def train(args, train_dataset, model, tokenizer):
""" Train the model """
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_sampler = SequentialSampler(
train_dataset) if args.local_rank == -1 else DistributedSampler(
train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size,
collate_fn=collate_fn)
if args.max_steps > 0:
num_training_steps = args.max_steps
args.num_train_epochs = args.max_steps // (
len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
num_training_steps = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
args.warmup_steps = int(num_training_steps * args.warmup_proportion)
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ['bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
'weight_decay':
args.weight_decay
}, {
'params': [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
'weight_decay':
0.0
}]
# optimizer = Lamb(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
optimizer = AdamW(params=optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=num_training_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)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d",
args.per_gpu_train_batch_size)
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", num_training_steps)
global_step = 0
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
seed_everything(
args.seed
) # Added here for reproductibility (even between python 2 and 3)
for _ in range(int(args.num_train_epochs)):
pbar = ProgressBar(n_total=len(train_dataloader), desc='Training')
for step, batch in enumerate(train_dataloader):
model.train()
batch = tuple(t.to(args.device) for t in batch)
inputs = {
'input_ids': batch[0],
'attention_mask': batch[1],
'labels': batch[3]
}
inputs['token_type_ids'] = batch[2]
outputs = model(**inputs)
loss = outputs[
0] # model outputs are always tuple in transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean(
) # mean() to average on multi-gpu parallel training
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()
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer),
args.max_grad_norm)
else:
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
scheduler.step() # Update learning rate schedule
model.zero_grad()
global_step += 1
if args.local_rank in [
-1, 0
] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
#Log metrics
if args.local_rank == -1: # Only evaluate when single GPU otherwise metrics may not average well
evaluate(args, model, tokenizer)
if args.local_rank in [
-1, 0
] and args.save_steps > 0 and global_step % args.save_steps == 0:
# Save model checkpoint
output_dir = os.path.join(args.output_dir,
'checkpoint-{}'.format(global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model_to_save = model.module if hasattr(
model, 'module'
) else model # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
torch.save(args, os.path.join(output_dir, 'training_args.bin'))
logger.info("Saving model checkpoint to %s", output_dir)
pbar(step, {'loss': loss.item()})
print(" ")
if 'cuda' in str(args.device):
torch.cuda.empty_cache()
return global_step, tr_loss / global_step
def main():
parser = ArgumentParser()
## Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument("--config_path", 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,
help=
"The output directory where the model predictions and checkpoints will be written."
)
parser.add_argument("--model_path", default='', type=str)
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(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
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("--weight_decay",
default=0.01,
type=float,
help="Weight deay if we apply some.")
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()
args.data_dir = Path(args.data_dir)
args.output_dir = Path(args.output_dir)
pregenerated_data = args.data_dir / "corpus/train"
init_logger(log_file=str(args.output_dir / "train_albert_model.log"))
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.from_pretrained(args.vocab_path,
do_lower_case=args.do_lower_case)
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 = AlbertConfig.from_pretrained(args.config_path,
share_type=args.share_type)
model = AlbertForPreTraining(config=bert_config)
if args.model_path:
model = AlbertForPreTraining.from_pretrained(args.model_path)
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':
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
}]
optimizer = AdamW(params=optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=num_train_optimization_steps)
# optimizer = Lamb(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
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)
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 = args.output_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 train(args, train_dataset, model, tokenizer):
""" Train the model """
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_sampler = RandomSampler(
train_dataset) if args.local_rank == -1 else DistributedSampler(
train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size,
collate_fn=collate_fn)
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = args.max_steps // (
len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
t_total = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0
},
]
args.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=args.warmup_steps,
num_training_steps=t_total)
# Check if saved optimizer or scheduler states exist
if os.path.isfile(os.path.join(
args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
os.path.join(args.model_name_or_path, "scheduler.pt")):
# Load in optimizer and scheduler states
optimizer.load_state_dict(
torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
scheduler.load_state_dict(
torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
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)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d",
args.per_gpu_train_batch_size)
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)
global_step = 0
steps_trained_in_current_epoch = 0
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path
) and "checkpoint" in args.model_name_or_path:
# set global_step to gobal_step of last saved checkpoint from model path
global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
epochs_trained = global_step // (len(train_dataloader) //
args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (
len(train_dataloader) // args.gradient_accumulation_steps)
logger.info(
" Continuing training from checkpoint, will skip to saved global_step"
)
logger.info(" Continuing training from epoch %d", epochs_trained)
logger.info(" Continuing training from global step %d", global_step)
logger.info(" Will skip the first %d steps in the first epoch",
steps_trained_in_current_epoch)
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
seed_everything(
args.seed
) # Added here for reproductibility (even between python 2 and 3)
for _ in range(int(args.num_train_epochs)):
pbar = ProgressBar(n_total=len(train_dataloader), desc='Training')
for step, batch in enumerate(train_dataloader):
# Skip past any already trained steps if resuming training
if steps_trained_in_current_epoch > 0:
steps_trained_in_current_epoch -= 1
continue
model.train()
batch = tuple(t.to(args.device) for t in batch)
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"start_positions": batch[3],
"end_positions": batch[4]
}
if args.model_type != "distilbert":
# XLM and RoBERTa don"t use segment_ids
inputs["token_type_ids"] = (batch[2] if args.model_type
in ["bert", "xlnet"] else None)
outputs = model(**inputs)
loss = outputs[
0] # model outputs are always tuple in pytorch-transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean(
) # mean() to average on multi-gpu parallel training
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()
pbar(step, {'loss': loss.item()})
tr_loss += loss.item()
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)
scheduler.step() # Update learning rate schedule
optimizer.step()
model.zero_grad()
global_step += 1
if args.local_rank in [
-1, 0
] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
# Log metrics
print(" ")
if args.local_rank == -1:
# Only evaluate when single GPU otherwise metrics may not average well
evaluate(args, model, tokenizer)
if args.local_rank in [
-1, 0
] and args.save_steps > 0 and global_step % args.save_steps == 0:
# Save model checkpoint
output_dir = os.path.join(
args.output_dir, "checkpoint-{}".format(global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
torch.save(args,
os.path.join(output_dir, "training_args.bin"))
tokenizer.save_vocabulary(output_dir)
logger.info("Saving model checkpoint to %s", output_dir)
torch.save(optimizer.state_dict(),
os.path.join(output_dir, "optimizer.pt"))
torch.save(scheduler.state_dict(),
os.path.join(output_dir, "scheduler.pt"))
logger.info("Saving optimizer and scheduler states to %s",
output_dir)
print(" ")
if 'cuda' in str(args.device):
torch.cuda.empty_cache()
return global_step, tr_loss / global_step
}, {
'params':
[p for n, p in other_params if any(nd in n for nd in no_decay)],
'weight_decay':
0.0,
'lr':
lr
}]
num_training_steps = (train_data.shape[0] // batch_size + 1) * epochs
# num_training_steps = 100000
# optimizer = RAdam(filter(lambda p: p.requires_grad, model.parameters()), lr=lr)
optimizer = RAdam(params=optimizer_grouped_params)
step_scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=num_training_steps // epochs,
num_training_steps=num_training_steps)
trainer = Trainer(model=model,
optimizer=optimizer,
trainloader=train_dataloader,
validloader=valid_dataloader,
epochs=epochs,
l2_reg=l2_reg,
step_scheduler=step_scheduler,
save_dir=f"albert_{name}_{data_type}",
checkpoint=checkpoint_file)
logger.info("开始训练模型")
trainer.run()