def do_train(self, data, truncated=False):
"""
训练模型, 数据集分成2部分，训练集和验证集, 默认比例9:1
:param data: 输入的数据，注意如果做truncated，那么输入的数据为 [(content,aspect,start_idx, end_idx, label),...,]
:param truncated: 是否要截断，截断按照 self.left_max_seq_len, self.right_max_seq_len进行
:return:
"""
if truncated:
data, locations = self.do_truncate_data(data)
train_data_len = int(len(data) * 0.9)
train_data = data[:train_data_len]
eval_data = data[train_data_len:]
train_dataset = load_examples(train_data, self.max_seq_length, self.train_tokenizer, self.label_list, self.reverse_truncate)
eval_dataset = load_examples(eval_data, self.max_seq_length, self.train_tokenizer, self.label_list, self.reverse_truncate)
logger.info("训练数据集已加载,开始训练")
num_train_steps = int(len(train_dataset) / self.train_batch_size) * self.num_train_epochs
forward_batch_size = int(self.train_batch_size / self.gradient_accumulation_steps)
# 开始训练
params = list(self.train_model.named_parameters())
all_trainable_params = divide_parameters(params, lr=self.learning_rate, weight_decay_rate=self.weight_decay_rate)
# 优化器设置
optimizer = BERTAdam(all_trainable_params, lr=self.learning_rate,
warmup=self.warmup_proportion, t_total=num_train_steps, schedule=self.schedule,
s_opt1=self.s_opt1, s_opt2=self.s_opt2, s_opt3=self.s_opt3)
logger.info("***** 开始训练 *****")
logger.info(" 训练样本数是 = %d", len(train_dataset))
logger.info(" 评估样本数是 = %d", len(eval_dataset))
logger.info(" 前向 batch size = %d", forward_batch_size)
logger.info(" 训练的steps = %d", num_train_steps)
########### 训练的配置 ###########
train_config = TrainingConfig(
gradient_accumulation_steps = self.gradient_accumulation_steps,
ckpt_frequency = self.ckpt_frequency,
log_dir = self.output_dir,
output_dir = self.output_dir,
device = self.device)
#初始化trainer，执行监督训练，而不是蒸馏。它可以把model_S模型训练成为teacher模型
distiller = BasicTrainer(train_config = train_config,
model = self.train_model,
adaptor = BertForGLUESimpleAdaptorTraining)
train_sampler = RandomSampler(train_dataset)
#训练的dataloader
train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=forward_batch_size, drop_last=True)
#执行callbakc函数，对eval数据集
callback_func = partial(self.do_predict, eval_dataset=eval_dataset)
with distiller:
#开始训练
distiller.train(optimizer, scheduler=None, dataloader=train_dataloader,
num_epochs=self.num_train_epochs, callback=callback_func)
logger.info(f"训练完成")
return "Done"
def main():
#parse arguments
config.parse()
args = config.args
for k, v in vars(args).items():
logger.info(f"{k}:{v}")
#set seeds
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed_all(args.random_seed)
np.random.seed(args.random_seed)
random.seed(args.random_seed)
#arguments check
device, n_gpu = args_check(args)
os.makedirs(args.output_dir, exist_ok=True)
forward_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
args.forward_batch_size = forward_batch_size
#load bert config
bert_config_S = BertConfig.from_json_file(args.bert_config_file_S)
assert args.max_seq_length <= bert_config_S.max_position_embeddings
#Prepare GLUE task
processor = processors[args.task_name]()
args.output_mode = output_modes[args.task_name]
# eg： MNLI，['contradiction', 'entailment', 'neutral'] --> [“矛盾”，“必然”，“中立”]
label_list = processor.get_labels()
num_labels = len(label_list)
#read data
train_dataset = None
eval_datasets = None
num_train_steps = None
tokenizer = BertTokenizer(vocab_file=args.vocab_file,
do_lower_case=args.do_lower_case)
# 加载数据集, 计算steps
if args.do_train:
train_dataset = load_and_cache_examples(args,
args.task_name,
tokenizer,
evaluate=False)
if args.aux_task_name:
aux_train_dataset = load_and_cache_examples(args,
args.aux_task_name,
tokenizer,
evaluate=False,
is_aux=True)
train_dataset = torch.utils.data.ConcatDataset(
[train_dataset, aux_train_dataset])
num_train_steps = int(
len(train_dataset) / args.train_batch_size) * args.num_train_epochs
if args.do_predict:
eval_datasets = []
eval_task_names = ("mnli",
"mnli-mm") if args.task_name == "mnli" else (
args.task_name, )
for eval_task in eval_task_names:
eval_datasets.append(
load_and_cache_examples(args,
eval_task,
tokenizer,
evaluate=True))
logger.info("数据集已加载")
#加载模型并初始化, 只用student模型，其实这里相当于在MNLI数据上训练教师模型，只训练一个模型
model_S = BertForGLUESimple(bert_config_S,
num_labels=num_labels,
args=args)
#初始化student模型
if args.load_model_type == 'bert':
assert args.init_checkpoint_S is not None
state_dict_S = torch.load(args.init_checkpoint_S, map_location='cpu')
if args.only_load_embedding:
state_weight = {
k[5:]: v
for k, v in state_dict_S.items()
if k.startswith('bert.embeddings')
}
missing_keys, _ = model_S.bert.load_state_dict(state_weight,
strict=False)
logger.info(f"Missing keys {list(missing_keys)}")
else:
state_weight = {
k[5:]: v
for k, v in state_dict_S.items() if k.startswith('bert.')
}
missing_keys, _ = model_S.bert.load_state_dict(state_weight,
strict=False)
assert len(missing_keys) == 0
logger.info("Model loaded")
elif args.load_model_type == 'all':
assert args.tuned_checkpoint_S is not None
state_dict_S = torch.load(args.tuned_checkpoint_S, map_location='cpu')
model_S.load_state_dict(state_dict_S)
logger.info("Model loaded")
else:
logger.info("Model is randomly initialized.")
model_S.to(device)
if args.local_rank != -1 or n_gpu > 1:
if args.local_rank != -1:
raise NotImplementedError
elif n_gpu > 1:
model_S = torch.nn.DataParallel(model_S) #,output_device=n_gpu-1)
if args.do_train:
#parameters
params = list(model_S.named_parameters())
all_trainable_params = divide_parameters(params, lr=args.learning_rate)
logger.info("Length of all_trainable_params: %d",
len(all_trainable_params))
# 优化器设置
optimizer = BERTAdam(all_trainable_params,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_steps,
schedule=args.schedule,
s_opt1=args.s_opt1,
s_opt2=args.s_opt2,
s_opt3=args.s_opt3)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Forward batch size = %d", forward_batch_size)
logger.info(" Num backward steps = %d", num_train_steps)
########### 蒸馏 ###########
train_config = TrainingConfig(
gradient_accumulation_steps=args.gradient_accumulation_steps,
ckpt_frequency=args.ckpt_frequency,
log_dir=args.output_dir,
output_dir=args.output_dir,
device=args.device)
#执行监督训练，而不是蒸馏。它可以用于训练teacher模型。初始化模型
distiller = BasicTrainer(train_config=train_config,
model=model_S,
adaptor=BertForGLUESimpleAdaptorTraining)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
raise NotImplementedError
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.forward_batch_size,
drop_last=True)
callback_func = partial(predict,
eval_datasets=eval_datasets,
args=args)
with distiller:
distiller.train(optimizer,
scheduler=None,
dataloader=train_dataloader,
num_epochs=args.num_train_epochs,
callback=callback_func)
if not args.do_train and args.do_predict:
res = predict(model_S, eval_datasets, step=0, args=args)
print(res)
def main():
#parse arguments
config.parse()
args = config.args
for k, v in vars(args).items():
logger.info(f"{k}:{v}")
#set seeds
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed_all(args.random_seed)
np.random.seed(args.random_seed)
random.seed(args.random_seed)
#arguments check
device, n_gpu = args_check(args)
os.makedirs(args.output_dir, exist_ok=True)
forward_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
args.forward_batch_size = forward_batch_size
#load bert config
bert_config_T = BertConfig.from_json_file(args.bert_config_file_T)
bert_config_S = BertConfig.from_json_file(args.bert_config_file_S)
assert args.max_seq_length <= bert_config_T.max_position_embeddings
assert args.max_seq_length <= bert_config_S.max_position_embeddings
#Prepare GLUE task
processor = processors[args.task_name]()
args.output_mode = output_modes[args.task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
#read data
train_dataset = None
eval_datasets = None
num_train_steps = None
tokenizer = BertTokenizer(vocab_file=args.vocab_file,
do_lower_case=args.do_lower_case)
# 加载数据集
if args.do_train:
train_dataset, examples = load_and_cache_examples(args,
args.task_name,
tokenizer,
evaluate=False)
if args.aux_task_name:
aux_train_dataset, examples = load_and_cache_examples(
args,
args.aux_task_name,
tokenizer,
evaluate=False,
is_aux=True)
train_dataset = torch.utils.data.ConcatDataset(
[train_dataset, aux_train_dataset])
num_train_steps = int(
len(train_dataset) / args.train_batch_size) * args.num_train_epochs
if args.do_predict:
eval_datasets = []
eval_task_names = ("mnli",
"mnli-mm") if args.task_name == "mnli" else (
args.task_name, )
for eval_task in eval_task_names:
eval_dataset, examples = load_and_cache_examples(args,
eval_task,
tokenizer,
evaluate=True)
eval_datasets.append(eval_dataset)
logger.info("数据集加载成功")
#加载模型，加载teacher和student模型
model_T = BertForGLUESimple(bert_config_T,
num_labels=num_labels,
args=args)
model_S = BertForGLUESimple(bert_config_S,
num_labels=num_labels,
args=args)
#加载teacher模型参数
if args.tuned_checkpoint_T is not None:
state_dict_T = torch.load(args.tuned_checkpoint_T, map_location='cpu')
model_T.load_state_dict(state_dict_T)
model_T.eval()
else:
assert args.do_predict is True
#Load student
if args.load_model_type == 'bert':
assert args.init_checkpoint_S is not None
state_dict_S = torch.load(args.init_checkpoint_S, map_location='cpu')
if args.only_load_embedding:
state_weight = {
k[5:]: v
for k, v in state_dict_S.items()
if k.startswith('bert.embeddings')
}
missing_keys, _ = model_S.bert.load_state_dict(state_weight,
strict=False)
logger.info(f"Missing keys {list(missing_keys)}")
else:
state_weight = {
k[5:]: v
for k, v in state_dict_S.items() if k.startswith('bert.')
}
missing_keys, _ = model_S.bert.load_state_dict(state_weight,
strict=False)
assert len(missing_keys) == 0
logger.info("Model loaded")
elif args.load_model_type == 'all':
assert args.tuned_checkpoint_S is not None
state_dict_S = torch.load(args.tuned_checkpoint_S, map_location='cpu')
model_S.load_state_dict(state_dict_S)
logger.info("Model loaded")
else:
logger.info("Student模型没有可加载参数，随机初始化参数 randomly initialized.")
model_T.to(device)
model_S.to(device)
if args.local_rank != -1 or n_gpu > 1:
if args.local_rank != -1:
raise NotImplementedError
elif n_gpu > 1:
model_T = torch.nn.DataParallel(model_T) #,output_device=n_gpu-1)
model_S = torch.nn.DataParallel(model_S) #,output_device=n_gpu-1)
if args.do_train:
#parameters
params = list(model_S.named_parameters())
all_trainable_params = divide_parameters(params, lr=args.learning_rate)
logger.info("Length of all_trainable_params: %d",
len(all_trainable_params))
#优化器配置
optimizer = BERTAdam(all_trainable_params,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_steps,
schedule=args.schedule,
s_opt1=args.s_opt1,
s_opt2=args.s_opt2,
s_opt3=args.s_opt3)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Forward batch size = %d", forward_batch_size)
logger.info(" Num backward steps = %d", num_train_steps)
########### DISTILLATION ###########
train_config = TrainingConfig(
gradient_accumulation_steps=args.gradient_accumulation_steps,
ckpt_frequency=args.ckpt_frequency,
log_dir=args.output_dir,
output_dir=args.output_dir,
device=args.device)
# 定义了一些固定的matches配置文件
from matches import matches
intermediate_matches = None
if isinstance(args.matches, (list, tuple)):
intermediate_matches = []
for match in args.matches:
intermediate_matches += matches[match]
logger.info(f"中间层match信息： {intermediate_matches}")
distill_config = DistillationConfig(
temperature=args.temperature,
intermediate_matches=intermediate_matches)
logger.info(f"训练配置： {train_config}")
logger.info(f"蒸馏配置： {distill_config}")
adaptor_T = partial(BertForGLUESimpleAdaptor,
no_logits=args.no_logits,
no_mask=args.no_inputs_mask)
adaptor_S = partial(BertForGLUESimpleAdaptor,
no_logits=args.no_logits,
no_mask=args.no_inputs_mask)
# 支持中间状态匹配的通用蒸馏模型
distiller = GeneralDistiller(train_config=train_config,
distill_config=distill_config,
model_T=model_T,
model_S=model_S,
adaptor_T=adaptor_T,
adaptor_S=adaptor_S)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
raise NotImplementedError
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.forward_batch_size,
drop_last=True)
callback_func = partial(predict,
eval_datasets=eval_datasets,
args=args,
examples=examples)
with distiller:
distiller.train(optimizer,
scheduler=None,
dataloader=train_dataloader,
num_epochs=args.num_train_epochs,
callback=callback_func)
if not args.do_train and args.do_predict:
res = predict(model_S,
eval_datasets,
step=0,
args=args,
examples=examples,
label_list=label_list)
print(res)
def main():
#parse arguments
config.parse()
args = config.args
for k, v in vars(args).items():
logger.info(f"{k}:{v}")
#set seeds
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed_all(args.random_seed)
np.random.seed(args.random_seed)
random.seed(args.random_seed)
#arguments check
device, n_gpu = args_check(args)
os.makedirs(args.output_dir, exist_ok=True)
forward_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
args.forward_batch_size = forward_batch_size
#load bert config
bert_config_S = BertConfig.from_json_file(args.bert_config_file_S)
assert args.max_seq_length <= bert_config_S.max_position_embeddings
#read data
train_examples = None
train_features = None
eval_examples = None
eval_features = None
num_train_steps = None
tokenizer = ChineseFullTokenizer(vocab_file=args.vocab_file,
do_lower_case=args.do_lower_case)
convert_fn = partial(convert_examples_to_features,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length)
if args.do_train:
train_examples, train_features = read_and_convert(
args.train_file,
is_training=True,
do_lower_case=args.do_lower_case,
read_fn=read_squad_examples,
convert_fn=convert_fn)
if args.fake_file_1:
fake_examples1, fake_features1 = read_and_convert(
args.fake_file_1,
is_training=True,
do_lower_case=args.do_lower_case,
read_fn=read_squad_examples,
convert_fn=convert_fn)
train_examples += fake_examples1
train_features += fake_features1
if args.fake_file_2:
fake_examples2, fake_features2 = read_and_convert(
args.fake_file_2,
is_training=True,
do_lower_case=args.do_lower_case,
read_fn=read_squad_examples,
convert_fn=convert_fn)
train_examples += fake_examples2
train_features += fake_features2
num_train_steps = int(len(train_features) /
args.train_batch_size) * args.num_train_epochs
if args.do_predict:
eval_examples, eval_features = read_and_convert(
args.predict_file,
is_training=False,
do_lower_case=args.do_lower_case,
read_fn=read_squad_examples,
convert_fn=convert_fn)
#Build Model and load checkpoint
model_S = BertForQASimple(bert_config_S, args)
#Load student
if args.load_model_type == 'bert':
assert args.init_checkpoint_S is not None
state_dict_S = torch.load(args.init_checkpoint_S, map_location='cpu')
state_weight = {
k[5:]: v
for k, v in state_dict_S.items() if k.startswith('bert.')
}
missing_keys, _ = model_S.bert.load_state_dict(state_weight,
strict=False)
assert len(missing_keys) == 0
elif args.load_model_type == 'all':
assert args.tuned_checkpoint_S is not None
state_dict_S = torch.load(args.tuned_checkpoint_S, map_location='cpu')
model_S.load_state_dict(state_dict_S)
else:
logger.info("Model is randomly initialized.")
model_S.to(device)
if args.local_rank != -1 or n_gpu > 1:
if args.local_rank != -1:
raise NotImplementedError
elif n_gpu > 1:
model_S = torch.nn.DataParallel(model_S) #,output_device=n_gpu-1)
if args.do_train:
#parameters
params = list(model_S.named_parameters())
all_trainable_params = divide_parameters(params, lr=args.learning_rate)
logger.info("Length of all_trainable_params: %d",
len(all_trainable_params))
optimizer = BERTAdam(all_trainable_params,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_steps,
schedule=args.schedule,
s_opt1=args.s_opt1,
s_opt2=args.s_opt2,
s_opt3=args.s_opt3)
logger.info("***** Running training *****")
logger.info(" Num orig examples = %d", len(train_examples))
logger.info(" Num split examples = %d", len(train_features))
logger.info(" Forward batch size = %d", forward_batch_size)
logger.info(" Num backward steps = %d", num_train_steps)
########### DISTILLATION ###########
train_config = TrainingConfig(
gradient_accumulation_steps=args.gradient_accumulation_steps,
ckpt_frequency=args.ckpt_frequency,
log_dir=args.output_dir,
output_dir=args.output_dir,
device=args.device)
distiller = BasicTrainer(train_config=train_config,
model=model_S,
adaptor=BertForQASimpleAdaptorTraining)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_doc_mask = torch.tensor([f.doc_mask for f in train_features],
dtype=torch.float)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_start_positions = torch.tensor(
[f.start_position for f in train_features], dtype=torch.long)
all_end_positions = torch.tensor(
[f.end_position for f in train_features], dtype=torch.long)
train_dataset = TensorDataset(all_input_ids, all_segment_ids,
all_input_mask, all_doc_mask,
all_start_positions, all_end_positions)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
raise NotImplementedError
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.forward_batch_size,
drop_last=True)
callback_func = partial(predict,
eval_examples=eval_examples,
eval_features=eval_features,
args=args)
with distiller:
distiller.train(optimizer,
scheduler=None,
dataloader=train_dataloader,
num_epochs=args.num_train_epochs,
callback=callback_func)
if not args.do_train and args.do_predict:
res = predict(model_S, eval_examples, eval_features, step=0, args=args)
print(res)
def main():
#parse arguments
config.parse()
args = config.args
for k, v in vars(args).items():
logger.info(f"{k}:{v}")
#set seeds
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed_all(args.random_seed)
np.random.seed(args.random_seed)
random.seed(args.random_seed)
#arguments check
device, n_gpu = args_check(args)
os.makedirs(args.output_dir, exist_ok=True)
forward_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
args.forward_batch_size = forward_batch_size
#load config
teachers_and_student = parse_model_config(args.model_config_json)
#Prepare GLUE task
processor = processors[args.task_name]()
args.output_mode = output_modes[args.task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
#read data
train_dataset = None
eval_datasets = None
num_train_steps = None
tokenizer_S = teachers_and_student['student']['tokenizer']
prefix_S = teachers_and_student['student']['prefix']
if args.do_train:
train_dataset = load_and_cache_examples(args,
args.task_name,
tokenizer_S,
prefix=prefix_S,
evaluate=False)
if args.do_predict:
eval_datasets = []
eval_task_names = ("mnli",
"mnli-mm") if args.task_name == "mnli" else (
args.task_name, )
for eval_task in eval_task_names:
eval_datasets.append(
load_and_cache_examples(args,
eval_task,
tokenizer_S,
prefix=prefix_S,
evaluate=True))
logger.info("Data loaded")
#Build Model and load checkpoint
if args.do_train:
model_Ts = []
for teacher in teachers_and_student['teachers']:
model_type_T = teacher['model_type']
model_config_T = teacher['config']
checkpoint_T = teacher['checkpoint']
_, _, model_class_T = MODEL_CLASSES[model_type_T]
model_T = model_class_T(model_config_T, num_labels=num_labels)
state_dict_T = torch.load(checkpoint_T, map_location='cpu')
missing_keys, un_keys = model_T.load_state_dict(state_dict_T,
strict=True)
logger.info(f"Teacher Model {model_type_T} loaded")
model_T.to(device)
model_T.eval()
model_Ts.append(model_T)
student = teachers_and_student['student']
model_type_S = student['model_type']
model_config_S = student['config']
checkpoint_S = student['checkpoint']
_, _, model_class_S = MODEL_CLASSES[model_type_S]
model_S = model_class_S(model_config_S, num_labels=num_labels)
if checkpoint_S is not None:
state_dict_S = torch.load(checkpoint_S, map_location='cpu')
missing_keys, un_keys = model_S.load_state_dict(state_dict_S,
strict=False)
logger.info(f"missing keys:{missing_keys}")
logger.info(f"unexpected keys:{un_keys}")
else:
logger.warning("Initializing student randomly")
logger.info("Student Model loaded")
model_S.to(device)
if args.local_rank != -1 or n_gpu > 1:
if args.local_rank != -1:
raise NotImplementedError
elif n_gpu > 1:
if args.do_train:
model_Ts = [
torch.nn.DataParallel(model_T) for model_T in model_Ts
]
model_S = torch.nn.DataParallel(model_S) #,output_device=n_gpu-1)
if args.do_train:
#parameters
params = list(model_S.named_parameters())
all_trainable_params = divide_parameters(params, lr=args.learning_rate)
logger.info("Length of all_trainable_params: %d",
len(all_trainable_params))
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
raise NotImplementedError
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.forward_batch_size,
drop_last=True)
num_train_steps = int(
len(train_dataloader) // args.gradient_accumulation_steps *
args.num_train_epochs)
########## DISTILLATION ###########
train_config = TrainingConfig(
gradient_accumulation_steps=args.gradient_accumulation_steps,
ckpt_frequency=args.ckpt_frequency,
log_dir=args.output_dir,
output_dir=args.output_dir,
fp16=args.fp16,
device=args.device)
distill_config = DistillationConfig(temperature=args.temperature,
kd_loss_type='ce')
logger.info(f"{train_config}")
logger.info(f"{distill_config}")
adaptor_T = BertForGLUESimpleAdaptor
adaptor_S = BertForGLUESimpleAdaptor
distiller = MultiTeacherDistiller(train_config=train_config,
distill_config=distill_config,
model_T=model_Ts,
model_S=model_S,
adaptor_T=adaptor_T,
adaptor_S=adaptor_S)
optimizer = AdamW(all_trainable_params, lr=args.learning_rate)
scheduler_class = get_linear_schedule_with_warmup
scheduler_args = {
'num_warmup_steps': int(args.warmup_proportion * num_train_steps),
'num_training_steps': num_train_steps
}
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Forward batch size = %d", forward_batch_size)
logger.info(" Num backward steps = %d", num_train_steps)
callback_func = partial(predict,
eval_datasets=eval_datasets,
args=args)
with distiller:
distiller.train(optimizer,
scheduler_class=scheduler_class,
scheduler_args=scheduler_args,
dataloader=train_dataloader,
num_epochs=args.num_train_epochs,
callback=callback_func,
max_grad_norm=1)
if not args.do_train and args.do_predict:
res = predict(model_S, eval_datasets, step=0, args=args)
print(res)
def train(args, train_dataset, model_T, model, tokenizer, labels,
pad_token_label_id, predict_callback):
""" 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)
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
}]
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=int(args.warmup_steps * t_total),
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)
# 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)
if args.do_train and args.do_distill:
from textbrewer import DistillationConfig, TrainingConfig, GeneralDistiller
distill_config = DistillationConfig(
temperature=8,
# intermediate_matches = [{'layer_T':10, 'layer_S':3, 'feature':'hidden','loss': 'hidden_mse', 'weight' : 1}]
)
train_config = TrainingConfig(device="cuda",
log_dir=args.output_dir,
output_dir=args.output_dir)
def adaptor_T(batch, model_output):
return {
"logits": (model_output[1], ),
'logits_mask': (batch['attention_mask'], )
}
def adaptor_S(batch, model_output):
return {
"logits": (model_output[1], ),
'logits_mask': (batch['attention_mask'], )
}
distiller = GeneralDistiller(
train_config,
distill_config,
model_T,
model,
adaptor_T,
adaptor_S,
)
distiller.train(optimizer,
scheduler,
train_dataloader,
args.num_train_epochs,
callback=predict_callback)
return
label_ids.append(labels[i])
model.train()
pred_logits = np.array(pred_logits)
label_ids = np.array(label_ids)
y_p = pred_logits.argmax(axis=-1)
accuracy = (y_p == label_ids).sum() / len(label_ids)
print("Number of examples: ", len(y_p))
print("Acc: ", accuracy)
from functools import partial
callback_fun = partial(predict, eval_dataset=eval_dataset,
device=device) # fill other arguments
# Initialize configurations and distiller
train_config = TrainingConfig(device=device)
distill_config = DistillationConfig(temperature=8,
hard_label_weight=0,
kd_loss_type='ce',
probability_shift=False,
intermediate_matches=[{
'layer_T': 0,
'layer_S': 0,
'feature': 'hidden',
'loss': 'hidden_mse',
'weight': 1
}, {
'layer_T': 8,
'layer_S': 2,
'feature': 'hidden',
'loss': 'hidden_mse',
def main():
#parse arguments
config.parse()
args = config.args
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%Y/%m/%d %H:%M:%S',
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger = logging.getLogger("Main")
#arguments check
device, n_gpu = args_check(logger, args)
if args.local_rank in [-1, 0]:
os.makedirs(args.output_dir, exist_ok=True)
if args.local_rank != -1:
logger.warning(
f"Process rank: {torch.distributed.get_rank()}, device : {args.device}, n_gpu : {args.n_gpu}, distributed training : {bool(args.local_rank!=-1)}"
)
for k, v in vars(args).items():
logger.info(f"{k}:{v}")
#set seeds
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed_all(args.random_seed)
np.random.seed(args.random_seed)
random.seed(args.random_seed)
forward_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
args.forward_batch_size = forward_batch_size
#load bert config
bert_config_S = ElectraConfig.from_json_file(args.bert_config_file_S)
bert_config_S.output_hidden_states = (args.output_encoded_layers == 'true')
bert_config_S.num_labels = len(label2id_dict)
assert args.max_seq_length <= bert_config_S.max_position_embeddings
#read data
train_examples = None
train_dataset = None
eval_examples = None
eval_dataset = None
num_train_steps = None
tokenizer = BertTokenizer(vocab_file=args.vocab_file,
do_lower_case=args.do_lower_case)
if args.local_rank not in [-1, 0]:
torch.distributed.barrier()
if args.do_train:
train_examples, train_dataset = read_features(
args.train_file,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length)
if args.do_predict:
eval_examples, eval_dataset = read_features(
args.predict_file,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length)
if args.local_rank == 0:
torch.distributed.barrier()
#Build Model and load checkpoint
model_S = ElectraForTokenClassification(bert_config_S)
#Load student
if args.load_model_type == 'bert':
assert args.init_checkpoint_S is not None
state_dict_S = torch.load(args.init_checkpoint_S, map_location='cpu')
#state_weight = {k[5:]:v for k,v in state_dict_S.items() if k.startswith('bert.')}
#missing_keys,_ = model_S.bert.load_state_dict(state_weight,strict=False)
missing_keys, unexpected_keys = model_S.load_state_dict(state_dict_S,
strict=False)
logger.info(f"missing keys:{missing_keys}")
logger.info(f"unexpected keys:{unexpected_keys}")
elif args.load_model_type == 'all':
assert args.tuned_checkpoint_S is not None
state_dict_S = torch.load(args.tuned_checkpoint_S, map_location='cpu')
model_S.load_state_dict(state_dict_S)
else:
logger.info("Model is randomly initialized.")
model_S.to(device)
if args.do_train:
#parameters
if args.lr_decay is not None:
outputs_params = list(model_S.classifier.named_parameters())
outputs_params = divide_parameters(outputs_params,
lr=args.learning_rate)
electra_params = []
n_layers = len(model_S.electra.encoder.layer)
assert n_layers == 12
for i, n in enumerate(reversed(range(n_layers))):
encoder_params = list(
model_S.electra.encoder.layer[n].named_parameters())
lr = args.learning_rate * args.lr_decay**(i + 1)
electra_params += divide_parameters(encoder_params, lr=lr)
logger.info(f"{i},{n},{lr}")
embed_params = [
(name, value)
for name, value in model_S.electra.named_parameters()
if 'embedding' in name
]
logger.info(f"{[name for name,value in embed_params]}")
lr = args.learning_rate * args.lr_decay**(n_layers + 1)
electra_params += divide_parameters(embed_params, lr=lr)
logger.info(f"embed lr:{lr}")
all_trainable_params = outputs_params + electra_params
assert sum(map(lambda x:len(x['params']), all_trainable_params))==len(list(model_S.parameters())),\
(sum(map(lambda x:len(x['params']), all_trainable_params)), len(list(model_S.parameters())))
else:
params = list(model_S.named_parameters())
all_trainable_params = divide_parameters(params,
lr=args.learning_rate)
logger.info("Length of all_trainable_params: %d",
len(all_trainable_params))
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
train_sampler = DistributedSampler(train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.forward_batch_size,
drop_last=True)
num_train_steps = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
optimizer = AdamW(all_trainable_params,
lr=args.learning_rate,
correct_bias=False)
if args.official_schedule == 'const':
scheduler_class = get_constant_schedule_with_warmup
scheduler_args = {
'num_warmup_steps':
int(args.warmup_proportion * num_train_steps)
}
#scheduler = get_constant_schedule_with_warmup(optimizer, num_warmup_steps=int(args.warmup_proportion*num_train_steps))
elif args.official_schedule == 'linear':
scheduler_class = get_linear_schedule_with_warmup
scheduler_args = {
'num_warmup_steps':
int(args.warmup_proportion * num_train_steps),
'num_training_steps': num_train_steps
}
#scheduler = get_linear_schedule_with_warmup(optimizer,num_warmup_steps=int(args.warmup_proportion*num_train_steps), num_training_steps = num_train_steps)
elif args.official_schedule == 'const_nowarmup':
scheduler_class = get_constant_schedule
scheduler_args = {}
else:
raise NotImplementedError
logger.warning("***** Running training *****")
logger.warning("local_rank %d Num orig examples = %d", args.local_rank,
len(train_examples))
logger.warning("local_rank %d Num split examples = %d",
args.local_rank, len(train_dataset))
logger.warning("local_rank %d Forward batch size = %d",
args.local_rank, forward_batch_size)
logger.warning("local_rank %d Num backward steps = %d",
args.local_rank, num_train_steps)
########### TRAINING ###########
train_config = TrainingConfig(
gradient_accumulation_steps=args.gradient_accumulation_steps,
ckpt_frequency=args.ckpt_frequency,
log_dir=args.output_dir,
output_dir=args.output_dir,
device=args.device,
fp16=args.fp16,
local_rank=args.local_rank)
logger.info(f"{train_config}")
distiller = BasicTrainer(
train_config=train_config,
model=model_S,
adaptor=ElectraForTokenClassificationAdaptorTraining)
# evluate the model in a single process in ddp_predict
callback_func = partial(ddp_predict,
eval_examples=eval_examples,
eval_dataset=eval_dataset,
args=args)
with distiller:
distiller.train(optimizer,
scheduler_class=scheduler_class,
scheduler_args=scheduler_args,
max_grad_norm=1.0,
dataloader=train_dataloader,
num_epochs=args.num_train_epochs,
callback=callback_func)
if not args.do_train and args.do_predict:
res = ddp_predict(model_S,
eval_examples,
eval_dataset,
step=0,
args=args)
print(res)
from textbrewer import GeneralDistiller
from textbrewer import TrainingConfig, DistillationConfig
# We omit the initialization of models, optimizer, and dataloader.
teacher_model : torch.nn.Module = ...
student_model : torch.nn.Module = ...
dataloader : torch.utils.data.DataLoader = ...
optimizer : torch.optim.Optimizer = ...
scheduler : torch.optim.lr_scheduler = ...
def simple_adaptor(batch, model_outputs):
# We assume that the first element of model_outputs
# is the logits before softmax
return {'logits': model_outputs[0]}
train_config = TrainingConfig()
distill_config = DistillationConfig()
distiller = GeneralDistiller(
train_config=train_config, distill_config = distill_config,
model_T = teacher_model, model_S = student_model,
adaptor_T = simple_adaptor, adaptor_S = simple_adaptor)
distiller.train(optimizer, scheduler,
dataloader, num_epochs, callback=None)
def predict(model, eval_dataset, step, args):
raise NotImplementedError
def main():
#解析参数
config.parse()
args = config.args
for k,v in vars(args).items():
logger.info(f"{k}:{v}")
#set seeds
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed_all(args.random_seed)
np.random.seed(args.random_seed)
random.seed(args.random_seed)
#arguments check
device, n_gpu = args_check(args)
os.makedirs(args.output_dir, exist_ok=True)
forward_batch_size = int(args.train_batch_size / args.gradient_accumulation_steps)
args.forward_batch_size = forward_batch_size
#准备任务
processor = processors[args.task_name]()
args.output_mode = output_modes[args.task_name]
# eg： MNLI，['contradiction', 'entailment', 'neutral'] --> [“矛盾”，“必然”，“中立”]
label_list = processor.get_labels()
num_labels = len(label_list)
# Student的配置
if args.model_architecture == "electra":
# 从transformers包中导入ElectraConfig, 并加载配置
bert_config_S = ElectraConfig.from_json_file(args.bert_config_file_S)
# (args.output_encoded_layers=='true') --> True, 默认输出隐藏层的状态
bert_config_S.output_hidden_states = (args.output_encoded_layers == 'true')
bert_config_S.output_attentions = (args.output_attention_layers=='true')
# num_labels；类别个数
bert_config_S.num_labels = num_labels
assert args.max_seq_length <= bert_config_S.max_position_embeddings
elif args.model_architecture == "albert":
# 从transformers包中导入AlbertConfig, 并加载配置
bert_config_S = AlbertConfig.from_json_file(args.bert_config_file_S)
# (args.output_encoded_layers=='true') --> True, 默认输出隐藏层的状态
bert_config_S.output_hidden_states = (args.output_encoded_layers == 'true')
bert_config_S.output_attentions = (args.output_attention_layers=='true')
# num_labels；类别个数
bert_config_S.num_labels = num_labels
assert args.max_seq_length <= bert_config_S.max_position_embeddings
else:
bert_config_S = BertConfig.from_json_file(args.bert_config_file_S)
assert args.max_seq_length <= bert_config_S.max_position_embeddings
#read data
train_dataset = None
eval_datasets = None
num_train_steps = None
# electra和bert都使用的bert的 tokenizer方式
tokenizer = BertTokenizer(vocab_file=args.vocab_file, do_lower_case=args.do_lower_case)
# 加载数据集, 计算steps
if args.do_train:
train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
if args.aux_task_name:
aux_train_dataset = load_and_cache_examples(args, args.aux_task_name, tokenizer, evaluate=False, is_aux=True)
train_dataset = torch.utils.data.ConcatDataset([train_dataset, aux_train_dataset])
num_train_steps = int(len(train_dataset)/args.train_batch_size) * args.num_train_epochs
logger.info("训练数据集已加载")
if args.do_predict:
eval_datasets = []
eval_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,)
for eval_task in eval_task_names:
eval_datasets.append(load_and_cache_examples(args, eval_task, tokenizer, evaluate=True))
logger.info("预测数据集已加载")
# Student的配置
if args.model_architecture == "electra":
#加载模型配置, 只用student模型，其实这里相当于训练教师模型，只训练一个模型
model_S = ElectraSPC(bert_config_S)
elif args.model_architecture == "albert":
model_S = AlbertSPC(bert_config_S)
else:
#加载模型配置, 只用student模型，其实这里相当于训练教师模型，只训练一个模型
model_S = BertSPCSimple(bert_config_S, num_labels=num_labels,args=args)
#对加载后的student模型的参数进行初始化, 使用student模型预测
if args.load_model_type=='bert' and args.model_architecture not in ["electra", "albert"]:
assert args.init_checkpoint_S is not None
state_dict_S = torch.load(args.init_checkpoint_S, map_location='cpu')
if args.only_load_embedding:
state_weight = {k[5:]:v for k,v in state_dict_S.items() if k.startswith('bert.embeddings')}
missing_keys,_ = model_S.bert.load_state_dict(state_weight,strict=False)
logger.info(f"Missing keys {list(missing_keys)}")
else:
state_weight = {k[5:]:v for k,v in state_dict_S.items() if k.startswith('bert.')}
missing_keys,_ = model_S.bert.load_state_dict(state_weight,strict=False)
print(f"missing_keys，注意丢失的参数{missing_keys}")
logger.info("Model loaded")
elif args.load_model_type=='all':
assert args.tuned_checkpoint_S is not None
state_dict_S = torch.load(args.tuned_checkpoint_S,map_location='cpu')
model_S.load_state_dict(state_dict_S)
logger.info("Model loaded")
elif args.model_architecture in ["electra", "albert"]:
assert args.init_checkpoint_S is not None
state_dict_S = torch.load(args.init_checkpoint_S, map_location='cpu')
missing_keys, unexpected_keys = model_S.load_state_dict(state_dict_S,strict=False)
logger.info(f"missing keys:{missing_keys}")
logger.info(f"unexpected keys:{unexpected_keys}")
else:
logger.info("Model is randomly initialized.")
#模型move to device
model_S.to(device)
if args.local_rank != -1 or n_gpu > 1:
if args.local_rank != -1:
raise NotImplementedError
elif n_gpu > 1:
model_S = torch.nn.DataParallel(model_S) #,output_device=n_gpu-1)
if args.do_train:
#parameters， params是模型的所有参数组成的列表
params = list(model_S.named_parameters())
all_trainable_params = divide_parameters(params, lr=args.learning_rate)
logger.info("要训练的模型参数量组是,包括decay_group和no_decay_group: %d", len(all_trainable_params))
# 优化器设置
optimizer = BERTAdam(all_trainable_params,lr=args.learning_rate,
warmup=args.warmup_proportion,t_total=num_train_steps,schedule=args.schedule,
s_opt1=args.s_opt1, s_opt2=args.s_opt2, s_opt3=args.s_opt3)
logger.info("***** 开始训练 *****")
logger.info(" 样本数是 = %d", len(train_dataset))
logger.info(" 前向 batch size = %d", forward_batch_size)
logger.info(" 训练的steps = %d", num_train_steps)
########### 训练的配置 ###########
train_config = TrainingConfig(
gradient_accumulation_steps = args.gradient_accumulation_steps,
ckpt_frequency = args.ckpt_frequency,
log_dir = args.output_dir,
output_dir = args.output_dir,
device = args.device)
#初始化trainer，执行监督训练，而不是蒸馏。它可以把model_S模型训练成为teacher模型
distiller = BasicTrainer(train_config = train_config,
model = model_S,
adaptor = BertForGLUESimpleAdaptorTraining)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
raise NotImplementedError
#训练的dataloader
train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.forward_batch_size,drop_last=True)
#执行callbakc函数，对eval数据集
callback_func = partial(predict, eval_datasets=eval_datasets, args=args)
with distiller:
#开始训练
distiller.train(optimizer, scheduler=None, dataloader=train_dataloader,
num_epochs=args.num_train_epochs, callback=callback_func)
if not args.do_train and args.do_predict:
res = predict(model_S,eval_datasets,step=0,args=args)
print(res)
def main():
config_distil.parse()
global args
args = config_distil.args
global logger
logger = create_logger(args.log_file)
for k, v in vars(args).items():
logger.info(f"{k}:{v}")
# set seeds
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed_all(args.random_seed)
np.random.seed(args.random_seed)
random.seed(args.random_seed)
logger.info('加载字典')
word2idx = load_chinese_base_vocab()
# 判断是否有可用GPU
args.device = torch.device(
"cuda" if torch.cuda.is_available() and args.is_cuda else "cpu")
logger.info('using device:{}'.format(args.device))
# 定义模型超参数
bertconfig_T = BertConfig(vocab_size=len(word2idx))
bertconfig_S = BertConfig(vocab_size=len(word2idx), num_hidden_layers=3)
logger.info('初始化BERT模型')
bert_model_T = Seq2SeqModel(config=bertconfig_T)
bert_model_S = Seq2SeqModel(config=bertconfig_S)
logger.info('加载Teacher模型～')
load_model(bert_model_T, args.tuned_checkpoint_T)
logger.info('将模型发送到计算设备(GPU或CPU)')
bert_model_T.to(args.device)
bert_model_T.eval()
logger.info('加载Student模型～')
if args.load_model_type == 'bert':
load_model(bert_model_S, args.init_checkpoint_S)
else:
logger.info(" Student Model is randomly initialized.")
logger.info('将模型发送到计算设备(GPU或CPU)')
bert_model_S.to(args.device)
# 声明自定义的数据加载器
logger.info('加载训练数据')
train = SelfDataset(args.train_path, args.max_length)
trainloader = DataLoader(train,
batch_size=args.train_batch_size,
shuffle=True,
collate_fn=collate_fn)
if args.do_train:
logger.info(' 声明需要优化的参数')
num_train_steps = int(
len(trainloader) / args.train_batch_size) * args.num_train_epochs
optim_parameters = list(bert_model_S.named_parameters())
all_trainable_params = divide_parameters(optim_parameters,
lr=args.learning_rate)
logger.info("Length of all_trainable_params: %d",
len(all_trainable_params))
optimizer = BERTAdam(all_trainable_params,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_steps,
schedule=args.schedule,
s_opt1=args.s_opt1,
s_opt2=args.s_opt2,
s_opt3=args.s_opt3)
train_config = TrainingConfig(
gradient_accumulation_steps=args.gradient_accumulation_steps,
ckpt_frequency=args.ckpt_frequency,
log_dir=args.output_dir,
output_dir=args.output_dir,
device=args.device)
from generative.matches import matches
intermediate_matches = None
if isinstance(args.matches, (list, tuple)):
intermediate_matches = []
for match in args.matches:
intermediate_matches += matches[match]
intermediate_matches = []
for match in args.matches:
intermediate_matches += matches[match]
logger.info(f"{intermediate_matches}")
distill_config = DistillationConfig(
temperature=args.temperature,
intermediate_matches=intermediate_matches)
def BertForS2SSimpleAdaptor(batch, model_outputs):
return {'hidden': model_outputs[0], 'logits': model_outputs[1], 'loss': model_outputs[2], 'attention': model_outputs[3]}
adaptor_T = partial(BertForS2SSimpleAdaptor)
adaptor_S = partial(BertForS2SSimpleAdaptor)
distiller = GeneralDistiller(train_config=train_config,
distill_config=distill_config,
model_T=bert_model_T,
model_S=bert_model_S,
adaptor_T=adaptor_T,
adaptor_S=adaptor_S)
callback_func = partial(predict, data_path=args.dev_path, args=args)
logger.info('Start distillation.')
with distiller:
distiller.train(optimizer,
scheduler=None,
dataloader=trainloader,
num_epochs=args.num_train_epochs,
callback=None)
if not args.do_train and args.do_predict:
res = predict(bert_model_S, args.test_path, step=0, args=args)
print(res)
def main():
#parse arguments
config.parse()
args = config.args
for k,v in vars(args).items():
logger.info(f"{k}:{v}")
#set seeds
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed_all(args.random_seed)
np.random.seed(args.random_seed)
random.seed(args.random_seed)
#arguments check
device, n_gpu = args_check(args)
os.makedirs(args.output_dir, exist_ok=True)
forward_batch_size = int(args.train_batch_size / args.gradient_accumulation_steps)
args.forward_batch_size = forward_batch_size
#load bert config
bert_config_T = BertConfig.from_json_file(args.bert_config_file_T)
bert_config_S = BertConfig.from_json_file(args.bert_config_file_S)
assert args.max_seq_length <= bert_config_T.max_position_embeddings
assert args.max_seq_length <= bert_config_S.max_position_embeddings
#Prepare GLUE task
processor = processors[args.task_name]()
args.output_mode = output_modes[args.task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
#read data
train_dataset = None
eval_datasets = None
num_train_steps = None
tokenizer = BertTokenizer(vocab_file=args.vocab_file, do_lower_case=args.do_lower_case)
if args.do_train:
train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
if args.aux_task_name:
aux_train_dataset = load_and_cache_examples(args, args.aux_task_name, tokenizer, evaluate=False, is_aux=True)
train_dataset = torch.utils.data.ConcatDataset([train_dataset, aux_train_dataset])
num_train_steps = int(len(train_dataset)/args.train_batch_size) * args.num_train_epochs
if args.do_predict:
eval_datasets = []
eval_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,)
for eval_task in eval_task_names:
eval_datasets.append(load_and_cache_examples(args, eval_task, tokenizer, evaluate=True))
logger.info("Data loaded")
#Build Model and load checkpoint
model_S = BertForGLUESimple(bert_config_S, num_labels=num_labels,args=args)
#Load teacher
if args.tuned_checkpoint_Ts:
model_Ts = [BertForGLUESimple(bert_config_T, num_labels=num_labels,args=args) for i in range(len(args.tuned_checkpoint_Ts))]
for model_T, ckpt_T in zip(model_Ts,args.tuned_checkpoint_Ts):
logger.info("Load state dict %s" % ckpt_T)
state_dict_T = torch.load(ckpt_T, map_location='cpu')
model_T.load_state_dict(state_dict_T)
model_T.eval()
else:
assert args.do_predict is True
#Load student
if args.load_model_type=='bert':
assert args.init_checkpoint_S is not None
state_dict_S = torch.load(args.init_checkpoint_S, map_location='cpu')
state_weight = {k[5:]:v for k,v in state_dict_S.items() if k.startswith('bert.')}
missing_keys,_ = model_S.bert.load_state_dict(state_weight,strict=False)
assert len(missing_keys)==0
elif args.load_model_type=='all':
assert args.tuned_checkpoint_S is not None
state_dict_S = torch.load(args.tuned_checkpoint_S,map_location='cpu')
model_S.load_state_dict(state_dict_S)
else:
logger.info("Model is randomly initialized.")
if args.do_train:
for model_T in model_Ts:
model_T.to(device)
model_S.to(device)
if args.local_rank != -1 or n_gpu > 1:
if args.local_rank != -1:
raise NotImplementedError
elif n_gpu > 1:
if args.do_train:
model_Ts = [torch.nn.DataParallel(model_T) for model_T in model_Ts]
model_S = torch.nn.DataParallel(model_S) #,output_device=n_gpu-1)
if args.do_train:
#parameters
params = list(model_S.named_parameters())
all_trainable_params = divide_parameters(params, lr=args.learning_rate)
logger.info("Length of all_trainable_params: %d", len(all_trainable_params))
optimizer = BERTAdam(all_trainable_params,lr=args.learning_rate,
warmup=args.warmup_proportion,t_total=num_train_steps,schedule=args.schedule,
s_opt1=args.s_opt1, s_opt2=args.s_opt2, s_opt3=args.s_opt3)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Forward batch size = %d", forward_batch_size)
logger.info(" Num backward steps = %d", num_train_steps)
########### DISTILLATION ###########
train_config = TrainingConfig(
gradient_accumulation_steps = args.gradient_accumulation_steps,
ckpt_frequency = args.ckpt_frequency,
log_dir = args.output_dir,
output_dir = args.output_dir,
device = args.device)
distill_config = DistillationConfig(
temperature = args.temperature,
kd_loss_type = 'ce')
logger.info(f"{train_config}")
logger.info(f"{distill_config}")
adaptor = partial(BertForGLUESimpleAdaptor, no_logits=False, no_mask = False)
distiller = MultiTeacherDistiller(train_config = train_config,
distill_config = distill_config,
model_T = model_Ts, model_S = model_S,
adaptor_T=adaptor,
adaptor_S=adaptor)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
raise NotImplementedError
train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.forward_batch_size,drop_last=True)
callback_func = partial(predict, eval_datasets=eval_datasets, args=args)
with distiller:
distiller.train(optimizer, scheduler=None, dataloader=train_dataloader,
num_epochs=args.num_train_epochs, callback=callback_func)
if not args.do_train and args.do_predict:
res = predict(model_S,eval_datasets,step=0,args=args)
print (res)
def main():
#parse arguments
config.parse()
args = config.args
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%Y/%m/%d %H:%M:%S',
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger = logging.getLogger("Main")
#arguments check
device, n_gpu = args_check(logger, args)
if args.local_rank in [-1, 0]:
os.makedirs(args.output_dir, exist_ok=True)
if args.local_rank != -1:
logger.warning(
f"Process rank: {torch.distributed.get_rank()}, device : {args.device}, n_gpu : {args.n_gpu}, distributed training : {bool(args.local_rank!=-1)}"
)
for k, v in vars(args).items():
logger.info(f"{k}:{v}")
#set seeds
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed_all(args.random_seed)
np.random.seed(args.random_seed)
random.seed(args.random_seed)
forward_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
args.forward_batch_size = forward_batch_size
#从transformers包中导入ElectraConfig, 并加载配置
bert_config_S = ElectraConfig.from_json_file(args.bert_config_file_S)
# (args.output_encoded_layers=='true') --> True, 默认输出隐藏层的状态
bert_config_S.output_hidden_states = (args.output_encoded_layers == 'true')
#num_labels；类别个数
bert_config_S.num_labels = len(label2id_dict)
assert args.max_seq_length <= bert_config_S.max_position_embeddings
#读取数据
train_examples = None
train_dataset = None
eval_examples = None
eval_dataset = None
num_train_steps = None
# 加载Tokenizer
tokenizer = BertTokenizer(vocab_file=args.vocab_file,
do_lower_case=args.do_lower_case)
if args.local_rank not in [-1, 0]:
torch.distributed.barrier()
if args.do_train:
# 返回所有的样本和样本形成的dataset格式, dataset的格式[all_token_ids,all_input_mask,all_label_ids]
print("开始加载训练集数据")
train_examples, train_dataset = read_features(
args.train_file,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length)
if args.do_predict:
print("开始加载测试集数据")
eval_examples, eval_dataset = read_features(
args.predict_file,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length)
if args.local_rank == 0:
torch.distributed.barrier()
#Build Model and load checkpoint
model_S = ElectraForTokenClassification(bert_config_S)
#加载student模型的参数, 默认是bert类型
if args.load_model_type == 'bert':
assert args.init_checkpoint_S is not None
state_dict_S = torch.load(args.init_checkpoint_S, map_location='cpu')
#state_weight = {k[5:]:v for k,v in state_dict_S.items() if k.startswith('bert.')}
#missing_keys,_ = model_S.bert.load_state_dict(state_weight,strict=False)
missing_keys, unexpected_keys = model_S.load_state_dict(state_dict_S,
strict=False)
logger.info(f"missing keys:{missing_keys}")
logger.info(f"unexpected keys:{unexpected_keys}")
elif args.load_model_type == 'all':
assert args.tuned_checkpoint_S is not None
state_dict_S = torch.load(args.tuned_checkpoint_S, map_location='cpu')
model_S.load_state_dict(state_dict_S)
else:
logger.info("Model is randomly initialized.")
#模型放到device
model_S.to(device)
if args.do_train:
#parameters
if args.lr_decay is not None:
# 分类器层的参数 weight, bias
outputs_params = list(model_S.classifier.named_parameters())
#拆分出做学习率衰减的参数和不衰减的参数
outputs_params = divide_parameters(outputs_params,
lr=args.learning_rate)
electra_params = []
# eg: 12, encoder层共12层
n_layers = len(model_S.electra.encoder.layer)
assert n_layers == 12
for i, n in enumerate(reversed(range(n_layers))):
encoder_params = list(
model_S.electra.encoder.layer[n].named_parameters())
lr = args.learning_rate * args.lr_decay**(i + 1)
electra_params += divide_parameters(encoder_params, lr=lr)
logger.info(f"{i}:第{n}层的学习率是:{lr}")
embed_params = [
(name, value)
for name, value in model_S.electra.named_parameters()
if 'embedding' in name
]
logger.info(f"{[name for name,value in embed_params]}")
lr = args.learning_rate * args.lr_decay**(n_layers + 1)
electra_params += divide_parameters(embed_params, lr=lr)
logger.info(f"embedding层的学习率 lr:{lr}")
all_trainable_params = outputs_params + electra_params
assert sum(map(lambda x:len(x['params']), all_trainable_params))==len(list(model_S.parameters())),\
(sum(map(lambda x:len(x['params']), all_trainable_params)), len(list(model_S.parameters())))
else:
params = list(model_S.named_parameters())
all_trainable_params = divide_parameters(params,
lr=args.learning_rate)
logger.info("可训练的参数all_trainable_params共有: %d",
len(all_trainable_params))
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
train_sampler = DistributedSampler(train_dataset)
#生成dataloader
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.forward_batch_size,
drop_last=True)
# 根据epoch计算出运行多少steps
num_train_steps = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
optimizer = AdamW(all_trainable_params,
lr=args.learning_rate,
correct_bias=False)
if args.official_schedule == 'const':
# 常数学习率
scheduler_class = get_constant_schedule_with_warmup
scheduler_args = {
'num_warmup_steps':
int(args.warmup_proportion * num_train_steps)
}
#scheduler = get_constant_schedule_with_warmup(optimizer, num_warmup_steps=int(args.warmup_proportion*num_train_steps))
elif args.official_schedule == 'linear':
# 线性学习率
scheduler_class = get_linear_schedule_with_warmup
# warmup多少步，10%的step是warmup的
scheduler_args = {
'num_warmup_steps':
int(args.warmup_proportion * num_train_steps),
'num_training_steps': num_train_steps
}
#scheduler = get_linear_schedule_with_warmup(optimizer,num_warmup_steps=int(args.warmup_proportion*num_train_steps), num_training_steps = num_train_steps)
elif args.official_schedule == 'const_nowarmup':
scheduler_class = get_constant_schedule
scheduler_args = {}
else:
raise NotImplementedError
logger.warning("***** 开始 训练 *****")
logger.warning("local_rank %d 原样本数 = %d", args.local_rank,
len(train_examples))
logger.warning("local_rank %d split之后的样本数 = %d", args.local_rank,
len(train_dataset))
logger.warning("local_rank %d 前向 batch size = %d", args.local_rank,
forward_batch_size)
logger.warning("local_rank %d 训练的steps = %d", args.local_rank,
num_train_steps)
########### TRAINING ###########
train_config = TrainingConfig(
gradient_accumulation_steps=args.gradient_accumulation_steps,
ckpt_frequency=args.ckpt_frequency, #保存频率
log_dir=args.output_dir,
output_dir=args.output_dir,
device=args.device,
fp16=args.fp16,
local_rank=args.local_rank)
logger.info(f"训练的配置文件:")
logger.info(f"{train_config}")
# 初始化训练器
distiller = BasicTrainer(
train_config=train_config,
model=model_S,
adaptor=ElectraForTokenClassificationAdaptorTraining)
# 初始化callback函数，使用的ddp_predict函数进行评估
callback_func = partial(ddp_predict,
eval_examples=eval_examples,
eval_dataset=eval_dataset,
args=args)
with distiller:
distiller.train(optimizer,
scheduler_class=scheduler_class,
scheduler_args=scheduler_args,
max_grad_norm=1.0,
dataloader=train_dataloader,
num_epochs=args.num_train_epochs,
callback=callback_func)
if not args.do_train and args.do_predict:
res = ddp_predict(model_S,
eval_examples,
eval_dataset,
step=0,
args=args)
print(res)