Langchain-Chatchat/embeddings/add_embedding_keywords.py

'''
该功能是为了将关键词加入到embedding模型中，以便于在embedding模型中进行关键词的embedding
该功能的实现是通过修改embedding模型的tokenizer来实现的
该功能仅仅对EMBEDDING_MODEL参数对应的的模型有效，输出后的模型保存在原本模型
感谢@CharlesJu1和@charlesyju的贡献提出了想法和最基础的PR

保存的模型的位置位于原本嵌入模型的目录下，模型的名称为原模型名称+Merge_Keywords_时间戳
'''
import sys
sys.path.append("..")
from datetime import datetime
from configs import (
    MODEL_PATH,
    EMBEDDING_MODEL,
    EMBEDDING_KEYWORD_FILE,
)
import os
import torch
from safetensors.torch import save_model
from sentence_transformers import SentenceTransformer


def get_keyword_embedding(bert_model, tokenizer, key_words):
    tokenizer_output = tokenizer(key_words, return_tensors="pt", padding=True, truncation=True)

    # No need to manually convert to tensor as we've set return_tensors="pt"
    input_ids = tokenizer_output['input_ids']

    # Remove the first and last token for each sequence in the batch
    input_ids = input_ids[:, 1:-1]

    keyword_embedding = bert_model.embeddings.word_embeddings(input_ids)
    keyword_embedding = torch.mean(keyword_embedding, 1)

    return keyword_embedding


def add_keyword_to_model(model_name=EMBEDDING_MODEL, keyword_file: str = "", output_model_path: str = None):
    key_words = []
    with open(keyword_file, "r") as f:
        for line in f:
            key_words.append(line.strip())

    st_model = SentenceTransformer(model_name)
    key_words_len = len(key_words)
    word_embedding_model = st_model._first_module()
    bert_model = word_embedding_model.auto_model
    tokenizer = word_embedding_model.tokenizer
    key_words_embedding = get_keyword_embedding(bert_model, tokenizer, key_words)
    # key_words_embedding = st_model.encode(key_words)

    embedding_weight = bert_model.embeddings.word_embeddings.weight
    embedding_weight_len = len(embedding_weight)
    tokenizer.add_tokens(key_words)
    bert_model.resize_token_embeddings(len(tokenizer), pad_to_multiple_of=32)

    # key_words_embedding_tensor = torch.from_numpy(key_words_embedding)
    embedding_weight = bert_model.embeddings.word_embeddings.weight
    with torch.no_grad():
        embedding_weight[embedding_weight_len:embedding_weight_len + key_words_len, :] = key_words_embedding

    if output_model_path:
        os.makedirs(output_model_path, exist_ok=True)
        word_embedding_model.save(output_model_path)
        safetensors_file = os.path.join(output_model_path, "model.safetensors")
        metadata = {'format': 'pt'}
        save_model(bert_model, safetensors_file, metadata)
        print("save model to {}".format(output_model_path))


def add_keyword_to_embedding_model(path: str = EMBEDDING_KEYWORD_FILE):
    keyword_file = os.path.join(path)
    model_name = MODEL_PATH["embed_model"][EMBEDDING_MODEL]
    model_parent_directory = os.path.dirname(model_name)
    current_time = datetime.now().strftime('%Y%m%d_%H%M%S')
    output_model_name = "{}_Merge_Keywords_{}".format(EMBEDDING_MODEL, current_time)
    output_model_path = os.path.join(model_parent_directory, output_model_name)
    add_keyword_to_model(model_name, keyword_file, output_model_path)


if __name__ == '__main__':
    add_keyword_to_embedding_model(EMBEDDING_KEYWORD_FILE)

    # input_model_name = ""
    # output_model_path = ""
    # # 以下为加入关键字前后tokenizer的测试用例对比
    # def print_token_ids(output, tokenizer, sentences):
    #     for idx, ids in enumerate(output['input_ids']):
    #         print(f'sentence={sentences[idx]}')
    #         print(f'ids={ids}')
    #         for id in ids:
    #             decoded_id = tokenizer.decode(id)
    #             print(f'    {decoded_id}->{id}')
    #
    # sentences = [
    #     '数据科学与大数据技术',
    #     'Langchain-Chatchat'
    # ]
    #
    # st_no_keywords = SentenceTransformer(input_model_name)
    # tokenizer_without_keywords = st_no_keywords.tokenizer
    # print("===== tokenizer with no keywords added =====")
    # output = tokenizer_without_keywords(sentences)
    # print_token_ids(output, tokenizer_without_keywords, sentences)
    # print(f'-------- embedding with no keywords added -----')
    # embeddings = st_no_keywords.encode(sentences)
    # print(embeddings)
    #
    # print("--------------------------------------------")
    # print("--------------------------------------------")
    # print("--------------------------------------------")
    #
    # st_with_keywords = SentenceTransformer(output_model_path)
    # tokenizer_with_keywords = st_with_keywords.tokenizer
    # print("===== tokenizer with keyword added =====")
    # output = tokenizer_with_keywords(sentences)
    # print_token_ids(output, tokenizer_with_keywords, sentences)
    #
    # print(f'-------- embedding with keywords added -----')
    # embeddings = st_with_keywords.encode(sentences)
    # print(embeddings)