def main():
# **************************** 基础信息 ***********************
logger = init_logger(log_name=config['model']['arch'],
log_dir=config['output']['log_dir'])
logger.info(f"seed is {config['train']['seed']}")
device = f"cuda: {config['train']['n_gpu'][0] if len(config['train']['n_gpu']) else 'cpu'}"
seed_everything(seed=config['train']['seed'], device=device)
logger.info('starting load data from disk')
id2label = {value: key for key, value in config['label2id'].items()}
# **************************** 数据生成 ***********************
DT = DataTransformer(logger=logger, seed=config['train']['seed'])
# 读取数据集以及数据划分
targets, sentences = DT.read_data(
raw_data_path=config['data']['raw_data_path'],
preprocessor=EnglishPreProcessor(),
is_train=True)
train, valid = DT.train_val_split(
X=sentences,
y=targets,
save=True,
shuffle=True,
stratify=False,
valid_size=config['train']['valid_size'],
train_path=config['data']['train_file_path'],
valid_path=config['data']['valid_file_path'])
tokenizer = BertTokenizer(
vocab_file=config['pretrained']['bert']['vocab_path'],
do_lower_case=config['train']['do_lower_case'])
# train
train_dataset = CreateDataset(data=train,
tokenizer=tokenizer,
max_seq_len=config['train']['max_seq_len'],
seed=config['train']['seed'],
example_type='train')
valid_dataset = CreateDataset(data=valid,
tokenizer=tokenizer,
max_seq_len=config['train']['max_seq_len'],
seed=config['train']['seed'],
example_type='valid')
#加载训练数据集
train_loader = DataLoader(dataset=train_dataset,
batch_size=config['train']['batch_size'],
num_workers=config['train']['num_workers'],
shuffle=True,
drop_last=False,
pin_memory=False)
# 验证数据集
valid_loader = DataLoader(dataset=valid_dataset,
batch_size=config['train']['batch_size'],
num_workers=config['train']['num_workers'],
shuffle=False,
drop_last=False,
pin_memory=False)
# **************************** 模型 ***********************
logger.info("initializing model")
model = BertFine.from_pretrained(
config['pretrained']['bert']['bert_model_dir'],
cache_dir=config['output']['cache_dir'],
num_classes=len(id2label))
# ************************** 优化器 *************************
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
}]
num_train_steps = int(
len(train_dataset.examples) / config['train']['batch_size'] /
config['train']['gradient_accumulation_steps'] *
config['train']['epochs'])
# t_total: total number of training steps for the learning rate schedule
# warmup: portion of t_total for the warmup
optimizer = BertAdam(optimizer_grouped_parameters,
lr=config['train']['learning_rate'],
warmup=config['train']['warmup_proportion'],
t_total=num_train_steps)
# **************************** callbacks ***********************
logger.info("initializing callbacks")
# 模型保存
model_checkpoint = ModelCheckpoint(
checkpoint_dir=config['output']['checkpoint_dir'],
mode=config['callbacks']['mode'],
monitor=config['callbacks']['monitor'],
save_best_only=config['callbacks']['save_best_only'],
arch=config['model']['arch'],
logger=logger)
# 监控训练过程
train_monitor = TrainingMonitor(file_dir=config['output']['figure_dir'],
arch=config['model']['arch'])
# 学习率机制
lr_scheduler = BertLR(optimizer=optimizer,
learning_rate=config['train']['learning_rate'],
t_total=num_train_steps,
warmup=config['train']['warmup_proportion'])
# **************************** training model ***********************
logger.info('training model....')
train_configs = {
'model':
model,
'logger':
logger,
'optimizer':
optimizer,
'resume':
config['train']['resume'],
'epochs':
config['train']['epochs'],
'n_gpu':
config['train']['n_gpu'],
'gradient_accumulation_steps':
config['train']['gradient_accumulation_steps'],
'epoch_metrics': [
F1Score(average='micro', task_type='binary'),
MultiLabelReport(id2label=id2label)
],
'batch_metrics': [AccuracyThresh(thresh=0.5)],
'criterion':
BCEWithLogLoss(),
'model_checkpoint':
model_checkpoint,
'training_monitor':
train_monitor,
'lr_scheduler':
lr_scheduler,
'early_stopping':
None,
'verbose':
1
}
trainer = Trainer(train_configs=train_configs)
# 拟合模型
trainer.train(train_data=train_loader, valid_data=valid_loader)
# 释放显存
if len(config['train']['n_gpu']) > 0:
torch.cuda.empty_cache()
def run_train(args):
# --------- data
model_to_use = "roberta-large"
processor = RobertaProcessor(model_type=model_to_use) #vocab_path=config['roberta_vocab_path'], merge_path=config['roberta_merge_path'])
label_list = processor.get_labels()
label2id = {label: i for i, label in enumerate(label_list)}
id2label = {i: label for i, label in enumerate(label_list)}
train_data = processor.get_train(config['data_dir'] / f"{args.data_name}.train.pkl")
train_examples = processor.create_examples(lines=train_data,
example_type='train',
cached_examples_file=config[
'data_dir'] / f"cached_train_examples_{args.arch}")
train_features = processor.create_features(examples=train_examples,
max_seq_len=args.train_max_seq_len,
cached_features_file=config[
'data_dir'] / "cached_train_features_{}_{}".format(
args.train_max_seq_len, args.arch
))
train_dataset = processor.create_dataset(train_features, is_sorted=args.sorted)
train_sampler = RandomSampler(train_dataset)
train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size,
collate_fn=collate_fn)
valid_data = processor.get_dev(config['data_dir'] / f"{args.data_name}.valid.pkl")
valid_examples = processor.create_examples(lines=valid_data,
example_type='valid',
cached_examples_file=config[
'data_dir'] / f"cached_valid_examples_{args.arch}")
valid_features = processor.create_features(examples=valid_examples,
max_seq_len=args.eval_max_seq_len,
cached_features_file=config[
'data_dir'] / "cached_valid_features_{}_{}".format(
args.eval_max_seq_len, args.arch
))
valid_dataset = processor.create_dataset(valid_features)
valid_sampler = SequentialSampler(valid_dataset)
valid_dataloader = DataLoader(valid_dataset, sampler=valid_sampler, batch_size=args.eval_batch_size,
collate_fn=collate_fn)
# ------- model
logger.info("initializing model")
if args.resume_path:
args.resume_path = Path(args.resume_path)
model = RobertaForMultiLable.from_pretrained(args.resume_path, num_labels=len(label_list))
else:
model = RobertaForMultiLable.from_pretrained(model_to_use, num_labels=len(label_list))
#config['roberta_model_dir']
print("""\n\nname module\n----------------------""")
for name, module in model.named_children():
if name == "roberta":
for n, _ in module.named_children():
print(f"{name}:{n}")
else:
print("{:15} {}".format(name, module))
print()
#return
# print("================= train dataloader length is", len(train_dataloader), "=================\n")
t_total = int(len(train_dataloader) / args.gradient_accumulation_steps * args.epochs)
param_optimizer = list(model.named_parameters())
no_decay = ['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) and 'classifier.weight' not in n], 'weight_decay': args.weight_decay}, #
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0},
{'params': model.classifier.weight, 'lr': 5e-4} # best: 5e-4
# {'params': model.classifier.bias, 'lr': 5e-4, 'weight_decay': 0.0}
]
# model.parameters()
warmup_steps = int(t_total * args.warmup_proportion)
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=warmup_steps,
num_training_steps=t_total)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
# ---- callbacks
logger.info("initializing callbacks")
train_monitor = TrainingMonitor(file_dir=config['figure_dir'], arch=args.arch)
model_checkpoint = ModelCheckpoint(checkpoint_dir=config['checkpoint_dir'],mode=args.mode,
monitor=args.monitor,arch=args.arch,
save_best_only=args.save_best)
# **************************** training model ***********************
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Num Epochs = %d", args.epochs)
logger.info(" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps * (
torch.distributed.get_world_size() if args.local_rank != -1 else 1))
logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
trainer = Trainer(args=args, model=model, logger=logger, criterion=BCEWithLogLoss(), optimizer=optimizer,
scheduler=scheduler, early_stopping=None, training_monitor=train_monitor,
model_checkpoint=model_checkpoint,
batch_metrics=[AccuracyThresh(thresh=0.5)],
epoch_metrics=[AUC(average='weighted', task_type='binary'), # average='micro'
MultiLabelReport(id2label=id2label)])
trainer.train(train_data=train_dataloader, valid_data=valid_dataloader)
def run_train(args):
# --------- data
processor = BertProcessor(vocab_path=config['bert_vocab_path'],
do_lower_case=args.do_lower_case)
label_list = processor.get_labels()
label2id = {label: i for i, label in enumerate(label_list)}
id2label = {i: label for i, label in enumerate(label_list)}
#print(config['data_dir'] / f"{args.data_name}.train.pkl")
#exit()
train_data = processor.get_train(config['data_dir'] /
f"{args.data_name}.train.pkl")
train_examples = processor.create_examples(
lines=train_data,
example_type='train',
cached_examples_file=config['data_dir'] /
f"cached_train_examples_{args.arch}")
train_features = processor.create_features(
examples=train_examples,
max_seq_len=args.train_max_seq_len,
cached_features_file=config['data_dir'] /
"cached_train_features_{}_{}".format(args.train_max_seq_len,
args.arch))
train_dataset = processor.create_dataset(train_features,
is_sorted=args.sorted)
if args.sorted:
train_sampler = SequentialSampler(train_dataset)
else:
train_sampler = RandomSampler(train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
valid_data = processor.get_dev(config['data_dir'] /
f"{args.data_name}.valid.pkl")
valid_examples = processor.create_examples(
lines=valid_data,
example_type='valid',
cached_examples_file=config['data_dir'] /
f"cached_valid_examples_{args.arch}")
valid_features = processor.create_features(
examples=valid_examples,
max_seq_len=args.eval_max_seq_len,
cached_features_file=config['data_dir'] /
"cached_valid_features_{}_{}".format(args.eval_max_seq_len, args.arch))
valid_dataset = processor.create_dataset(valid_features)
valid_sampler = SequentialSampler(valid_dataset)
valid_dataloader = DataLoader(valid_dataset,
sampler=valid_sampler,
batch_size=args.eval_batch_size)
# ------- model
logger.info("initializing model")
if args.resume_path:
args.resume_path = Path(args.resume_path)
model = BertForMultiLable.from_pretrained(args.resume_path,
num_labels=len(label_list))
else:
model = BertForMultiLable.from_pretrained(config['bert_model_dir'],
num_labels=len(label_list))
t_total = int(
len(train_dataloader) / args.gradient_accumulation_steps * args.epochs)
param_optimizer = list(model.named_parameters())
no_decay = ['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':
args.weight_decay
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
warmup_steps = int(t_total * args.warmup_proportion)
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
lr_scheduler = WarmupLinearSchedule(optimizer,
warmup_steps=warmup_steps,
t_total=t_total)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
# ---- callbacks
logger.info("initializing callbacks")
train_monitor = TrainingMonitor(file_dir=config['figure_dir'],
arch=args.arch)
model_checkpoint = ModelCheckpoint(checkpoint_dir=config['checkpoint_dir'],
mode=args.mode,
monitor=args.monitor,
arch=args.arch,
save_best_only=args.save_best)
# **************************** training model ***********************
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Num Epochs = %d", args.epochs)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps *
(torch.distributed.get_world_size() if args.local_rank != -1 else 1))
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
trainer = Trainer(
n_gpu=args.n_gpu,
model=model,
epochs=args.epochs,
logger=logger,
criterion=BCEWithLogLoss(),
optimizer=optimizer,
lr_scheduler=lr_scheduler,
early_stopping=None,
training_monitor=train_monitor,
fp16=args.fp16,
resume_path=args.resume_path,
grad_clip=args.grad_clip,
model_checkpoint=model_checkpoint,
gradient_accumulation_steps=args.gradient_accumulation_steps,
batch_metrics=[AccuracyThresh(thresh=0.5)],
epoch_metrics=[
AUC(average='micro', task_type='binary'),
MultiLabelReport(id2label=id2label)
])
trainer.train(train_data=train_dataloader,
valid_data=valid_dataloader,
seed=args.seed)