initial commit

.github/workflows/pre-commit.yaml

@@ -0,0 +1,22 @@
+name: pre-commit
+
+on:
+  pull_request:
+  push:
+    branches: [main]
+
+jobs:
+  check_and_test:
+    runs-on: [self-hosted, linux, x64, cpu-only]
+
+    steps:
+    - uses: actions/checkout@v3
+    - uses: actions/setup-python@v4
+      id: ko-sentence-transformers
+      with:
+        python-version: '3.10'
+    - name: pre-commit #  don't use in self-hosted `- uses: pre-commit/action@v2.0.3`
+      run: |
+        pip install -U pre-commit
+        pre-commit install --install-hooks
+        pre-commit run -a

.pre-commit-config.yaml

@@ -0,0 +1,30 @@
+exclude: ^(legacy|bin)
+repos:
+  - repo: https://github.com/pre-commit/pre-commit-hooks
+    rev: v4.0.1
+    hooks:
+      - id: end-of-file-fixer
+        types: [python]
+      - id: trailing-whitespace
+        types: [python]
+      - id: mixed-line-ending
+        types: [python]
+      - id: check-added-large-files
+        args: [--maxkb=4096]
+  - repo: https://github.com/psf/black
+    rev: 22.3.0
+    hooks:
+      - id: black
+        args: ["--line-length", "120"]
+  - repo: https://github.com/pycqa/isort
+    rev: 5.12.0
+    hooks:
+      - id: isort
+        name: isort (python)
+        args: ["--profile", "black", "-l", "120"]
+  - repo: https://github.com/pycqa/flake8.git
+    rev: 6.0.0
+    hooks:
+      - id: flake8
+        types: [python]
+        args: ["--max-line-length", "120", "--ignore", "F811,F841,E203,E402,E712,W503"]

README.md

@@ -0,0 +1,242 @@
+# kf-deberta-multitask
+
+kakaobank의 [kf-deberta-base](https://huggingface.co/kakaobank/kf-deberta-base) 모델을 KorNLI, KorSTS 데이터셋으로 파인튜닝한 모델입니다.
+[jhgan00/ko-sentence-transformers](https://github.com/jhgan00/ko-sentence-transformers) 코드를 기반으로 일부 수정하여 진행하였습니다.
+
+<br>
+
+## KorSTS Benchmark
+
+- [jhgan00/ko-sentence-transformers](https://github.com/jhgan00/ko-sentence-transformers#korsts-benchmarks)의 결과를 참고하여 재작성하였습니다.
+- 학습 및 성능 평가 과정은 `training_*.py`, `benchmark.py` 에서 확인할 수 있습니다.
+- 학습된 모델은 허깅페이스 모델 허브에 공개되어 있습니다.
+
+<br>
+
+|model|cosine_pearson|cosine_spearman|euclidean_pearson|euclidean_spearman|manhattan_pearson|manhattan_spearman|dot_pearson|dot_spearman|
+|:-------------------------|-----------------:|------------------:|--------------------:|---------------------:|--------------------:|---------------------:|--------------:|---------------:|
+|[kf-deberta-multitask](https://huggingface.co/upskyy/kf-deberta-multitask)|**85.75**|**86.25**|**84.79**|**85.25**|**84.80**|**85.27**|**82.93**|**82.86**|
+|[ko-sroberta-multitask](https://huggingface.co/jhgan/ko-sroberta-multitask)|84.77|85.6|83.71|84.40|83.70|84.38|82.42|82.33|
+|[ko-sbert-multitask](https://huggingface.co/jhgan/ko-sbert-multitask)|84.13|84.71|82.42|82.66|82.41|82.69|80.05|79.69|
+|[ko-sroberta-base-nli](https://huggingface.co/jhgan/ko-sroberta-nli)|82.83|83.85|82.87|83.29|82.88|83.28|80.34|79.69|
+|[ko-sbert-nli](https://huggingface.co/jhgan/ko-sbert-multitask)|82.24|83.16|82.19|82.31|82.18|82.3|79.3|78.78|
+|[ko-sroberta-sts](https://huggingface.co/jhgan/ko-sroberta-sts)|81.84|81.82|81.15|81.25|81.14|81.25|79.09|78.54|
+|[ko-sbert-sts](https://huggingface.co/jhgan/ko-sbert-sts)|81.55|81.23|79.94|79.79|79.9|79.75|76.02|75.31|
+
+<br>
+
+## Examples
+
+- 예시 출처: <https://github.com/BM-K/KoSentenceBERT-SKT>
+
+<br>
+
+아래는 임베딩 벡터를 통해 가장 유사한 문장을 찾는 예시입니다.
+더 많은 예시는 [sentence-transformers 문서](https://www.sbert.net/index.html)를 참고해주세요.
+
+```python
+from transformers import AutoTokenizer, AutoModel
+import torch
+
+
+# Mean Pooling - Take attention mask into account for correct averaging
+def mean_pooling(model_output, attention_mask):
+    token_embeddings = model_output[0] # First element of model_output contains all token embeddings
+    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
+    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)
+
+
+# Sentences we want sentence embeddings for
+sentences = ["안녕하세요?", "한국어 문장 임베딩을 위한 버트 모델입니다."]
+
+# Load model from HuggingFace Hub
+tokenizer = AutoTokenizer.from_pretrained("upskyy/kf-deberta-multitask")
+model = AutoModel.from_pretrained("upskyy/kf-deberta-multitask")
+
+# Tokenize sentences
+encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
+
+# Compute token embeddings
+with torch.no_grad():
+    model_output = model(**encoded_input)
+
+# Perform pooling. In this case, mean pooling.
+sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
+
+print("Sentence embeddings:")
+print(sentence_embeddings)
+```
+
+```python
+from sentence_transformers import SentenceTransformer, util
+import numpy as np
+
+
+embedder = SentenceTransformer("upskyy/kf-deberta-multitask")
+
+# Corpus with example sentences
+corpus = [
+    "한 남자가 음식을 먹는다.",
+    "한 남자가 빵 한 조각을 먹는다.",
+    "그 여자가 아이를 돌본다.",
+    "한 남자가 말을 탄다.",
+    "한 여자가 바이올린을 연주한다.",
+    "두 남자가 수레를 숲 속으로 밀었다.",
+    "한 남자가 담으로 싸인 땅에서 백마를 타고 있다.",
+    "원숭이 한 마리가 드럼을 연주한다.",
+    "치타 한 마리가 먹이 뒤에서 달리고 있다.",
+]
+
+corpus_embeddings = embedder.encode(corpus, convert_to_tensor=True)
+
+# Query sentences:
+queries = [
+    "한 남자가 파스타를 먹는다.",
+    "고릴라 의상을 입은 누군가가 드럼을 연주하고 있다.",
+    "치타가 들판을 가로 질러 먹이를 쫓는다."
+]
+
+# Find the closest 5 sentences of the corpus for each query sentence based on cosine similarity
+top_k = 5
+for query in queries:
+    query_embedding = embedder.encode(query, convert_to_tensor=True)
+    cos_scores = util.pytorch_cos_sim(query_embedding, corpus_embeddings)[0]
+    cos_scores = cos_scores.cpu()
+
+    # We use np.argpartition, to only partially sort the top_k results
+    top_results = np.argpartition(-cos_scores, range(top_k))[0:top_k]
+
+    print("\n\n======================\n\n")
+    print("Query:", query)
+    print("\nTop 5 most similar sentences in corpus:")
+
+    for idx in top_results[0:top_k]:
+        print(corpus[idx].strip(), "(Score: %.4f)" % (cos_scores[idx]))
+```
+
+<br>
+
+```
+======================
+
+
+Query: 한 남자가 파스타를 먹는다.
+
+Top 5 most similar sentences in corpus:
+한 남자가 음식을 먹는다. (Score: 0.5826)
+한 남자가 빵 한 조각을 먹는다. (Score: 0.5507)
+한 남자가 말을 탄다. (Score: 0.1767)
+한 남자가 담으로 싸인 땅에서 백마를 타고 있다. (Score: 0.0965)
+치타 한 마리가 먹이 뒤에서 달리고 있다. (Score: 0.0429)
+
+
+======================
+
+
+Query: 고릴라 의상을 입은 누군가가 드럼을 연주하고 있다.
+
+Top 5 most similar sentences in corpus:
+원숭이 한 마리가 드럼을 연주한다. (Score: 0.7093)
+한 여자가 바이올린을 연주한다. (Score: 0.2374)
+치타 한 마리가 먹이 뒤에서 달리고 있다. (Score: 0.1872)
+그 여자가 아이를 돌본다. (Score: 0.1574)
+한 남자가 말을 탄다. (Score: 0.0883)
+
+
+======================
+
+
+Query: 치타가 들판을 가로 질러 먹이를 쫓는다.
+
+Top 5 most similar sentences in corpus:
+치타 한 마리가 먹이 뒤에서 달리고 있다. (Score: 0.7740)
+한 남자가 담으로 싸인 땅에서 백마를 타고 있다. (Score: 0.2161)
+두 남자가 수레를 숲 속으로 밀었다. (Score: 0.1806)
+한 남자가 음식을 먹는다. (Score: 0.1651)
+한 남자가 말을 탄다. (Score: 0.1352)
+```
+
+<br>
+
+## Training
+
+직접 모델을 파인튜닝하려면 [`kor-nlu-datasets`](https://github.com/kakaobrain/kor-nlu-datasets) 저장소를 clone 하고 `training_*.py` 스크립트를 실행시키면 됩니다.
+
+`train.sh` 파일에서 학습 예시를 확인할 수 있습니다.
+
+```bash
+git clone https://github.com/upskyy/kf-deberta-multitask.git
+cd kf-deberta-multitask
+
+pip install -r requirements.txt
+
+git clone https://github.com/kakaobrain/kor-nlu-datasets.git
+
+python training_multi_task.py --model_name_or_path kakaobank/kf-deberta-base
+./bin/train.sh
+```
+
+<br>
+
+## Evaluation
+
+KorSTS Benchmark를 평가하는 방법입니다.
+
+```bash
+git clone https://github.com/upskyy/kf-deberta-multitask.git
+cd kf-deberta-multitask
+
+pip install -r requirements.txt
+
+git clone https://github.com/kakaobrain/kor-nlu-datasets.git
+python bin/benchmark.py
+```
+
+<br>
+
+## Export ONNX
+
+`requirements.txt` 설치 후 `bin` 디렉토리에서 `export_onnx.py` 스크립트를 실행시키면 됩니다.
+
+```bash
+git clone https://github.com/upskyy/kf-deberta-multitask.git
+cd kf-deberta-multitask
+
+pip install -r requirements.txt
+
+python bin/export_onnx.py
+```
+
+<br>
+
+## Acknowledgements
+
+- [kakaobank/kf-deberta-base](https://huggingface.co/kakaobank/kf-deberta-base) for pretrained model
+- [jhgan00/ko-sentence-transformers](https://github.com/jhgan00/ko-sentence-transformers) for original codebase
+- [kor-nlu-datasets](https://github.com/kakaobrain/kor-nlu-datasets) for training data
+
+<br>
+
+## Citation
+
+```bibtex
+@proceedings{jeon-etal-2023-kfdeberta,
+  title         = {KF-DeBERTa: Financial Domain-specific Pre-trained Language Model},
+  author        = {Eunkwang Jeon, Jungdae Kim, Minsang Song, and Joohyun Ryu},
+  booktitle     = {Proceedings of the 35th Annual Conference on Human and Cognitive Language Technology},
+  moth          = {oct},
+  year          = {2023},
+  publisher     = {Korean Institute of Information Scientists and Engineers},
+  url           = {http://www.hclt.kr/symp/?lnb=conference},
+  pages         = {143--148},
+}
+```
+
+```bibtex
+@article{ham2020kornli,
+  title={KorNLI and KorSTS: New Benchmark Datasets for Korean Natural Language Understanding},
+  author={Ham, Jiyeon and Choe, Yo Joong and Park, Kyubyong and Choi, Ilji and Soh, Hyungjoon},
+  journal={arXiv preprint arXiv:2004.03289},
+  year={2020}
+}
+```

bin/benchmark.py

@@ -0,0 +1,33 @@
+import argparse
+import csv
+import logging
+import os
+
+from sentence_transformers import InputExample, LoggingHandler, SentenceTransformer
+from sentence_transformers.evaluation import EmbeddingSimilarityEvaluator
+
+logging.basicConfig(
+    format="%(asctime)s - %(message)s", datefmt="%Y-%m-%d %H:%M:%S", level=logging.INFO, handlers=[LoggingHandler()]
+)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--sts_dataset_path", type=str, default="kor-nlu-datasets/KorSTS")
+    parser.add_argument("--model_name_or_path", type=str, required=True)
+    args = parser.parse_args()
+
+    # Read STSbenchmark dataset and use it as development set
+    test_samples = []
+    test_file = os.path.join(args.sts_dataset_path, "sts-test.tsv")
+
+    with open(test_file, "rt", encoding="utf8") as fIn:
+        reader = csv.DictReader(fIn, delimiter="\t", quoting=csv.QUOTE_NONE)
+        for row in reader:
+            score = float(row["score"]) / 5.0  # Normalize score to range 0 ... 1
+            test_samples.append(InputExample(texts=[row["sentence1"], row["sentence2"]], label=score))
+
+    test_evaluator = EmbeddingSimilarityEvaluator.from_input_examples(test_samples, name="sts-test")
+
+    model = SentenceTransformer(args.model_name_or_path)
+    test_evaluator(model)

bin/export_onnx.py

@@ -0,0 +1,13 @@
+import os
+from pathlib import Path
+from transformers.convert_graph_to_onnx import convert
+
+
+if __name__ == "__main__":
+    output_dir = "models"
+
+    if not os.path.exists(output_dir):
+        os.makedirs(output_dir, exist_ok=False)
+
+    output_fpath = os.path.join(output_dir, "kf-deberta-multitask.onnx")
+    convert(framework="pt", model="upskyy/kf-deberta-multitask", output=Path(output_fpath), opset=15)

bin/train.sh

@@ -0,0 +1,23 @@
+# To start training, you need to download the KorNLUDatasets first.
+# git clone https://github.com/kakaobrain/kor-nlu-datasets.git
+
+# train on STS dataset only
+# python training_sts.py --model_name_or_path klue/bert-base
+# python training_sts.py --model_name_or_path klue/roberta-base
+# python training_sts.py --model_name_or_path klue/roberta-small
+# python training_sts.py --model_name_or_path klue/roberta-large
+python training_sts.py --model_name_or_path kakaobank/kf-deberta-base
+
+# train on both NLI and STS dataset (multi-task)
+# python training_multi_task.py --model_name_or_path klue/bert-base
+# python training_multi_task.py --model_name_or_path klue/roberta-base
+# python training_multi_task.py --model_name_or_path klue/roberta-small
+# python training_multi_task.py --model_name_or_path klue/roberta-large
+python training_multi_task.py --model_name_or_path kakaobank/kf-deberta-base
+
+# train on NLI dataset only
+# python training_nli.py --model_name_or_path klue/bert-base
+# python training_nli.py --model_name_or_path klue/roberta-base
+# python training_nli.py --model_name_or_path klue/roberta-small
+# python training_nli.py --model_name_or_path klue/roberta-large
+python training_nli.py --model_name_or_path kakaobank/kf-deberta-base