def create_examples(self, lines, example_type, cached_examples_file):
'''
Creates examples for data
'''
pbar = ProgressBar(n_total=len(lines), desc='create examples')
if cached_examples_file.exists():
logger.info("Loading examples from cached file %s",
cached_examples_file)
examples = torch.load(cached_examples_file)
else:
examples = []
for i, line in enumerate(lines):
guid = '%s-%d' % (example_type, i)
text_a = line[0]
text_b = line[1]
label = line[2]
label = int(label)
example = InputExample(guid=guid,
text_a=text_a,
text_b=text_b,
label=label)
examples.append(example)
pbar(step=i)
logger.info("Saving examples into cached file %s",
cached_examples_file)
torch.save(examples, cached_examples_file)
return examples
def evaluate(args, model, tokenizer, prefix=""):
eval_task_names = (args.task_name,)
eval_outputs_dirs = (args.output_dir,)
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, data_type='dev')
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,
collate_fn=xlnet_collate_fn if args.model_type in ['xlnet'] else collate_fn)
# Eval!
logger.info("********* Running evaluation {} ********".format(prefix))
eval_loss = 0.0
nb_eval_steps = 0
preds = None
out_label_ids = None
pbar = ProgressBar(n_total=len(eval_dataloader), desc="Evaluating")
for step, batch in enumerate(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],
'labels': batch[3]}
if args.model_type != 'distilbert':
inputs['token_type_ids'] = batch[2] if args.model_type in ['bert', 'xlnet', 'albert',
'roberta'] else None # XLM, DistilBERT and RoBERTa don't use segment_ids
outputs = model(**inputs)
tmp_eval_loss, logits = outputs[:2]
eval_loss += tmp_eval_loss.mean().item()
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)
pbar(step)
print(' ')
if 'cuda' in str(args.device):
torch.cuda.empty_cache()
eval_loss = eval_loss / nb_eval_steps
if args.output_mode == "classification":
preds = np.argmax(preds, axis=1)
elif args.output_mode == "regression":
preds = np.squeeze(preds)
result = compute_metrics(eval_task, preds, out_label_ids)
results.update(result)
logger.info(" Num examples = %d", len(eval_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
logger.info("******** Eval results {} ********".format(prefix))
for key in sorted(result.keys()):
logger.info(" dev: %s = %s", key, str(result[key]))
return results
def load_and_cache_examples(args, task, tokenizer, data_type='train'):
if args.local_rank not in [-1, 0] and not evaluate:
torch.distributed.barrier(
) # Make sure only the first process in distributed training process the dataset, and the others will use the cache
processor = processors[task]()
output_mode = output_modes[task]
# Load data features from cache or dataset file
cached_features_file = os.path.join(
args.data_dir, 'cached_{}_{}_{}_{}'.format(
data_type,
list(filter(None, args.model_name_or_path.split('/'))).pop(),
str(args.max_seq_length), str(task)))
if os.path.exists(cached_features_file):
logger.info("Loading features from cached file %s",
cached_features_file)
features = torch.load(cached_features_file)
else:
logger.info("Creating features from dataset file at %s", args.data_dir)
# os.makedirs(cached_features_file, exist_ok=True, mode=0o777)
if task == "ner":
label_list = processor.get_labels_ner(args.data_dir,
args.label_with_bi)
else:
label_list = processor.get_labels()
if task in ['mnli', 'mnli-mm'] and 'roberta' in args.model_type:
# HACK(label indices are swapped in RoBERTa pretrained model)
label_list[1], label_list[2] = label_list[2], label_list[1]
if data_type == 'train':
examples = processor.get_train_examples(args.data_dir)
elif data_type == 'dev':
examples = processor.get_dev_examples(args.data_dir)
else:
examples = processor.get_test_examples(args.data_dir)
features = convert_examples_to_features(
examples,
tokenizer,
label_list=label_list,
max_seq_length=args.max_seq_length,
output_mode=output_mode)
if args.local_rank == 0 and not evaluate:
torch.distributed.barrier(
) # Make sure only the first process in distributed training process the dataset, and the others will use the cache
# Convert to Tensors and build dataset
all_input_ids = torch.tensor([f.input_ids for f in features],
dtype=torch.long)
all_attention_mask = torch.tensor([f.attention_mask for f in features],
dtype=torch.long)
all_token_type_ids = torch.tensor([f.token_type_ids for f in features],
dtype=torch.long)
all_lens = torch.tensor([f.input_len for f in features], dtype=torch.long)
all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
dataset = TensorDataset(all_input_ids, all_attention_mask,
all_token_type_ids, all_lens, all_labels)
return dataset
def load_and_cache_examples(args, task, tokenizer, data_type='train'):
if args.local_rank not in [-1, 0] and not evaluate:
torch.distributed.barrier(
) # Make sure only the first process in distributed training process the dataset, and the others will use the cache
processor = processors[task]()
# Load data features from dataset file
logger.info("Creating features from dataset file at %s", args.data_dir)
label_list = processor.get_labels()
if data_type == 'train':
examples = processor.get_train_examples(args.data_dir)
elif data_type == 'dev':
examples = processor.get_dev_examples(args.data_dir)
else:
examples = processor.get_test_examples(args.data_dir)
if "bert" in args.model_type:
# 数据处理成bert所需的格式
features = convert_examples_to_features_for_bert(
examples=examples,
tokenizer=tokenizer,
label_list=label_list,
max_seq_length=args.train_max_seq_length
if data_type == 'train' else args.eval_max_seq_length,
cls_token_at_end=bool(args.model_type in ["xlnet"]),
pad_on_left=bool(args.model_type in ['xlnet']),
cls_token=tokenizer.cls_token,
cls_token_segment_id=2 if args.model_type in ["xlnet"] else 0,
sep_token=tokenizer.sep_token,
# pad on the left for xlnet
pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token
])[0],
pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0,
)
else:
# 一般格式
features = convert_examples_to_features_for_lstm(
examples=examples,
label_list=label_list,
max_seq_length=args.train_max_seq_length
if data_type == 'train' else args.eval_max_seq_length,
tokenizer=tokenizer)
if args.local_rank == 0 and not evaluate:
torch.distributed.barrier(
) # Make sure only the first process in distributed training process the dataset, and the others will use the cache
# Convert to Tensors and build dataset
all_input_ids = torch.tensor([f.input_ids for f in features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_ids for f in features],
dtype=torch.long)
all_lens = torch.tensor([f.input_len for f in features], dtype=torch.long)
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
all_lens, all_label_ids)
return dataset
def load_and_cache_examples(args, task, tokenizer, data_type='train'):
if args.local_rank not in [-1, 0] and not evaluate:
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
processor = processors[task]()
# Load data features from cache or dataset file
cached_features_file = os.path.join(args.data_dir, 'cached_crf-{}_{}_{}_{}'.format(
data_type,
list(filter(None, args.model_name_or_path.split('/'))).pop(),
str(args.train_max_seq_length if data_type=='train' else args.eval_max_seq_length),
str(task)))
if os.path.exists(cached_features_file) and not args.overwrite_cache:
logger.info("Loading features from cached file %s", cached_features_file)
features = torch.load(cached_features_file)
else:
label_list = processor.get_labels()
if data_type == 'train':
logger.info("Creating features from dataset file at %s", args.train_data_path)
examples = processor.get_train_examples(args.train_data_path)
elif data_type == 'dev':
logger.info("Creating features from dataset file at %s", args.dev_data_path)
examples = processor.get_dev_examples(args.dev_data_path)
else:
logger.info("Creating features from dataset file at %s", args.test_data_path)
examples = processor.get_test_examples(args.test_data_path)
features = convert_examples_to_features(examples=examples,
tokenizer=tokenizer,
label_list=label_list,
max_seq_length=args.train_max_seq_length if data_type=='train' \
else args.eval_max_seq_length,
max_attr_length = args.max_attr_length
)
if args.local_rank in [-1, 0]:
logger.info("Saving features into cached file %s", cached_features_file)
torch.save(features, cached_features_file)
if args.local_rank == 0 and not evaluate:
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
# Convert to Tensors and build dataset
t_all_input_ids = torch.tensor([f.t_input_ids for f in features], dtype=torch.long)
t_all_input_mask = torch.tensor([f.t_input_mask for f in features], dtype=torch.long)
t_all_segment_ids = torch.tensor([f.t_segment_ids for f in features], dtype=torch.long)
t_all_lens = torch.tensor([f.t_input_len for f in features], dtype=torch.long)
t_all_orig_to_tok_index = torch.tensor([f.t_orig_to_tok_index for f in features], dtype=torch.long)
t_all_word_lens = torch.tensor([f.t_word_len for f in features], dtype=torch.long)
a_all_input_ids = torch.tensor([f.a_input_ids for f in features], dtype=torch.long)
a_all_input_mask = torch.tensor([f.a_input_mask for f in features], dtype=torch.long)
a_all_segment_ids = torch.tensor([f.a_segment_ids for f in features], dtype=torch.long)
a_all_lens = torch.tensor([f.a_input_len for f in features], dtype=torch.long)
a_all_orig_to_tok_index = torch.tensor([f.a_orig_to_tok_index for f in features], dtype=torch.long)
a_all_word_lens = torch.tensor([f.a_word_len for f in features], dtype=torch.long)
all_label_ids = torch.tensor([f.label_ids for f in features], dtype=torch.long)
dataset = TensorDataset(t_all_input_ids, t_all_input_mask, t_all_segment_ids, t_all_lens, t_all_orig_to_tok_index, t_all_word_lens,
a_all_input_ids, a_all_input_mask, a_all_segment_ids, a_all_lens, a_all_orig_to_tok_index, a_all_word_lens,
all_label_ids)
return dataset
def evaluate(args, model, tokenizer, prefix=""):
# Loop to handle MNLI double evaluation (matched, mis-matched)
eval_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,)
eval_outputs_dirs = (args.output_dir, args.output_dir + '-MM') if args.task_name == "mnli" else (args.output_dir,)
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, data_type='dev')
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,
collate_fn=collate_fn)
# 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
preds = None
out_label_ids = None
pbar = ProgressBar(n_total=len(eval_dataloader), desc="Evaluating")
for step, batch in enumerate(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],
'labels': batch[3]}
inputs['token_type_ids'] = batch[2]
outputs = model(**inputs)
tmp_eval_loss, logits = outputs[:2]
eval_loss += tmp_eval_loss.mean().item()
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)
pbar(step)
print(' ')
if 'cuda' in str(args.device):
torch.cuda.empty_cache()
eval_loss = eval_loss / nb_eval_steps
preds = np.argmax(preds, axis=1)
result = metrics(eval_task, preds, out_label_ids)
results.update(result)
logger.info("***** Eval results {} *****".format(prefix))
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
return results
def is_valid(self, proxy):
proxies = {"http": "http://" + proxy, "https": "http://" + proxy, }
try:
requests.get(self.TEST_URL, proxies=proxies, timeout=self.TIMEOUT)
except Exception as e:
logger.error(repr(e))
return False
else:
logger.info("it's works {0}".format(proxy))
return True
def refresh_train_data(big_file):
cmd = f"python prepare_lm_data_ngram.py --do_split --big_file={big_file} "
command = cmd
logger.info(f"command:{command}")
os.system(f"{command}")
cmd = f"python prepare_lm_data_ngram.py --do_mdata"
command = cmd
logger.info(f"command:{command}")
os.system(f"{command}")
def is_valid_batch(self, proxys):
"""批量测试代理是否可用"""
i = 1
for proxy in proxys:
logger.info('正在发送第 {0} 个请求。\n\r'.format(i))
i += 1
self.is_valid_batch(proxy)
logger.info('InvalidProxiesList: {0}'.format(len(self.InvalidProxiesList)))
logger.info('***已经有 {0} 个代理被淘汰***'.format(len(self.InvalidProxiesList)))
logger.info('UseProxiesList: {0}'.format(self.UseProxiesList))
logger.info('可用代理数量 {0}'.format(len(self.UseProxiesList)))
def __init__(self, training_path, file_id, tokenizer, data_name, reduce_memory=False):
self.tokenizer = tokenizer
self.file_id = file_id
data_file = training_path / f"{data_name}_file_{self.file_id}.json"
metrics_file = training_path / f"{data_name}_file_{self.file_id}_metrics.json"
assert data_file.is_file() and metrics_file.is_file()
metrics = json.loads(metrics_file.read_text())
num_samples = metrics['num_training_examples']
seq_len = metrics['max_seq_len']
self.temp_dir = None
self.working_dir = None
if reduce_memory:
self.temp_dir = TemporaryDirectory()
self.working_dir = Path(self.temp_dir.name)
input_ids = np.memmap(filename=self.working_dir / 'input_ids.memmap',
mode='w+', dtype=np.int32, shape=(num_samples, seq_len))
input_masks = np.memmap(filename=self.working_dir / 'input_masks.memmap',
shape=(num_samples, seq_len), mode='w+', dtype=np.bool)
segment_ids = np.memmap(filename=self.working_dir / 'segment_ids.memmap',
shape=(num_samples, seq_len), mode='w+', dtype=np.bool)
lm_label_ids = np.memmap(filename=self.working_dir / 'lm_label_ids.memmap',
shape=(num_samples, seq_len), mode='w+', dtype=np.int32)
lm_label_ids[:] = -1
is_nexts = np.memmap(filename=self.working_dir / 'is_nexts.memmap',
shape=(num_samples,), mode='w+', dtype=np.bool)
else:
input_ids = np.zeros(shape=(num_samples, seq_len), dtype=np.int32)
input_masks = np.zeros(shape=(num_samples, seq_len), dtype=np.bool)
segment_ids = np.zeros(shape=(num_samples, seq_len), dtype=np.bool)
lm_label_ids = np.full(shape=(num_samples, seq_len), dtype=np.int32, fill_value=-1)
is_nexts = np.zeros(shape=(num_samples,), dtype=np.bool)
logger.info(f"Loading training examples for {str(data_file)}")
with data_file.open() as f:
for i, line in enumerate(f):
line = line.strip()
example = json.loads(line)
features = convert_example_to_features(example, tokenizer, seq_len)
input_ids[i] = features.input_ids
segment_ids[i] = features.segment_ids
input_masks[i] = features.input_mask
lm_label_ids[i] = features.lm_label_ids
is_nexts[i] = features.is_next
assert i == num_samples - 1 # Assert that the sample count metric was true
logger.info("Loading complete!")
self.num_samples = num_samples
self.seq_len = seq_len
self.input_ids = input_ids
self.input_masks = input_masks
self.segment_ids = segment_ids
self.lm_label_ids = lm_label_ids
self.is_nexts = is_nexts
def predict(args, model, tokenizer, prefix=""):
'''模型预测'''
pred_output_dir = args.output_dir
if not os.path.exists(pred_output_dir) and args.local_rank in [-1, 0]:
os.makedirs(pred_output_dir)
test_dataset = load_and_cache_examples(args,
args.task_name,
tokenizer,
data_type='test')
# Note that DistributedSampler samples randomly
test_sampler = SequentialSampler(
test_dataset) if args.local_rank == -1 else DistributedSampler(
test_dataset)
test_dataloader = DataLoader(test_dataset,
sampler=test_sampler,
batch_size=1,
collate_fn=collate_fn)
# Eval!
logger.info("***** Running prediction %s *****", prefix)
logger.info(" Num examples = %d", len(test_dataset))
logger.info(" Batch size = %d", 1)
results = []
output_submit_file = os.path.join(pred_output_dir, prefix,
"test_prediction.json")
pbar = ProgressBar(n_total=len(test_dataloader), desc="Predicting")
for step, batch in enumerate(test_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],
"labels": None
}
if args.model_type != "distilbert":
# XLM and RoBERTa don"t use segment_ids
inputs["token_type_ids"] = (batch[2] if args.model_type
in ["bert", "xlnet"] else None)
outputs = model(**inputs)
logits = outputs[0]
preds = logits.detach().cpu().numpy()
preds = np.argmax(preds, axis=2).tolist()
preds = preds[0][1:-1] # [CLS]XXXX[SEP]
tags = [args.id2label[x] for x in preds]
label_entities = get_entities(preds, args.id2label,
args.markup) # 得到实体
json_d = {}
json_d['id'] = step
json_d['tag_seq'] = " ".join(tags)
json_d['entities'] = label_entities
results.append(json_d)
pbar(step)
logger.info("\n")
with open(output_submit_file, "w") as writer:
for record in results:
writer.write(json.dumps(record) + '\n')
def create_training_instances(input_file, tokenizer, max_seq_len,
short_seq_prob):
# def create_training_instances(input_file, tokenizer, max_seq_len, short_seq_prob,
# max_ngram, masked_lm_prob, max_predictions_per_seq):
"""Create `TrainingInstance`s from raw text."""
all_documents = [[]]
# Input file format:
# (1) One sentence per line. These should ideally be actual sentences, not
# entire paragraphs or arbitrary spans of text. (Because we use the
# sentence boundaries for the "next sentence prediction" task).
# (2) Blank lines between documents. Document boundaries are needed so
# that the "next sentence prediction" task doesn't span between documents.
with open(input_file, 'r') as f:
lines = f.readlines()
pbar = ProgressBar(n_total=len(lines), desc='read data')
for line_cnt, line in enumerate(lines):
line = line.strip()
# Empty lines are used as document delimiters
if not line:
all_documents.append([])
tokens = tokenizer.tokenize(line)
if tokens:
all_documents[-1].append(tokens)
pbar(step=line_cnt)
print(' ')
# Remove empty documents
all_documents = [x for x in all_documents if x]
random.shuffle(all_documents)
# vocab_words = list(tokenizer.vocab.keys())
instances = []
pbar = ProgressBar(n_total=len(all_documents), desc='create instances')
for document_index in range(len(all_documents)):
instances.extend(
create_instances_from_document(all_documents, document_index,
max_seq_len, short_seq_prob))
pbar(step=document_index)
# def create_instances_from_document(all_documents, document_index, max_seq_length, short_seq_prob):
print(' ')
ex_idx = 0
while ex_idx < 2:
instance = instances[ex_idx]
logger.info("-------------------------Example-----------------------")
logger.info(f"id: {ex_idx}")
logger.info(
f"tokens: {' '.join([str(x) for x in instance['tokens']])}")
# logger.info(f"masked_lm_labels: {' '.join([str(x) for x in instance['masked_lm_labels']])}")
logger.info(
f"segment_ids: {' '.join([str(x) for x in instance['segment_ids']])}"
)
# logger.info(f"masked_lm_positions: {' '.join([str(x) for x in instance['masked_lm_positions']])}")
# logger.info(f"is_random_next : {instance['is_random_next']}")
ex_idx += 1
random.shuffle(instances)
return instances
def main():
parser = ArgumentParser()
## Required parameters
parser.add_argument("--do_data", default=True, action='store_true')
parser.add_argument('--data_name', default='albert', type=str)
parser.add_argument('--max_ngram', default=3, type=int)
parser.add_argument("--do_lower_case", default=False, action='store_true')
parser.add_argument('--seed', default=42, type=int)
# parser.add_argument("--file_num", type=int, default=10, help="Number of dynamic masking to pregenerate (with different masks)")
parser.add_argument("--max_seq_len", type=int, default=128)
parser.add_argument(
"--short_seq_prob",
type=float,
default=0.1,
help="Probability of making a short sentence as a training example")
parser.add_argument(
"--masked_lm_prob",
type=float,
default=0.15,
help="Probability of masking each token for the LM task")
# 128 * 0.15
parser.add_argument(
"--max_predictions_per_seq",
type=int,
default=20,
help="Maximum number of tokens to mask in each sequence")
args = parser.parse_args()
seed_everything(args.seed)
from configs.base import config
args.vocab_path = config['albert_vocab_path']
args.data_dir = config['data_dir']
logger.info("pregenerate training data parameters:\n %s", args)
tokenizer = BertTokenizer(vocab_file=args.vocab_path,
do_lower_case=args.do_lower_case)
small_path = config['data_dir'] / "corpus/small"
files = sorted(
[f for f in small_path.iterdir() if f.exists() and '.txt' in str(f)])
file_path = files[0].absolute()
max_seq_len = args.max_seq_len
train(file_path, tokenizer, max_seq_len)
print(" dataloader ok! ")
sys.exit(0)
def load_and_cache_examples(args, task, tokenizer, data_type='train'):
'''加载数据'''
if args.local_rank not in [-1, 0] and not evaluate:
torch.distributed.barrier(
) # Make sure only the first process in distributed training process the dataset, and the others will use the cache
processor = processors[task]()
# Load data features from cache or dataset file
cached_features_file = os.path.join(
args.data_dir, 'cached_soft-{}_{}_{}_{}'.format(
data_type,
list(filter(None, args.model_name_or_path.split('/'))).pop(),
str(args.train_max_seq_length if data_type ==
'train' else args.eval_max_seq_length), str(task)))
if os.path.exists(cached_features_file) and not args.overwrite_cache:
logger.info("Loading features from cached file %s",
cached_features_file)
features = torch.load(cached_features_file)
else:
logger.info("Creating features from dataset file at %s", args.data_dir)
label_list = processor.get_labels()
if data_type == 'train':
examples = processor.get_train_examples(args.data_dir)
elif data_type == 'dev':
examples = processor.get_dev_examples(args.data_dir)
else:
examples = processor.get_test_examples(args.data_dir)
features = convert_examples_to_features(examples=examples,
tokenizer=tokenizer,
label_list=label_list,
max_seq_length=args.train_max_seq_length if data_type=='train' \
else args.eval_max_seq_length,
cls_token_at_end=bool(args.model_type in ["xlnet"]),
pad_on_left=bool(args.model_type in ['xlnet']),
cls_token = tokenizer.cls_token,
cls_token_segment_id=2 if args.model_type in ["xlnet"] else 0,
sep_token=tokenizer.sep_token,
# pad on the left for xlnet
pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0,
)
if args.local_rank in [-1, 0]:
logger.info("Saving features into cached file %s",
cached_features_file)
torch.save(features, cached_features_file)
if args.local_rank == 0 and not evaluate:
torch.distributed.barrier(
) # Make sure only the first process in distributed training process the dataset, and the others will use the cache
# Convert to Tensors and build dataset
all_input_ids = torch.tensor([f.input_ids for f in features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_ids for f in features],
dtype=torch.long)
all_lens = torch.tensor([f.input_len for f in features], dtype=torch.long)
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
all_lens, all_label_ids)
return dataset
def apicall():
logger.info('<<albert_cls')
try:
start = time.time()
req_data = request.get_data(as_text=True)
logger.info(req_data)
if req_data:
req_data = json.loads(req_data)
text = req_data['text']
task = req_data['task']
dict_res = for_server(text=text, task=task)
logger.info("时间{}".format(time.time() - start))
dict_ = {}
dict_['status'] = 'success'
dict_['results'] = dict_res
return jsonify(dict_)
else:
res = {'status': 'failed', 'results': '没有收到request消息'}
return jsonify(res)
except BadRequestKeyError as e:
logger.error(e)
res = {'status': 'failed', 'results': str(e)}
return jsonify(res)
except FileNotFoundError as e:
logger.error(e)
res = {'status': 'failed', 'results': e.strerror}
return jsonify(res)
except Exception as e:
logger.error(e)
res = {'status': 'failed', 'results': str(e)}
return jsonify(res)
def predict(args, model, tokenizer, prefix=""):
pred_output_dir = args.output_dir
if not os.path.exists(pred_output_dir) and args.local_rank in [-1, 0]:
os.makedirs(pred_output_dir)
test_dataset = load_and_cache_examples(args,
args.task_name,
tokenizer,
data_type='test')
print(len(test_dataset))
# Note that DistributedSampler samples randomly
test_sampler = SequentialSampler(
test_dataset) if args.local_rank == -1 else DistributedSampler(
test_dataset)
test_dataloader = DataLoader(test_dataset,
sampler=test_sampler,
batch_size=1,
collate_fn=collate_fn)
# Eval!
logger.info("***** Running prediction %s *****", prefix)
logger.info(" Num examples = %d", len(test_dataset))
logger.info(" Batch size = %d", 1)
results = []
output_submit_file = os.path.join(pred_output_dir, prefix,
"test_prediction.json")
pbar = ProgressBar(n_total=len(test_dataloader), desc="Predicting")
for step, batch in enumerate(test_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],
"start_positions": None,
"end_positions": None
}
if args.model_type != "distilbert":
# XLM and RoBERTa don"t use segment_ids
inputs["token_type_ids"] = (batch[2] if args.model_type
in ["bert", "xlnet"] else None)
outputs = model(**inputs)
start_logits, end_logits = outputs[:2]
R = bert_extract_item(start_logits, end_logits)
if R:
label_entities = [[args.id2label[x[0]], x[1], x[2]] for x in R]
else:
label_entities = []
json_d = {}
json_d['id'] = step
json_d['entities'] = label_entities
results.append(json_d)
pbar(step)
print(" ")
with open(output_submit_file, "w") as writer:
for record in results:
writer.write(json.dumps(record) + '\n')
def load_and_cache_examples(args, task, tokenizer, data_type='train'):
if args.local_rank not in [-1, 0] and not evaluate:
torch.distributed.barrier(
) # Make sure only the first process in distributed training process the dataset, and the others will use the cache
processor = processors[task]()
logger.info("Creating features from dataset file at %s", args.data_dir)
label_list = processor.get_labels() #[B I O]
if data_type == 'train':
examples = processor.get_train_examples(args.data_dir)
elif data_type == 'dev':
examples = processor.get_dev_examples(args.data_dir)
else:
examples = processor.get_test_examples(args.data_dir)
features = convert_examples_to_features(examples=examples,
tokenizer=tokenizer,
label_list=label_list,
max_seq_length=args.max_seq_length if data_type == 'train' \
else args.max_seq_length,
pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0,
)
return features
def predict(args, model, tokenizer, lines, prefix=""):
pred_output_dir = args.output_dir
if not os.path.exists(pred_output_dir) and args.local_rank in [-1, 0]:
os.makedirs(pred_output_dir)
test_dataset = load_and_cache_examples(args,
args.task_name,
tokenizer,
lines,
data_type='test')
# Note that DistributedSampler samples randomly
test_sampler = SequentialSampler(
test_dataset) if args.local_rank == -1 else DistributedSampler(
test_dataset)
test_dataloader = DataLoader(test_dataset,
sampler=test_sampler,
batch_size=1,
collate_fn=collate_fn)
# Eval!
logger.info("***** Running prediction %s *****", prefix)
logger.info(" Num examples = %d", len(test_dataset))
logger.info(" Batch size = %d", 1)
results = []
pbar = ProgressBar(n_total=len(test_dataloader), desc="Predicting")
if isinstance(model, nn.DataParallel):
model = model.module
for step, batch in enumerate(test_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],
"labels": None,
'input_lens': batch[4]
}
if args.model_type != "distilbert":
# XLM and RoBERTa don"t use segment_ids
inputs["token_type_ids"] = (batch[2] if args.model_type
in ["bert", "xlnet"] else None)
outputs = model(**inputs)
logits = outputs[0]
preds, _ = model.crf._obtain_labels(logits, args.id2label,
inputs['input_lens'])
preds = preds[0][1:-1] # [CLS]XXXX[SEP]
label_entities = get_entities(preds, args.id2label, args.markup)
json_d = {}
json_d['id'] = step
json_d['tag_seq'] = " ".join(preds)
json_d['entities'] = label_entities
results.append(json_d)
pbar(step)
print(results[:3])
def predict(args, model, tokenizer, prefix=""):
# Loop to handle MNLI double evaluation (matched, mis-matched)
pred_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,)
pred_outputs_dirs = (args.output_dir, args.output_dir + '-MM') if args.task_name == "mnli" else (args.output_dir,)
results = {}
for pred_task, pred_output_dir in zip(pred_task_names, pred_outputs_dirs):
pred_dataset = load_and_cache_examples(args, pred_task, tokenizer, data_type='test')
if not os.path.exists(pred_output_dir) and args.local_rank in [-1, 0]:
os.makedirs(pred_output_dir)
args.pred_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
# Note that DistributedSampler samples randomly
pred_sampler = SequentialSampler(pred_dataset) if args.local_rank == -1 else DistributedSampler(pred_dataset)
pred_dataloader = DataLoader(pred_dataset, sampler=pred_sampler, batch_size=args.pred_batch_size,
collate_fn=collate_fn)
logger.info("***** Running prediction {} *****".format(prefix))
logger.info(" Num examples = %d", len(pred_dataset))
logger.info(" Batch size = %d", args.pred_batch_size)
nb_pred_steps = 0
preds = None
pbar = ProgressBar(n_total=len(pred_dataloader), desc="Predicting")
for step, batch in enumerate(pred_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],
'labels': batch[3]}
if args.model_type != 'distilbert':
inputs['token_type_ids'] = batch[2] if (
'bert' in args.model_type or 'xlnet' in args.model_type) else None # XLM, DistilBERT and RoBERTa don't use segment_ids
outputs = model(**inputs)
_, logits = outputs[:2]
nb_pred_steps += 1
if preds is None:
preds = logits.detach().cpu().numpy()
else:
preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
pbar(step)
print(' ')
if args.output_mode == "classification":
preds = np.argmax(preds, axis=1)
elif args.output_mode == "regression":
preds = np.squeeze(preds)
output_pred_file = os.path.join(pred_output_dir, prefix, "test_prediction.txt")
with open(output_pred_file, "w") as writer:
for pred in preds:
writer.write(str(pred) + '\n')
return results
def train(args, train_dataset, model, tokenizer, config):
""" Train the model """
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_sampler = RandomSampler(
train_dataset) if args.local_rank == -1 else DistributedSampler(
train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size,
collate_fn=collate_fn)
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = args.max_steps // (
len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
t_total = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ["bias", "LayerNorm.weight"]
bert_param_optimizer = list(model.bert.named_parameters())
crf_param_optimizer = list(model.crf.named_parameters())
linear_param_optimizer = list(model.classifier.named_parameters())
optimizer_grouped_parameters = [{
'params': [
p for n, p in bert_param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
args.weight_decay,
'lr':
args.learning_rate
}, {
'params': [
p for n, p in bert_param_optimizer
if any(nd in n for nd in no_decay)
],
'weight_decay':
0.0,
'lr':
args.learning_rate
}, {
'params': [
p for n, p in crf_param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
args.weight_decay,
'lr':
args.crf_learning_rate
}, {
'params':
[p for n, p in crf_param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0,
'lr':
args.crf_learning_rate
}, {
'params': [
p for n, p in linear_param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
args.weight_decay,
'lr':
args.crf_learning_rate
}, {
'params': [
p for n, p in linear_param_optimizer
if any(nd in n for nd in no_decay)
],
'weight_decay':
0.0,
'lr':
args.crf_learning_rate
}]
args.warmup_steps = int(t_total * args.warmup_proportion)
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=t_total)
# Check if saved optimizer or scheduler states exist
if os.path.isfile(os.path.join(
args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
os.path.join(args.model_name_or_path, "scheduler.pt")):
# Load in optimizer and scheduler states
optimizer.load_state_dict(
torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
scheduler.load_state_dict(
torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d",
args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps *
(torch.distributed.get_world_size() if args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
global_step = 0
steps_trained_in_current_epoch = 0
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path
) and "checkpoint" in args.model_name_or_path:
# set global_step to gobal_step of last saved checkpoint from model path
global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
epochs_trained = global_step // (len(train_dataloader) //
args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (
len(train_dataloader) // args.gradient_accumulation_steps)
logger.info(
" Continuing training from checkpoint, will skip to saved global_step"
)
logger.info(" Continuing training from epoch %d", epochs_trained)
logger.info(" Continuing training from global step %d", global_step)
logger.info(" Will skip the first %d steps in the first epoch",
steps_trained_in_current_epoch)
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
seed_everything(
args.seed
) # Added here for reproductibility (even between python 2 and 3)
# train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
best_eval_p = 0.0
best_eval_r = 0.0
best_eval_f1 = 0.0
for _ in range(int(args.num_train_epochs)):
# pbar = ProgressBar(n_total=len(train_dataloader), desc='Training')
epoch_iterator = tqdm(train_dataloader,
desc="Iteration",
disable=args.local_rank not in [-1, 0])
for step, batch in enumerate(epoch_iterator):
# Skip past any already trained steps if resuming training
if steps_trained_in_current_epoch > 0:
steps_trained_in_current_epoch -= 1
continue
model.train()
batch = tuple(t.to(args.device) for t in batch)
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"labels": batch[3],
'input_lens': batch[4]
}
if args.model_type != "distilbert":
# XLM and RoBERTa don"t use segment_ids
inputs["token_type_ids"] = (batch[2] if args.model_type
in ["bert", "xlnet"] else None)
outputs = model(**inputs)
loss = outputs[
0] # model outputs are always tuple in pytorch-transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean(
) # mean() to average on multi-gpu parallel training
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# pbar(step, {'loss': loss.item()})
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
scheduler.step() # Update learning rate schedule
optimizer.step()
model.zero_grad()
global_step += 1
# if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
# # Log metrics
# print(" ")
# if args.local_rank == -1:
# # Only evaluate when single GPU otherwise metrics may not average well
# evaluate(args, model, tokenizer)
# if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
# # Save model checkpoint
# output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
# if not os.path.exists(output_dir):
# os.makedirs(output_dir)
# model_to_save = (
# model.module if hasattr(model, "module") else model
# ) # Take care of distributed/parallel training
# model_to_save.save_pretrained(output_dir)
# torch.save(args, os.path.join(output_dir, "training_args.bin"))
# logger.info("Saving model checkpoint to %s", output_dir)
# tokenizer.save_vocabulary(output_dir)
# torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
# torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
# logger.info("Saving optimizer and scheduler states to %s", output_dir)
# best eval
results = evaluate(args, model, tokenizer)
if results["f1"] > best_eval_f1:
best_eval_p = results['acc']
best_eval_r = results['recall']
best_eval_f1 = results["f1"]
## 保存最好模型
args.best_eval_output_dir = os.path.join(args.output_dir)
if not os.path.exists(args.best_eval_output_dir):
os.makedirs(args.best_eval_output_dir)
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(args.best_eval_output_dir)
torch.save(
args,
os.path.join(args.best_eval_output_dir, "training_args.bin"))
logger.info("eval results: p:{:.4f} r:{:.4f} f1:{:.4f}".format(
best_eval_p, best_eval_r, best_eval_f1))
logger.info("Saving step:{} as best model to {}".format(
global_step, args.best_eval_output_dir))
tokenizer.save_vocabulary(args.best_eval_output_dir)
torch.save(optimizer.state_dict(),
os.path.join(args.best_eval_output_dir, "optimizer.pt"))
torch.save(scheduler.state_dict(),
os.path.join(args.best_eval_output_dir, "scheduler.pt"))
# logger.info("Saving optimizer and scheduler states to %s", args.best_eval_output_dir)
# config_file = os.path.join(args.best_eval_output_dir, "best_config.json")
# json.dump(config, config_file)
logger.info("\n")
if 'cuda' in str(args.device):
torch.cuda.empty_cache()
if args.do_predict:
config_class, model_class, tokenizer_class = MODEL_CLASSES[
args.model_type]
model = model_class.from_pretrained(args.best_eval_output_dir,
config=config)
model.to(args.device)
predict(args, model, tokenizer)
return global_step, tr_loss / global_step
def evaluate(args, model, tokenizer, prefix=""):
metric = SeqEntityScore(args.id2label, markup=args.markup)
eval_output_dir = args.output_dir
if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
os.makedirs(eval_output_dir)
eval_dataset = load_and_cache_examples(args,
args.task_name,
tokenizer,
data_type='dev')
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,
collate_fn=collate_fn)
# Eval!
logger.info("***** Running evaluation %s *****", 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
# pbar = ProgressBar(n_total=len(eval_dataloader), desc="Evaluating")
eval_iterator = tqdm(eval_dataloader,
desc="Iteration",
disable=args.local_rank not in [-1, 0])
if isinstance(model, nn.DataParallel):
model = model.module
for step, batch in enumerate(eval_iterator):
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],
"labels": batch[3],
'input_lens': batch[4]
}
if args.model_type != "distilbert":
# XLM and RoBERTa don"t use segment_ids
inputs["token_type_ids"] = (batch[2] if args.model_type
in ["bert", "xlnet"] else None)
outputs = model(**inputs)
tmp_eval_loss, logits = outputs[:2]
tags = model.crf.decode(logits, inputs['attention_mask'])
if args.n_gpu > 1:
tmp_eval_loss = tmp_eval_loss.mean(
) # mean() to average on multi-gpu parallel evaluating
eval_loss += tmp_eval_loss.item()
nb_eval_steps += 1
out_label_ids = inputs['labels'].cpu().numpy().tolist()
input_lens = inputs['input_lens'].cpu().numpy().tolist()
tags = tags.squeeze(0).cpu().numpy().tolist()
for i, label in enumerate(out_label_ids):
temp_1 = []
temp_2 = []
for j, m in enumerate(label):
if j == 0:
continue
elif j == input_lens[i] - 1:
metric.update(pred_paths=[temp_2], label_paths=[temp_1])
break
else:
temp_1.append(args.id2label[out_label_ids[i][j]])
temp_2.append(args.id2label[tags[i][j]])
# pbar(step)
logger.info("\n")
eval_loss = eval_loss / nb_eval_steps
eval_info, entity_info = metric.result()
results = {f'{key}': value for key, value in eval_info.items()}
results['loss'] = eval_loss
logger.info("***** Eval results %s *****", prefix)
info = "-".join(
[f' {key}: {value:.4f} ' for key, value in results.items()])
logger.info(info)
logger.info("***** Entity results %s *****", prefix)
for key in sorted(entity_info.keys()):
logger.info("******* %s results ********" % key)
info = "-".join([
f' {key}: {value:.4f} ' for key, value in entity_info[key].items()
])
logger.info(info)
return results
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train selected in the list: " +
", ".join(processors.keys()))
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the training files for the CoNLL-2003 NER task.",
)
parser.add_argument(
"--model_type",
default=None,
type=str,
required=True,
help="Model type selected in the list: " +
", ".join(MODEL_CLASSES.keys()),
)
parser.add_argument(
"--model_name_or_path",
default=None,
type=str,
required=True,
help="Path to pre-trained model or shortcut name selected in the list: "
+ ", ".join(ALL_MODELS),
)
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('--markup',
default='bios',
type=str,
choices=['bios', 'bio'])
parser.add_argument('--loss_type',
default='ce',
type=str,
choices=['lsr', 'focal', 'ce'])
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 s3",
)
parser.add_argument(
"--train_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(
"--eval_max_seq_length",
default=512,
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",
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_predict",
action="store_true",
help="Whether to run predictions on the test set.")
parser.add_argument(
"--evaluate_during_training",
action="store_true",
help="Whether to 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("--do_adv",
action="store_true",
help="Whether to adversarial training.")
parser.add_argument('--adv_epsilon',
default=1.0,
type=float,
help="Epsilon for adversarial.")
parser.add_argument('--adv_name',
default='word_embeddings',
type=str,
help="name for adversarial layer.")
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("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay",
default=0.01,
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_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("--logging_steps",
type=int,
default=50,
help="Log every X updates steps.")
parser.add_argument("--save_steps",
type=int,
default=50,
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(
'--predict_checkpoints',
action="store_true",
help=
"Predict 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("--server_ip",
type=str,
default="",
help="For distant debugging.")
parser.add_argument("--server_port",
type=str,
default="",
help="For distant debugging.")
args = parser.parse_args()
if not os.path.exists(args.output_dir):
os.mkdir(args.output_dir)
args.output_dir = args.output_dir + '{}'.format(args.model_type)
if not os.path.exists(args.output_dir):
os.mkdir(args.output_dir)
time_ = time.strftime("%Y-%m-%d-%H:%M:%S", time.localtime())
init_logger(log_file=args.output_dir +
f'/{args.model_type}-{args.task_name}-{time_}.log')
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 = 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
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
seed_everything(args.seed)
# Prepare NER task
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]()
label_list = processor.get_labels()
args.id2label = {i: label for i, label in enumerate(label_list)}
args.label2id = {label: i for i, label in enumerate(label_list)}
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]
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_labels,
loss_type=args.loss_type,
cache_dir=args.cache_dir if args.cache_dir else None,
)
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,
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,
)
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,
data_type='train')
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step,
tr_loss)
# 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):
# Create output directory if needed
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
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()`
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_vocabulary(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"))
# Evaluation
results = {}
if args.do_eval and args.local_rank in [-1, 0]:
tokenizer = tokenizer_class.from_pretrained(
args.output_dir, do_lower_case=args.do_lower_case)
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("pytorch_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 = model_class.from_pretrained(checkpoint)
model.to(args.device)
result = evaluate(args, model, tokenizer, prefix=prefix)
if global_step:
result = {
"{}_{}".format(global_step, k): v
for k, v in result.items()
}
results.update(result)
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
for key in sorted(results.keys()):
writer.write("{} = {}\n".format(key, str(results[key])))
if args.do_predict and args.local_rank in [-1, 0]:
tokenizer = tokenizer_class.from_pretrained(
args.output_dir, do_lower_case=args.do_lower_case)
checkpoints = [args.output_dir]
if args.predict_checkpoints > 0:
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
checkpoints = [
x for x in checkpoints
if x.split('-')[-1] == str(args.predict_checkpoints)
]
logger.info("Predict the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
prefix = checkpoint.split(
'/')[-1] if checkpoint.find('checkpoint') != -1 else ""
model = model_class.from_pretrained(checkpoint)
model.to(args.device)
predict(args, model, tokenizer, prefix=prefix)
def main():
args = deal_parser()
if not os.path.exists(args.output_dir):
os.mkdir(args.output_dir)
args.output_dir = args.output_dir + '{}'.format(args.model_type)
if not os.path.exists(args.output_dir):
os.mkdir(args.output_dir)
init_logger(log_file=args.output_dir +
'/{}-{}.log'.format(args.model_type, args.task_name))
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 = 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
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
seed_everything(args.seed)
# Prepare GLUE task
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()
if args.local_rank == 0:
torch.distributed.barrier(
) # Make sure only the first process in distributed training will download model & vocab
config = BertConfig.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_labels,
finetuning_task=args.task_name)
# albert model
# tokenizer = tokenization_albert.FullTokenizer(vocab_file=args.vocab_file, do_lower_case=args.do_lower_case,spm_model_file=args.spm_model_file)
# model = AlbertForSequenceClassification.from_pretrained(args.model_name_or_path,from_tf=bool('.ckpt' in args.model_name_or_path),config=config)
# bert model
tokenizer = tokenization_bert.BertTokenizer(
vocab_file=args.vocab_file,
do_lower_case=args.do_lower_case,
spm_model_file=args.spm_model_file)
model = BertForSequenceClassification.from_pretrained(
args.model_name_or_path,
from_tf=bool('.ckpt' in args.model_name_or_path),
config=config)
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,
data_type='train')
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step,
tr_loss)
# 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):
# Create output directory if needed
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
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()`
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)
# Good practice: save your training arguments together with the trained model
torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
# Evaluation
results = []
if args.do_eval and args.local_rank in [-1, 0]:
# albert model
#tokenizer = tokenization_albert.FullTokenizer(vocab_file=args.vocab_file,do_lower_case=args.do_lower_case,spm_model_file=args.spm_model_file)
# bert model
tokenizer = tokenization_bert.BertTokenizer(
vocab_file=args.vocab_file,
do_lower_case=args.do_lower_case,
spm_model_file=args.spm_model_file)
checkpoints = [(0, 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)))
checkpoints = [(int(checkpoint.split('-')[-1]), checkpoint)
for checkpoint in checkpoints
if checkpoint.find('checkpoint') != -1]
checkpoints = sorted(checkpoints, key=lambda x: x[0])
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 ""
# albert model
# model = AlbertForSequenceClassification.from_pretrained(checkpoint)
# bert model
model = BertForSequenceClassification.from_pretrained(checkpoint)
model.to(args.device)
result = evaluate(args, model, tokenizer, prefix=prefix)
results.extend([(k + '_{}'.format(global_step), v)
for k, v in result.items()])
output_eval_file = os.path.join(args.output_dir,
"checkpoint_eval_results.txt")
with open(output_eval_file, "w") as writer:
for key, value in results:
writer.write("%s = %s\n" % (key, str(value)))
# Test
results = []
if args.do_predict and args.local_rank in [-1, 0]:
# albert model
# tokenizer = tokenization_albert.FullTokenizer(vocab_file=args.vocab_file,do_lower_case=args.do_lower_case,spm_model_file=args.spm_model_file)
# bert model
tokenizer = tokenization_bert.BertTokenizer(
vocab_file=args.vocab_file,
do_lower_case=args.do_lower_case,
spm_model_file=args.spm_model_file)
checkpoints = [(0, args.output_dir)]
if args.predict_all_checkpoints:
checkpoints = list(
os.path.dirname(c) for c in sorted(
glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME,
recursive=True)))
checkpoints = [(int(checkpoint.split('-')[-1]), checkpoint)
for checkpoint in checkpoints
if checkpoint.find('checkpoint') != -1]
checkpoints = sorted(checkpoints, key=lambda x: x[0])
logger.info("Test 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 ""
# albert model
# model = AlbertForSequenceClassification.from_pretrained(checkpoint)
# bert model
model = BertForSequenceClassification.from_pretrained(checkpoint)
model.to(args.device)
result = test(args, model, tokenizer, prefix=prefix)
results.extend([(k + '_{}'.format(global_step), v)
for k, v in result.items()])
output_test_file = os.path.join(args.output_dir,
"checkpoint_test_results.txt")
with open(output_test_file, "w") as writer:
for key, value in results:
writer.write("%s = %s\n" % (key, str(value)))
def train(args, train_dataset, model, tokenizer):
""" Train the model """
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_sampler = RandomSampler(
train_dataset) if args.local_rank == -1 else DistributedSampler(
train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size,
collate_fn=collate_fn)
if args.max_steps > 0:
num_training_steps = args.max_steps
args.num_train_epochs = args.max_steps // (
len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
num_training_steps = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
args.warmup_steps = int(num_training_steps * args.warmup_proportion)
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ['bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
'weight_decay':
args.weight_decay
}, {
'params': [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
'weight_decay':
0.0
}]
# optimizer = Lamb(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
optimizer = AdamW(params=optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=num_training_steps)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d",
args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps *
(torch.distributed.get_world_size() if args.local_rank != -1 else 1))
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", num_training_steps)
global_step = 0
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
seed_everything(
args.seed
) # Added here for reproductibility (even between python 2 and 3)
for _ in range(int(args.num_train_epochs)):
pbar = ProgressBar(n_total=len(train_dataloader), desc='Training')
for step, batch in enumerate(train_dataloader):
model.train()
batch = tuple(t.to(args.device) for t in batch)
inputs = {
'input_ids': batch[0],
'attention_mask': batch[1],
'labels': batch[3],
'token_type_ids': batch[2]
}
outputs = model(**inputs)
loss = outputs[
0] # model outputs are always tuple in transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean(
) # mean() to average on multi-gpu parallel training
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer),
args.max_grad_norm)
else:
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
scheduler.step() # Update learning rate schedule
model.zero_grad()
global_step += 1
if args.local_rank in [
-1, 0
] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
# Log metrics
if args.local_rank == -1: # Only evaluate when single GPU otherwise metrics may not average well
evaluate(args, model, tokenizer)
if args.local_rank in [
-1, 0
] and args.save_steps > 0 and global_step % args.save_steps == 0:
# Save model checkpoint
output_dir = os.path.join(args.output_dir,
'checkpoint-{}'.format(global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model_to_save = model.module if hasattr(
model, 'module'
) else model # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
torch.save(args, os.path.join(output_dir, 'training_args.bin'))
logger.info("Saving model checkpoint to %s", output_dir)
pbar(step, {'loss': loss.item()})
print(" ")
if 'cuda' in str(args.device):
torch.cuda.empty_cache()
return global_step, tr_loss / global_step
def test(args, model, tokenizer, prefix=""):
# Loop to handle MNLI double evaluation (matched, mis-matched)
test_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (
args.task_name, )
test_outputs_dirs = (args.output_dir, args.output_dir +
'-MM') if args.task_name == "mnli" else (
args.output_dir, )
results = {}
for test_task, test_output_dir in zip(test_task_names, test_outputs_dirs):
test_dataset = load_and_cache_examples(args,
test_task,
tokenizer,
data_type='test')
if not os.path.exists(test_output_dir) and args.local_rank in [-1, 0]:
os.makedirs(test_output_dir)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(
1, args.n_gpu)
# Note that DistributedSampler samples randomly
test_sampler = SequentialSampler(
test_dataset) if args.local_rank == -1 else DistributedSampler(
test_dataset)
test_dataloader = DataLoader(test_dataset,
sampler=test_sampler,
batch_size=args.eval_batch_size,
collate_fn=collate_fn)
# Test!
logger.info("***** Running test {} *****".format(prefix))
logger.info(" Num examples = %d", len(test_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
eval_loss = 0.0
nb_eval_steps = 0
preds = None
out_label_ids = None
pbar = ProgressBar(n_total=len(test_dataloader), desc="Testing")
for step, batch in enumerate(test_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],
'labels': batch[3],
'token_type_ids': batch[2]
}
outputs = model(**inputs)
tmp_eval_loss, logits = outputs[:2]
eval_loss += tmp_eval_loss.mean().item()
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)
pbar(step)
print(' ')
if 'cuda' in str(args.device):
torch.cuda.empty_cache()
eval_loss = eval_loss / nb_eval_steps
if args.output_mode == "classification":
preds = np.argmax(preds, axis=1)
elif args.output_mode == "regression":
preds = np.squeeze(preds)
result = compute_metrics(test_task, preds, out_label_ids)
results.update(result)
logger.info("***** Test results {} *****".format(prefix))
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
classreport = ClassReport([
'Joint', 'Sequence', 'Progression', "Contrast", "Supplement",
"Cause-Result", "Result-Cause", "Background", "Behavior-Purpose",
"Purpose-Behavior", "Elaboration", "Summary", "Evaluation",
"Statement-Illustration", "Illustration-Statement"
])
classreport(preds, out_label_ids)
logger.info("%s : %s", classreport.name(), classreport.value())
return 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_type", default=None, type=str, required=True,
help="Model type selected in the list: ")
parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
help="Path to pre-trained model or shortcut name selected in the list")
parser.add_argument("--task_name", default=None, type=str, required=True,
help="The name of the task to train selected in the list: " + ", ".join(processors.keys()))
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("--vocab_file",default='', type=str)
parser.add_argument("--spm_model_file",default='',type=str)
## Other parameters
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 s3")
parser.add_argument("--max_seq_length", default=512, 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", 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("--output_eval", action='store_true',
help="Whether to write output result.")
parser.add_argument("--do_predict", action='store_true',
help="Whether to run the model in inference mode on the test set.")
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('--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-5, 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-6, 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_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('--logging_steps', type=int, default=10,
help="Log every X updates steps.")
parser.add_argument('--save_steps', type=int, default=1000,
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('--server_ip', type=str, default='', help="For distant debugging.")
parser.add_argument('--server_port', type=str, default='', help="For distant debugging.")
parser.add_argument("--label_with_bi", action='store_true', help="Label with B/I")
args = parser.parse_args()
if not os.path.exists(args.output_dir):
os.mkdir(args.output_dir)
args.output_dir = args.output_dir + '{}'.format(args.model_type)
if not os.path.exists(args.output_dir):
os.mkdir(args.output_dir)
init_logger(log_file=args.output_dir + '/{}-{}.log'.format(args.model_type, args.task_name))
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 = 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
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
seed_everything(args.seed)
# Prepare GLUE task
args.task_name = args.task_name.lower()
if args.task_name != "ner":
raise ValueError("Task error: %s, must be ner" % (args.task_name))
processor = processors[args.task_name]()
args.output_mode = output_modes[args.task_name]
label_list = processor.get_labels_ner(args.data_dir, args.label_with_bi)
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 = AlbertConfig.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 = tokenization_albert.FullTokenizer(vocab_file=args.vocab_file, do_lower_case=args.do_lower_case,
spm_model_file=args.spm_model_file)
model =AlbertFocalLossForNer.from_pretrained(args.model_name_or_path,
from_tf=bool('.ckpt' in args.model_name_or_path),
config=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, data_type='train')
global_step, tr_loss = train(args, train_dataset, label_list, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
# 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):
# Create output directory if needed
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
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()`
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)
# Good practice: save your training arguments together with the trained model
torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
# Evaluation
results = []
if args.do_eval and args.local_rank in [-1, 0]:
tokenizer = tokenization_albert.FullTokenizer(vocab_file=args.vocab_file,
do_lower_case=args.do_lower_case,
spm_model_file=args.spm_model_file)
checkpoints = [(0,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)))
checkpoints = [(int(checkpoint.split('-')[-1]),checkpoint) for checkpoint in checkpoints if checkpoint.find('checkpoint') != -1]
checkpoints = sorted(checkpoints,key =lambda x:x[0])
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 = AlbertFocalLossForNer.from_pretrained(checkpoint)
model.to(args.device)
result = evaluate(args, model, tokenizer, label_list, prefix=prefix)
results.extend([(k + '_{}'.format(global_step), v) for k, v in result.items()])
output_eval_file = os.path.join(args.output_dir, "checkpoint_eval_results.txt")
with open(output_eval_file, "w") as writer:
for key,value in results:
writer.write("%s = %s\n" % (key, str(value)))
if args.do_predict and args.local_rank in [-1, 0]:
tokenizer = tokenization_albert.FullTokenizer(vocab_file=args.vocab_file,
do_lower_case=args.do_lower_case,
spm_model_file=args.spm_model_file)
result = evaluate(args, model, tokenizer, label_list, prefix="")
output_eval_file = os.path.join(args.output_dir, "checkpoint_eval_results.txt")
with open(output_eval_file, "w") as writer:
for key,value in result.items():
writer.write("%s = %s\n" % (key, str(value)))
def main():
parser = ArgumentParser()
## Required parameters
parser.add_argument(
"--data_dir",
default="dataset",
type=str,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument("--config_path",
default="prev_trained_model/electra_small/config.json",
type=str)
parser.add_argument("--vocab_path",
default="prev_trained_model/electra_small/vocab.txt",
type=str)
parser.add_argument(
"--output_dir",
default="outputs",
type=str,
help=
"The output directory where the model predictions and checkpoints will be written."
)
parser.add_argument("--model_path",
default='prev_trained_model/electra_small',
type=str)
parser.add_argument('--data_name', default='electra', type=str)
parser.add_argument(
"--file_num",
type=int,
default=10,
help="Number of dynamic masking to pregenerate (with different masks)")
parser.add_argument(
"--reduce_memory",
action="store_true",
help=
"Store training data as on-disc memmaps to massively reduce memory usage"
)
parser.add_argument("--epochs",
type=int,
default=4,
help="Number of epochs to train for")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument('--num_eval_steps', default=100)
parser.add_argument('--num_save_steps', default=2000)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument("--weight_decay",
default=0.01,
type=float,
help="Weight deay if we apply some.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument("--train_batch_size",
default=128,
type=int,
help="Total batch size for training.")
parser.add_argument("--gen_weight",
default=1.0,
type=float,
help='masked language modeling / generator loss')
parser.add_argument("--disc_weight",
default=50,
type=float,
help='discriminator loss')
parser.add_argument('--untied_generator',
action='store_true',
help='tie all generator/discriminator weights?')
parser.add_argument('--temperature',
default=0,
type=float,
help='temperature for sampling from generator')
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument("--warmup_proportion",
default=0.1,
type=float,
help="Linear warmup over warmup_steps.")
parser.add_argument("--adam_epsilon",
default=1e-8,
type=float,
help="Epsilon for Adam optimizer.")
parser.add_argument('--max_grad_norm', default=1.0, type=float)
parser.add_argument("--learning_rate",
default=0.000176,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--fp16_opt_level',
type=str,
default='O2',
help=
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html")
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--continue_train',
default='',
help="continue train path")
args = parser.parse_args()
args.data_dir = Path(args.data_dir)
args.output_dir = Path(args.output_dir)
pregenerated_data = args.data_dir / "corpus/train"
init_logger(log_file=str(args.output_dir / "train_albert_model.log"))
assert pregenerated_data.is_dir(), \
"--pregenerated_data should point to the folder of files made by prepare_lm_data_mask.py!"
samples_per_epoch = 0
for i in range(args.file_num):
data_file = pregenerated_data / f"{args.data_name}_file_{i}.json"
metrics_file = pregenerated_data / f"{args.data_name}_file_{i}_metrics.json"
if data_file.is_file() and metrics_file.is_file():
metrics = json.loads(metrics_file.read_text())
samples_per_epoch += metrics['num_training_examples']
else:
if i == 0:
exit("No training data was found!")
print(
f"Warning! There are fewer epochs of pregenerated data ({i}) than training epochs ({args.epochs})."
)
print(
"This script will loop over the available data, but training diversity may be negatively impacted."
)
break
logger.info(f"samples_per_epoch: {samples_per_epoch}")
if args.local_rank == -1 or args.no_cuda:
device = torch.device(f"cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
args.n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
f"device: {device} , distributed training: {bool(args.local_rank != -1)}, 16-bits training: {args.fp16}"
)
if args.gradient_accumulation_steps < 1:
raise ValueError(
f"Invalid gradient_accumulation_steps parameter: {args.gradient_accumulation_steps}, should be >= 1"
)
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
seed_everything(args.seed)
tokenizer = BertTokenizer.from_pretrained(args.vocab_path,
do_lower_case=args.do_lower_case)
total_train_examples = samples_per_epoch * args.epochs
num_train_optimization_steps = int(total_train_examples /
args.train_batch_size /
args.gradient_accumulation_steps)
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
args.warmup_steps = int(num_train_optimization_steps *
args.warmup_proportion)
bert_config = ElectraConfig.from_pretrained(args.config_path,
gen_weight=args.gen_weight,
temperature=args.temperature,
disc_weight=args.disc_weight)
model = ElectraForPreTraining(config=bert_config)
if args.continue_train:
print(f"Continue train from {args.continue_train}")
model = model.from_pretrained(args.continue_train)
elif args.model_path:
print("载入预训练模型")
model.generator = AutoModel.from_pretrained(args.model_path + "/G")
model.electra = AutoModel.from_pretrained(args.model_path + "/D")
# print(model)
model.to(device)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
args.weight_decay
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
optimizer = AdamW(params=optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=num_train_optimization_steps)
# optimizer = Lamb(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
# if args.model_path:
# optimizer.load_state_dict(torch.load(args.model_path + "/optimizer.bin"))
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
if args.n_gpu > 1:
# model = BalancedDataParallel(gpu0_bsz=32,dim=0,model).to(device)
model = torch.nn.DataParallel(model)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank)
global_step = 0
g_metric = LMAccuracy()
d_metric = AccuracyThresh()
tr_g_acc = AverageMeter()
tr_d_acc = AverageMeter()
tr_loss = AverageMeter()
tr_g_loss = AverageMeter()
tr_d_loss = AverageMeter()
train_logs = {}
logger.info("***** Running training *****")
logger.info(f" Num examples = {total_train_examples}")
logger.info(f" Batch size = {args.train_batch_size}")
logger.info(f" Num steps = {num_train_optimization_steps}")
logger.info(f" warmup_steps = {args.warmup_steps}")
logger.info(f" Num workable gpus = {args.n_gpu}")
start_time = time.time()
seed_everything(args.seed) # Added here for reproducibility
for epoch in range(args.epochs):
for idx in range(args.file_num):
epoch_dataset = PregeneratedDataset(
file_id=idx,
training_path=pregenerated_data,
tokenizer=tokenizer,
reduce_memory=args.reduce_memory,
data_name=args.data_name)
if args.local_rank == -1:
train_sampler = RandomSampler(epoch_dataset)
else:
train_sampler = DistributedSampler(epoch_dataset)
train_dataloader = DataLoader(epoch_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(train_dataloader):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, lm_label_ids = batch
outputs = model(input_ids=input_ids,
token_type_ids=segment_ids,
attention_mask=input_mask,
masked_lm_labels=lm_label_ids)
loss, g_loss, d_loss, d_logits, g_logits, is_replaced_label = outputs
active_indices = input_mask.view(-1) == 1
active_logits = d_logits.view(-1)[active_indices]
active_labels = is_replaced_label.view(-1)[active_indices]
g_metric(logits=g_logits.view(-1, bert_config.vocab_size),
target=lm_label_ids.view(-1))
d_metric(logits=active_logits.view(-1, 1),
target=active_labels)
if args.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
g_loss = g_loss.mean()
d_loss = d_loss.mean()
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
nb_tr_steps += 1
tr_g_acc.update(g_metric.value(), n=input_ids.size(0))
tr_d_acc.update(d_metric.value(), n=input_ids.size(0))
tr_loss.update(loss.item(), n=1)
tr_g_loss.update(g_loss.item(), n=1)
tr_d_loss.update(d_loss.item(), n=1)
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
scheduler.step()
optimizer.step()
optimizer.zero_grad()
global_step += 1
if global_step % args.num_eval_steps == 0:
now = time.time()
eta = now - start_time
if eta > 3600:
eta_format = ('%d:%02d:%02d' %
(eta // 3600,
(eta % 3600) // 60, eta % 60))
elif eta > 60:
eta_format = '%d:%02d' % (eta // 60, eta % 60)
else:
eta_format = '%ds' % eta
train_logs['loss'] = tr_loss.avg
train_logs['g_acc'] = tr_g_acc.avg
train_logs['d_acc'] = tr_d_acc.avg
train_logs['g_loss'] = tr_g_loss.avg
train_logs['d_loss'] = tr_d_loss.avg
show_info = f'[Training]:[{epoch}/{args.epochs}]{global_step}/{num_train_optimization_steps} ' \
f'- ETA: {eta_format}' + "-".join(
[f' {key}: {value:.4f} ' for key, value in train_logs.items()])
logger.info(show_info)
tr_g_acc.reset()
tr_d_acc.reset()
tr_loss.reset()
tr_g_loss.reset()
tr_d_loss.reset()
start_time = now
if global_step % args.num_save_steps == 0:
if args.local_rank in [-1, 0] and args.num_save_steps > 0:
# Save model checkpoint
output_dir = args.output_dir / f'lm-checkpoint-{global_step}'
if not output_dir.exists():
output_dir.mkdir()
# save model
model_to_save = model.module if hasattr(
model, 'module'
) else model # Take care of distributed/parallel training
model_to_save.save_pretrained(str(output_dir))
torch.save(args, str(output_dir / 'training_args.bin'))
logger.info("Saving model checkpoint to %s",
output_dir)
model.module.generator.save_pretrained(
str(output_dir / "G"))
logger.info("Saving generator model checkpoint to %s",
output_dir / "G")
model.module.electra.save_pretrained(
str(output_dir / "D"))
logger.info("Saving electra model checkpoint to %s",
output_dir / "D")
torch.save(optimizer.state_dict(),
str(output_dir / "optimizer.bin"))
# save config
output_config_file = output_dir / CONFIG_NAME
output_config_file_D = output_dir / "D" / CONFIG_NAME
output_config_file_G = output_dir / "G" / CONFIG_NAME
with open(str(output_config_file), 'w') as f:
f.write(model_to_save.config.to_json_string())
with open(str(output_config_file_D), 'w') as f:
f.write(
model.module.electra.config.to_json_string())
with open(str(output_config_file_G), 'w') as f:
f.write(
model.module.generator.config.to_json_string())
# save vocab
tokenizer.save_vocabulary(output_dir)
def predict(args, model, tokenizer, prefix=""):
metric = SeqEntityScore(args.id2label, markup=args.markup)
pred_output_dir = args.output_dir
if not os.path.exists(pred_output_dir) and args.local_rank in [-1, 0]:
os.makedirs(pred_output_dir)
test_dataset = load_and_cache_examples(args,
args.task_name,
tokenizer,
data_type='test')
# Note that DistributedSampler samples randomly
test_sampler = SequentialSampler(
test_dataset) if args.local_rank == -1 else DistributedSampler(
test_dataset)
test_dataloader = DataLoader(test_dataset,
sampler=test_sampler,
batch_size=1,
collate_fn=collate_fn)
# Eval!
logger.info("***** Running prediction %s *****", prefix)
logger.info(" Num examples = %d", len(test_dataset))
logger.info(" Batch size = %d", 1)
f_results = []
output_predict_file = os.path.join(pred_output_dir, prefix,
"crf_test_prediction.json")
# pbar = ProgressBar(n_total=len(test_dataloader), desc="Predicting")
test_iterator = tqdm(test_dataloader,
desc="Iteration",
disable=args.local_rank not in [-1, 0])
if isinstance(model, nn.DataParallel):
model = model.module
test_loss = 0.0
nb_test_steps = 0
for step, batch in enumerate(test_iterator):
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],
"labels": batch[3],
'input_lens': batch[4]
}
if args.model_type != "distilbert":
# XLM and RoBERTa don"t use segment_ids
inputs["token_type_ids"] = (batch[2] if args.model_type
in ["bert", "xlnet"] else None)
outputs = model(**inputs)
tmp_test_loss, logits = outputs[:2]
if args.n_gpu > 1:
tmp_eval_loss = tmp_test_loss.mean()
# print(logits.shape)
# print(logits)
tags = model.crf.decode(logits, inputs['attention_mask'])
tags = tags.squeeze(0).cpu().numpy().tolist()
### 测试集结果保存到文件
preds = tags[0][1:-1] # [CLS]XXXX[SEP]
label_entities = get_entities(preds, args.id2label, args.markup)
json_d = {}
json_d['id'] = step
json_d['tag_seq'] = " ".join([args.id2label[x] for x in preds])
json_d['entities'] = label_entities
f_results.append(json_d)
test_loss += tmp_test_loss
nb_test_steps += 1
out_label_ids = inputs['labels'].cpu().numpy().tolist()
input_lens = inputs['input_lens'].cpu().numpy().tolist()
for i, label in enumerate(out_label_ids):
temp_1 = []
temp_2 = []
for j, m in enumerate(label):
if j == 0:
continue
elif j == input_lens[i] - 1:
metric.update(pred_paths=[temp_2], label_paths=[temp_1])
break
else:
temp_1.append(args.id2label[out_label_ids[i][j]])
temp_2.append(args.id2label[tags[i][j]])
### 打印测试集的最终结果 ###
logger.info("\n")
test_loss = test_loss / nb_test_steps
test_info, entity_info = metric.result()
results = {f'{key}': value for key, value in test_info.items()}
results['loss'] = test_loss
logger.info("***** Test results %s *****", prefix)
info = "-".join(
[f' {key}: {value:.4f} ' for key, value in results.items()])
logger.info(info)
logger.info("\n")
logger.info("***** Entity results %s *****", prefix)
for key in sorted(entity_info.keys()):
logger.info("******* %s results ********" % key)
info = "-".join([
f' {key}: {value:.4f} ' for key, value in entity_info[key].items()
])
logger.info(info)
# pbar(step)
logger.info("\n")
with open(output_predict_file, "w") as writer:
for record in f_results:
writer.write(json.dumps(record) + '\n')
if args.task_name == 'cluener':
output_submit_file = os.path.join(pred_output_dir, prefix,
"test_submit.json")
test_text = []
with open(os.path.join(args.data_dir, "test.json"), 'r') as fr:
for line in fr:
test_text.append(json.loads(line))
test_submit = []
for x, y in zip(test_text, results):
json_d = {}
json_d['id'] = x['id']
json_d['label'] = {}
entities = y['entities']
words = list(x['text'])
if len(entities) != 0:
for subject in entities:
tag = subject[0]
start = subject[1]
end = subject[2]
word = "".join(words[start:end + 1])
if tag in json_d['label']:
if word in json_d['label'][tag]:
json_d['label'][tag][word].append([start, end])
else:
json_d['label'][tag][word] = [[start, end]]
else:
json_d['label'][tag] = {}
json_d['label'][tag][word] = [[start, end]]
test_submit.append(json_d)
json_to_text(output_submit_file, test_submit)
def main():
args = get_argparse().parse_args()
if not os.path.exists(args.output_dir):
os.mkdir(args.output_dir)
args.output_dir = args.output_dir + '{}'.format(args.model_type)
if not os.path.exists(args.output_dir):
os.mkdir(args.output_dir)
time_ = time.strftime("%Y-%m-%d-%H:%M:%S", time.localtime())
init_logger(log_file=args.output_dir +
f'/{args.model_type}-{args.task_name}-{time_}.log')
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 = 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
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
seed_everything(args.seed)
# Prepare NER task
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]()
label_list = processor.get_labels()
args.id2label = {i: label for i, label in enumerate(label_list)}
args.label2id = {label: i for i, label in enumerate(label_list)}
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]
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_labels,
cache_dir=args.cache_dir if args.cache_dir else None)
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,
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)
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,
data_type='train')
global_step, tr_loss = train(args, train_dataset, model, tokenizer,
config)
logger.info(" global_step = %s, average loss = %s", global_step,
tr_loss)
def predict(args, model, tokenizer, label_list, prefix=""):
pred_task_names = (args.task_name, )
pred_outputs_dirs = (args.output_dir, )
label_map = {i: label for i, label in enumerate(label_list)}
for pred_task, pred_output_dir in zip(pred_task_names, pred_outputs_dirs):
pred_dataset = load_and_cache_examples(args,
pred_task,
tokenizer,
data_type='test')
if not os.path.exists(pred_output_dir) and args.local_rank in [-1, 0]:
os.makedirs(pred_output_dir)
args.pred_batch_size = args.per_gpu_eval_batch_size * max(
1, args.n_gpu)
# Note that DistributedSampler samples randomly
pred_sampler = SequentialSampler(
pred_dataset) if args.local_rank == -1 else DistributedSampler(
pred_dataset)
pred_dataloader = DataLoader(
pred_dataset,
sampler=pred_sampler,
batch_size=args.pred_batch_size,
collate_fn=xlnet_collate_fn
if args.model_type in ['xlnet'] else collate_fn)
logger.info("******** Running prediction {} ********".format(prefix))
logger.info(" Num examples = %d", len(pred_dataset))
logger.info(" Batch size = %d", args.pred_batch_size)
nb_pred_steps = 0
preds = None
pbar = ProgressBar(n_total=len(pred_dataloader), desc="Predicting")
for step, batch in enumerate(pred_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],
'labels': batch[3]
}
if args.model_type != 'distilbert':
inputs['token_type_ids'] = batch[2] if (
'bert' in args.model_type or 'xlnet' in args.model_type
) else None # XLM, DistilBERT and RoBERTa don't use segment_ids
outputs = model(**inputs)
_, logits = outputs[:2]
nb_pred_steps += 1
if preds is None:
if pred_task == 'copa':
preds = logits.softmax(-1).detach().cpu().numpy()
else:
preds = logits.detach().cpu().numpy()
else:
if pred_task == 'copa':
preds = np.append(
preds,
logits.softmax(-1).detach().cpu().numpy(),
axis=0)
else:
preds = np.append(preds,
logits.detach().cpu().numpy(),
axis=0)
pbar(step)
print(' ')
if args.output_mode == "classification":
predict_label = np.argmax(preds, axis=1)
elif args.output_mode == "regression":
predict_label = np.squeeze(preds)
if pred_task == 'copa':
predict_label = []
pred_logits = preds[:, 1]
i = 0
while (i < len(pred_logits) - 1):
if pred_logits[i] >= pred_logits[i + 1]:
predict_label.append(0)
else:
predict_label.append(1)
i += 2
output_submit_file = os.path.join(pred_output_dir, prefix,
"test_prediction.json")
output_logits_file = os.path.join(pred_output_dir, prefix,
"test_logits")
# 保存标签结果
with open(output_submit_file, "w") as writer:
for i, pred in enumerate(predict_label):
json_d = {}
json_d['id'] = i
json_d['label'] = str(label_map[pred])
writer.write(json.dumps(json_d) + '\n')
# 保存中间预测结果
save_numpy(file_path=output_logits_file, data=preds)
