moco-sentencedistilbertV2.1

This is a sentence-transformers model: It maps sentences & paragraphs to a 768 dimensional dense vector space and can be used for tasks like clustering or semantic search.

• 이 모델은 bongsoo/mdistilbertV2.1 MLM 모델을 sentencebert로 만든 후,추가적으로 STS Tearch-student 증류 학습 시켜 만든 모델 입니다.
• vocab: 152,537 개(기존 119,548 vocab 에 32,989 신규 vocab 추가)

Usage (Sentence-Transformers)

Using this model becomes easy when you have sentence-transformers installed:

pip install -U sentence_transformers

Then you can use the model like this:

from sentence_transformers import SentenceTransformer
sentences = ["서울은 한국이 수도이다", "The capital of Korea is Seoul"]

model = SentenceTransformer('bongsoo/moco-sentencedistilbertV2.1')
embeddings = model.encode(sentences)
print(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(embeddings[0].reshape(1,-1), embeddings[1].reshape(1,-1)))

print(f'*cosine_score:{cosine_scores[0]}')

출력(Outputs)

[[ 0.27124503 -0.5836643   0.00736023 ... -0.0038319   0.01802095 -0.09652182]
 [ 0.2765149  -0.5754248   0.00788184 ...  0.07659392 -0.07825544 -0.06120609]]
*cosine_score:0.9513546228408813

Usage (HuggingFace Transformers)

Without sentence-transformers, you can use the model like this: First, you pass your input through the transformer model, then you have to apply the right pooling-operation on-top of the contextualized word embeddings.

pip install transformers[torch]

• 평균 폴링(mean_pooling) 방식 사용. (cls 폴링, max 폴링)

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 = ["서울은 한국이 수도이다", "The capital of Korea is Seoul"]

# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained('bongsoo/moco-sentencedistilbertV2.1')
model = AutoModel.from_pretrained('bongsoo/moco-sentencedistilbertV2.1')

# 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]}')

출력(Outputs)

Sentence embeddings:
tensor([[ 0.2712, -0.5837,  0.0074,  ..., -0.0038,  0.0180, -0.0965],
        [ 0.2765, -0.5754,  0.0079,  ...,  0.0766, -0.0783, -0.0612]])
*cosine_score:0.9513546228408813

Evaluation Results

• 성능 측정을 위한 말뭉치는, 아래 한국어 (kor), 영어(en) 평가 말뭉치를 이용함 한국어 : korsts(1,379쌍문장) 와 klue-sts(519쌍문장) 영어 : stsb_multi_mt(1,376쌍문장) 와 glue:stsb (1,500쌍문장)
• 성능 지표는 cosin.spearman/max(cosine,eculidean,manhatten,doc중 max값)
• 평가 측정 코드는 여기 참조

For an automated evaluation of this model, see the Sentence Embeddings Benchmark: https://seb.sbert.net

Training(훈련 과정)

The model was trained with the parameters:

1. MLM 훈련

• 입력 모델 : distilbert-base-multilingual-cased
• 말뭉치 : 훈련 : bongsoo/moco-corpus-kowiki2022(7.6M) , 평가: bongsoo/bongevalsmall
• HyperParameter : LearningRate : 5e-5, epochs: 8, batchsize: 32, max_token_len : 128
• vocab : 152,537개 (기존 119,548 에 32,989 신규 vocab 추가)
• 출력 모델 : mdistilbertV2.1 (size: 643MB)
• 훈련시간 : 63h/1GPU (24GB/23.9 use)
• 평가: 훈련loss: 2.203400, 평가loss: 2.972835, perplexity: 23.43(bong_eval:1,500)
• 훈련코드 여기 참조

2. STS 훈련 =>bert를 sentencebert로 만듬.

• 입력 모델 : mdistilbertV2.1 (size: 643MB)
• 말뭉치 : korsts(5,749) + kluestsV1.1(11,668) + stsb_multi_mt(5,749) + mteb/sickr-sts(9,927) + glue stsb(5,749) (총:38,842)
• HyperParameter : LearningRate : 3e-5, epochs: 800, batchsize: 128, max_token_len : 256
• 출력 모델 : sbert-mdistilbertV2.1 (size: 640MB)
• 훈련시간 : 13h/1GPU (24GB/16.1GB use)
• 평가(cosin_spearman) : 0.790(말뭉치:korsts(tune_test.tsv))
• 훈련코드 여기 참조

3.증류(distilation) 훈련

• 학생 모델 : sbert-mdistilbertV2.1
• 교사 모델 : paraphrase-multilingual-mpnet-base-v2(max_token_len:128)
• 말뭉치 : news_talk_en_ko_train.tsv (영어-한국어 대화-뉴스 병렬 말뭉치 : 1.38M)
• HyperParameter : LearningRate : 5e-5, epochs: 40, batchsize: 128, max_token_len : 128(교사모델이 128이므로 맟춰줌)
• 출력 모델 : sbert-mdistilbertV2.1-distil
• 훈련시간 : 17h/1GPU (24GB/9GB use)
• 훈련코드 여기 참조

4.STS 훈련 => sentencebert 모델을 sts 훈련시킴

• 입력 모델 : sbert-mdistilbertV2.1-distil
• 말뭉치 : korsts(5,749) + kluestsV1.1(11,668) + stsb_multi_mt(5,749) + mteb/sickr-sts(9,927) + glue stsb(5,749) (총:38,842)
• HyperParameter : LearningRate : 3e-5, epochs: 1200, batchsize: 128, max_token_len : 256
• 출력 모델 : moco-sentencedistilbertV2.1
• 훈련시간 : 12/1GPU (24GB/16.1GB use)
• 평가(cosin_spearman) : 0.839(말뭉치:korsts(tune_test.tsv))
• 훈련코드 여기 참조

모델 제작 과정에 대한 자세한 내용은 여기를 참조 하세요.

Config:

{
  "_name_or_path": "../../data11/model/sbert/sbert-mdistilbertV2.1-distil",
  "activation": "gelu",
  "architectures": [
    "DistilBertModel"
  ],
  "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": 152537
}

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 256, 'do_lower_case': False}) with Transformer model: DistilBertModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False})
)

tokenizer_config

{
  "cls_token": "[CLS]",
  "do_basic_tokenize": true,
  "do_lower_case": false,
  "mask_token": "[MASK]",
  "max_len": 128,
  "name_or_path": "../../data11/model/sbert/sbert-mdistilbertV2.1-distil",
  "never_split": null,
  "pad_token": "[PAD]",
  "sep_token": "[SEP]",
  "special_tokens_map_file": "../../data11/model/distilbert/mdistilbertV2.1-4/special_tokens_map.json",
  "strip_accents": false,
  "tokenize_chinese_chars": true,
  "tokenizer_class": "DistilBertTokenizer",
  "unk_token": "[UNK]"
}

sentence_bert_config

{
  "max_seq_length": 256,
  "do_lower_case": false
}

config_sentence_transformers

{
  "__version__": {
    "sentence_transformers": "2.2.0",
    "transformers": "4.21.2",
    "pytorch": "1.10.1"
  }
}

Citing & Authors

bongsoo