fuyundian / helsinki-nlp-ai

helsinki-nlp-ai

翻译模型安装

1. Helsinki-NLP/opus-mt-zh-en

   1. 安装python3编译器

      sudo apt update
      sudo apt install python3-pip

   2. python 安装transformers torch

      pip3 install transformers torch

   3. python 安装 sentencepiece

       sudo apt-get install cmake build-essential pkg-config libgoogle-perftools-dev
       pip install sentencepiece
       pip install sacremoses

   4. 创建python调用翻译文件（translate.py）

      from transformers import MarianMTModel, MarianTokenizer
      import sys
      
      def translate(text, src_lang, tgt_lang):
          model_name = f'Helsinki-NLP/opus-mt-{src_lang}-{tgt_lang}'
          # 'Helsinki-NLP/opus-mt-en-vi'
          tokenizer = MarianTokenizer.from_pretrained(model_name)
          model = MarianMTModel.from_pretrained(model_name)
      
          translated = model.generate(**tokenizer(text, return_tensors="pt", padding=True))
          result = [tokenizer.decode(t, skip_special_tokens=True) for t in translated]
          return result[0]
      
      if __name__ == "__main__":
          src_lang = sys.argv[1]
          tgt_lang = sys.argv[2]
          text = sys.argv[3]
          translated_text = translate(text,src_lang,tgt_lang)
          print(translated_text)

   5. 执行命令

      python3 [translate.py](http://translate.py/) en vi  "Poetry and distant places are both an infinite yearning for a peaceful rural life and an unremitting pursuit of an ideal life. They are like a spiritual haven, allowing us to find tranquility and peace in the busy world; they are like a lighthouse in the distance, guiding us to move forward bravely and pursue a better future. In the poetic words, we feel the expression of emotions, and among the green mountains and clear waters in the distance, we explore the true meaning of life and ultimately achieve a fulfilled and rich self."

   6. 进阶安装服务

      	pip install gunicorn
      	pip install Flask
      	pip install fastapi uvicorn
      

   7. 下载离线模型代码

      from transformers import MarianMTModel, MarianTokenizer
      
      def download_model(model_name,model_project):
          tokenizer = MarianTokenizer.from_pretrained(model_name)
          model = MarianMTModel.from_pretrained(model_name)
          tokenizer.save_pretrained(f'./model/{model_project}')
          model.save_pretrained(f'./model/{model_project}')
      
      if __name__ == "__main__":
          model_name = "Helsinki-NLP/opus-mt-en-vi"
          download_model(model_name,'en-vi')
      
      

   8. 离线模型下载

      python3 /mnt/d/project/by_token/download_model.py
      

   9. 编写服务端代码

      from fastapi import FastAPI, HTTPException
      from pydantic import BaseModel
      from transformers import MarianMTModel, MarianTokenizer
      import torch
      from concurrent.futures import ThreadPoolExecutor
      import asyncio
      
      app = FastAPI()
      
      device = 'cuda' if torch.cuda.is_available() else 'cpu'
      
      # Load models and tokenizers once
      models = {
          'zh_vi': MarianMTModel.from_pretrained('Helsinki-NLP/opus-mt-zh-vi').to(device),
          'zh_en': MarianMTModel.from_pretrained('Helsinki-NLP/opus-mt-zh-en').to(device),
          'en_vi': MarianMTModel.from_pretrained('Helsinki-NLP/opus-mt-en-vi').to(device),
          'en_zh': MarianMTModel.from_pretrained('Helsinki-NLP/opus-mt-en-zh').to(device)
      }
      tokenizers = {
          'zh_vi': MarianTokenizer.from_pretrained('Helsinki-NLP/opus-mt-zh-vi'),
          'zh_en': MarianTokenizer.from_pretrained('Helsinki-NLP/opus-mt-zh-en'),
          'en_vi': MarianTokenizer.from_pretrained('Helsinki-NLP/opus-mt-en-vi'),
          'en_zh': MarianTokenizer.from_pretrained('Helsinki-NLP/opus-mt-en-zh')
      }
      
      class TranslationRequest(BaseModel):
          src_lang: str
          tag_lang: str
          texts: list
          num_beams: int = 2
          max_length: int = 50
          batch_size: int = 16
      
      def translate_batch(batch, src_lang, tag_lang, num_beams, max_length):
          key = f"{src_lang}_{tag_lang}"
          tokenizer = tokenizers[key]
          model = models[key]
      
          inputs = tokenizer(batch, return_tensors="pt", padding=True, truncation=True).to(device)
          with torch.no_grad():
              translated = model.generate(
                  **inputs,
                  num_beams=num_beams,
                  max_length=max_length,
                  early_stopping=True
              )
          return [tokenizer.decode(t, skip_special_tokens=True) for t in translated]
      
      @app.post("/translate")
      async def translate(request: TranslationRequest):
          src_lang = request.src_lang
          tag_lang = request.tag_lang
          texts = request.texts
          num_beams = request.num_beams
          max_length = request.max_length
          batch_size = request.batch_size
      
          if f"{src_lang}_{tag_lang}" not in models or f"{src_lang}_{tag_lang}" not in tokenizers:
              raise HTTPException(status_code=400, detail="Unsupported language pair")
      
          batches = [texts[i:i + batch_size] for i in range(0, len(texts), batch_size)]
          results = []
      
          loop = asyncio.get_event_loop()
          with ThreadPoolExecutor(max_workers=4) as executor:
              tasks = [loop.run_in_executor(executor, translate_batch, batch, src_lang, tag_lang, num_beams, max_length) for batch in batches]
              results = await asyncio.gather(*tasks)
      
          flattened_results = [item for sublist in results for item in sublist]
          return { "status_code" : 200 ,"data": "".join(flattened_results)}
      
      if __name__ == "__main__":
          import uvicorn
          uvicorn.run(app, host="0.0.0.0", port=9380, workers=4)

   10. 执行服务端代码

        uvicorn src/main/translate_server_v2:app --host 0.0.0.0 --port 9380 --workers 8

   11. 客户端调用

       curl -X POST "http://localhost:9380/translate" -H "Content-Type: application/json" -d '{
         "src_lang": "zh",
         "tgt_lang": "en",
         "texts": "你好,
       }'