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"]
bert_param_optimizer = list(model.bert.named_parameters())
crf_param_optimizer = list(model.crf.named_parameters())
linear_param_optimizer = list(model.classifier.named_parameters())
optimizer_grouped_parameters = [
{'params': [p for n, p in bert_param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay': args.weight_decay, 'lr': args.learning_rate},
{'params': [p for n, p in bert_param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0,
'lr': args.learning_rate},
{'params': [p for n, p in crf_param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay': args.weight_decay, 'lr': args.crf_learning_rate},
{'params': [p for n, p in crf_param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0,
'lr': args.crf_learning_rate},
{'params': [p for n, p in linear_param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay': args.weight_decay, 'lr': args.crf_learning_rate},
{'params': [p for n, p in linear_param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0,
'lr': args.crf_learning_rate}
]
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], "labels": batch[3], 'input_lens': 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)
optimizer.step()
scheduler.step() # Update learning rate schedule
# lr_scheduler.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"))
logger.info("Saving model checkpoint to %s", output_dir)
tokenizer.save_vocabulary(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)
logger.info("\n")
if 'cuda' in str(args.device):
torch.cuda.empty_cache()
return global_step, tr_loss / global_step
def train(args, train_dataset, model, tokenizer):
""" Train the model """
# 总batch size = GPU数量 * 每个GPU上的batch size
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
# 如果无指定GPU就随机采样，如果指定了GPU就分布式采样
train_sampler = RandomSampler(
train_dataset) if args.local_rank == -1 else DistributedSampler(
train_dataset)
# 准备dataloader
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size,
collate_fn=collate_fn)
# 训练steps
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_dacay内的参数不参与权重衰减
# BN是固定C,[B,H,W]进行归一化处理（处理为均值0,方差1的正太分布上），适用于CNN
# LN是固定N,[C,H,W]进行归一化处理，适用于RNN（BN适用于固定深度的前向神经网络，而RNN因输入序列长度不一致而深度不固定，因此BN不合适，而LN不依赖于batch的大小和输入sequence的深度，因此可以用于batchsize为1和RNN中对边长的输入sequence的normalize操作）
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
},
]
# 配置优化器和warmup机制
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
# 检查是否有已保存的optimizer或scheduler
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")))
# 开启FP16
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)
# # 如果有多个GPU，就直接用torch.nn.DataParallel，会自动调用当前可用的多个GPU
# 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! 日志记录
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 检查是否从断点开始重训练
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
# 是否开启adversarial
if args.do_adv:
fgm = FGM(model, emb_name=args.adv_name, epsilon=args.adv_epsilon)
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],
"labels": batch[3]
}
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)
# 训练模型
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()
if args.do_adv:
fgm.attack()
loss_adv = model(**inputs)[0]
if args.n_gpu > 1:
loss_adv = loss_adv.mean()
loss_adv.backward()
fgm.restore()
pbar(step, {'loss': loss.item()})
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# 梯度裁剪
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 学习率更新
optimizer.step() # 优化器更新
model.zero_grad() # 清空现有梯度，避免累计
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)
# Take care of distributed/parallel training
model_to_save = (model.module
if hasattr(model, "module") else model)
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)
logger.info("\n")
if 'cuda' in str(args.device):
torch.cuda.empty_cache()
return global_step, tr_loss / global_step