def main():
parser = ArgumentParser()
## Required parameters
parser.add_argument("--do_data", default=True, action='store_true')
parser.add_argument('--data_name', default='albert', type=str)
parser.add_argument('--max_ngram', default=3, type=int)
parser.add_argument("--do_lower_case", default=False, action='store_true')
parser.add_argument('--seed', default=42, type=int)
# parser.add_argument("--file_num", type=int, default=10, help="Number of dynamic masking to pregenerate (with different masks)")
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")
# 128 * 0.15
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)
from configs.base import config
args.vocab_path = config['albert_vocab_path']
args.data_dir = config['data_dir']
logger.info("pregenerate training data parameters:\n %s", args)
tokenizer = BertTokenizer(vocab_file=args.vocab_path,
do_lower_case=args.do_lower_case)
small_path = config['data_dir'] / "corpus/small"
files = sorted(
[f for f in small_path.iterdir() if f.exists() and '.txt' in str(f)])
file_path = files[0].absolute()
max_seq_len = args.max_seq_len
train(file_path, tokenizer, max_seq_len)
print(" dataloader ok! ")
sys.exit(0)
def main():
parser = ArgumentParser()
## Required parameters
parser.add_argument("--data_dir", default=None, type=str, required=True)
parser.add_argument("--vocab_path", default=None, type=str, required=True)
parser.add_argument("--output_dir", default=None, type=str, required=True)
parser.add_argument('--data_name', default='albert', type=str)
parser.add_argument('--max_ngram', default=3, type=int)
parser.add_argument("--do_data", default=False, action='store_true')
parser.add_argument("--do_split", default=False, action='store_true')
parser.add_argument("--do_lower_case", 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 (with different masks)")
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, # 128 * 0.15
help="Maximum number of tokens to mask in each sequence")
args = parser.parse_args()
seed_everything(args.seed)
args.data_dir = Path(args.data_dir)
if not os.path.exists(args.output_dir):
os.mkdir(args.output_dir)
init_logger(log_file=args.output_dir +"pregenerate_training_data_ngram.log")
logger.info("pregenerate training data parameters:\n %s", args)
tokenizer = BertTokenizer(vocab_file=args.vocab_path, do_lower_case=args.do_lower_case)
# split big file
if args.do_split:
corpus_path =args.data_dir / "corpus/corpus.txt"
split_save_path = args.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 {corpus_path} {split_save_path}/shard_'
os.system(f"{command}")
# generator train data
if args.do_data:
data_path = args.data_dir / "corpus/train"
files = sorted([f for f in data_path.parent.iterdir() if f.exists() and '.txt' in str(f)])
for idx in range(args.file_num):
logger.info(f"pregenetate {args.data_name}_file_{idx}.json")
save_filename = data_path / f"{args.data_name}_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 = create_training_instances(input_file=file_path,
tokenizer=tokenizer,
max_seq_len=args.max_seq_len,
max_ngram=args.max_ngram,
short_seq_prob=args.short_seq_prob,
masked_lm_prob=args.masked_lm_prob,
max_predictions_per_seq=args.max_predictions_per_seq)
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"{args.data_name}_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('--data_name', default='albert', type=str)
parser.add_argument(
"--file_num",
type=int,
default=10,
help="Number of dynamic masking to pregenerate (with different masks)")
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=4,
help="Number of epochs to train for")
parser.add_argument('--share_type',
default='all',
type=str,
choices=['all', 'attention', 'ffn', 'None'])
parser.add_argument('--num_eval_steps', default=100)
parser.add_argument('--num_save_steps', default=200)
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=4,
type=int,
help="Total batch size for training.")
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('--max_grad_norm', default=1.0, type=float)
parser.add_argument("--learning_rate",
default=0.00176,
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='O2',
help=
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html")
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
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"{args.data_name}_file_{i}.json"
metrics_file = pregenerated_data / f"{args.data_name}_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(f"cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
args.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} , distributed training: {bool(args.local_rank != -1)}, 16-bits training: {args.fp16}, "
f"share_type: {args.share_type}")
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 = BertTokenizer(vocab_file=config['albert_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)
bert_config = BertConfig.from_pretrained(str(config['albert_config_path']),
share_type=args.share_type)
model = BertForPreTraining(config=bert_config)
# model = BertForMaskedLM.from_pretrained(config['checkpoint_dir'] / 'checkpoint-580000')
model.to(device)
# 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)
# optimizer = Lamb(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
lr_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.n_gpu > 1:
model = torch.nn.DataParallel(model)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank)
global_step = 0
mask_metric = LMAccuracy()
sop_metric = LMAccuracy()
tr_mask_acc = AverageMeter()
tr_sop_acc = AverageMeter()
tr_loss = AverageMeter()
tr_mask_loss = AverageMeter()
tr_sop_loss = AverageMeter()
loss_fct = CrossEntropyLoss(ignore_index=-1)
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}")
start_time = time.time()
seed_everything(args.seed) # Added here for reproducibility
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,
data_name=args.data_name)
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)
model.train()
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, is_next = batch
outputs = model(input_ids=input_ids,
token_type_ids=segment_ids,
attention_mask=input_mask)
prediction_scores = outputs[0]
seq_relationship_score = outputs[1]
masked_lm_loss = loss_fct(
prediction_scores.view(-1, bert_config.vocab_size),
lm_label_ids.view(-1))
next_sentence_loss = loss_fct(
seq_relationship_score.view(-1, 2), is_next.view(-1))
loss = masked_lm_loss + next_sentence_loss
mask_metric(logits=prediction_scores.view(
-1, bert_config.vocab_size),
target=lm_label_ids.view(-1))
sop_metric(logits=seq_relationship_score.view(-1, 2),
target=is_next.view(-1))
if args.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_mask_acc.update(mask_metric.value(), n=input_ids.size(0))
tr_sop_acc.update(sop_metric.value(), n=input_ids.size(0))
tr_loss.update(loss.item(), n=1)
tr_mask_loss.update(masked_lm_loss.item(), n=1)
tr_sop_loss.update(next_sentence_loss.item(), n=1)
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
lr_scheduler.step()
optimizer.step()
optimizer.zero_grad()
global_step += 1
if global_step % args.num_eval_steps == 0:
now = time.time()
eta = now - start_time
if eta > 3600:
eta_format = ('%d:%02d:%02d' %
(eta // 3600,
(eta % 3600) // 60, eta % 60))
elif eta > 60:
eta_format = '%d:%02d' % (eta // 60, eta % 60)
else:
eta_format = '%ds' % eta
train_logs['loss'] = tr_loss.avg
train_logs['mask_acc'] = tr_mask_acc.avg
train_logs['sop_acc'] = tr_sop_acc.avg
train_logs['mask_loss'] = tr_mask_loss.avg
train_logs['sop_loss'] = tr_sop_loss.avg
show_info = f'[Training]:[{epoch}/{args.epochs}]{global_step}/{num_train_optimization_steps} ' \
f'- ETA: {eta_format}' + "-".join(
[f' {key}: {value:.4f} ' for key, value in train_logs.items()])
logger.info(show_info)
tr_mask_acc.reset()
tr_sop_acc.reset()
tr_loss.reset()
tr_mask_loss.reset()
tr_sop_loss.reset()
start_time = now
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'lm-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)
# 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)
def __init__(self, vocab_path, do_lower_case):
self.tokenizer = BertTokenizer(vocab_path, do_lower_case)
def main():
parser = ArgumentParser()
parser.add_argument('--data_name', default='albert', type=str)
parser.add_argument("--do_data", default=False, action='store_true')
parser.add_argument("--do_split", default=False, action='store_true')
parser.add_argument("--do_lower_case", 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 (with different masks)")
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)
tokenizer = BertTokenizer(vocab_file=config['checkpoint_dir'] /
'vocab.txt',
do_lower_case=args.do_lower_case)
if args.do_split:
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 {corpus_path} {split_save_path}/shard_'
os.system(f"{command}")
if args.do_data:
data_path = config['data_dir'] / "corpus/train"
files = sorted([
f for f in config['data_dir'].iterdir()
if f.exists() and '.txt' in str(f)
])
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"mask file num : {args.file_num}")
logger.info(f"train file num : {len(files)}")
for idx in range(args.file_num):
logger.info(f"pregenetate file_{idx}.json")
save_filename = data_path / f"{args.data_name}_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 = create_training_instances(
input_file=file_path,
tokenizer=tokenizer,
max_seq_len=args.max_seq_len,
short_seq_prob=args.short_seq_prob,
masked_lm_prob=args.masked_lm_prob,
max_predictions_per_seq=args.max_predictions_per_seq)
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"{args.data_name}_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))
