korscideberta-colab.py

# -*- coding: utf-8 -*-
"""korscideberta.ipynb

Automatically generated by Colaboratory.

Original file is located at
 https://colab.research.google.com/drive/1vJNUG_F5El5LY8xmmwRVXo66bYBfXtdz
"""

#!git clone https://huggingface.co/kisti/korscideberta; cd korscideberta

# Commented out IPython magic to ensure Python compatibility.
#!pwd
#%cd ..
#!pip install konlpy
# %cd korscideberta

# Commented out IPython magic to ensure Python compatibility.
'''
! git clone https://github.com/SOMJANG/Mecab-ko-for-Google-Colab.git
# %cd Mecab-ko-for-Google-Colab/
! bash install_mecab-ko_on_colab_light_220429.sh
# %cd ..
!pip install datasets transformers[sentencepiece]
'''

# Commented out IPython magic to ensure Python compatibility.
'''
!pip install -U accelerate; pip install -U transformers; pip install pydantic==1.8
'''

!pwd
# %cd /content/korscideberta

'''
#[필수]리눅스 터미널에서 본 코드 및 토크나이저 다운로드
#git clone https://huggingface.co/kisti/korscideberta
#cd korscideberta

#[필수]라이브러리 설치(Mecab 등 자세한 설치 방법은 KorSciDeBERTa환경설치+파인튜닝.pdf 참조)
!apt install git-lfs

'''

from datasets import load_dataset
import datasets
from huggingface_hub import notebook_login

notebook_login() #Huggingface 로그인
#토큰 예시: hf_jRjLZcSBibYHwUaTjiNUEeoJlFxhFkGM

model_repository = "kisti/korscideberta" #Huggingface 모델명 설정
#model_repository = "./"
from transformers import AutoTokenizer
from tokenization_korscideberta_v2 import DebertaV2Tokenizer
tokenizer = DebertaV2Tokenizer.from_pretrained(model_repository)
out = tokenizer.tokenize("<cls> 한국어 모델을 <s> 한국어 모델을 공유합니다. <s>")
print(str(out))

#데이터셋 로드
#data_files = {"train": "문장의미-균등저널/test.json", "test": "문장의미-균등저널/train.json", 'dev':'문장의미-균등저널/dev.json'}
#dataset = load_dataset('json', data_files=data_files)
dataset = load_dataset('csv', data_files='data/Abstract_Annotation_Data_tagsentence.csv', split='train')
dataset = dataset.shuffle(seed=42)
dataset = dataset.train_test_split(test_size=0.1)
print("dataset:", str(dataset))

#데이터셋을 토크나이징 후 저장
from datasets import ClassLabel
labels = [x for x in dataset['train']['tag']]
labels = list(set(labels))
labels.sort()
num_labels = len(labels)
print('Labels: '+str(labels)[:200])
ClassLabels = ClassLabel(num_classes=len(labels), names=labels)

def preprocess_function(example):
 output_dict = tokenizer('<cls>'+example["sentence"]+'<s>', max_length=512, truncation=True)
 output_dict['labels'] = ClassLabels.str2int(example['tag'])
 return output_dict
#tokenized_datasets = dataset.map(preprocess_function, batched=False, remove_columns=dataset["train"].column_names)
tokenized_datasets = dataset.map(preprocess_function, batched=False)
tokenized_datasets = tokenized_datasets.cast_column("labels", ClassLabel(names=labels))

#데이터셋 토크나이징 확인
random_id = 1
print("Input IDS:", tokenized_datasets["train"][random_id]["input_ids"])
print("Labels:", tokenized_datasets["train"][random_id]["labels"])
tokenized_datasets.save_to_disk('data/tok')

#KorSciDeBERTa 모델 로딩
from transformers import AutoModelForSequenceClassification

num_labels = len(labels)
def model_init():
 #return AutoModelForSequenceClassification.from_pretrained(model_checkpoint, num_labels=7)
 #return AutoModelForSequenceClassification.from_pretrained(model_repository, num_labels=num_labels, hidden_dropout_prob=0.3, attention_probs_dropout_prob=0.25)
 return AutoModelForSequenceClassification.from_pretrained(model_repository, num_labels=num_labels, hidden_dropout_prob=0.1, attention_probs_dropout_prob=0.1)
model = model_init()

#DataCollator 확인
from transformers import DataCollatorWithPadding
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
from collections import Counter
print("Test:", Counter(tokenized_datasets["test"]["labels"]))

#정확도 척도
from datasets import load_metric
accuracy = load_metric("accuracy")

import numpy as np
def compute_metrics(pred):
 pred_logits = pred.predictions
 pred_classes = np.argmax(pred_logits, axis=-1)
 labels = np.asarray(pred.label_ids)
 acc = accuracy.compute(predictions=pred_classes, references=labels)
 return {"accuracy": acc["accuracy"]}

#training_args 설정
#다음 에러 발생시 output_dir을 변경하여 다시 시도
#MlflowException: Changing param values is not allowed. Param with key=

import gc
gc.collect()
from transformers import TrainingArguments
training_args = TrainingArguments(
 output_dir="deberta_sent4455",
 num_train_epochs=4,
 #learning_rate=5e-5,
 learning_rate=1.5e-5,
 per_device_train_batch_size=16,
 per_device_eval_batch_size=8,
 weight_decay=0.01,
 fp16=True, # Use mixed precision
 fp16_opt_level="01", # mixed precision mode
 warmup_steps=500,
 logging_steps=200,
 save_steps=2000,
 eval_steps=500,
 push_to_hub=True,
 evaluation_strategy="steps",
)

#Trainer 설정 후 학습 시작
import gc
gc.collect()

from transformers import Trainer
trainer = Trainer(
 args=training_args,
 compute_metrics=compute_metrics,
 model=model,
 #tokenizer=tokenizer, #에러 유발: TypeError: save_vocabulary() got an unexpected keyword argument 'filename_prefix'
 data_collator=data_collator,
 train_dataset=tokenized_datasets["train"],
 eval_dataset=tokenized_datasets["test"]
)
train_metrics = trainer.train().metrics
trainer.save_metrics("train", train_metrics)
trainer.push_to_hub()
#### 파인튜닝 및 모델 업로드 완료

# Commented out IPython magic to ensure Python compatibility.
# %cd mecab
!bash <(curl -s https://raw.githubusercontent.com/konlpy/konlpy/master/scripts/mecab.sh); cd mecab-0.996-ko-0.9.2;
!chmod 775 ./configure; ./configure; make; chmod 775 tests/*.sh; make check; make install

# Commented out IPython magic to ensure Python compatibility.
!pwd
# %cd mecab
!cd mecab-ko-dic-2.1.1-20180720; chmod 775 ./autogen.sh; ./autogen.sh; ./configure; make

#!mecab -d /usr/local/lib/mecab/dic/mecab-ko-dic

# Commented out IPython magic to ensure Python compatibility.
!pwd
!ls
# %cd korscideberta

! unzip korscideberta.zip -d korscideberta; cd korscideberta

# Commented out IPython magic to ensure Python compatibility.
!pwd
# %cd korscideberta

! pip3 install -r requirements.txt; pip install --upgrade nltk;
!pip uninstall -y torch torchtext torch-tensorrt; pip install --upgrade pip; pip install torch==1.10.1+cu111 torchvision==0.11.2+cu111 torchaudio==0.10.1 -f https://download.pytorch.org/whl/cu111/torch_stable.html --default-timeout=100; pip install setuptools_scm six mlflow; pip install "numpy<1.24.0"; pip install .