def main(args):
init_logger()
set_seed(args)
tokenizer = load_tokenizer(args)
train_dataset = load_and_cache_examples(args, tokenizer, mode="train")
dev_dataset = load_and_cache_examples(args, tokenizer, mode="dev")
test_dataset = load_and_cache_examples(args, tokenizer, mode="test")
trainer = Trainer(args, train_dataset, dev_dataset, test_dataset)
if args.do_train:
trainer.train()
if args.do_eval:
trainer.load_model()
print('Test Result:')
trainer.eval("test")
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 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 main():
parser = argparse.ArgumentParser()
parser.add_argument("--model_type",
default='bert',
type=str,
help="Model type selected in the list: " +
", ".join(MODEL_CLASSES.keys()))
parser.add_argument(
"--model_name_or_path",
default='bert-base-uncased',
type=str,
help="Path to pre-trained model or shortcut name selected in the list: "
+ ", ".join(ALL_MODELS))
parser.add_argument(
"--output_dir",
default='../output_mc',
type=str,
help=
"The output directory where the model checkpoints and predictions will be written."
)
parser.add_argument("--raw_data_dir", default='../data_mc', type=str)
parser.add_argument(
"--config_name",
default="",
type=str,
help="Pretrained config name or path if not the same as model_name")
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help="Pretrained tokenizer name or path if not the same as model_name")
parser.add_argument(
"--max_seq_length",
default=384,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. Sequences "
"longer than this will be truncated, and sequences shorter than this will be padded."
)
parser.add_argument("--task_name", default='DREAM')
parser.add_argument("--pre_model_dir",
default='2020-03-12-10-58-checkpoint-3048')
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--do_test",
action='store_true',
help='Whether to run test on the test set')
parser.add_argument(
"--evaluate_during_training",
action='store_true',
help="Rul evaluation during training at each logging step.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--per_gpu_train_batch_size",
default=8,
type=int,
help="Batch size per GPU/CPU for training.")
parser.add_argument("--per_gpu_eval_batch_size",
default=8,
type=int,
help="Batch size per GPU/CPU for evaluation.")
parser.add_argument("--learning_rate",
default=3e-5,
type=float,
help="The initial learning rate for Adam.")
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("--weight_decay",
default=0.0,
type=float,
help="Weight deay if we apply some.")
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,
help="Max gradient norm.")
parser.add_argument("--num_train_epochs",
default=2.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--max_steps",
default=-1,
type=int,
help=
"If > 0: set total number of training steps to perform. Override num_train_epochs."
)
parser.add_argument("--warmup_proportion",
default=0.1,
type=float,
help="Linear warmup over warmup_steps.")
parser.add_argument(
"--verbose_logging",
action='store_true',
help=
"If true, all of the warnings related to data processing will be printed. "
"A number of warnings are expected for a normal SQuAD evaluation.")
parser.add_argument(
"--eval_all_checkpoints",
action='store_true',
help=
"Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number"
)
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--overwrite_output_dir',
action='store_true',
help="Overwrite the content of the output directory")
parser.add_argument(
'--overwrite_cache',
action='store_true',
help="Overwrite the cached training and evaluation sets")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument(
'--fp16',
action='store_true',
help=
"Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit"
)
parser.add_argument(
'--fp16_opt_level',
type=str,
default='O1',
help=
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html")
args = parser.parse_args()
args.checkpoint = os.path.join(args.output_dir, args.pre_model_dir)
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir
) and args.do_train and not args.overwrite_output_dir:
logger.info(
"Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome."
.format(args.output_dir))
# Setup CUDA, GPU & distributed training
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend='nccl')
args.n_gpu = 1
args.device = device
# Set seed
set_seed(args)
args.model_type = args.model_type.lower()
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path)
config.output_hidden_states = True
config.num_options = int(
MULTIPLE_CHOICE_TASKS_NUM_LABELS[args.task_name.lower()])
tokenizer = tokenizer_class.from_pretrained(
args.tokenizer_name
if args.tokenizer_name else args.model_name_or_path,
do_lower_case=args.do_lower_case)
post_model = Post_MV(args, config)
post_model.to(args.device)
logger.info("Training/evaluation parameters %s", args)
if args.fp16:
try:
import apex
apex.amp.register_half_function(torch, 'einsum')
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
if args.do_train:
logging.getLogger("transformers.tokenization_utils").setLevel(
logging.ERROR) # Reduce logging
train_dataset = load_and_cache_examples(args,
task=args.task_name,
tokenizer=tokenizer,
evaluate=False)
global_step, tr_loss = train_process(args, train_dataset, post_model,
tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step,
tr_loss)
if args.do_test:
logging.getLogger("transformers.modeling_utils").setLevel(
logging.WARN) # Reduce logging
logging.getLogger("transformers.configuration_utils").setLevel(
logging.WARN) # Reduce logging
logging.getLogger("transformers.tokenization_utils").setLevel(
logging.ERROR) # Reduce logging
checkpoint = os.path.join(args.output_dir, args.pre_model_dir + '-TIE')
logger.info(" Load model from %s", checkpoint)
post_model.load_state_dict(
torch.load(os.path.join(checkpoint, 'pytorch_model.bin')))
post_model.to(args.device)
task_string = [
'', '-Add1OtherTruth2Opt', '-Add2OtherTruth2Opt',
'-Add1PasSent2Opt', '-Add1NER2Pass'
]
task_string = [args.task_name + item for item in task_string]
result = evaluate(args, task_string, post_model, tokenizer, test=True)
def evaluate(args, eval_task_names, model, tokenizer, test=False):
results = []
table = PrettyTable()
table.add_column(' ', ['Accuracy'])
for eval_task in eval_task_names:
eval_dataset = load_and_cache_examples(args,
eval_task,
tokenizer,
evaluate=not test,
test=test)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(
1, args.n_gpu)
# Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(eval_dataset)
eval_dataloader = DataLoader(eval_dataset,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
nb_eval_steps = 0
preds = None
out_label_ids = None
for batch in eval_dataloader:
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {
'input_ids':
batch[0],
'attention_mask':
batch[1],
'token_type_ids':
batch[2] if args.model_type in ['bert', 'xlnet'] else None,
# XLM don't use segment_ids
'input_ids_np':
batch[3],
'attention_mask_np':
batch[4],
'token_type_ids_np':
batch[5],
'labels':
batch[6]
}
output = model(**inputs)
logits = output[0]
nb_eval_steps += 1
if preds is None:
preds = logits.detach().cpu().numpy()
out_label_ids = inputs['labels'].detach().cpu().numpy()
else:
preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
out_label_ids = np.append(
out_label_ids,
inputs['labels'].detach().cpu().numpy(),
axis=0)
preds = np.argmax(preds, axis=1)
acc = simple_accuracy(preds, out_label_ids)
result = {"task_name": eval_task, "eval_acc": acc}
results.append(result)
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
table.add_column(eval_task, [round(acc * 100, 2)])
print(table)
return results
def main(path_to_train_data, path_to_validation_data):
args = Args()
df_trn, df_val = dataset.make_dataset(path_to_train_data,
path_to_validation_data)
if args.should_continue:
sorted_checkpoints = _sorted_checkpoints(args)
if len(sorted_checkpoints) == 0:
raise ValueError(
"Used --should_continue but no checkpoint was found in --output_dir."
)
else:
args.model_name_or_path = sorted_checkpoints[-1]
if (os.path.exists(args.output_dir) and os.listdir(args.output_dir)
and args.do_train and not args.overwrite_output_dir
and not args.should_continue):
raise ValueError(
"Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome."
.format(args.output_dir))
# Setup CUDA, GPU & distributed training
device = torch.device("cuda")
args.n_gpu = torch.cuda.device_count()
args.device = device
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
args.local_rank,
device,
args.n_gpu,
bool(args.local_rank != -1),
args.fp16,
)
# Set seed
utils.set_seed(args)
config = AutoConfig.from_pretrained(args.config_name,
cache_dir=args.cache_dir)
tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name,
cache_dir=args.cache_dir)
model = AutoModelWithLMHead.from_pretrained(
args.model_name_or_path,
from_tf=False,
config=config,
cache_dir=args.cache_dir,
)
model.to(args.device)
logger.info("Training/evaluation parameters %s", args)
# Training
if args.do_train:
train_dataset = utils.load_and_cache_examples(args,
tokenizer,
df_trn,
df_val,
evaluate=False)
global_step, tr_loss = train.train(args, train_dataset, model,
tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step,
tr_loss)
# Saving best-practices: if you use save_pretrained for the model and tokenizer, you can reload them using from_pretrained()
if args.do_train:
# Create output directory if needed
os.makedirs(args.output_dir, exist_ok=True)
logger.info("Saving model checkpoint to %s", args.output_dir)
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
# They can then be reloaded using `from_pretrained()`
model_to_save = (model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(args.output_dir)
tokenizer.save_pretrained(args.output_dir)
# Good practice: save your training arguments together with the trained model
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
# Load a trained model and vocabulary that you have fine-tuned
model = AutoModelWithLMHead.from_pretrained(args.output_dir)
tokenizer = AutoTokenizer.from_pretrained(args.output_dir)
model.to(args.device)
# Evaluation
results = {}
if args.do_eval and args.local_rank in [-1, 0]:
checkpoints = [args.output_dir]
if args.eval_all_checkpoints:
checkpoints = list(
os.path.dirname(c) for c in sorted(
glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME,
recursive=True)))
logging.getLogger("transformers.modeling_utils").setLevel(
logging.WARN) # Reduce logging
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
global_step = checkpoint.split(
"-")[-1] if len(checkpoints) > 1 else ""
prefix = checkpoint.split(
"/")[-1] if checkpoint.find("checkpoint") != -1 else ""
model = AutoModelWithLMHead.from_pretrained(checkpoint)
model.to(args.device)
result = evaluate.evaluate(args,
model,
tokenizer,
df_trn,
df_val,
prefix=prefix)
result = dict(
(k + "_{}".format(global_step), v) for k, v in result.items())
results.update(result)
return results
def main():
args = parse_args()
if args.doc_stride >= args.max_seq_length - args.max_query_length:
logger.warning(
"WARNING - You've set a doc stride which may be superior to the document length in some "
"examples. This could result in errors when building features from the examples. Please reduce the doc "
"stride or increase the maximum length to ensure the features are correctly built."
)
if (os.path.exists(args.output_dir) and os.listdir(args.output_dir)
and args.do_train and not args.overwrite_output_dir):
raise ValueError(
"Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome."
.format(args.output_dir))
device = torch.device(
"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
args.device = device
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.warning(
"Device: %s, n_gpu: %s",
device,
args.n_gpu,
)
# Set seed
set_seed(args)
args.model_type = args.model_type.lower()
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
config = config_class.from_pretrained(
args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None,
# [A] output attentions
output_attentions=True,
# [A] output hidden states
output_hidden_states=True)
tokenizer = tokenizer_class.from_pretrained(
args.model_name_or_path,
do_lower_case=args.do_lower_case,
cache_dir=args.cache_dir if args.cache_dir else None,
)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None,
)
model.to(args.device)
logger.info("Training/evaluation parameters %s", args)
# Training
if args.do_train:
train_dataset = load_and_cache_examples(args,
tokenizer,
evaluate=False,
output_examples=False)
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step,
tr_loss)
# Save the trained model and the tokenizer
if args.do_train:
# Create output directory if needed
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
logger.info("Saving model checkpoint to %s", args.output_dir)
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
# They can then be reloaded using `from_pretrained()`
# Take care of distributed/parallel training
model_to_save = model.module if hasattr(model, "module") else model
model_to_save.save_pretrained(args.output_dir)
tokenizer.save_pretrained(args.output_dir)
# Good practice: save your training arguments together with the trained model
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
# Load a trained model and vocabulary that you have fine-tuned
model = model_class.from_pretrained(
args.output_dir) # , force_download=True)
tokenizer = tokenizer_class.from_pretrained(
args.output_dir, do_lower_case=args.do_lower_case)
model.to(args.device)
# Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
results = {}
if args.do_eval:
if args.do_train:
logger.info(
"Loading checkpoints saved during training for evaluation")
checkpoints = [args.output_dir]
if args.eval_all_checkpoints:
checkpoints = list(
os.path.dirname(c) for c in sorted(
glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME,
recursive=True)))
logging.getLogger("transformers.modeling_utils").setLevel(
logging.WARN) # Reduce model loading logs
else:
logger.info("Loading checkpoint %s for evaluation",
args.model_name_or_path)
checkpoints = [args.model_name_or_path]
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
# Reload the model
global_step = checkpoint.split(
"-")[-1] if len(checkpoints) > 1 else ""
model = model_class.from_pretrained(
checkpoint) # , force_download=True)
model.to(args.device)
# Evaluate
result = evaluate(args, model, tokenizer, prefix=global_step)
result = dict(
(k + ("_{}".format(global_step) if global_step else ""), v)
for k, v in result.items())
results.update(result)
logger.info("Results: {}".format(results))
return results
def evaluate(args, model, tokenizer, prefix=""):
dataset, examples, features = load_and_cache_examples(args,
tokenizer,
evaluate=True,
output_examples=True)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
eval_sampler = SequentialSampler(dataset)
eval_dataloader = DataLoader(dataset,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
# multi-gpu evaluate
if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
model = torch.nn.DataParallel(model)
# Eval!
logger.info("***** Running evaluation {} *****".format(prefix))
logger.info(" Num examples = %d", len(dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
all_results = []
start_time = timeit.default_timer()
for batch in tqdm(eval_dataloader, desc="Evaluating"):
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"token_type_ids": batch[2],
}
example_indices = batch[3]
outputs = model(**inputs)
for i, example_index in enumerate(example_indices):
eval_feature = features[example_index.item()]
unique_id = int(eval_feature.unique_id)
# [A] bug fix
output = [
to_list(output[i]) for output in (outputs[0], outputs[1])
]
start_logits, end_logits = output
result = SquadResult(unique_id, start_logits, end_logits)
all_results.append(result)
evalTime = timeit.default_timer() - start_time
logger.info(" Evaluation done in total %f secs (%f sec per example)",
evalTime, evalTime / len(dataset))
# Compute predictions
output_prediction_file = os.path.join(args.output_dir,
"predictions_{}.json".format(prefix))
output_nbest_file = os.path.join(
args.output_dir, "nbest_predictions_{}.json".format(prefix))
if args.version_2_with_negative:
output_null_log_odds_file = os.path.join(
args.output_dir, "null_odds_{}.json".format(prefix))
else:
output_null_log_odds_file = None
predictions = compute_predictions_logits(
examples,
features,
all_results,
args.n_best_size,
args.max_answer_length,
args.do_lower_case,
output_prediction_file,
output_nbest_file,
output_null_log_odds_file,
args.verbose_logging,
args.version_2_with_negative,
args.null_score_diff_threshold,
tokenizer,
)
# Compute the F1 and exact scores.
results = squad_evaluate(examples, predictions)
return results
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():
#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():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--data_dir",
default="./data",
type=str,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument("--model_type",
default="bert",
type=str,
help="Model type selected in the list: " +
", ".join(MODEL_CLASSES.keys()))
parser.add_argument(
"--model_name_or_path",
default="./roberta-base/roberta-base-pytorch_model.bin",
type=str,
help="Path to pre-trained model or shortcut name selected in the list: "
+ ", ".join(ALL_MODELS))
parser.add_argument(
"--task_name",
default="multi_task",
type=str,
help="The name of the task to train selected in the list: " +
", ".join(processors.keys()))
parser.add_argument(
"--output_dir",
default='./check_points',
type=str,
help=
"The output directory where the model predictions and checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--config_name",
default="./roberta-base/roberta-base-config.json",
type=str,
help="Pretrained config name or path if not the same as model_name")
parser.add_argument(
"--tokenizer_name",
default="./roberta-base/roberta-base-vocab.json",
type=str,
help="Pretrained tokenizer name or path if not the same as model_name")
parser.add_argument(
"--cache_dir",
default="",
type=str,
help=
"Where do you want to store the pre-trained models downloaded from s3")
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after tokenization. Sequences longer "
"than this will be truncated, sequences shorter will be padded.")
parser.add_argument("--do_train",
default=True,
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
default=True,
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--do_test",
default=False,
action='store_true',
help="Whether to run test on the test set.")
parser.add_argument(
"--evaluate_during_training",
default=True,
action='store_true',
help="Rul evaluation during training at each logging step.")
parser.add_argument(
"--do_lower_case",
default=True,
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--per_gpu_train_batch_size",
default=16,
type=int,
help="Batch size per GPU/CPU for training.")
parser.add_argument("--per_gpu_eval_batch_size",
default=16,
type=int,
help="Batch size per GPU/CPU for evaluation.")
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("--learning_rate",
default=5e-6,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay",
default=0.0,
type=float,
help="Weight deay if we apply some.")
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,
help="Max gradient norm.")
parser.add_argument("--num_train_epochs",
default=10.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--max_steps",
default=-1,
type=int,
help=
"If > 0: set total number of training steps to perform. Override num_train_epochs."
)
parser.add_argument("--warmup_steps",
default=0,
type=int,
help="Linear warmup over warmup_steps.")
parser.add_argument('--logging_steps',
type=int,
default=2975,
help="Log every X updates steps.")
parser.add_argument('--save_steps',
type=int,
default=2975,
help="Save checkpoint every X updates steps.")
parser.add_argument(
"--eval_all_checkpoints",
action='store_true',
help=
"Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number"
)
parser.add_argument("--no_cuda",
action='store_true',
help="Avoid using CUDA when available")
parser.add_argument('--overwrite_output_dir',
action='store_true',
help="Overwrite the content of the output directory")
parser.add_argument(
'--overwrite_cache',
action='store_true',
help="Overwrite the cached training and evaluation sets")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--fp16',
action='store_true',
help=
"Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit"
)
parser.add_argument(
'--fp16_opt_level',
type=str,
default='O1',
help=
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="For distributed training: local_rank")
parser.add_argument("--mPos", default=1.5, type=float)
parser.add_argument("--mNeg", default=1.5, type=float)
parser.add_argument("--gamma", default=0.05, type=float)
parser.add_argument("--all_task", default=True)
parser.add_argument("--aggression_attack_task", default=False)
parser.add_argument("--aggression_toxicity_task", default=False)
parser.add_argument("--attack_toxicity_task", default=False)
args = parser.parse_args()
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir
) and args.do_train and not args.overwrite_output_dir:
raise ValueError(
"Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome."
.format(args.output_dir))
# Setup CUDA, GPU & distributed training
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda:0" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
#args.n_gpu = torch.cuda.device_count()
args.n_gpu = 1
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda:0", args.local_rank)
torch.distributed.init_process_group(backend='nccl')
args.n_gpu = 1
args.device = device
# Setup logging
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
args.local_rank, device, args.n_gpu, bool(args.local_rank != -1),
args.fp16)
# Set seed
set_seed(args)
args.task_name = args.task_name.lower()
if args.task_name not in processors:
raise ValueError("Task not found: %s" % (args.task_name))
processor = processors[args.task_name]()
args.output_mode = output_modes[args.task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
# Load pretrained model and tokenizer
if args.local_rank not in [-1, 0]:
torch.distributed.barrier(
) # Make sure only the first process in distributed training will download model & vocab
args.model_type = args.model_type.lower()
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
'''
bert_config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_labels, finetuning_task=args.task_name)
bert_model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path),
config=bert_config)
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.bert_vocab,
do_lower_case=args.do_lower_case)
'''
if args.model_type == "bert":
args.config_name = "./bert-base-uncased/bert-base-uncased-config.json"
args.tokenizer_name = "./bert-base-uncased/bert-base-uncased-vocab.txt"
args.model_name_or_path = "./bert-base-uncased/bert-base-uncased-pytorch_model.bin"
elif args.model_type == "xlnet":
args.config_name = "./xlnet-base-cased/xlnet-base-cased-config.json"
args.tokenizer_name = "./xlnet-base-cased/xlnet-base-cased-spiece.model"
args.model_name_or_path = "./xlnet-base-cased/xlnet-base-cased-pytorch_model.bin"
elif args.model_type == "roberta":
args.config_name = "./roberta-base/roberta-base-config.json"
args.tokenizer_name = "./roberta-base/roberta-base-vocab.json"
args.model_name_or_path = "./roberta-base/roberta-base-pytorch_model.bin"
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_labels,
finetuning_task=args.task_name)
tokenizer = tokenizer_class.from_pretrained(
args.tokenizer_name if args.tokenizer_name else args.bert_vocab)
model = Multi_Model(args, config)
if args.local_rank == 0:
torch.distributed.barrier(
) # Make sure only the first process in distributed training will download model & vocab
model.to(args.device)
logger.info("Training/evaluation parameters %s", args)
# Training
if args.do_train:
train_dataset = load_and_cache_examples(args,
args.task_name,
tokenizer,
evaluate=False)
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step,
tr_loss)
# test
results = {}
if args.do_test and args.local_rank in [-1, 0]:
tokenizer = tokenizer_class.from_pretrained(
args.tokenizer_name, do_lower_case=args.do_lower_case)
logger.info("test the model")
checkpoint = './check_points/checkpoint-200/model.pt'
prefix = checkpoint.split(
'/')[-1] if checkpoint.find('checkpoint') != -1 else ""
model.load_state_dict(torch.load(checkpoint))
model.to(args.device)
aggression_results, attack_results, toxicity_results = evaluate(
args, model, tokenizer, prefix=prefix)
def evaluate(args, model, tokenizer, prefix=""):
# Loop to handle MNLI double evaluation (matched, mis-matched)
eval_task_names = (args.task_name, )
eval_outputs_dirs = (args.output_dir, )
aggression_results = {}
attack_results = {}
toxicity_results = {}
for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
eval_dataset = load_and_cache_examples(args,
eval_task,
tokenizer,
evaluate=True)
if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
os.makedirs(eval_output_dir)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(
1, args.n_gpu)
# Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(
eval_dataset) if args.local_rank == -1 else DistributedSampler(
eval_dataset)
eval_dataloader = DataLoader(eval_dataset,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
if args.model_type in ["bert", "xlnet"]:
aggression_tensor = torch.tensor(tokenizer.encode("aggression"),
dtype=torch.long).to(args.device)
attack_tensor = torch.tensor(tokenizer.encode("attack"),
dtype=torch.long).to(args.device)
toxicity_tensor = torch.tensor(tokenizer.encode("toxicity"),
dtype=torch.long).to(args.device)
elif args.model_type == "roberta":
aggression_tensor = torch.tensor([0],
dtype=torch.long).to(args.device)
attack_tensor = torch.tensor([1], dtype=torch.long).to(args.device)
toxicity_tensor = torch.tensor([2],
dtype=torch.long).to(args.device)
char_vocab = get_char_vocab()
aggression_char_ids = char2ids("aggression", char_vocab)
attack_char_ids = char2ids("attack", char_vocab)
toxicity_char_ids = char2ids("toxicity", char_vocab)
aggression_char_tenor = torch.tensor(aggression_char_ids,
dtype=torch.long).to(args.device)
attack_char_tenor = torch.tensor(attack_char_ids,
dtype=torch.long).to(args.device)
toxicity_char_tenor = torch.tensor(toxicity_char_ids,
dtype=torch.long).to(args.device)
# Eval!
logger.info("***** Running evaluation {} *****".format(prefix))
logger.info(" Num examples = %d", len(eval_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
eval_loss = 0.0
nb_eval_steps = 0
aggression_preds = None
attack_preds = None
toxicity_preds = None
aggression_out_label_ids = None
attack_out_label_ids = None
toxicity_out_label_ids = None
for batch in tqdm(eval_dataloader, desc="Evaluating"):
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {
'input_ids': batch[0],
'attention_mask': batch[1],
'aggression_labels': batch[3],
'attack_labels': batch[4],
'toxicity_labels': batch[5],
'aggression_tensor': aggression_tensor,
'attack_tensor': attack_tensor,
'toxicity_tensor': toxicity_tensor,
'aggression_char_tensor': aggression_char_tenor,
'attack_char_tensor': attack_char_tenor,
'toxicity_char_tensor': toxicity_char_tenor
}
if args.model_type != 'distilbert':
inputs['token_type_ids'] = batch[2] if args.model_type in [
'bert', 'xlnet'
] else None # XLM, DistilBERT and RoBERTa don't use segment_ids
if args.all_task:
aggression_logits, attack_logits, toxicity_logits, tmp_eval_loss, _, _, _ = model(
**inputs)
if args.aggression_attack_task:
aggression_logits, attack_logits, tmp_eval_loss = model(
**inputs)
if args.aggression_toxicity_task:
aggression_logits, toxicity_logits, tmp_eval_loss = model(
**inputs)
if args.attack_toxicity_task:
attack_logits, toxicity_logits, tmp_eval_loss = model(
**inputs)
eval_loss += tmp_eval_loss.mean().item()
nb_eval_steps += 1
if aggression_preds is None and attack_preds is None and toxicity_preds is None:
if args.all_task:
aggression_preds = aggression_logits.detach().cpu().numpy()
aggression_out_label_ids = inputs[
'aggression_labels'].detach().cpu().numpy()
attack_preds = attack_logits.detach().cpu().numpy()
attack_out_label_ids = inputs['attack_labels'].detach(
).cpu().numpy()
toxicity_preds = toxicity_logits.detach().cpu().numpy()
toxicity_out_label_ids = inputs['toxicity_labels'].detach(
).cpu().numpy()
elif args.aggression_attack_task:
aggression_preds = aggression_logits.detach().cpu().numpy()
aggression_out_label_ids = inputs[
'aggression_labels'].detach().cpu().numpy()
attack_preds = attack_logits.detach().cpu().numpy()
attack_out_label_ids = inputs['attack_labels'].detach(
).cpu().numpy()
elif args.aggression_toxicity_task:
aggression_preds = aggression_logits.detach().cpu().numpy()
aggression_out_label_ids = inputs[
'aggression_labels'].detach().cpu().numpy()
toxicity_preds = toxicity_logits.detach().cpu().numpy()
toxicity_out_label_ids = inputs['toxicity_labels'].detach(
).cpu().numpy()
elif args.attack_toxicity_task:
attack_preds = attack_logits.detach().cpu().numpy()
attack_out_label_ids = inputs['attack_labels'].detach(
).cpu().numpy()
toxicity_preds = toxicity_logits.detach().cpu().numpy()
toxicity_out_label_ids = inputs['toxicity_labels'].detach(
).cpu().numpy()
else:
if args.all_task:
aggression_preds = np.append(
aggression_preds,
aggression_logits.detach().cpu().numpy(),
axis=0)
aggression_out_label_ids = np.append(
aggression_out_label_ids,
inputs['aggression_labels'].detach().cpu().numpy(),
axis=0)
attack_preds = np.append(
attack_preds,
attack_logits.detach().cpu().numpy(),
axis=0)
attack_out_label_ids = np.append(
attack_out_label_ids,
inputs['attack_labels'].detach().cpu().numpy(),
axis=0)
toxicity_preds = np.append(
toxicity_preds,
toxicity_logits.detach().cpu().numpy(),
axis=0)
toxicity_out_label_ids = np.append(
toxicity_out_label_ids,
inputs['toxicity_labels'].detach().cpu().numpy(),
axis=0)
elif args.aggression_attack_task:
aggression_preds = np.append(
aggression_preds,
aggression_logits.detach().cpu().numpy(),
axis=0)
aggression_out_label_ids = np.append(
aggression_out_label_ids,
inputs['aggression_labels'].detach().cpu().numpy(),
axis=0)
attack_preds = np.append(
attack_preds,
attack_logits.detach().cpu().numpy(),
axis=0)
attack_out_label_ids = np.append(
attack_out_label_ids,
inputs['attack_labels'].detach().cpu().numpy(),
axis=0)
elif args.aggression_toxicity_task:
aggression_preds = np.append(
aggression_preds,
aggression_logits.detach().cpu().numpy(),
axis=0)
aggression_out_label_ids = np.append(
aggression_out_label_ids,
inputs['aggression_labels'].detach().cpu().numpy(),
axis=0)
toxicity_preds = np.append(
toxicity_preds,
toxicity_logits.detach().cpu().numpy(),
axis=0)
toxicity_out_label_ids = np.append(
toxicity_out_label_ids,
inputs['toxicity_labels'].detach().cpu().numpy(),
axis=0)
elif args.attack_toxicity_task:
attack_preds = np.append(
attack_preds,
attack_logits.detach().cpu().numpy(),
axis=0)
attack_out_label_ids = np.append(
attack_out_label_ids,
inputs['attack_labels'].detach().cpu().numpy(),
axis=0)
toxicity_preds = np.append(
toxicity_preds,
toxicity_logits.detach().cpu().numpy(),
axis=0)
toxicity_out_label_ids = np.append(
toxicity_out_label_ids,
inputs['toxicity_labels'].detach().cpu().numpy(),
axis=0)
eval_loss = eval_loss / nb_eval_steps
if args.all_task:
aggression_preds = np.argmax(aggression_preds, axis=1)
attack_preds = np.argmax(attack_preds, axis=1)
toxicity_preds = np.argmax(toxicity_preds, axis=1)
aggression_result = compute_metrics(eval_task, aggression_preds,
aggression_out_label_ids)
aggression_results.update(aggression_result)
attack_result = compute_metrics(eval_task, attack_preds,
attack_out_label_ids)
attack_results.update(attack_result)
toxicity_result = compute_metrics(eval_task, toxicity_preds,
toxicity_out_label_ids)
toxicity_results.update(toxicity_result)
logger.info(
"***** Eval aggression results {} *****".format(prefix))
for key in sorted(aggression_result.keys()):
logger.info(" %s = %s", key, str(aggression_result[key]))
logger.info("***** Eval attack results {} *****".format(prefix))
for key in sorted(attack_result.keys()):
logger.info(" %s = %s", key, str(attack_result[key]))
logger.info("***** Eval toxicity results {} *****".format(prefix))
for key in sorted(toxicity_result.keys()):
logger.info(" %s = %s", key, str(toxicity_result[key]))
elif args.aggression_attack_task:
aggression_preds = np.argmax(aggression_preds, axis=1)
attack_preds = np.argmax(attack_preds, axis=1)
aggression_result = compute_metrics(eval_task, aggression_preds,
aggression_out_label_ids)
aggression_results.update(aggression_result)
attack_result = compute_metrics(eval_task, attack_preds,
attack_out_label_ids)
attack_results.update(attack_result)
logger.info(
"***** Eval aggression results {} *****".format(prefix))
for key in sorted(aggression_result.keys()):
logger.info(" %s = %s", key, str(aggression_result[key]))
logger.info("***** Eval attack results {} *****".format(prefix))
for key in sorted(attack_result.keys()):
logger.info(" %s = %s", key, str(attack_result[key]))
elif args.aggression_toxicity_task:
aggression_preds = np.argmax(aggression_preds, axis=1)
toxicity_preds = np.argmax(toxicity_preds, axis=1)
aggression_result = compute_metrics(eval_task, aggression_preds,
aggression_out_label_ids)
aggression_results.update(aggression_result)
toxicity_result = compute_metrics(eval_task, toxicity_preds,
toxicity_out_label_ids)
toxicity_results.update(toxicity_result)
logger.info(
"***** Eval aggression results {} *****".format(prefix))
for key in sorted(aggression_result.keys()):
logger.info(" %s = %s", key, str(aggression_result[key]))
logger.info("***** Eval toxicity results {} *****".format(prefix))
for key in sorted(toxicity_result.keys()):
logger.info(" %s = %s", key, str(toxicity_result[key]))
elif args.attack_toxicity_task:
attack_preds = np.argmax(attack_preds, axis=1)
toxicity_preds = np.argmax(toxicity_preds, axis=1)
attack_result = compute_metrics(eval_task, attack_preds,
attack_out_label_ids)
attack_results.update(attack_result)
toxicity_result = compute_metrics(eval_task, toxicity_preds,
toxicity_out_label_ids)
toxicity_results.update(toxicity_result)
logger.info("***** Eval attack results {} *****".format(prefix))
for key in sorted(attack_result.keys()):
logger.info(" %s = %s", key, str(attack_result[key]))
logger.info("***** Eval toxicity results {} *****".format(prefix))
for key in sorted(toxicity_result.keys()):
logger.info(" %s = %s", key, str(toxicity_result[key]))
if args.all_task:
return aggression_results, attack_results, toxicity_results
elif args.aggression_attack_task:
return aggression_results, attack_results
elif args.aggression_toxicity_task:
return aggression_results, toxicity_results
elif args.attack_toxicity_task:
return attack_results, toxicity_results
def main():
parser = argparse.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("--model_name_or_path",
default=None,
type=str,
required=True,
help="")
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("--train_language",
default=None,
type=str,
required=True)
parser.add_argument("--model_type", type=str, required=True)
## Other parameters
parser.add_argument(
"--max_seq_length",
default=256,
type=int,
help=
"The maximum total input sequence length after tokenization. Sequences longer than this will be truncated, sequences shorter will be padded."
)
parser.add_argument("--num_labels", default=2, type=int)
parser.add_argument("--overwrite_cache", action='store_true')
parser.add_argument("--config_name", default=None, type=str)
parser.add_argument("--tokenizer_name", default=None, type=str)
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_lower_case", action='store_true')
parser.add_argument("--do_test",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument('--from_tf',
action='store_true',
help='whether load tensorflow weights')
parser.add_argument("--do_eval_train",
action='store_true',
help="Whether to run eval on the train set.")
parser.add_argument("--per_gpu_train_batch_size",
default=8,
type=int,
help="Batch size per GPU/CPU for training.")
parser.add_argument("--per_gpu_eval_batch_size",
default=8,
type=int,
help="Batch size per GPU/CPU for evaluation.")
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("--local_rank", default=-1, type=int)
parser.add_argument("--learning_rate",
default=1e-4,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay",
default=0.0,
type=float,
help="Weight deay if we apply some.")
parser.add_argument("--adam_epsilon",
default=1e-8,
type=float,
help="Epsilon for Adam optimizer.")
parser.add_argument("--eval_steps", default=-1, type=int, help="")
parser.add_argument("--train_steps", default=-1, type=int, help="")
parser.add_argument("--warmup_steps",
default=0,
type=int,
help="Linear warmup over warmup_steps.")
parser.add_argument('--seed',
type=int,
default=1,
help="random seed for initialization")
args = parser.parse_args()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
args.n_gpu = torch.cuda.device_count()
args.device = device
# Set seed
set_seed(args)
args.model_type = args.model_type.lower()
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
num_labels=args.num_labels)
tokenizer = tokenizer_class(tokens_folder=args.model_name_or_path)
# tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
logger.info("Training/evaluation parameters %s", args)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
# Training
if args.do_train:
# Prepare model
model = model_class.from_pretrained(args.model_name_or_path,
from_tf=args.from_tf,
config=config)
fgm = FGM(model)
model.to(args.device)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
args.train_batch_size = args.per_gpu_train_batch_size * max(
1, args.n_gpu)
train_dataset = load_and_cache_examples(args, tokenizer, is_training=1)
train_sampler = RandomSampler(train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size //
args.gradient_accumulation_steps)
num_train_optimization_steps = args.train_steps
# Prepare optimizer
param_optimizer = list(model.named_parameters())
# hack to remove pooler, which is not used
# thus it produce None grad that break apex
param_optimizer = [n for n in param_optimizer]
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(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
global_step = 0
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
logger.info(" Num steps = %d", num_train_optimization_steps)
best_acc = 0
min_loss = math.inf
model.train()
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
bar = tqdm(range(num_train_optimization_steps),
total=num_train_optimization_steps)
train_dataloader = cycle(train_dataloader)
output_dir = args.output_dir + "eval_results_{}_{}_{}_{}_{}".format(
list(filter(None, args.model_name_or_path.split("/"))).pop(),
str(args.max_seq_length),
str(args.learning_rate),
str(args.train_batch_size),
# str(args.train_language),
str(args.train_steps))
try:
os.makedirs(output_dir)
except:
pass
output_eval_file = os.path.join(output_dir, 'eval_result.txt')
with open(output_eval_file, "w") as writer:
writer.write('*' * 80 + '\n')
for step in bar:
batch = next(train_dataloader)
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, output_ids = batch
# 正常训练
outputs = model(input_ids=input_ids,
token_type_ids=segment_ids,
attention_mask=input_mask,
masked_lm_labels=output_ids)
loss = outputs[0]
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
tr_loss += loss.item()
train_loss = round(
tr_loss * args.gradient_accumulation_steps / (nb_tr_steps + 1),
4)
bar.set_description("loss {}".format(train_loss))
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
loss.backward()
# 对抗训练
fgm.attack() # 在embedding上添加对抗扰动
loss_adv = model(input_ids=input_ids,
token_type_ids=segment_ids,
attention_mask=input_mask,
masked_lm_labels=output_ids)[0]
if args.n_gpu > 1:
loss_adv = loss_adv.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss_adv = loss_adv / args.gradient_accumulation_steps
loss_adv.backward() # 反向传播，并在正常的grad基础上，累加对抗训练的梯度
fgm.restore() # 恢复embedding参数
if (nb_tr_steps + 1) % args.gradient_accumulation_steps == 0:
# scheduler.step()
optimizer.step()
optimizer.zero_grad()
global_step += 1
if (step + 1) % (args.eval_steps *
args.gradient_accumulation_steps) == 0:
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
logger.info("***** Report result *****")
logger.info(" %s = %s", 'global_step', str(global_step))
logger.info(" %s = %s", 'train loss', str(train_loss))
if args.do_eval and (step + 1) % (
args.eval_steps * args.gradient_accumulation_steps) == 0:
if args.do_eval_train:
file_list = ['train.csv', 'dev.csv']
else:
file_list = ['dev.csv']
for file in file_list:
inference_labels = []
gold_labels = []
inference_logits = []
dev_dataset = load_and_cache_examples(args,
tokenizer,
is_training=2)
dev_sampler = SequentialSampler(dev_dataset)
dev_dataloader = DataLoader(
dev_dataset,
sampler=dev_sampler,
batch_size=args.eval_batch_size)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(dev_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
for input_ids, input_mask, segment_ids, output_ids in tqdm(
dev_dataloader):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
output_ids = output_ids.to(device)
with torch.no_grad():
tmp_eval_loss, logits = model(
input_ids=input_ids,
token_type_ids=segment_ids,
attention_mask=input_mask,
masked_lm_labels=output_ids)
if args.n_gpu > 1:
tmp_eval_loss = tmp_eval_loss.mean(
) # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
tmp_eval_loss = tmp_eval_loss / args.gradient_accumulation_steps
eval_loss += tmp_eval_loss.item()
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
model.train()
eval_loss = eval_loss / nb_eval_steps
result = {
'eval_loss': eval_loss,
'global_step': global_step,
'train_loss': train_loss
}
if 'dev' in file:
with open(output_eval_file, "a") as writer:
writer.write(file + '\n')
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" %
(key, str(result[key])))
writer.write('*' * 80)
writer.write('\n')
if eval_loss < min_loss and 'dev' in file:
print("=" * 80)
print("Min Loss", eval_loss)
print("Saving Model......")
min_loss = eval_loss
# Save a trained model
model_to_save = model.module if hasattr(
model,
'module') else model # Only save the model it-self
output_model_file = os.path.join(
output_dir, "pytorch_model.bin")
torch.save(model_to_save.state_dict(),
output_model_file)
print("=" * 80)
else:
print("=" * 80)
with open(output_eval_file, "a") as writer:
writer.write('Min Loss: %f' % min_loss)
if args.do_test:
if args.do_train == False:
output_dir = args.output_dir
model = model_class.from_pretrained(output_dir, config=config)
model.to(args.device)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
dev_dataset = load_and_cache_examples(args, tokenizer, is_training=3)
dev_sampler = SequentialSampler(dev_dataset)
dev_dataloader = DataLoader(dev_dataset,
sampler=dev_sampler,
batch_size=args.eval_batch_size)
logger.info(" *** Run Prediction ***")
logger.info(" Num examples = %d", len(dev_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
model.eval()
test_loss = 0
for input_ids, input_mask, segment_ids, output_ids in tqdm(
dev_dataloader):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
output_ids = output_ids.to(device)
with torch.no_grad():
tmp_eval_loss = model(
input_ids=input_ids,
token_type_ids=segment_ids,
attention_mask=input_mask,
masked_lm_labels=output_ids)[0].detach().cpu().numpy()
if args.n_gpu > 1:
tmp_eval_loss = tmp_eval_loss.mean(
) # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
tmp_eval_loss = tmp_eval_loss / args.gradient_accumulation_steps
test_loss += tmp_eval_loss.item()
avg_test_loss = round(test_loss / len(dev_dataloader), 4)
logger.info('predict loss:{}'.format(str(avg_test_loss)))
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--model_type",
default=None,
type=str,
required=True,
help="Model type selected in the list: " + ", ".join(MODEL_TYPES),
)
parser.add_argument(
"--model_name_or_path",
default=None,
type=str,
required=True,
help="Path to pretrained model or model identifier from huggingface.co/models",
)
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model checkpoints and predictions will be written.",
)
# Other parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
help="The input data dir. Should contain the .json files for the task."
+ "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
)
parser.add_argument(
"--train_file",
default=None,
type=str,
help="The input training file. If a data dir is specified, will look for the file there"
+ "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
)
parser.add_argument(
"--predict_file",
default=None,
type=str,
help="The input evaluation file. If a data dir is specified, will look for the file there"
+ "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
)
parser.add_argument(
"--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name"
)
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help="Pretrained tokenizer name or path if not the same as model_name",
)
parser.add_argument(
"--cache_dir",
default="",
type=str,
help="Where do you want to store the pre-trained models downloaded from huggingface.co",
)
parser.add_argument(
"--version_2_with_negative",
action="store_true",
help="If true, the SQuAD examples contain some that do not have an answer.",
)
parser.add_argument(
"--null_score_diff_threshold",
type=float,
default=0.0,
help="If null_score - best_non_null is greater than the threshold predict null.",
)
parser.add_argument(
"--max_seq_length",
default=384,
type=int,
help="The maximum total input sequence length after WordPiece tokenization. Sequences "
"longer than this will be truncated, and sequences shorter than this will be padded.",
)
parser.add_argument(
"--doc_stride",
default=128,
type=int,
help="When splitting up a long document into chunks, how much stride to take between chunks.",
)
parser.add_argument(
"--max_query_length",
default=64,
type=int,
help="The maximum number of tokens for the question. Questions longer than this will "
"be truncated to this length.",
)
parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
parser.add_argument(
"--evaluate_during_training", action="store_true", help="Run evaluation during training at each logging step."
)
parser.add_argument(
"--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model."
)
parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.")
parser.add_argument(
"--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for evaluation."
)
parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
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("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
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, help="Max gradient norm.")
parser.add_argument(
"--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform."
)
parser.add_argument(
"--max_steps",
default=-1,
type=int,
help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
)
parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
parser.add_argument(
"--n_best_size",
default=20,
type=int,
help="The total number of n-best predictions to generate in the nbest_predictions.json output file.",
)
parser.add_argument(
"--max_answer_length",
default=30,
type=int,
help="The maximum length of an answer that can be generated. This is needed because the start "
"and end predictions are not conditioned on one another.",
)
parser.add_argument(
"--verbose_logging",
action="store_true",
help="If true, all of the warnings related to data processing will be printed. "
"A number of warnings are expected for a normal SQuAD evaluation.",
)
parser.add_argument(
"--lang_id",
default=0,
type=int,
help="language id of input for language-specific xlm models (see tokenization_xlm.PRETRAINED_INIT_CONFIGURATION)",
)
parser.add_argument("--logging_steps", type=int, default=500, help="Log every X updates steps.")
parser.add_argument("--save_steps", type=int, default=500, help="Save checkpoint every X updates steps.")
parser.add_argument(
"--eval_all_checkpoints",
action="store_true",
help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
)
parser.add_argument("--no_cuda", action="store_true", help="Whether not to use CUDA when available")
parser.add_argument(
"--overwrite_output_dir", action="store_true", help="Overwrite the content of the output directory"
)
parser.add_argument(
"--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets"
)
parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus")
parser.add_argument(
"--fp16",
action="store_true",
help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
)
parser.add_argument(
"--fp16_opt_level",
type=str,
default="O1",
help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html",
)
parser.add_argument("--server_ip", type=str, default="", help="Can be used for distant debugging.")
parser.add_argument("--server_port", type=str, default="", help="Can be used for distant debugging.")
parser.add_argument("--threads", type=int, default=1, help="multiple threads for converting example to features")
args = parser.parse_args()
if args.doc_stride >= args.max_seq_length - args.max_query_length:
logger.warning(
"WARNING - You've set a doc stride which may be superior to the document length in some "
"examples. This could result in errors when building features from the examples. Please reduce the doc "
"stride or increase the maximum length to ensure the features are correctly built."
)
if (
os.path.exists(args.output_dir)
and os.listdir(args.output_dir)
and args.do_train
and not args.overwrite_output_dir
):
raise ValueError(
"Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
args.output_dir
)
)
# Setup distant debugging if needed
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
ptvsd.wait_for_attach()
# Setup CUDA, GPU & distributed training
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend="nccl")
args.n_gpu = 1
args.device = device
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
args.local_rank,
device,
args.n_gpu,
bool(args.local_rank != -1),
args.fp16,
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# Set seed
set_seed(args)
# Load pretrained model and tokenizer
if args.local_rank not in [-1, 0]:
# Make sure only the first process in distributed training will download model & vocab
torch.distributed.barrier()
args.model_type = args.model_type.lower()
config = AutoConfig.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None,
)
tokenizer = AutoTokenizer.from_pretrained(
args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
do_lower_case=args.do_lower_case,
cache_dir=args.cache_dir if args.cache_dir else None,
use_fast=False, # SquadDataset is not compatible with Fast tokenizers which have a smarter overflow handeling
)
model = AutoModelForQuestionAnswering.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None,
)
if args.local_rank == 0:
# Make sure only the first process in distributed training will download model & vocab
torch.distributed.barrier()
model.to(args.device)
logger.info("Training/evaluation parameters %s", args)
# Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is set.
# Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running `--fp16_opt_level="O2"` will
# remove the need for this code, but it is still valid.
if args.fp16:
try:
import apex
apex.amp.register_half_function(torch, "einsum")
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
# Training
if args.do_train:
# train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
feature_reader = load_and_cache_examples(args, tokenizer, evaluate=False)
train_dataset = HDF5Dataset(feature_reader)
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
# Save the trained model and the tokenizer
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
logger.info("Saving model checkpoint to %s", args.output_dir)
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
# They can then be reloaded using `from_pretrained()`
# Take care of distributed/parallel training
model_to_save = model.module if hasattr(model, "module") else model
model_to_save.save_pretrained(args.output_dir)
tokenizer.save_pretrained(args.output_dir)
# Good practice: save your training arguments together with the trained model
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
# Load a trained model and vocabulary that you have fine-tuned
model = AutoModelForQuestionAnswering.from_pretrained(args.output_dir) # , force_download=True)
# SquadDataset is not compatible with Fast tokenizers which have a smarter overflow handeling
# So we use use_fast=False here for now until Fast-tokenizer-compatible-examples are out
tokenizer = AutoTokenizer.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case, use_fast=False)
model.to(args.device)
# Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
results = {}
if args.do_eval and args.local_rank in [-1, 0]:
if args.do_train:
logger.info("Loading checkpoints saved during training for evaluation")
checkpoints = [args.output_dir]
if args.eval_all_checkpoints:
checkpoints = list(
os.path.dirname(c)
for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
)
else:
logger.info("Loading checkpoint %s for evaluation", args.model_name_or_path)
checkpoints = [args.model_name_or_path]
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
# Reload the model
global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
model = AutoModelForQuestionAnswering.from_pretrained(checkpoint) # , force_download=True)
model.to(args.device)
# Evaluate
result = evaluate(args, model, tokenizer, prefix=global_step)
result = dict((k + ("_{}".format(global_step) if global_step else ""), v) for k, v in result.items())
results.update(result)
logger.info("Results: {}".format(results))
return results
def evaluate(args, model, tokenizer, prefix=""):
# dataset, examples, features = load_and_cache_examples(args, tokenizer, evaluate=True, output_examples=True)
feature_reader = load_and_cache_examples(args, tokenizer, evaluate=True)
dataset = HDF5Dataset(feature_reader)
examples = feature_reader.load_examples()
features = feature_reader.get_features()
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
os.makedirs(args.output_dir)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
# Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(dataset)
eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
# multi-gpu evaluate
if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
model = torch.nn.DataParallel(model)
# Eval!
logger.info("***** Running evaluation {} *****".format(prefix))
logger.info(" Num examples = %d", len(dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
all_results = []
start_time = timeit.default_timer()
for batch in tqdm(eval_dataloader, desc="Evaluating"):
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"token_type_ids": batch[2],
}
if args.model_type in ["xlm", "roberta", "distilbert", "camembert", "bart", "longformer"]:
del inputs["token_type_ids"]
feature_indices = batch[3]
# XLNet and XLM use more arguments for their predictions
if args.model_type in ["xlnet", "xlm"]:
inputs.update({"cls_index": batch[4], "p_mask": batch[5]})
# for lang_id-sensitive xlm models
if hasattr(model, "config") and hasattr(model.config, "lang2id"):
inputs.update(
{"langs": (torch.ones(batch[0].shape, dtype=torch.int64) * args.lang_id).to(args.device)}
)
outputs = model(**inputs)
for i, feature_index in enumerate(feature_indices):
eval_feature = features[feature_index.item()]
unique_id = int(eval_feature.unique_id)
output = [to_list(output[i]) for output in outputs.to_tuple()]
# Some models (XLNet, XLM) use 5 arguments for their predictions, while the other "simpler"
# models only use two.
if len(output) >= 5:
start_logits = output[0]
start_top_index = output[1]
end_logits = output[2]
end_top_index = output[3]
cls_logits = output[4]
result = SquadResult(
unique_id,
start_logits,
end_logits,
start_top_index=start_top_index,
end_top_index=end_top_index,
cls_logits=cls_logits,
)
else:
start_logits, end_logits = output
result = SquadResult(unique_id, start_logits, end_logits)
all_results.append(result)
evalTime = timeit.default_timer() - start_time
logger.info(" Evaluation done in total %f secs (%f sec per example)", evalTime, evalTime / len(dataset))
# Compute predictions
output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
if args.version_2_with_negative:
output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
else:
output_null_log_odds_file = None
# XLNet and XLM use a more complex post-processing procedure
if args.model_type in ["xlnet", "xlm"]:
start_n_top = model.config.start_n_top if hasattr(model, "config") else model.module.config.start_n_top
end_n_top = model.config.end_n_top if hasattr(model, "config") else model.module.config.end_n_top
predictions = compute_predictions_log_probs(
examples,
features,
all_results,
args.n_best_size,
args.max_answer_length,
output_prediction_file,
output_nbest_file,
output_null_log_odds_file,
start_n_top,
end_n_top,
args.version_2_with_negative,
tokenizer,
args.verbose_logging,
)
else:
predictions = compute_predictions_logits(
examples,
features,
all_results,
args.n_best_size,
args.max_answer_length,
args.do_lower_case,
output_prediction_file,
output_nbest_file,
output_null_log_odds_file,
args.verbose_logging,
args.version_2_with_negative,
args.null_score_diff_threshold,
tokenizer,
)
# Compute the F1 and exact scores.
results = squad_evaluate(examples, predictions)
return results
