def run_test(args):
from pybert.io.task_data import TaskData
from pybert.test.predictor import Predictor
data = TaskData()
targets, sentences = data.read_data(raw_data_path=config['test_path'],
preprocessor=EnglishPreProcessor(),
is_train=False)
lines = list(zip(sentences, targets))
processor = BertProcessor(vocab_path=config['bert_vocab_path'], do_lower_case=args.do_lower_case)
label_list = processor.get_labels()
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='test',
cached_examples_file=config[
'data_dir'] / f"cached_test_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.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)
model = BertForMultiLable.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)
print(result)
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 = 'cuda:%d' % 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']['test_file_path'],
preprocessor=EnglishPreProcessor(),
is_train=False)
tokenizer = BertTokenizer(
vocab_file=config['pretrained']['bert']['vocab_path'],
do_lower_case=config['train']['do_lower_case'])
# train
test_dataset = CreateDataset(data=list(zip(sentences, targets)),
tokenizer=tokenizer,
max_seq_len=config['train']['max_seq_len'],
seed=config['train']['seed'],
example_type='test')
# 验证数据集
test_loader = DataLoader(dataset=test_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))
# **************************** training model ***********************
logger.info('model predicting....')
predicter = Predicter(
model=model,
logger=logger,
n_gpu=config['train']['n_gpu'],
model_path=config['output']['checkpoint_dir'] /
f"best_{config['model']['arch']}_model.pth",
)
# 拟合模型
result = predicter.predict(data=test_loader)
print(result)
# 释放显存
if len(config['train']['n_gpu']) > 0:
torch.cuda.empty_cache()
def main():
parser = ArgumentParser()
parser.add_argument("--arch", default='bert', type=str)
parser.add_argument("--do_data", action='store_true')
parser.add_argument("--do_train", action='store_true')
parser.add_argument("--do_test", action='store_true')
parser.add_argument("--save_best", action='store_true')
parser.add_argument("--do_lower_case", action='store_true')
parser.add_argument('--data_name', default='kaggle', type=str)
parser.add_argument("--epochs", default=6, type=int)
parser.add_argument("--resume_path", default='', type=str)
parser.add_argument("--mode", default='min', type=str)
parser.add_argument("--monitor", default='valid_loss', type=str)
parser.add_argument("--valid_size", default=0.2, type=float)
parser.add_argument("--local_rank", type=int, default=-1)
parser.add_argument("--sorted", default=1, type=int, help='1 : True 0:False ')
parser.add_argument("--n_gpu", type=str, default='0', help='"0,1,.." or "0" or "" ')
parser.add_argument('--gradient_accumulation_steps', type=int, default=1)
parser.add_argument("--train_batch_size", default=8, type=int)
parser.add_argument('--eval_batch_size', default=8, type=int)
parser.add_argument("--train_max_seq_len", default=256, type=int)
parser.add_argument("--eval_max_seq_len", default=256, type=int)
parser.add_argument('--loss_scale', type=float, default=0)
parser.add_argument("--warmup_proportion", default=0.1, type=int, )
parser.add_argument("--weight_decay", default=0.01, type=float)
parser.add_argument("--adam_epsilon", default=1e-8, type=float)
parser.add_argument("--grad_clip", default=1.0, type=float)
parser.add_argument("--learning_rate", default=2e-5, type=float)
parser.add_argument('--seed', type=int, default=42)
parser.add_argument('--fp16', action='store_true')
parser.add_argument('--fp16_opt_level', type=str, default='O1')
args = parser.parse_args()
config['checkpoint_dir'] = config['checkpoint_dir'] / args.arch
config['checkpoint_dir'].mkdir(exist_ok=True)
# Good practice: save your training arguments together with the trained model
torch.save(args, config['checkpoint_dir'] / 'training_args.bin')
seed_everything(args.seed)
init_logger(log_file=config['log_dir'] / f"{args.arch}.log")
logger.info("Training/evaluation parameters %s", args)
if args.do_data:
from pybert.io.task_data import TaskData
data = TaskData()
targets, sentences = data.read_data(raw_data_path=config['raw_data_path'],
preprocessor=EnglishPreProcessor(),
is_train=True)
data.train_val_split(X=sentences, y=targets, shuffle=True, stratify=False,
valid_size=args.valid_size, data_dir=config['data_dir'],
data_name=args.data_name)
if args.do_train:
run_train(args)
if args.do_test:
run_test(args)
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_test(args, test=False, k=7, med_map='pybert/dataset/med_map.csv'):
from pybert.io.task_data import TaskData
from pybert.test.predictor import Predictor
data = TaskData()
targets, sentences = data.read_data(raw_data_path=config['test_path'],
preprocessor=EnglishPreProcessor(),
is_train=test)
print(
f'-----------------------------------------\ntargets {targets}\n---------------------------------------------------'
)
lines = list(zip(sentences, targets))
processor = BertProcessor(vocab_path=config['bert_vocab_path'],
do_lower_case=args.do_lower_case)
label_list = processor.get_labels()
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='test',
cached_examples_file=config['data_dir'] /
f"cached_test_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.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)
model = BertForMultiLable.from_pretrained(config['checkpoint_dir'])
# ----------- predicting
logger.info('model predicting....')
predictor = Predictor(model=model,
logger=logger,
n_gpu=args.n_gpu,
test=test)
if test:
results, targets = predictor.predict(data=test_dataloader)
#print(f'results {results.shape}')
#print(f'targets {targets.shape}')
result = dict()
metrics = [Recall(), Acc()]
for metric in metrics:
metric.reset()
metric(logits=results, target=targets)
value = metric.value()
if value is not None:
result[f'valid_{metric.name()}'] = value
return result
else:
results = predictor.predict(data=test_dataloader)
pred = np.argsort(results)[:, -k:][:, ::-1]
with open('pybert/dataset/med_map.csv', mode='r') as infile:
reader = csv.reader(infile)
med_dict = {int(rows[0]): rows[1] for rows in reader}
pred = np.vectorize(med_dict.get)(pred)
return pred
def main():
# **************************** log ***********************
logger = init_logger(log_name=config['model']['arch'], log_dir=config['output']['log_dir'])
logger.info(f"seed is {config['train']['seed']}")
device = 'cuda:%d' % 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()}
#**************************** data input ***********************
DT = DataTransformer(logger = logger,seed = config['train']['seed'])
# read test data
targets, sentences = DT.read_data(raw_data_path=config['data']['test_file_path'],
preprocessor=EnglishPreProcessor(),
is_train=False)
tokenizer = BertTokenizer(vocab_file=config['pretrained']['bert']['vocab_path'],
do_lower_case=config['train']['do_lower_case'])
# prepare test dataset
test_dataset = CreateDataset(data = list(zip(sentences,targets)),
tokenizer = tokenizer,
max_seq_len = config['train']['max_seq_len'],
seed = config['train']['seed'],
example_type = 'test')
# pytorch dataloader
test_loader = DataLoader(dataset = test_dataset,
batch_size = config['train']['batch_size'],
num_workers = config['train']['num_workers'],
shuffle = False,
drop_last = False,
pin_memory = False)
# **************************** start 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))
# **************************** training model ***********************
logger.info('model predicting....')
'''predicter = Predicter(model = model,
logger = logger,
n_gpu=config['train']['n_gpu'],
model_path = config['output']['checkpoint_dir'] / f"best_{config['model']['arch']}_model.pth",
)'''
predicter = Predicter(model = model,
logger = logger,
n_gpu=config['train']['n_gpu'],
model_path = config['output']['checkpoint_dir'] / f"best_{config['model']['arch']}_model.pth"
)
# predict results
result = predicter.predict(data = test_loader)
result = np.where(result > 0.5, 1, 0)
print('accuracy score', accuracy_score(targets, result))
print('\nF1 score', f1_score(targets, result))
print('\nclassification report', classification_report(targets, result))
# empty cache after testing
if len(config['train']['n_gpu']) > 0:
torch.cuda.empty_cache()
def main():
logger = init_logger(log_name=config['model']['arch'],
log_dir=config['output']['log_dir'])
logger.info(f"seed is {config['train']['seed']}")
device = 'cuda:%d' % 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, ids = DT.read_data(
raw_data_path=config['data']['test_file_path'],
preprocessor=EnglishPreProcessor(),
is_train=False)
tokenizer = BertTokenizer(
vocab_file=config['pretrained']['bert']['vocab_path'],
do_lower_case=config['train']['do_lower_case'])
# test dataset
test_dataset = CreateDataset(data=list(zip(sentences, targets)),
tokenizer=tokenizer,
max_seq_len=config['train']['max_seq_len'],
seed=config['train']['seed'],
example_type='test')
test_loader = DataLoader(dataset=test_dataset,
batch_size=config['train']['batch_size'],
num_workers=config['train']['num_workers'],
shuffle=False,
drop_last=False,
pin_memory=False)
# **************************** load pretrained model from cache ***********************
logger.info("initializing model")
model = BertFine.from_pretrained(
config['pretrained']['bert']['bert_model_dir'],
cache_dir=config['output']['cache_dir'],
num_classes=len(id2label))
# **************************** inference ***********************
logger.info('model predicting....')
predicter = Predicter(
model=model,
logger=logger,
n_gpu=config['train']['n_gpu'],
model_path=config['output']['checkpoint_dir'] /
f"best_{config['model']['arch']}_model.pth",
)
# predict
result = predicter.predict(data=test_loader)
file = open(config['output']['inference_output_dir'], 'w')
for index, line, score in zip(ids, sentences, result):
file.write(str(index) + '\t' + line + '\t' + str(score[0]))
file.write('\n')
file.close()
if len(config['train']['n_gpu']) > 0:
torch.cuda.empty_cache()
def main():
# **************************** SETUP/READ FROM CONFIG ***********************
logger = init_logger(log_name=config['model']['arch'],
log_dir=config['output']['log_dir'])
logger.info(f"seed is {config['train']['seed']}")
device = 'cuda:%d' % 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'])
# Preprocessing
targets, sentences = DT.read_data(
raw_data_path=config['data']['test_file_path'],
preprocessor=EnglishPreProcessor(),
is_train=False)
tokenizer = BertTokenizer(
vocab_file=config['pretrained']['bert']['vocab_path'],
do_lower_case=config['train']['do_lower_case'])
#**************************** TOKENIZING *********************************
test_dataset = CreateDataset(data=list(zip(sentences, targets)),
tokenizer=tokenizer,
max_seq_len=config['train']['max_seq_len'],
seed=config['train']['seed'],
example_type='test')
#*************************** DATALOADER ******************************
test_loader = DataLoader(dataset=test_dataset,
batch_size=config['train']['batch_size'],
num_workers=config['train']['num_workers'],
shuffle=False,
drop_last=False,
pin_memory=False)
# **************************** LOAD 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))
# **************************** RUNNING PREDICTIONS ***********************
logger.info('model predicting....')
predicter = Predicter(
model=model,
logger=logger,
n_gpu=config['train']['n_gpu'],
model_path=config['output']['checkpoint_dir'] /
f"best_{config['model']['arch']}_model.pth",
)
# *************************OUTPUT RESULTS TO CSV*************************
result = predicter.predict(data=test_loader)
print(result)
df = pd.DataFrame(result)
cols = [
'toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate'
]
df.columns = cols
print(df.head())
df.to_csv('pybert/output/result/result.csv')
# ******************************EMPTY GPU CACHE************************************
if len(config['train']['n_gpu']) > 0:
torch.cuda.empty_cache()
html.unescape(abstract)).lstrip(';\n\n')
description = line[3]
description = html2text.html2text(
html.unescape(description)).lstrip(';\n\n')
input = title + ';' + abstract + ';' + description
labels = line[4:]
onehot_label = np.zeros(645)
for label in labels:
for i, subclass in enumerate(subclass_list):
if subclass == label:
onehot_label[i] = 1
break
if preprocessor:
input = preprocessor(input)
if is_train:
train_sentences.append(input)
train_targets.append(onehot_label)
else:
val_sentences.append(input)
val_targets.append(onehot_label)
return train_targets, train_sentences, val_targets, val_sentences
if __name__ == "__main__":
data = TaskData()
train_targets, train_sentences, val_targets, val_sentences = data.read_data(
config,
raw_data_path=
"/Users/xiaohan/Desktop/bert_HMC/data/summary/summary_1574.csv",
preprocessor=EnglishPreProcessor())