README.md

---
license: apache-2.0
pipeline_tag: fill-mask
tags:
- fill-mask
- transformers
- en
- ko
widget:
  - text: 대한민국의 수도는 [MASK] 입니다.
---
# mdistilbertV3.1

- distilbert-base-multilingual-cased 모델에 [moco-corpus-kowiki2022 말뭉치](https://huggingface.co/datasets/bongsoo/moco-corpus-kowiki2022)(kowiki202206 + MOCOMSYS 추출 3.2M 문장)로  vocab 추가하여 학습 시킨 모델
- **vocab: 159,552개 (기존 bert 모델 vocab(119,548개)에 40,004개 (한글단어30,000개+영문10,000개+수동 4개)vocab 추가**
- mdistilbertV2.1 보다 약 **7,000개** 단어가 더 많고, 한글단어는 **mecab를 이용하여 추출**함.
- **epoch은 12**번 학습함(mdistilbertV2.1은 8번)

## Usage (HuggingFace Transformers)

### 1. MASK 예시
```python
from transformers import AutoTokenizer, AutoModel, DistilBertForMaskedLM
import torch
import torch.nn.functional as F

tokenizer = AutoTokenizer.from_pretrained('bongsoo/mdistilbertV3.1', do_lower_case=False)
model = DistilBertForMaskedLM.from_pretrained('bongsoo/mdistilbertV3.1')

text = ['한국의 수도는 [MASK] 이다', '에펠탑은 [MASK]에 있다', '충무공 이순신은 [MASK]에 최고의 장수였다']
tokenized_input = tokenizer(text, max_length=128, truncation=True, padding='max_length', return_tensors='pt')

outputs = model(**tokenized_input)
logits = outputs.logits

mask_idx_list = []
for tokens in tokenized_input['input_ids'].tolist():
    token_str = [tokenizer.convert_ids_to_tokens(s) for s in tokens]
    
    # **위 token_str리스트에서 [MASK] 인덱스를 구함
    # => **해당 [MASK] 안덱스 값 mask_idx 에서는 아래 출력하는데 사용됨
    mask_idx = token_str.index('[MASK]')
    mask_idx_list.append(mask_idx)
    
for idx, mask_idx in enumerate(mask_idx_list):
    
    logits_pred=torch.argmax(F.softmax(logits[idx]), dim=1)
    mask_logits_idx = int(logits_pred[mask_idx])
    # [MASK]에 해당하는 token 구함
    mask_logits_token = tokenizer.convert_ids_to_tokens(mask_logits_idx)
    # 결과 출력 
    print('\n')
    print('*Input: {}'.format(text[idx]))
    print('*[MASK] : {} ({})'.format(mask_logits_token, mask_logits_idx))
```
- 결과
```
*Input: 한국의 수도는 [MASK] 이다
*[MASK] : 서울 (48253)


*Input: 에펠탑은 [MASK]에 있다
*[MASK] : 프랑스 (47364)


*Input: 충무공 이순신은 [MASK]에 최고의 장수였다
*[MASK] : 임진왜란 (121990)
```
### 2. 임베딩 예시
- 평균 폴링(mean_pooling) 방식 사용. ([cls 폴링](https://huggingface.co/sentence-transformers/bert-base-nli-cls-token), [max 폴링](https://huggingface.co/sentence-transformers/bert-base-nli-max-tokens))

```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 = ['This is an example sentence', 'Each sentence is converted']

# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained('bongsoo/mdistilbertV3.1')
model = AutoModel.from_pretrained('bongsoo/mdistilbertV3.1')

# 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)

# sklearn 을 이용하여 cosine_scores를 구함
# => 입력값 embeddings 은 (1,768) 처럼 2D 여야 함.
from sklearn.metrics.pairwise import paired_cosine_distances, paired_euclidean_distances, paired_manhattan_distances
cosine_scores = 1 - (paired_cosine_distances(sentence_embeddings[0].reshape(1,-1), sentence_embeddings[1].reshape(1,-1)))

print(f'*cosine_score:{cosine_scores[0]}')
```
- 결과
```
Sentence embeddings:
tensor([[-0.1137,  0.1491,  0.6711,  ..., -0.0217,  0.1839, -0.6143],
        [ 0.0482, -0.0649,  0.5333,  ...,  0.1424, -0.0982, -0.3414]])
*cosine_score:0.4784715175628662
```
## Training

**MLM(Masked Langeuage Model) 훈련**
- 입력 모델 : distilbert-base-multilingual-cased
- 말뭉치 : 훈련 : bongsoo/moco-corpus-kowiki2022(7.6M) , 평가: **bongsoo/moco_eval**
- HyperParameter : **LearningRate : 5e-5, epochs: 12 , batchsize: 32, max_token_len : 128**
- vocab : **159,552개** (기존 bert 모델 vocab(119,548개)에 40,004개 (한글단어30,000개+영문10,000개+수동 4개)vocab 추가
- 출력 모델 : mdistilbertV3.1 (size: 634MB)
- 훈련시간 : 90h/1GPU (24GB/16.5 use)
- **훈련loss: 2.1154, 평가loss: 2.5275**
- 훈련코드 [여기](https://github.com/kobongsoo/BERT/blob/master/distilbert/distilbert-MLM-Trainer-V1.2.ipynb) 참조
<br>perplexity 평가 코드는 [여기](https://github.com/kobongsoo/BERT/blob/master/distilbert/distilbert-perplexity-eval-V1.2.ipynb) 참조

## Model Config
```
{
  "_name_or_path": "",
  "activation": "gelu",
  "architectures": [
    "DistilBertForMaskedLM"
  ],
  "attention_dropout": 0.1,
  "dim": 768,
  "dropout": 0.1,
  "hidden_dim": 3072,
  "initializer_range": 0.02,
  "max_position_embeddings": 512,
  "model_type": "distilbert",
  "n_heads": 12,
  "n_layers": 6,
  "output_past": true,
  "pad_token_id": 0,
  "qa_dropout": 0.1,
  "seq_classif_dropout": 0.2,
  "sinusoidal_pos_embds": false,
  "tie_weights_": true,
  "torch_dtype": "float32",
  "transformers_version": "4.21.2",
  "vocab_size": 159552
}
```
## Citing & Authors

bongsoo