-
Notifications
You must be signed in to change notification settings - Fork 2
/
gpt2_pretrain.py
200 lines (166 loc) · 7.85 KB
/
gpt2_pretrain.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
199
200
"""
Copyright (c) VisualJoyce.
Licensed under the MIT license.
"""
import logging
import os
import torch
from torch.cuda.amp import autocast
from tqdm import tqdm
from transformers import HfArgumentParser, set_seed, AdamW, get_linear_schedule_with_warmup, BertTokenizer, \
GPT2Config, GPT2LMHeadModel
from transformers4ime.data.arguments import MMModelArguments, MMDataTrainingArguments, MMTrainingArguments
from transformers4ime.data.loaders import MM_LOADERS
from transformers4ime.utils.misc import NoOp
logger = logging.getLogger(__name__)
BUFSIZE = 40960000
def update_lr(optimizer, lr):
for param_group in optimizer.param_groups:
param_group['lr'] = lr
return lr
def concat_all_gather(tensor):
"""
Performs all_gather operation on the provided tensors.
*** Warning ***: torch.distributed.all_gather has no gradient.
"""
tensors_gather = [torch.ones_like(tensor)
for _ in range(torch.distributed.get_world_size())]
torch.distributed.all_gather(tensors_gather, tensor, async_op=False)
output = torch.cat(tensors_gather, dim=0)
return output
# light
# @light_init(params={"training_framework": "pytorch_ddp"})
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((MMModelArguments, MMDataTrainingArguments, MMTrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if training_args.process_index in [-1, 0]:
from transformers4ime.utils.logger import TensorboardLogger
TB_LOGGER = TensorboardLogger()
TB_LOGGER.create(training_args.logging_dir)
pbar = tqdm(total=training_args.max_steps)
else:
pbar = NoOp()
TB_LOGGER = NoOp()
# training_args.local_rank = 0 # for debug
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
logger.info("Data parameters %s", data_args)
# set_seed(training_args.seed)
set_seed(training_args.seed + training_args.process_index)
config = GPT2Config.from_pretrained(model_args.model_name_or_path)
logger.info("Model configurations %s", config)
tokenizer = BertTokenizer.from_pretrained(model_args.model_name_or_path)
model = GPT2LMHeadModel.from_pretrained(model_args.model_name_or_path, config=config)
model.resize_token_embeddings(len(tokenizer))
device = training_args.device
model = model.to(device)
best_pt = model_args.best_pt
if best_pt:
logger.info(f"Loading best checkpoint from: {best_pt}")
model.load_state_dict(torch.load(best_pt, map_location=device), strict=True)
logger.info("getting data")
train_data = MM_LOADERS[training_args.interchange_mode](tokenizer, model_args, training_args, data_args,
config) # infinite data generator
assert len(train_data.modalities) == 1 and 'text_only' in train_data.modalities
logger.info("init trainer")
# Initialize our Trainer
logger.info("start training")
# Training
if training_args.do_train:
# do train:
model.train()
scaler = torch.cuda.amp.GradScaler()
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": training_args.weight_decay,
},
{
"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)],
"weight_decay": 0.0,
},
]
optimizer = AdamW(
optimizer_grouped_parameters,
lr=training_args.learning_rate,
betas=(training_args.adam_beta1, training_args.adam_beta2),
eps=training_args.adam_epsilon,
)
lr_scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=training_args.warmup_steps,
num_training_steps=training_args.max_steps)
if os.path.isfile(os.path.join(model_args.model_name_or_path, "scheduler.pt")):
optimizer.load_state_dict(model_args.model_name_or_path + '')
# os.environ['MASTER_ADDR'] = 'localhost' # for debug
# os.environ['MASTER_PORT'] = '8888'
# torch.distributed.init_process_group(backend='nccl',init_method='env://',
# world_size=1, rank=training_args.local_rank) # for debug
if training_args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[training_args.local_rank],
output_device=training_args.local_rank,
find_unused_parameters=True)
model.zero_grad()
global_step = 0
logger.info("start iterate")
# do train
for step, batch in enumerate(train_data):
should_grad_sync_and_apply = batch.pop('should_grad_sync_and_apply')
gradient_accumulation_steps = batch.pop('gradient_accumulation_steps')
with autocast():
if not should_grad_sync_and_apply:
if training_args.local_rank != -1:
with model.no_sync():
outputs = model(**batch, return_dict=True)
loss = outputs.loss / gradient_accumulation_steps
scaler.scale(loss).backward()
else:
outputs = model(**batch, return_dict=True)
loss = outputs.loss / gradient_accumulation_steps
scaler.scale(loss).backward()
else:
global_step += 1
outputs = model(**batch, return_dict=True)
loss = outputs.loss / gradient_accumulation_steps
scaler.scale(loss).backward()
scaler.unscale_(optimizer)
total_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), training_args.max_grad_norm)
scaler.step(optimizer)
scaler.update()
lr_scheduler.step()
model.zero_grad()
pbar.update(1)
TB_LOGGER.add_scalar('train/grad_norm', total_norm, global_step)
TB_LOGGER.add_scalar('train/loss', outputs.loss.item(), global_step)
for k, v in train_data.all_epochs.items():
TB_LOGGER.add_scalar(f'train_epoch/{k}', v, global_step)
for gid, group in enumerate(optimizer.param_groups):
TB_LOGGER.add_scalar(f'train/lr_{gid}', group['lr'], global_step)
TB_LOGGER.step()
if global_step % training_args.save_steps == 0 and training_args.should_save:
ckpt_output_dir = os.path.join(training_args.output_dir, 'ckpt' + str(global_step))
if training_args.local_rank != -1:
model.module.save_pretrained(ckpt_output_dir)
else:
model.save_pretrained(ckpt_output_dir)
if global_step > training_args.max_steps:
break
if __name__ == "__main__":
main()