def save_vocabs_and_config(self,
idx2label_path=None,
idx2cls_path=None,
config_path=None):
idx2label_path = if_none(idx2label_path, self.idx2label_path)
idx2cls_path = if_none(idx2cls_path, self.idx2cls_path)
config_path = if_none(config_path, self.config_path)
logging.info("Saving vocabs...")
save_json(self.idx2label, idx2label_path)
if self.idx2cls:
save_json(self.idx2cls, idx2cls_path)
save_json(self.get_config(), config_path)
def load(self, df_path=None, df=None):
df_path = if_none(df_path, self.config["df_path"])
if df is None:
self.df = pd.read_csv(df_path, sep='\t')
self.label2idx = {}
self.idx2label = []
with open(self.config["idx2labels_path"], "r", encoding="utf-8") as f:
for idx, label in enumerate(f.readlines()):
label = label.strip()
self.label2idx[label] = idx
self.idx2label.append(label)
self.pos2idx = {}
self.idx2pos = []
with open(self.config["idx2pos_path"], "r", encoding="utf-8") as f:
for idx, posl in enumerate(f.readlines()):
pos = posl.strip()
self.pos2idx[pos] = idx
self.idx2pos.append(pos)
if self.config["is_cls"]:
self.idx2cls = []
self.cls2idx = {}
with open(self.config["idx2cls_path"], "r", encoding="utf-8") as f:
for idx, label in enumerate(f.readlines()):
label = label.strip()
self.cls2idx[label] = idx
self.idx2cls.append(label)
def load(self, df_path=None, df=None):
'''
# Objective
- load label2idx from self.config["idx2labels_path"],
- load idx2label
- load **data** and save to instance 'df' attribute
df is normal pandas dataframe
- olunlah add char2idx and idx2char
# Parameter
- df_path(O) :
# Return
- None, label2idx and idx2label is setted to instance attribute
'''
df_path = if_none(df_path, self.config["df_path"])
if df is None:
self.df = pd.read_csv(df_path, sep='\t')
self.label2idx = {}
self.idx2label = []
with open(self.config["idx2labels_path"], "r", encoding="utf-8") as f:
for idx, label in enumerate(f.readlines()):
label = label.strip()
self.label2idx[label] = idx
self.idx2label.append(label)
if self.config["is_cls"]:
self.idx2cls = []
self.cls2idx = {}
with open(self.config["idx2cls_path"], "r", encoding="utf-8") as f:
for idx, label in enumerate(f.readlines()):
label = label.strip()
self.cls2idx[label] = idx
self.idx2cls.append(label)
#----Olunlah----
#------- Char2idx Model ------------
import json
#CHARIDX_PATH = '/content/drive/My Drive/Colab Notebooks/IS_NER/data/model/Supanut/charidx/char2idx.txt'
CHARIDX_PATH = '/content/drive/My Drive/Colab Notebooks/IS_NER/data/03_BERT_Thai_NER/in/LST20/char2idx.txt'
with open(CHARIDX_PATH) as json_file:
char2idx = json.load(json_file)
idx2char = {char2idx[c]: c for i , c in enumerate(char2idx)}
self.char2idx = char2idx
self.idx2char = idx2char
def save(self, df_path=None):
df_path = if_none(df_path, self.config["df_path"])
self.df.to_csv(df_path, sep='\t', index=False)
def create(cls,
bert_vocab_file,
config_path=None,
train_path=None,
valid_path=None,
idx2label=None,
bert_model_type="bert_cased",
idx2cls=None,
max_seq_len=424,
batch_size=16,
is_cls=False,
idx2label_path=None,
idx2cls_path=None,
pad="<pad>",
device="cuda:0",
clear_cache=True,
data_columns=["0", "1", "2"],
shuffle=True,
dir_config=None,
prc_text=preprocess_text):
"""
Create or skip data loaders, load or create vocabs.
DataFrame should has 2 or 3 columns. Structure see in data_columns description.
Parameters
----------
bert_vocab_file : str
Path of vocabulary for BERT tokenizer.
config_path : str, or None, optional (default=None)
Path of config of BertNerData.
train_path : str or None, optional (default=None)
Path of train data frame. If not None update idx2label, idx2cls, idx2meta.
valid_path : str or None, optional (default=None)
Path of valid data frame. If not None update idx2label, idx2cls, idx2meta.
idx2label : list or None, optional (default=None)
Map form index to label.
bert_model_type : str, optional (default="bert_cased")
Mode of BERT model (CASED or UNCASED).
idx2cls : list or None, optional (default=None)
Map form index to cls.
max_seq_len : int, optional (default=424)
Max sequence length.
batch_size : int, optional (default=16)
Batch size.
is_cls : bool, optional (default=False)
Use joint model or single.
idx2label_path : str or None, optional (default=None)
Path to idx2label map. If not None and idx2label is None load idx2label.
idx2cls_path : str or None, optional (default=None)
Path to idx2cls map. If not None and idx2cls is None load idx2cls.
pad : str, optional (default="<pad>")
Padding token.
device : str, optional (default="cuda:0")
Run model on gpu or cpu. If "cpu" don't pin tensors in data loaders to gpu.
Notation similar as torch.cuda.device.
clear_cache : bool, optional (default=True)
If True, rewrite all vocabs and BertNerData config.
data_columns : list[str]
Columns if pandas.DataFrame.
data_columns[0] - represent labels column. Each label should be joined by space;
data_columns[1] - represent tokens column. Input sequence should be tokenized and joined by space;
data_columns[2] - represent cls column (if is_cls is not None).
shuffle : bool, optional (default=True)
Is shuffle data.
dir_config : str or None, optional (default=None)
Dir for store vocabs if paths is not set.
prc_text : callable, optional (default=preprocess_text)
Function for preprocess text. By default remove some bad unicode words.
Note. don't see in word. Remove only full match bad symbol with word.
Returns
----------
data : BertNerData
Created object of BertNerData.
"""
idx2label_path = if_none(
idx2label_path,
os.path.join(dir_config, "idx2label.json")
if dir_config is not None else None)
if idx2label is None and idx2label_path is None:
raise ValueError("Must set idx2label_path.")
if bert_model_type == "bert_cased":
do_lower_case = False
elif bert_model_type == "bert_uncased":
do_lower_case = True
else:
raise NotImplementedError("No requested mode :(.")
tokenizer = tokenization.FullTokenizer(vocab_file=bert_vocab_file,
do_lower_case=do_lower_case)
if idx2label is None and os.path.exists(
str(idx2label_path)) and not clear_cache:
idx2label = read_json(idx2label_path)
if is_cls:
idx2cls_path = if_none(
idx2cls_path,
os.path.join(dir_config, "idx2cls.json")
if dir_config is not None else None)
if is_cls and idx2cls is None and os.path.exists(
str(idx2cls_path)) and not clear_cache:
idx2cls = read_json(idx2cls_path)
config_path = if_none(
config_path,
os.path.join(dir_config, "data_ner.json")
if dir_config is not None else None)
data = cls(bert_vocab_file=bert_vocab_file,
train_path=train_path,
valid_path=valid_path,
idx2label=idx2label,
config_path=config_path,
tokenizer=tokenizer,
bert_model_type=bert_model_type,
idx2cls=idx2cls,
max_seq_len=max_seq_len,
batch_size=batch_size,
is_cls=is_cls,
idx2label_path=idx2label_path,
idx2cls_path=idx2cls_path,
pad=pad,
device=device,
data_columns=data_columns,
shuffle=shuffle,
prc_text=prc_text)
if train_path is not None:
_ = data.load_train_dl(train_path)
if valid_path is not None:
_ = data.load_valid_dl(valid_path)
if clear_cache:
data.save_vocabs_and_config()
return data
