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)}
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 = BertForMultiClass.from_pretrained(args.resume_path, num_labels=len(label_list))
else:
model = BertForMultiClass.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=CrossEntropy(),
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,
evaluate=F1Score(),
class_report=ClassReport(target_names=[id2label[x] for x in range(len(label2id))]))
trainer.train(train_data=train_dataloader, valid_data=valid_dataloader, seed=args.seed)
def main():
# **************************** Log initial data ***********************
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()}
DT = DataTransformer(logger=logger, seed=config['train']['seed'])
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
valid_dataset = CreateDataset(data=valid,
tokenizer=tokenizer,
max_seq_len=config['train']['max_seq_len'],
seed=config['train']['seed'],
example_type='valid')
# train loader
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)
# validation set loader
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)
# **************************** initialize model ***********************
logger.info("initializing model")
model = BertFine.from_pretrained(
config['pretrained']['bert']['bert_model_dir'],
cache_dir=config['output']['cache_dir'],
num_classes=len(id2label))
# ************************** set params *************************
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")
# model checkpoint
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)
# monitor
train_monitor = TrainingMonitor(file_dir=config['output']['figure_dir'],
arch=config['model']['arch'])
# learning rate scheduler
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')],
'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)
trainer.train(train_data=train_loader, valid_data=valid_loader)
if len(config['train']['n_gpu']) > 0:
torch.cuda.empty_cache()
def run_train(args, data_names):
# --------- data
# processor = BertProcessor(vocab_path=config['bert_vocab_path'], do_lower_case=args.do_lower_case)
processor = BertProcessor()
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"{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,
# collate_fn=collate_fn)
#
# valid_data = processor.get_dev(config['data_dir'] / f"{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 = 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))
model = BertForMultiLable.from_pretrained("bert-base-multilingual-cased", num_labels=len(label_list))
#t_total = int(len(train_dataloader) / args.gradient_accumulation_steps * args.epochs)
t_total = 200000
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)
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='micro', task_type='binary'),
MultiLabelReport(id2label=id2label),
F1Score(average='micro', task_type='binary')])
trainer.model.zero_grad()
seed_everything(trainer.args.seed) # Added here for reproductibility (even between python 2 a
iter_num = 0
valid_dataloader = get_valid_dataloader(args)
for epoch in range(trainer.start_epoch, trainer.start_epoch + trainer.args.epochs):
trainer.logger.info(f"Epoch {epoch}/{trainer.args.epochs}")
update_epoch = True
for i, data_name in enumerate(data_names):
filename_int = int(data_name)
if filename_int > 3500:
continue
trainer.logger.info(f"Epoch {epoch} - summary {i+1}/{len(data_names)}"+ f": summary_{data_name}")
# train_dataloader, valid_dataloader = get_dataloader(args, data_name)
train_dataloader = get_dataloader(args, data_name)
# train_log, valid_log = trainer.train(train_data=train_dataloader, valid_data=valid_dataloader, epoch=update_epoch)
train_log = trainer.train(train_data=train_dataloader, epoch=update_epoch)
update_epoch = False
# if train_log == None:
# continue
iter_num += 1
# logs = dict(train_log)
# show_info = f'\nEpoch: {epoch} - ' + "-".join([f' {key}: {value:.4f} ' for key, value in logs.items()])
# trainer.logger.info(show_info)
if iter_num % 50 == 0:
valid_log = trainer.valid_epoch(valid_dataloader)
logs = dict(valid_log)
show_info = f'\nEpoch: {epoch} - ' + "-".join([f' {key}: {value:.4f} ' for key, value in logs.items()])
trainer.logger.info(show_info)
# save
if trainer.training_monitor:
trainer.training_monitor.epoch_step(logs)
# save model
if trainer.model_checkpoint:
if iter_num % 50 == 0:
# state = trainer.save_info(epoch, best=logs[trainer.model_checkpoint.monitor])
state = trainer.save_info(iter_num, best=logs[trainer.model_checkpoint.monitor])
trainer.model_checkpoint.bert_epoch_step(current=logs[trainer.model_checkpoint.monitor], state=state)
# early_stopping
if trainer.early_stopping:
trainer.early_stopping.epoch_step(epoch=epoch, current=logs[trainer.early_stopping.monitor])
if trainer.early_stopping.stop_training:
break
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(args.task_type)
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.{args.task_type}.pkl")
train_examples = processor.create_examples(
lines=train_data,
example_type=f'train_{args.task_type}',
cached_examples_file=config['data_dir'] /
f"cached_train_{args.task_type}_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.task_type, 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,
collate_fn=collate_fn)
valid_data = processor.get_dev(
config['data_dir'] / f"{args.data_name}.valid.{args.task_type}.pkl")
valid_examples = processor.create_examples(
lines=valid_data,
example_type=f'valid_{args.task_type}',
cached_examples_file=config['data_dir'] /
f"cached_valid_{args.task_type}_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.task_type, 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 = BertForMultiLable.from_pretrained(args.resume_path,
num_labels=len(label_list))
else:
if args.task_type == 'trans':
model = BertForMultiLable_Fewshot.from_pretrained(
Path('pybert/output/checkpoints/bert/base'),
num_labels=len(label_list))
#model = BertForMultiLable.from_pretrained(config['bert_model_dir'], 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)
# 下面是optimizer和scheduler的设计
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)
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
) # TODO: 理解train_monitor的作用，感觉就是一个用来绘图的东西，用于记录每一个epoch中得到的结果
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)
], # 作用于batch之上的metrics，在每次loss.backward()之后都会执行计算，记得区分它与loss
epoch_metrics=[
AUC(average='micro', task_type='binary'), # 作用于epoch之上的metrics
MultiLabelReport(id2label=id2label),
F1Score(task_type='binary', average='micro', search_thresh=True)
]) # TODO: 考虑是否应该使用F1-score替代指标
trainer.train(train_data=train_dataloader, valid_data=valid_dataloader)
def run_test(args):
# TODO: 对训练集使用micro F1-score进行结果评测
from pybert.io.task_data import TaskData
from pybert.test.predictor import Predictor
data = TaskData()
ids, targets, sentences = data.read_data(
raw_data_path=config['test_path'],
preprocessor=ChinesePreProcessor(),
is_train=True) # 设置为True
lines = list(zip(sentences, targets))
#print(ids,sentences)
processor = BertProcessor(vocab_path=config['bert_vocab_path'],
do_lower_case=args.do_lower_case)
label_list = processor.get_labels(args.task_type)
id2label = {i: label for i, label in enumerate(label_list)}
test_data = processor.get_test(lines=lines)
test_examples = processor.create_examples(
lines=test_data,
example_type=f'test_{args.task_type}',
cached_examples_file=config['data_dir'] /
f"cached_test_{args.task_type}_examples_{args.arch}")
test_features = processor.create_features(
examples=test_examples,
max_seq_len=args.eval_max_seq_len,
cached_features_file=config['data_dir'] /
"cached_test_{}_features_{}_{}".format(
args.task_type, args.eval_max_seq_len, args.arch))
test_dataset = processor.create_dataset(test_features)
test_sampler = SequentialSampler(test_dataset)
test_dataloader = DataLoader(test_dataset,
sampler=test_sampler,
batch_size=args.train_batch_size,
collate_fn=collate_fn)
model = None
if args.task_type == 'base':
model = BertForMultiLable.from_pretrained(config['checkpoint_dir'],
num_labels=len(label_list))
else:
# model = BertForMultiLable.from_pretrained(config['checkpoint_dir'], num_labels=len(label_list))
model = BertForMultiLable_Fewshot.from_pretrained(
config['checkpoint_dir'], num_labels=len(label_list))
# ----------- predicting
logger.info('model predicting....')
predictor = Predictor(model=model, logger=logger, n_gpu=args.n_gpu)
result = predictor.predict(data=test_dataloader) # 感觉这个变量名叫all_logits可能更好
# TODO: 计算F1-score，这个功能模块需要用代码测试一下~
f1_metric = F1Score(task_type='binary',
average='micro',
search_thresh=True)
all_logits = torch.tensor(result, dtype=torch.float) # 转换成tensor
all_labels = torch.tensor(targets, dtype=torch.long) # 转换成tensor
f1_metric(all_logits, all_labels) # 会自动打印结果
print(f1_metric.value())
# 将结果写入一个文件之中
with open('test_output/test.log', 'a+') as f:
f.write(str(f1_metric.value()) + "\n")
thresh = f1_metric.thresh
ids = np.array(ids)
df1 = pd.DataFrame(ids, index=None)
df2 = pd.DataFrame(result, index=None)
all_df = pd.concat([df1, df2], axis=1)
if args.task_type == 'base':
all_df.columns = ['id', 'zy', 'gfgqzr', 'qs', 'tz', 'ggjc']
else:
all_df.columns = ['id', 'sg', 'pj', 'zb', 'qsht', 'db']
for column in all_df.columns[1:]:
all_df[column] = all_df[column].apply(lambda x: 1 if x > thresh else 0)
# all_df['zy'] = all_df['zy'].apply(lambda x: 1 if x>thresh else 0)
# all_df['gfgqzr'] = all_df['gfgqzr'].apply(lambda x: 1 if x>thresh else 0)
# all_df['qs'] = all_df['qs'].apply(lambda x: 1 if x>thresh else 0)
# all_df['tz'] = all_df['tz'].apply(lambda x: 1 if x>thresh else 0)
# all_df['ggjc'] = all_df['ggjc'].apply(lambda x: 1 if x>thresh else 0)
all_df.to_csv(f"test_output/{args.task_type}/cls_out.csv", index=False)
def main():
# **************************** Basic Info ***********************
logger = init_logger(log_name=config['arch'], log_dir=config['log_dir'])
logger.info("seed is %d" % config['seed'])
device = 'cuda:%d' % config['n_gpus'][0] if len(
config['n_gpus']) else 'cpu'
seed_everything(seed=config['seed'], device=device)
logger.info('starting load data from disk')
# split the reports
if config['resume']:
split_reports = SplitReports(raw_reports_dir=config['raw_reports_dir'],
raw_data_path=config['raw_data_path'])
split_reports.split()
df = pd.read_csv(config['raw_data_path'])
label_list = df.columns.values[2:].tolist()
config['label_to_id'] = {label: i for i, label in enumerate(label_list)}
config['id_to_label'] = {i: label for i, label in enumerate(label_list)}
config['vocab_path'] = path.sep.join(
[config['bert_model_dir'], 'vocab.txt'])
# **************************** Data ***********************
data_transformer = DataTransformer(logger=logger,
raw_data_path=config['raw_data_path'],
label_to_id=config['label_to_id'],
train_file=config['train_file_path'],
valid_file=config['valid_file_path'],
valid_size=config['valid_size'],
seed=config['seed'],
preprocess=Preprocessor(),
shuffle=config['shuffle'],
skip_header=True,
stratify=False)
# dataloader and pre-processing
data_transformer.read_data()
tokenizer = BertTokenizer(vocab_file=config['vocab_path'],
do_lower_case=config['do_lower_case'])
# train
train_dataset = CreateDataset(data_path=config['train_file_path'],
tokenizer=tokenizer,
max_seq_len=config['max_seq_len'],
seed=config['seed'],
example_type='train')
# valid
valid_dataset = CreateDataset(data_path=config['valid_file_path'],
tokenizer=tokenizer,
max_seq_len=config['max_seq_len'],
seed=config['seed'],
example_type='valid')
# resume best model
if config['resume']:
train_loader = [0]
else:
train_loader = DataLoader(dataset=train_dataset,
batch_size=config['batch_size'],
num_workers=config['num_workers'],
shuffle=True,
drop_last=False,
pin_memory=False)
# valid
valid_loader = DataLoader(dataset=valid_dataset,
batch_size=config['batch_size'],
num_workers=config['num_workers'],
shuffle=False,
drop_last=False,
pin_memory=False)
# **************************** Model ***********************
logger.info("initializing model")
if config['resume']:
with open(config['lab_dir'], 'r') as f:
config['label_to_id'] = load(f)
model = BertFine.from_pretrained(config['bert_model_dir'],
cache_dir=config['cache_dir'],
num_classes=len(config['label_to_id']))
# ************************** 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
}]
num_train_steps = int(
len(train_dataset.examples) / config['batch_size'] /
config['gradient_accumulation_steps'] * config['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['learning_rate'],
warmup=config['warmup_proportion'],
t_total=num_train_steps)
# **************************** callbacks ***********************
logger.info("initializing callbacks")
# save model
model_checkpoint = ModelCheckpoint(
checkpoint_dir=config['checkpoint_dir'],
mode=config['mode'],
monitor=config['monitor'],
save_best_only=config['save_best_only'],
best_model_name=config['best_model_name'],
epoch_model_name=config['epoch_model_name'],
arch=config['arch'],
logger=logger)
# monitor
train_monitor = TrainingMonitor(fig_dir=config['figure_dir'],
json_dir=config['log_dir'],
arch=config['arch'])
# TensorBoard
start_time = datetime.datetime.now().strftime('%m%d_%H%M%S')
writer_dir = os.path.join(config['writer_dir'], config['feature-based'],
start_time)
TSBoard = WriterTensorboardX(writer_dir=writer_dir,
logger=logger,
enable=True)
# learning rate
lr_scheduler = BertLr(optimizer=optimizer,
lr=config['learning_rate'],
t_total=num_train_steps,
warmup=config['warmup_proportion'])
# **************************** training model ***********************
logger.info('training model....')
trainer = Trainer(model=model,
train_data=train_loader,
val_data=valid_loader,
optimizer=optimizer,
epochs=config['epochs'],
criterion=BCEWithLogLoss(),
logger=logger,
model_checkpoint=model_checkpoint,
training_monitor=train_monitor,
TSBoard=TSBoard,
resume=config['resume'],
lr_scheduler=lr_scheduler,
n_gpu=config['n_gpus'],
label_to_id=config['label_to_id'],
evaluate_auc=AUC(sigmoid=True),
evaluate_f1=F1Score(sigmoid=True),
incorrect=Incorrect(sigmoid=True))
trainer.summary()
trainer.train()
# release cache
if len(config['n_gpus']) > 0:
torch.cuda.empty_cache()
def main():
# **************************** 基础信息 ***********************
logger = init_logger(log_name=config['arch'], log_dir=config['log_dir'])
logger.info("seed is %d" % config['seed'])
device = 'cuda:%d' % config['n_gpus'][0] if len(
config['n_gpus']) else 'cpu'
seed_everything(seed=config['seed'], device=device)
logger.info('starting load data from disk')
config['id_to_label'] = {v: k for k, v in config['label_to_id'].items()}
# **************************** 数据生成 ***********************
data_transformer = DataTransformer(logger=logger,
label_to_id=config['label_to_id'],
train_file=config['train_file_path'],
valid_file=config['valid_file_path'],
valid_size=config['valid_size'],
seed=config['seed'],
shuffle=True,
skip_header=False,
preprocess=None,
raw_data_path=config['raw_data_path'])
# 读取数据集以及数据划分
data_transformer.read_data()
# train
train_dataset = CreateDataset(data_path=config['train_file_path'],
vocab_path=config['vocab_path'],
max_seq_len=config['max_seq_len'],
seed=config['seed'],
example_type='train')
# valid
valid_dataset = CreateDataset(data_path=config['valid_file_path'],
vocab_path=config['vocab_path'],
max_seq_len=config['max_seq_len'],
seed=config['seed'],
example_type='valid')
#加载训练数据集
train_loader = DataLoader(dataset=train_dataset,
batch_size=config['batch_size'],
num_workers=config['num_workers'],
shuffle=True,
drop_last=False,
pin_memory=False)
# 验证数据集
valid_loader = DataLoader(dataset=valid_dataset,
batch_size=config['batch_size'],
num_workers=config['num_workers'],
shuffle=False,
drop_last=False,
pin_memory=False)
# **************************** 模型 ***********************
logger.info("initializing model")
model = BertFine.from_pretrained(config['bert_model_dir'],
cache_dir=config['cache_dir'],
num_classes=len(config['label_to_id']))
# ************************** 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['batch_size'] /
config['gradient_accumulation_steps'] * config['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['learning_rate'],
warmup=config['warmup_proportion'],
t_total=num_train_steps)
# **************************** callbacks ***********************
logger.info("initializing callbacks")
# 模型保存
model_checkpoint = ModelCheckpoint(
checkpoint_dir=config['checkpoint_dir'],
mode=config['mode'],
monitor=config['monitor'],
save_best_only=config['save_best_only'],
best_model_name=config['best_model_name'],
epoch_model_name=config['epoch_model_name'],
arch=config['arch'],
logger=logger)
# 监控训练过程
train_monitor = TrainingMonitor(fig_dir=config['figure_dir'],
json_dir=config['log_dir'],
arch=config['arch'])
# 学习率机制
lr_scheduler = BertLr(optimizer=optimizer,
lr=config['learning_rate'],
t_total=num_train_steps,
warmup=config['warmup_proportion'])
# **************************** training model ***********************
logger.info('training model....')
trainer = Trainer(model=model,
train_data=train_loader,
val_data=valid_loader,
optimizer=optimizer,
epochs=config['epochs'],
criterion=CrossEntropy(),
logger=logger,
model_checkpoint=model_checkpoint,
training_monitor=train_monitor,
resume=config['resume'],
lr_scheduler=lr_scheduler,
n_gpu=config['n_gpus'],
evaluate=F1Score(),
class_report=ClassReport(target_names=[
config['id_to_label'][x]
for x in range(len(config['label_to_id']))
]))
# 查看模型结构
trainer.summary()
# 拟合模型
trainer.train()
# 释放显存
if len(config['n_gpus']) > 0:
torch.cuda.empty_cache()