Prollama Stage 1

🚀 ProLLaMA：用於多任務蛋白質語言處理的蛋白質大語言模型

ProLLaMA 是一款專門為多任務蛋白質語言處理設計的蛋白質大語言模型，能夠有效處理蛋白質相關的多種任務，為蛋白質研究提供強大的語言處理能力。

🚀 快速開始

輸入格式

Seq=<
# 你也可以指定蛋白質序列的前幾個氨基酸：
Seq=<MAPGGMPRE

快速使用

# 你可以將 model_path 替換為你本地的路徑
CUDA_VISIBLE_DEVICES=0 python main.py --model "GreatCaptainNemo/ProLLaMA_Stage_1" --interactive
# main.py 代碼如下 👇:

import argparse
import json, os
import torch
from transformers import LlamaForCausalLM, LlamaTokenizer
from transformers import GenerationConfig
from tqdm import tqdm

generation_config = GenerationConfig(
    temperature=0.2,
    top_k=40,
    top_p=0.9,
    do_sample=True,
    num_beams=1,
    repetition_penalty=1.2,
    max_new_tokens=400
)

parser = argparse.ArgumentParser()
parser.add_argument('--model', default=None, type=str,help="The local path of the model. If None, the model will be downloaded from HuggingFace")
parser.add_argument('--interactive', action='store_true',help="If True, you can input instructions interactively. If False, the input instructions should be in the input_file.")
parser.add_argument('--input_file', default=None, help="You can put all your input instructions in this file (one instruction per line).")
parser.add_argument('--output_file', default=None, help="All the outputs will be saved in this file.")
args = parser.parse_args()

if __name__ == '__main__':
    if args.interactive and args.input_file:
        raise ValueError("interactive is True, but input_file is not None.")
    if (not args.interactive) and (args.input_file is None):
        raise ValueError("interactive is False, but input_file is None.")
    if args.input_file and (args.output_file is None):
        raise ValueError("input_file is not None, but output_file is None.")

    load_type = torch.bfloat16
    if torch.cuda.is_available():
        device = torch.device(0)
    else:
        raise ValueError("No GPU available.")


    model = LlamaForCausalLM.from_pretrained(
        args.model,
        torch_dtype=load_type,
        low_cpu_mem_usage=True,
        device_map='auto',
        quantization_config=None
    )
    tokenizer = LlamaTokenizer.from_pretrained(args.model)

    model.eval()
    with torch.no_grad():
        if args.interactive:
            while True:
                raw_input_text = input("Input:")
                if len(raw_input_text.strip())==0:
                    break
                input_text = raw_input_text
                input_text = tokenizer(input_text,return_tensors="pt")  

                generation_output = model.generate(
                                input_ids = input_text["input_ids"].to(device),
                                attention_mask = input_text['attention_mask'].to(device),
                                eos_token_id=tokenizer.eos_token_id,
                                pad_token_id=tokenizer.pad_token_id,
                                generation_config = generation_config,
                                output_attentions=False
                            )
                s = generation_output[0]
                output = tokenizer.decode(s,skip_special_tokens=True)
                print("Output:",output)
                print("\n")
        else:
            outputs=[]
            with open(args.input_file, 'r') as f:
                examples =f.read().splitlines()
            print("Start generating...")
            for index, example in tqdm(enumerate(examples),total=len(examples)):
                input_text = tokenizer(example,return_tensors="pt")  #add_special_tokens=False ?

                generation_output = model.generate(
                    input_ids = input_text["input_ids"].to(device),
                    attention_mask = input_text['attention_mask'].to(device),
                    eos_token_id=tokenizer.eos_token_id,
                    pad_token_id=tokenizer.pad_token_id,
                    generation_config = generation_config
                )
                s = generation_output[0]
                output = tokenizer.decode(s,skip_special_tokens=True)
                outputs.append(output)
            with open(args.output_file,'w') as f:
                f.write("\n".join(outputs))
            print("All the outputs have been saved in",args.output_file)

📄 許可證

本項目採用 Apache-2.0 許可證。此外，ProLLaMA_Stage_1 基於 Llama-2-7b，因此請遵循 Llama2 的許可證。

📚 詳細文檔

• arXiv 論文 可獲取更多信息。
• GitHub 倉庫 可獲取更多信息。

📖 引用

如果你使用了本項目，請引用以下論文：

@article{lv2024prollama,
  title={ProLLaMA: A Protein Large Language Model for Multi-Task Protein Language Processing},
  author={Lv, Liuzhenghao and Lin, Zongying and Li, Hao and Liu, Yuyang and Cui, Jiaxi and Chen, Calvin Yu-Chian and Yuan, Li and Tian, Yonghong},
  journal={arXiv preprint arXiv:2402.16445},
  year={2024}
}