfg.local_rank = local_rank
cfg.world_size = world_size
if local_rank == -1:
device = "cuda" if torch.cuda.is_available() else "cpu"
else:
torch.cuda.set_device(local_rank)
device = torch.device('cuda', local_rank)
cfg.device = device
if not hasattr(cfg.TRAIN, "SEED"):
cfg.TRAIN.SEED = 42
set_seed(cfg.TRAIN.SEED)
# initialize distributed process group
if local_rank !=- 1:
torch.distributed.init_process_group(backend='nccl')
handlers = [logging.StreamHandler()]
logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO,
handlers=handlers)
logging.getLogger("transformers.tokenization_utils").setLevel(logging.ERROR)
logger = logging.getLogger(__name__)
cfg.logger = logger
cfg.logger.info(args)
# load model
model_name = cfg.MODEL.PATH
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoAdapterModel.from_pretrained(model_name)
if not cfg.TRAIN.TWO_STAGE:
model, adapter_name, head_name, best_f1, valid_f1s = train_single(cfg, model, tokenizer)
else:
model, adapter_name, head_name, best_f1, valid_f1s = train_two_stage(cfg, model, tokenizer)
if cfg.local_rank in [-1, 0]:
if cfg.ADAPTER.ENABLE:
os.makedirs(os.path.join(cfg.OUTPUT.ADAPTER_SAVE_DIR, adapter_name), exist_ok=True)
os.makedirs(os.path.join(cfg.OUTPUT.HEAD_SAVE_DIR, head_name), exist_ok=True)
model.save_adapter(os.path.join(cfg.OUTPUT.ADAPTER_SAVE_DIR, adapter_name), adapter_name)
model.save_head(os.path.join(cfg.OUTPUT.HEAD_SAVE_DIR, head_name), head_name)
else:
model.save_pretrained(cfg.OUTPUT.ADAPTER_SAVE_DIR)
cfg.logger.info("Best model saved")