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 run_test(args):
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=False)
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()
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,
collate_fn=collate_fn)
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)
ids = np.array(ids)
df1 = pd.DataFrame(ids, index=None)
df2 = pd.DataFrame(result, index=None)
all_df = pd.concat([df1, df2], axis=1)
all_df.columns = ['id', 'sg', 'pj']
all_df['sg'] = all_df['sg'].apply(lambda x: 1 if x > 0.5 else 0)
all_df['pj'] = all_df['pj'].apply(lambda x: 1 if x > 0.5 else 0)
#all_df['qs'] = all_df['qs'].apply(lambda x: 1 if x>0.5 else 0)
#all_df['tz'] = all_df['tz'].apply(lambda x: 1 if x>0.5 else 0)
#all_df['ggjc'] = all_df['ggjc'].apply(lambda x: 1 if x>0.5 else 0)
#all_df.columns = ['id','zy','gfgqzr','qs','tz','ggjc']
#all_df['zy'] = all_df['zy'].apply(lambda x: 1 if x>0.5 else 0)
#all_df['gfgqzr'] = all_df['gfgqzr'].apply(lambda x: 1 if x>0.5 else 0)
#all_df['qs'] = all_df['qs'].apply(lambda x: 1 if x>0.5 else 0)
#all_df['tz'] = all_df['tz'].apply(lambda x: 1 if x>0.5 else 0)
#all_df['ggjc'] = all_df['ggjc'].apply(lambda x: 1 if x>0.5 else 0)
all_df.to_csv(
"/home/LAB/liqian/test/game/Fin/CCKS-Cls/test_output/cls_out.csv",
index=False)
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='train', type=str)
parser.add_argument("--epochs", default=4, type=int)
parser.add_argument("--resume_path", default='', type=str)
parser.add_argument("--mode", default='max', type=str)
parser.add_argument("--monitor", default='valid_f1', 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
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)}
data = TaskData()
targets, sentences = data.read_data(raw_data_path=config['raw_data_path'],
preprocessor=None, is_train=True,label2id=label2id)
data.train_val_split(X=sentences, y=targets, shuffle=True, stratify=targets,
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 run_test(args):
from pybert.io.task_data import TaskData
from pybert.test.predictor import Predictor
import pickle
import os
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)}
test_data = processor.get_train(config['data_dir'] / f"{args.data_name}.test.pkl")
print ("Test data is:")
print (test_data)
print ("Label list is:")
print (label_list)
print ("----------------------------------------")
# test_data = processor.get_test(lines=lines)
test_examples = processor.create_examples(lines=test_data,
example_type='test',
cached_examples_file=config[
'data_cache'] / 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_cache'] / "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,
batch_metrics=[AccuracyThresh(thresh=0.5)],
epoch_metrics=[AUC(average='micro', task_type='binary'),
MultiLabelReport(id2label=id2label)])
result, test_predicted, test_true = predictor.predict(data=test_dataloader)
pickle.dump(test_true, open(os.path.join(config["test/checkpoint_dir"], "test_true.p"), "wb"))
pickle.dump(test_predicted, open(os.path.join(config["test/checkpoint_dir"], "test_predicted.p"), "wb"))
pickle.dump(id2label, open(os.path.join(config["test/checkpoint_dir"], "id2label.p"), "wb"))
print ("Predictor results:")
print(result)
print ("-----------------------------------------------")
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)
_, _, targets, sentences = data.read_data(config, raw_data_path=config['test_path'],
is_train=False)
lines = list(zip(sentences, targets))
# processor = BertProcessor(vocab_path=config['bert_vocab_path'], do_lower_case=args.do_lower_case)
processor = BertProcessor()
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,
collate_fn=collate_fn)
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)
result[result<0.5] = 0
result[result>=0.5] = 1
labels = []
for i in range(result.shape[0]):
ids = np.where(result[i]==1)[0]
each_patent_label = [id2label[id] for id in ids]
labels.append(each_patent_label)
if os.path.exists(config['predictions']):
os.remove(config['predictions'])
with open(config['test_path'], 'r') as f:
reader = csv.reader(f)
for j, line in enumerate(reader):
id = line[0]
with open(config['predictions'], 'a+') as g:
g.write("{}\t".format(id))
for label in labels[j]:
g.write("{}\t".format(label))
g.write("\n")
def run_test(args):
from pybert.test.predictor import Predictor
processor = BertProcessor(vocab_path=config['bert_vocab_path'],
do_lower_case=args.do_lower_case)
test_data = processor.get_test(config['test_path'])
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.eval_batch_size)
idx2word = {}
for (w, i) in processor.tokenizer.vocab.items():
idx2word[i] = w
label_list = processor.get_labels(label_path=config['data_label_path'])
idx2label = {i: label for i, label in enumerate(label_list)}
if args.test_path:
args.test_path = Path(args.test_path)
model = BertForMultiLable.from_pretrained(args.test_path,
num_labels=len(label_list))
else:
model = BertForMultiLable.from_pretrained(config['bert_model_dir'],
num_labels=len(label_list))
for p in model.bert.parameters():
p.require_grad = False
# ----------- predicting -----------
writer = SummaryWriter()
logger.info('model predicting....')
predictor = Predictor(model=model,
logger=logger,
n_gpu=args.n_gpu,
i2w=idx2word,
i2l=idx2label)
result = predictor.predict(data=test_dataloader)
if args.predict_labels:
predictor.labels(result, args.predict_idx)
def main(text,arch,max_seq_length,do_lower_case):
processor = BertProcessor(vocab_path=config['bert_vocab_path'], do_lower_case=do_lower_case)
label_list = processor.get_labels()
id2label = {i: label for i, label in enumerate(label_list)}
model = BertForMultiLable.from_pretrained(config['checkpoint_dir'] /f'{arch}', num_labels=len(label_list))
tokens = processor.tokenizer.tokenize(text)
if len(tokens) > max_seq_length - 2:
tokens = tokens[:max_seq_length - 2]
tokens = ['[CLS]'] + tokens + ['[SEP]']
input_ids = processor.tokenizer.convert_tokens_to_ids(tokens)
input_ids = torch.tensor(input_ids).unsqueeze(0) # Batch size 1, 2 choices
logits = model(input_ids)
probs = logits.sigmoid()
return probs.cpu().detach().numpy()[0]
def run_test(args):
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=None,
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 = BertForMultiClass.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)
import numpy as np
result=np.argmax(result,axis=1)
with open('submit1.csv','w',encoding='utf-8') as f:
for id,pre in zip(ids,result):
f.write(id+','+str(pre)+'\n')
print(result)
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 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 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)
