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 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("--mode", default='min', type=str)
parser.add_argument("--monitor", default='valid_loss', type=str)
parser.add_argument("--epochs", default=20, type=int)
parser.add_argument("--resume_path", default='', type=str)
parser.add_argument("--predict_checkpoints", type=int, default=0)
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=float)
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()
init_logger(
log_file=config['log_dir'] /
f'{args.arch}-{time.strftime("%Y-%m-%d-%H:%M:%S", time.localtime())}.log'
)
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)
logger.info("Training/evaluation parameters %s", args)
args.save_best = False
args.do_train = True
args.resume_path = 'pybert/output/checkpoints/bert/checkpoint-epoch-3'
args.do_lower_case = True
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():
parser = ArgumentParser()
parser.add_argument("--do_data", default=False, action='store_true')
parser.add_argument("--do_corpus", default=False, action='store_true')
parser.add_argument("--do_vocab", default=False, action='store_true')
parser.add_argument("--do_split", default=False, action='store_true')
parser.add_argument('--seed',default=42,type=int)
parser.add_argument("--line_per_file", default=1000000000, type=int)
parser.add_argument("--file_num", type=int, default=10,
help="Number of dynamic masking to pregenerate")
parser.add_argument("--max_seq_len", type=int, default=128)
parser.add_argument("--short_seq_prob", type=float, default=0.1,
help="Probability of making a short sentence as a training example")
parser.add_argument("--masked_lm_prob", type=float, default=0.15,
help="Probability of masking each token for the LM task")
parser.add_argument("--max_predictions_per_seq", type=int, default=20,
help="Maximum number of tokens to mask in each sequence")
args = parser.parse_args()
seed_everything(args.seed)
if args.do_corpus:
corpus = []
train_path = str(config['data_dir'] / 'train.txt')
with open(train_path, 'r') as fr:
for ex_id, line in enumerate(fr):
line = line.strip("\n")
lines = [" ".join(x.split("/")[0].split("_")) for x in line.split(" ")]
if ex_id == 0:
logger.info(f"Train example: {' '.join(lines)}")
corpus.append(" ".join(lines))
test_path = str(config['data_dir'] / 'test.txt')
with open(test_path, 'r') as fr:
for ex_id, line in enumerate(fr):
line = line.strip("\n")
lines = line.split("_")
if ex_id == 0:
logger.info(f"Test example: {' '.join(lines)}")
corpus.append(" ".join(lines))
corpus_path = str(config['data_dir'] / 'corpus.txt')
with open(corpus_path, 'r') as fr:
for ex_id, line in enumerate(fr):
line = line.strip("\n")
lines = line.split("_")
if ex_id == 0:
logger.info(f"Corpus example: {' '.join(lines)}")
corpus.append(" ".join(lines))
corpus = list(set(corpus))
logger.info(f"corpus size: {len(corpus)}")
random_order = list(range(len(corpus)))
np.random.shuffle(random_order)
corpus = [corpus[i] for i in random_order]
new_corpus_path = config['data_dir'] / "corpus/corpus.txt"
if not new_corpus_path.exists():
new_corpus_path.parent.mkdir(exist_ok = True)
with open(new_corpus_path, 'w') as fr:
for line in corpus:
fr.write(line + "\n")
if args.do_split:
new_corpus_path = config['data_dir'] / "corpus/corpus.txt"
split_save_path = config['data_dir'] / "corpus/train"
if not split_save_path.exists():
split_save_path.mkdir(exist_ok=True)
line_per_file = args.line_per_file
command = f'split -a 4 -l {line_per_file} -d {new_corpus_path} {split_save_path}/shard_'
os.system(f"{command}")
if args.do_vocab:
vocab = Vocabulary(min_freq=0,add_unused=True)
vocab.read_data(data_path=config['data_dir'] / "corpus/train")
vocab.build_vocab()
vocab.save(file_path=config['data_dir'] / 'corpus/vocab_mapping.pkl')
vocab.save_bert_vocab(file_path=config['bert_vocab_path'])
logger.info(f"vocab size: {len(vocab)}")
bert_base_config['vocab_size'] = len(vocab)
save_json(data=bert_base_config, file_path=config['bert_config_file'])
if args.do_data:
vocab_list = load_vocab(config['bert_vocab_path'])
data_path = config['data_dir'] / "corpus/train"
files = sorted([f for f in data_path.iterdir() if f.exists() and "." not in str(f)])
print(files)
logger.info("--- pregenerate training data parameters ---")
logger.info(f'max_seq_len: {args.max_seq_len}')
logger.info(f"max_predictions_per_seq: {args.max_predictions_per_seq}")
logger.info(f"masked_lm_prob: {args.masked_lm_prob}")
logger.info(f"seed: {args.seed}")
logger.info(f"file num : {args.file_num}")
for idx in range(args.file_num):
logger.info(f"pregenetate file_{idx}.json")
save_filename = data_path / f"file_{idx}.json"
num_instances = 0
with save_filename.open('w') as fw:
for file_idx in range(len(files)):
file_path = files[file_idx]
file_examples = build_examples(file_path,max_seq_len=args.max_seq_len,
masked_lm_prob=args.masked_lm_prob,
max_predictions_per_seq=args.max_predictions_per_seq,
vocab_list = vocab_list)
file_examples = [json.dumps(instance) for instance in file_examples]
for instance in file_examples:
fw.write(instance + '\n')
num_instances += 1
metrics_file = data_path / f"file_{idx}_metrics.json"
print(f"num_instances: {num_instances}")
with metrics_file.open('w') as metrics_file:
metrics = {
"num_training_examples": num_instances,
"max_seq_len": args.max_seq_len
}
metrics_file.write(json.dumps(metrics))
def main():
parser = ArgumentParser()
parser.add_argument("--arch", default='bert', type=str)
parser.add_argument("--do_data", action='store_true')
parser.add_argument("--train", action='store_true')
parser.add_argument("--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='job_dataset', type=str)
parser.add_argument("--epochs", default=10, type=int)
parser.add_argument("--resume_path", default='', type=str)
parser.add_argument("--test_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=0.05, 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=4, type=int)
parser.add_argument('--eval_batch_size', default=4, 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=1.0e-4, 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')
parser.add_argument('--predict_labels', type=bool, default=False)
parser.add_argument('--predict_idx',
type=str,
default="0",
help=' "idx" or "start-end" or "all" ')
args = parser.parse_args()
config['checkpoint_dir'] = config['checkpoint_dir'] / args.arch
config['checkpoint_dir'].mkdir(exist_ok=True)
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=False,
stratify=False,
valid_size=args.valid_size,
data_dir=config['data_dir'],
data_name=args.data_name)
if args.train:
run_train(args)
if args.test:
run_test(args)
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')
parser.add_argument("--prob_thresh", default=0.5, type=float)
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_label import TaskData
data = TaskData()
print("Train data path:")
print(config['raw_data_path'])
targets, sentences_char = data.read_data(
raw_data_path=config['raw_data_path'],
preprocessor=EnglishPreProcessor(),
is_train=True)
print("Target:")
print(targets)
print(" ")
print("Sentence:")
print(sentences_char)
print(" ")
data.train_val_split(X=sentences_char,
y=targets,
valid_size=args.valid_size,
data_dir=config['data_dir'],
data_name=args.data_name)
##Get the test data
targets_test, sentences_char_test = data.read_data(
raw_data_path=config['test_path'],
preprocessor=EnglishPreProcessor(),
is_train=True)
print(targets_test)
data.save_test_data(X=sentences_char_test,
y=targets_test,
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():
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='HPC', type=str)
parser.add_argument("--mode", default='min', type=str)
parser.add_argument("--monitor", default='valid_loss', type=str)
parser.add_argument("--epochs", default=10, type=int)
parser.add_argument("--resume_path", default='', type=str)
parser.add_argument("--predict_checkpoints", type=int, default=0)
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=float)
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()
init_logger(log_file=config['log_dir'] / f'{args.arch}-{time.strftime("%Y-%m-%d-%H:%M:%S", time.localtime())}.log')
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)
logger.info("Training/evaluation parameters %s", args)
if args.do_data:
data_names = []
train_sentenses_all = []
train_target_all = []
from pybert.io.task_data import TaskData
data = TaskData()
total_valid = 0
for filename in os.listdir(config['summary_path']):
if filename == ".DS_Store" or filename == "summary":
continue
filename_int = int(filename.split('.')[0].split('_')[-1])
if filename_int > 3500:
try:
raw_data_path = os.path.join(config['summary_path'], filename)
# train_targets, train_sentences, val_targets, val_sentences = data.read_data(config,
# raw_data_path=raw_data_path,
# preprocessor=EnglishPreProcessor())
train_targets, train_sentences, val_targets, val_sentences = data.read_data(config,
raw_data_path=raw_data_path)
train_sentenses_all = train_sentenses_all + train_sentences
train_target_all = train_target_all + train_targets
total_valid = len(train_target_all)
print("valid number: ", total_valid)
# data.save_pickle(train_sentences, train_targets, data_dir=config['data_dir'],
# data_name=filename.split('.')[0].split('_')[-1], is_train=True)
# data.save_pickle(val_sentences, val_targets, data_dir=config['data_dir'],
# data_name=filename.split('.')[0].split('_')[-1], is_train=False)
# data_names.append(filename.split('.')[0].split('_')[-1])
except:
pass
total_valid = len(train_target_all)
print("valid number: ", total_valid)
data.save_pickle(train_sentenses_all, train_target_all, data_dir=config['data_dir'],
data_name="all_valid", is_train=False)
# with open(config['data_name'], 'w') as f:
# json.dump(data_names, f)
with open(config['data_name'], 'r') as f:
data_names = json.load(f)
if args.do_train:
run_train(args, data_names)
if args.do_test:
run_test(args)
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 main():
parser = ArgumentParser()
parser.add_argument("--arch", default='bert', type=str) # 使用的预训练语言模型
parser.add_argument("--do_data", action='store_true') # 进行数据切分
parser.add_argument("--do_train", action='store_true') # 进行模型训练
parser.add_argument("--do_test", action='store_true') # 进行模型推断
parser.add_argument("--save_best", action='store_true') # 保留最好的模型
parser.add_argument("--do_lower_case", action='store_true')
parser.add_argument('--data_name', default='ccks', type=str) # 数据集的名字
parser.add_argument("--mode", default='min', type=str) # 设置monitor关注的角度
parser.add_argument("--monitor", default='valid_loss', type=str)
parser.add_argument("--task_type", default='base', type=str)
parser.add_argument("--epochs", default=4, type=int)
parser.add_argument("--resume_path", default='',
type=str) # 恢复路径，从pretrained model中载入模型
parser.add_argument("--predict_checkpoints", type=int, default=0)
parser.add_argument("--valid_size", default=0.2, type=float) # 验证集的大小
parser.add_argument("--local_rank", type=int, default=-1)
parser.add_argument("--sorted",
default=1,
type=int,
help='1 : True 0:False ') # 表示是否按照序列的长度排序
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) # gradient_accumulation_steps的大小，用于解决内存小，无法使用大batch_size的问题
parser.add_argument("--train_batch_size", default=8,
type=int) # 训练集batch_size
parser.add_argument('--eval_batch_size', default=8,
type=int) # 测试集batch_size
parser.add_argument("--train_max_seq_len", default=256,
type=int) # 训练集sequence的最大长度
parser.add_argument("--eval_max_seq_len", default=256,
type=int) # 测试集sequence的最大长度
parser.add_argument('--loss_scale', type=float,
default=0) # TODO: 理解loss scale的作用
parser.add_argument("--warmup_proportion", default=0.1,
type=float) # 用于learning rate上的warmup proportion
parser.add_argument("--weight_decay", default=0.01,
type=float) # TODO: 理解weight decay的含义
parser.add_argument("--adam_epsilon", default=1e-8,
type=float) # adam优化器的参数
parser.add_argument("--grad_clip", default=1.0,
type=float) # TODO: 理解grad clip的含义
parser.add_argument("--learning_rate", default=2e-5, type=float) # 学习率
parser.add_argument('--seed', type=int, default=42) # 随机数种子
parser.add_argument('--fp16', action='store_true') # TODO: 理解fp16是什么
parser.add_argument('--fp16_opt_level', type=str, default='O1')
args = parser.parse_args()
# 初始化日志记录器logger
config['log_dir'].mkdir(exist_ok=True) # 源代码没有写这句代码
init_logger(
log_file=config['log_dir'] /
f'{args.arch}-{time.strftime("%Y-%m-%d-%H:%M:%S", time.localtime())}.log'
)
config['checkpoint_dir'] = config[
'checkpoint_dir'] / args.arch / args.task_type # 重新调整输出的位置
config['checkpoint_dir'].mkdir(exist_ok=True)
BASE_DIR = Path('pybert')
config[
'raw_data_path'] = BASE_DIR / f'dataset/train_{args.task_type}_sample.csv'
config['test_path'] = BASE_DIR / f'dataset/test_{args.task_type}.csv'
config['figure_dir'] = config['figure_dir'] / f'{args.task_type}'
config['figure_dir'].mkdir(exist_ok=True)
# 动态修改文件路径
# BASE_DIR = Path('pybert')
# if args.task_type == 'trans':
# config['raw_data_path'] = BASE_DIR / 'dataset/train_trans_sample.csv'
# config['test_path'] = BASE_DIR / 'dataset/test_trans.csv'
# config['figure_dir'] = config['figure_dir'] / f'{args.task_type}'
# config['figure_dir'].mkdir(exist_ok=True)
# elif args.task_type == 'base':
# config['raw_data_path'] = BASE_DIR / 'dataset/train_base_sample.csv'
# config['test_path'] = BASE_DIR / 'dataset/test_base.csv'
# config['figure_dir'] = config['figure_dir'] / f'{args.task_type}'
# config['figure_dir'].mkdir(exist_ok=True)
# else:
# raise ValueError(f"Invalid task_type {args.task_type}")
# Good practice: save your training arguments together with the trained model
torch.save(args, config['checkpoint_dir'] / 'training_args.bin')
seed_everything(args.seed) # 一个方法设置所有的seed
logger.info("Training/evaluation parameters %s", args)
if args.do_data:
from pybert.io.task_data import TaskData
data = TaskData()
ids, targets, sentences = data.read_data(
raw_data_path=config['raw_data_path'],
preprocessor=ChinesePreProcessor(),
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,
task_type=args.task_type) # 增加了task_type参数
if args.do_train:
run_train(args)
if args.do_test:
run_test(args)
def main():
parser = ArgumentParser()
parser.add_argument("--file_num", type=int, default=10,
help="Number of pregenerate file")
parser.add_argument("--reduce_memory", action="store_true",
help="Store training data as on-disc memmaps to massively reduce memory usage")
parser.add_argument("--epochs", type=int, default=2,
help="Number of epochs to train for")
parser.add_argument('--num_eval_steps', default=200)
parser.add_argument('--num_save_steps', default=5000)
parser.add_argument("--local_rank", type=int, default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument("--no_cuda", action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.")
parser.add_argument("--train_batch_size", default=24, type=int,
help="Total batch size for training.")
parser.add_argument('--fp16', action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--loss_scale', type=float, default=0,
help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument("--warmup_proportion",default=0.1,type=float,
help="Linear warmup over warmup_steps.")
parser.add_argument("--adam_epsilon", default=1e-8, type=float,
help="Epsilon for Adam optimizer.")
parser.add_argument("--learning_rate", default=1e-4, type=float,
help="The initial learning rate for Adam.")
parser.add_argument('--seed', type=int, default=42,
help="random seed for initialization")
parser.add_argument('--fp16_opt_level', type=str, default='O1',
help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html")
args = parser.parse_args()
pregenerated_data = config['data_dir'] / "corpus/train"
assert pregenerated_data.is_dir(), \
"--pregenerated_data should point to the folder of files made by prepare_lm_data_mask.py!"
samples_per_epoch = 0
for i in range(args.file_num):
data_file = pregenerated_data / f"file_{i}.json"
metrics_file = pregenerated_data / f"file_{i}_metrics.json"
if data_file.is_file() and metrics_file.is_file():
metrics = json.loads(metrics_file.read_text())
samples_per_epoch += metrics['num_training_examples']
else:
if i == 0:
exit("No training data was found!")
print(f"Warning! There are fewer epochs of pregenerated data ({i}) than training epochs ({args.epochs}).")
print("This script will loop over the available data, but training diversity may be negatively impacted.")
break
logger.info(f"samples_per_epoch: {samples_per_epoch}")
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info( f"device: {device} n_gpu: {n_gpu}, distributed training: {bool(args.local_rank != -1)}, 16-bits training: {args.fp16}")
if args.gradient_accumulation_steps < 1:
raise ValueError(f"Invalid gradient_accumulation_steps parameter: {args.gradient_accumulation_steps}, should be >= 1")
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
seed_everything(args.seed)
tokenizer = CustomTokenizer(vocab_file=config['bert_vocab_path'])
total_train_examples = samples_per_epoch * args.epochs
num_train_optimization_steps = int(
total_train_examples / args.train_batch_size / args.gradient_accumulation_steps)
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
args.warmup_steps = int(num_train_optimization_steps * args.warmup_proportion)
# Prepare model
with open(str(config['bert_config_file']), "r", encoding='utf-8') as reader:
json_config = json.loads(reader.read())
print(json_config)
bert_config = BertConfig.from_json_file(str(config['bert_config_file']))
model = BertForMaskedLM(config=bert_config)
# model = BertForMaskedLM.from_pretrained(config['checkpoint_dir'] / 'checkpoint-580000')
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare 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}
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=num_train_optimization_steps)
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)
# if args.fp16:
# try:
# from apex.optimizers import FP16_Optimizer
# from apex.optimizers import FusedAdam
# except ImportError:
# raise ImportError(
# "Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
#
# optimizer = FusedAdam(optimizer_grouped_parameters,
# lr=args.learning_rate,
# bias_correction=False,
# max_grad_norm=1.0)
# if args.loss_scale == 0:
# optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
# else:
# optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
# else:
# optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
# scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=num_train_optimization_steps)
global_step = 0
metric = LMAccuracy()
tr_acc = AverageMeter()
tr_loss = AverageMeter()
train_logs = {}
logger.info("***** Running training *****")
logger.info(f" Num examples = {total_train_examples}")
logger.info(f" Batch size = {args.train_batch_size}" )
logger.info(f" Num steps = {num_train_optimization_steps}" )
logger.info(f" warmup_steps = {args.warmup_steps}")
model.train()
for epoch in range(args.epochs):
for idx in range(args.file_num):
epoch_dataset = PregeneratedDataset(file_id=idx, training_path=pregenerated_data, tokenizer=tokenizer,
reduce_memory=args.reduce_memory)
if args.local_rank == -1:
train_sampler = RandomSampler(epoch_dataset)
else:
train_sampler = DistributedSampler(epoch_dataset)
train_dataloader = DataLoader(epoch_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(train_dataloader):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, lm_label_ids = batch
outputs = model(input_ids, segment_ids, input_mask, lm_label_ids)
pred_output = outputs[1]
loss = outputs[0]
metric(logits=pred_output.view(-1, bert_config.vocab_size), target=lm_label_ids.view(-1))
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
nb_tr_steps += 1
tr_acc.update(metric.value(), n=input_ids.size(0))
tr_loss.update(loss.item(), n=1)
if (step + 1) % args.gradient_accumulation_steps == 0:
# if args.fp16:
# # modify learning rate with special warm up BERT uses
# # if args.fp16 is False, BertAdam is used that handles this automatically
# lr_this_step = args.learning_rate * warmup_linear.get_lr(global_step, args.warmup_proportion)
# for param_group in optimizer.param_groups:
# param_group['lr'] = lr_this_step
scheduler.step() # Update learning rate schedule
optimizer.step()
optimizer.zero_grad()
global_step += 1
if global_step % args.num_eval_steps == 0:
train_logs['loss'] = tr_loss.avg
train_logs['acc'] = tr_acc.avg
show_info = f'\n[Training]:[{epoch}/{args.epochs}]{global_step}/{num_train_optimization_steps} ' + "-".join(
[f' {key}: {value:.4f} ' for key, value in train_logs.items()])
logger.info(show_info)
tr_acc.reset()
tr_loss.reset()
if global_step % args.num_save_steps == 0:
if args.local_rank in [-1, 0] and args.num_save_steps > 0:
# Save model checkpoint
output_dir = config['checkpoint_dir'] / f'checkpoint-{global_step}'
if not output_dir.exists():
output_dir.mkdir()
# save model
model_to_save = model.module if hasattr(model,'module') else model # Take care of distributed/parallel training
model_to_save.save_pretrained(str(output_dir))
torch.save(args, str(output_dir / 'training_args.bin'))
logger.info("Saving model checkpoint to %s", output_dir)
torch.save(args, str(output_dir / 'training_args.bin'))
#save config
output_config_file = output_dir / CONFIG_NAME
with open(str(output_config_file), 'w') as f:
f.write(model_to_save.config.to_json_string())
#save vocab
tokenizer.save_vocabulary(output_dir)
