def __init__(self, mask_rate, bert_model, do_lower_case, max_seq_length):
super(RandMask, self).__init__()
self.mask_rate = mask_rate
self.max_seq_length = max_seq_length
self.tokenizer = BertTokenizer.from_pretrained(
bert_model, do_lower_case=do_lower_case)
self.vocab = list(self.tokenizer.vocab.keys())
def __init__(self,
mask_rate,
top_sen_rate,
threshold,
bert_model,
do_lower_case,
max_seq_length,
label_list,
sen_batch_size,
use_gpu=True):
super(SC, self).__init__()
self.mask_rate = mask_rate
self.top_sen_rate = top_sen_rate
self.threshold = threshold
self.label_list = label_list
self.num_labels = len(self.label_list)
self.max_seq_length = max_seq_length
self.tokenizer = BertTokenizer.from_pretrained(
bert_model, do_lower_case=do_lower_case)
self.model = BertForSequenceClassification.from_pretrained(
bert_model, num_labels=self.num_labels)
self.device = torch.device(
"cuda" if torch.cuda.is_available() and use_gpu else "cpu")
self.model.to(self.device)
self.n_gpu = torch.cuda.device_count()
self.sen_batch_size = sen_batch_size
self.vocab = list(self.tokenizer.vocab.keys())
if self.n_gpu > 1:
self.model = torch.nn.DataParallel(self.model)
def get_preprocessor(ptr_config_info, model_config):
with open(ptr_config_info.vocab, mode='rb') as io:
vocab = pickle.load(io)
if model_config.type == 'etri':
ptr_tokenizer = ETRITokenizer.from_pretrained(ptr_config_info.tokenizer, do_lower_case=False)
pad_sequence = PadSequence(length=model_config.length, pad_val=vocab.to_indices(vocab.padding_token))
preprocessor = PreProcessor(vocab=vocab, split_fn=ptr_tokenizer.tokenize, pad_fn=pad_sequence)
elif model_config.type == 'skt':
ptr_tokenizer = SentencepieceTokenizer(ptr_config_info.tokenizer)
pad_sequence = PadSequence(length=model_config.length, pad_val=vocab.to_indices(vocab.padding_token))
preprocessor = PreProcessor(vocab=vocab, split_fn=ptr_tokenizer, pad_fn=pad_sequence)
return preprocessor
def __init__(self,
mask_rate,
bert_model,
do_lower_case,
max_seq_length,
sen_batch_size,
with_rand=False,
use_gpu=True):
super(ModelGen, self).__init__()
self.mask_rate = mask_rate
self.max_seq_length = max_seq_length
self.tokenizer = BertTokenizer.from_pretrained(
bert_model, do_lower_case=do_lower_case)
self.model = BertForTokenClassification.from_pretrained(bert_model,
num_labels=2)
self.device = torch.device(
"cuda" if torch.cuda.is_available() and use_gpu else "cpu")
self.model.to(self.device)
self.n_gpu = torch.cuda.device_count()
self.sen_batch_size = sen_batch_size
self.vocab = list(self.tokenizer.vocab.keys())
self.with_rand = with_rand
if self.n_gpu > 1:
self.model = torch.nn.DataParallel(self.model)
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
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("--train_file",
default=None,
type=str,
help="SQuAD json for training. E.g., train-v1.1.json")
parser.add_argument(
"--predict_file",
default=None,
type=str,
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json"
)
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_predict",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--train_batch_size",
default=8,
type=int,
help="Total batch size for training.")
parser.add_argument("--predict_batch_size",
default=8,
type=int,
help="Total batch size for predictions.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. E.g., 0.1 = 10%% "
"of training.")
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("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
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(
"--do_lower_case",
action='store_true',
help=
"Whether to lower case the input text. True for uncased models, False for cased models."
)
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 float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument(
'--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('--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.")
args = parser.parse_args()
print(args)
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")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
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.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_predict:
raise ValueError(
"At least one of `do_train` or `do_predict` must be True.")
if args.do_train:
if not args.train_file:
raise ValueError(
"If `do_train` is True, then `train_file` must be specified.")
if args.do_predict:
if not args.predict_file:
raise ValueError(
"If `do_predict` is True, then `predict_file` must be specified."
)
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
raise ValueError(
"Output directory () already exists and is not empty.")
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = read_squad_examples(
input_file=args.train_file,
is_training=True,
version_2_with_negative=args.version_2_with_negative)
train_dataloader = prepare_data(train_examples,
tokenizer,
args,
task_name="train")
num_train_optimization_steps = int(
len(train_dataloader) /
args.gradient_accumulation_steps) * args.num_train_epochs
# Prepare model
model = BertForQuestionAnswering.from_pretrained(args.bert_model)
model.to(device)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
def eval_model(model, data_loader):
eval_loss = 0
model.eval()
for step, batch in enumerate(
tqdm(data_loader,
desc="Evaluating",
disable=args.local_rank not in [-1, 0])):
if n_gpu == 1:
batch = tuple(
t.to(device)
for t in batch) # multi-gpu does scattering it-self
input_ids, input_mask, segment_ids, start_positions, end_positions = batch
loss = model(input_ids, segment_ids, input_mask, start_positions,
end_positions)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
eval_loss += loss.detach().cpu()
return eval_loss / (step + 1)
eval_examples = read_squad_examples(
input_file=args.predict_file,
is_training=False,
version_2_with_negative=args.version_2_with_negative)
eval_dataloader = prepare_data(eval_examples,
tokenizer,
args,
task_name='eval')
if args.do_train:
param_optimizer = list(model.named_parameters())
param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
model.train()
train_loss = 0
min_loss = 100000
for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
for step, batch in enumerate(
tqdm(train_dataloader,
desc="Training",
disable=args.local_rank not in [-1, 0])):
if n_gpu == 1:
batch = tuple(
t.to(device)
for t in batch) # multi-gpu does scattering it-self
input_ids, input_mask, segment_ids, start_positions, end_positions = batch
# print(f'start_positions = {start_positions}')
# print(f'end_positions = {end_positions}')
loss = model(input_ids, segment_ids, input_mask,
start_positions, end_positions)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
train_loss += loss.detach().cpu()
eval_loss = eval_model(model, eval_dataloader)
print(
f'epoch {epoch} train loss {train_loss / (step + 1)} eval_loss {eval_loss}'
)
if eval_loss < min_loss:
min_loss = eval_loss
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
# If we save using the predefined names, we can load using `from_pretrained`
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
model_to_save.config.to_json_file(output_config_file)
tokenizer.save_vocabulary(args.output_dir)
elif args.do_predict:
predict_dataloader, predict_features = prepare_data(
eval_examples, tokenizer, args, task_name='predict')
all_results = []
logger.info("Start Predicting")
for input_ids, input_mask, segment_ids, example_indices in tqdm(
predict_dataloader,
desc="Predicting",
disable=args.local_rank not in [-1, 0]):
if len(all_results) % 1000 == 0:
logger.info("Processing example: %d" % (len(all_results)))
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
with torch.no_grad():
batch_start_logits, batch_end_logits = model(
input_ids, segment_ids, input_mask)
for i, example_index in enumerate(example_indices):
start_logits = batch_start_logits[i].detach().cpu().tolist(
) # 某一句话的答案起始位置, 长度为len
end_logits = batch_end_logits[i].detach().cpu().tolist()
predict_features = predict_features[
example_index.item()] # 某句话的特征
unique_id = int(predict_features.unique_id) # 某句话的id
all_results.append(
RawResult(unique_id=unique_id,
start_logits=start_logits,
end_logits=end_logits))
output_prediction_file = os.path.join(args.output_dir,
"predictions.json")
output_nbest_file = os.path.join(args.output_dir,
"nbest_predictions.json")
output_null_log_odds_file = os.path.join(args.output_dir,
"null_odds.json")
write_predictions(eval_examples, predict_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)
else:
raise ValueError(
'please confirm at least one task mode, such as ’train‘ or ’predict‘.'
)
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(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
## Other parameters
parser.add_argument("--ckpt", default="", type=str)
parser.add_argument(
"--vocab_file",
default="",
type=str,
)
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 WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
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_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
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(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision 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(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument('--sample_weight', type=float, default=1)
parser.add_argument("--save_all", action="store_true")
args = parser.parse_args()
processors = {"maskgen": MaskGenProcessor}
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")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
# raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
logger.warning(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
label_list = processor.get_labels()
num_labels = len(label_list)
if args.local_rank not in [-1, 0]:
# Make sure only the first process in distributed training will download model & vocab
torch.distributed.barrier()
if args.vocab_file:
tokenizer = BertTokenizer(args.vocab_file, args.do_lower_case)
else:
tokenizer = BertTokenizer.from_pretrained(
args.bert_model, do_lower_case=args.do_lower_case)
# Prepare model
model = BertForTokenClassification.from_pretrained(args.bert_model,
num_labels=num_labels)
if args.ckpt:
print("load from", args.ckpt)
model_dict = model.state_dict()
ckpt = torch.load(args.ckpt)
pretrained_dict = ckpt['model']
new_dict = {
k: v
for k, v in pretrained_dict.items() if k in model_dict.keys()
}
model_dict.update(new_dict)
print('Total : {}, update: {}'.format(len(pretrained_dict),
len(new_dict)))
model.load_state_dict(model_dict)
if args.local_rank == 0:
torch.distributed.barrier()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
global_step = 0
nb_tr_steps = 0
tr_loss = 0
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
if args.fp16:
sample_weight = torch.HalfTensor([1.0, args.sample_weight]).cuda()
else:
sample_weight = torch.FloatTensor([1.0, args.sample_weight]).cuda()
cached_train_features_file = os.path.join(
args.data_dir, 'train_{}_{}_{}'.format(
list(filter(None, args.bert_model.split('/'))).pop(),
str(args.max_seq_length), str(task_name)))
try:
with open(cached_train_features_file, "rb") as reader:
logger.info("Load from cache dir: {}".format(
cached_train_features_file))
train_features = pickle.load(reader)
except:
train_features = convert_examples_to_features(
train_examples, label_list, args.max_seq_length, tokenizer)
if args.local_rank == -1 or torch.distributed.get_rank() == 0:
logger.info("Saving train features into cached file {}".format(
cached_train_features_file))
with open(cached_train_features_file, "wb") as writer:
pickle.dump(train_features, writer)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
num_train_optimization_steps = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
label_map = {i: label for i, label in enumerate(label_list, 1)}
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
os.makedirs(os.path.join(args.output_dir, "all_models"), exist_ok=True)
model.train()
for e in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_steps = 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss = model(input_ids,
segment_ids,
input_mask,
label_ids,
weight=sample_weight)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
tr_loss += loss.item()
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
global_step += 1
# save each epoch
model_to_save = model.module if hasattr(model, 'module') else model
output_model_file = os.path.join(args.output_dir, "all_models",
"e{}_{}".format(e, WEIGHTS_NAME))
torch.save(model_to_save.state_dict(), output_model_file)
if args.do_train and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
output_args_file = os.path.join(args.output_dir, 'training_args.bin')
torch.save(args, output_args_file)
else:
model = BertForTokenClassification.from_pretrained(
args.bert_model, num_labels=num_labels)
### Evaluation
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
best_f1 = 0
best_epoch = 0
val_res_file = os.path.join(args.output_dir, "valid_results.txt")
val_f = open(val_res_file, "w")
logger.info("***** Dev Eval results *****")
for e in range(int(args.num_train_epochs)):
weight_path = os.path.join(args.output_dir, "all_models",
"e{}_{}".format(e, WEIGHTS_NAME))
model.load_state_dict(torch.load(weight_path))
model.to(device)
eval_examples = processor.get_dev_examples(args.data_dir)
cached_eval_features_file = os.path.join(
args.data_dir, 'dev_{0}_{1}_{2}'.format(
list(filter(None, args.bert_model.split('/'))).pop(),
str(args.max_seq_length), str(task_name)))
try:
with open(cached_eval_features_file, "rb") as reader:
eval_features = pickle.load(reader)
except:
eval_features = convert_examples_to_features(
eval_examples, label_list, args.max_seq_length, tokenizer)
if args.local_rank == -1 or torch.distributed.get_rank() == 0:
logger.info(" Saving eval features into cached file %s",
cached_eval_features_file)
with open(cached_eval_features_file, "wb") as writer:
pickle.dump(eval_features, writer)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor(
[f.input_mask for f in eval_features], dtype=torch.long)
all_segment_ids = torch.tensor(
[f.segment_ids for f in eval_features], dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
if args.local_rank == -1:
eval_sampler = SequentialSampler(eval_data)
else:
eval_sampler = DistributedSampler(
eval_data) # Note that this sampler samples randomly
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
y_true_L = []
y_pred_L = []
for input_ids, input_mask, segment_ids, label_ids in tqdm(
eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
logits = model(input_ids, segment_ids, input_mask)
logits = torch.argmax(F.log_softmax(logits, dim=2), dim=2)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
input_mask = input_mask.to('cpu').numpy()
y_true = [[str(x) for x in L] for L in label_ids]
y_pred = [[str(x) for x in L] for L in logits]
for (m, t, p) in zip(input_mask, y_true, y_pred):
for mm, tt, pp in zip(m, t, p):
if mm == 1:
y_true_L.append(int(tt))
y_pred_L.append(int(pp))
acc = accuracy_score(y_true_L, y_pred_L)
f1 = f1_score(y_true_L, y_pred_L)
recall = recall_score(y_true_L, y_pred_L)
prec = precision_score(y_true_L, y_pred_L)
if f1 > best_f1:
best_f1 = f1
best_epoch = e
result = {"acc": acc, "f1": f1, "recall": recall, "prec": prec}
logger.info("Epoch {}".format(e))
val_f.write("Epoch {}\n".format(e))
for key in sorted(result.keys()):
logger.info("{} = {}".format(key, str(result[key])))
val_f.write("{} = {}\n".format(key, str(result[key])))
val_f.write("\n")
logger.info("\nBest epoch: {}. Best val f1: {}".format(
best_epoch, best_f1))
val_f.write("Best epoch: {}. Best val f1: {}\n".format(
best_epoch, best_f1))
val_f.close()
best_weight_path = os.path.join(
args.output_dir, "all_models",
"e{}_{}".format(best_epoch, WEIGHTS_NAME))
best_model_dir = os.path.join(args.output_dir, "best_model")
os.makedirs(best_model_dir, exist_ok=True)
os.system("cp {} {}/{}".format(best_weight_path, best_model_dir,
WEIGHTS_NAME))
with open(os.path.join(best_model_dir, CONFIG_NAME), 'w') as f:
f.write(model_to_save.config.to_json_string())
tokenizer.save_vocab(os.path.join(best_model_dir, VOCAB_NAME))
if not args.save_all:
os.system("rm -r {}".format(
os.path.join(args.output_dir, "all_models")))
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(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
## Other parameters
parser.add_argument("--vocab_file", default="", type=str)
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 WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
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_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
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("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
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(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision 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(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument("--ckpt", type=str, help="ckpt position")
parser.add_argument("--save_all", action="store_true")
parser.add_argument("--output_dev_detail", action="store_true")
args = parser.parse_args()
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")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
args.device = device
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.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
os.makedirs(args.output_dir)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
output_mode = output_modes[task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
if args.local_rank not in [-1, 0]:
torch.distributed.barrier(
) # Make sure only the first process in distributed training will download model & vocab
if args.vocab_file:
tokenizer = BertTokenizer(args.vocab_file, args.do_lower_case)
else:
tokenizer = BertTokenizer.from_pretrained(
args.bert_model, do_lower_case=args.do_lower_case)
model = BertForSequenceClassification.from_pretrained(
args.bert_model, num_labels=num_labels)
if args.ckpt:
print("load from", args.ckpt)
model_dict = model.state_dict()
ckpt = torch.load(args.ckpt)
if "model" in ckpt:
pretrained_dict = ckpt['model']
else:
pretrained_dict = ckpt
new_dict = {
k: v
for k, v in pretrained_dict.items() if k in model_dict.keys()
and k not in ["classifier.weight", "classifier.bias"]
}
model_dict.update(new_dict)
print('Total : {}, update: {}'.format(len(pretrained_dict),
len(new_dict)))
model.load_state_dict(model_dict)
if args.local_rank == 0:
torch.distributed.barrier()
model.to(device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
global_step = 0
nb_tr_steps = 0
tr_loss = 0
if args.do_train:
# Prepare data loader
train_examples = processor.get_train_examples(args.data_dir)
print(len(train_examples))
cached_train_features_file = os.path.join(
args.data_dir, 'train_{0}_{1}_{2}'.format(
list(filter(None, args.bert_model.split('/'))).pop(),
str(args.max_seq_length), str(task_name)))
try:
with open(cached_train_features_file, "rb") as reader:
train_features = pickle.load(reader)
except:
train_features = convert_examples_to_features(
train_examples, label_list, args.max_seq_length, tokenizer,
output_mode)
if args.local_rank == -1 or torch.distributed.get_rank() == 0:
logger.info(" Saving train features into cached file %s",
cached_train_features_file)
with open(cached_train_features_file, "wb") as writer:
pickle.dump(train_features, writer)
all_input_ids = [f.input_ids for f in train_features]
all_input_mask = [f.input_mask for f in train_features]
all_segment_ids = [f.segment_ids for f in train_features]
if output_mode == "classification":
all_label_ids = [f.label_id for f in train_features]
elif output_mode == "regression":
all_label_ids = [f.label_id for f in train_features]
train_data = InputDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size,
collate_fn=train_data.collate)
num_train_optimization_steps = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
os.makedirs(os.path.join(args.output_dir, "all_models"), exist_ok=True)
model.train()
for e in trange(int(args.num_train_epochs),
desc="Epoch",
disable=args.local_rank not in [-1, 0]):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader,
desc="Iteration",
disable=args.local_rank not in [-1, 0])):
inputs, labels = batch
for key in inputs.keys():
inputs[key] = inputs[key].to(args.device)
for key in labels.keys():
labels[key] = labels[key].to(args.device)
# define a new function to compute loss values for both output_modes
label_ids = labels["labels"]
logits = model(**inputs)
if output_mode == "classification":
loss_fct = CrossEntropyLoss()
loss = loss_fct(logits.view(-1, num_labels),
label_ids.view(-1))
elif output_mode == "regression":
loss_fct = MSELoss()
loss = loss_fct(logits.view(-1), label_ids.view(-1))
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
tr_loss += loss.item()
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
global_step += 1
# save each epoch
model_to_save = model.module if hasattr(model, 'module') else model
output_model_file = os.path.join(args.output_dir, "all_models",
"e{}_{}".format(e, WEIGHTS_NAME))
torch.save(model_to_save.state_dict(), output_model_file)
### Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
if args.do_train and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
output_args_file = os.path.join(args.output_dir, 'training_args.bin')
torch.save(args, output_args_file)
else:
model = BertForSequenceClassification.from_pretrained(
args.bert_model, num_labels=num_labels)
### Evaluation
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
best_acc = 0
best_epoch = 0
val_res_file = os.path.join(args.output_dir, "valid_results.txt")
val_f = open(val_res_file, "w")
if args.output_dev_detail:
logger.info("***** Dev Eval results *****")
for e in tqdm(range(int(args.num_train_epochs)), desc="Epoch on dev"):
weight_path = os.path.join(args.output_dir, "all_models",
"e{}_{}".format(e, WEIGHTS_NAME))
result = evaluate(args,
model,
weight_path,
processor,
device,
task_name,
"dev",
label_list,
tokenizer,
output_mode,
num_labels,
show_detail=False)
if result["acc"] > best_acc:
best_acc = result["acc"]
best_epoch = e
if args.output_dev_detail:
logger.info("Epoch {}".format(e))
val_f.write("Epoch {}\n".format(e))
for key in sorted(result.keys()):
if args.output_dev_detail:
logger.info("{} = {}".format(key, str(result[key])))
val_f.write("{} = {}\n".format(key, str(result[key])))
val_f.write("\n")
logger.info("\nBest epoch: {}. Best val acc: {}".format(
best_epoch, best_acc))
val_f.write("Best epoch: {}. Best val acc: {}\n".format(
best_epoch, best_acc))
val_f.close()
test_weight_path = os.path.join(
args.output_dir, "all_models",
"e{}_{}".format(best_epoch, WEIGHTS_NAME))
test_result = evaluate(args, model, test_weight_path, processor,
device, task_name, "test", label_list,
tokenizer, output_mode, num_labels)
test_res_file = os.path.join(args.output_dir, "test_results.txt")
logger.info("***** Test Eval results *****")
with open(test_res_file, "w") as test_f:
for key in sorted(test_result.keys()):
logger.info("{} = {}".format(key, str(test_result[key])))
test_f.write("{} = {}\n".format(key, str(test_result[key])))
best_model_dir = os.path.join(args.output_dir, "best_model")
os.makedirs(best_model_dir, exist_ok=True)
os.system("cp {} {}/{}".format(test_weight_path, best_model_dir,
WEIGHTS_NAME))
with open(os.path.join(best_model_dir, CONFIG_NAME), 'w') as f:
f.write(model_to_save.config.to_json_string())
tokenizer.save_vocab(os.path.join(best_model_dir, VOCAB_NAME))
if not args.save_all:
os.system("rm -r {}".format(
os.path.join(args.output_dir, "all_models")))
zipfile_path = ptr_dir / "etri.zip"
if not zipfile_path.exists():
url = "https://drive.google.com/uc?id=1qVY-zZc2O2OliGNUwWClhcqJkLG_6uoD"
gdown.download(url, output=str(zipfile_path))
with zipfile.ZipFile(str(zipfile_path)) as unzip:
unzip.extractall(str(ptr_dir))
from model.tokenization import BertTokenizer as ETRITokenizer
# loading BertTokenizer
ptr_config_path = ptr_dir / 'bert_config_etri.json'
ptr_tokenizer_path = ptr_dir / "vocab.korean.rawtext.list"
ptr_bert_path = ptr_dir / "pytorch_model_etri.bin"
ptr_tokenizer = ETRITokenizer.from_pretrained(ptr_tokenizer_path,
do_lower_case=False)
# generate vocab
idx_to_token = list(ptr_tokenizer.vocab.keys())
token_to_idx = {
token: idx
for idx, token in enumerate(idx_to_token)
}
vocab = Vocab(
idx_to_token,
padding_token="[PAD]",
unknown_token="[UNK]",
bos_token=None,
eos_token=None,
reserved_tokens=["[CLS]", "[SEP]", "[MASK]"],
token_to_idx=token_to_idx,