scrl_compressor.py

from SCRL_new.scrl.model import load_model
from transformers import AutoTokenizer
import re
from abs_compressor import AbstractCompressor


class SCRLCompressor(AbstractCompressor):

    def __init__(self, model_dir: str, device: str = "cpu", tokenizer_dir: str = "sentence-transformers/paraphrase-distilroberta-base-v2"):
        self.model_dir = model_dir
        self.device = device
        self.model = load_model(self.model_dir, self.device)
        self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_dir)

    def compress(self, original_prompt: str, ratio: float = 0.5, max_length: int = 256) -> dict:
        original_tokens = len(self.gpt_tokenizer.encode(original_prompt))

        # sources = [original_prompt.strip()]
        sources = re.findall(r'.{%d}' % max_length, original_prompt.strip())
        # print(sources)
        if sources:
            summaries = self.model.predict(sources, self.tokenizer, self.device)
            # print(sources)
            # print(summaries)

            compressed_prompt = ""
            for s in summaries:
                compressed_prompt += s

            compressed_tokens = len(self.gpt_tokenizer.encode(compressed_prompt))

            result = {
                'compressed_prompt': compressed_prompt,
                'ratio': compressed_tokens / original_tokens,
                'original_tokens': original_tokens,
                'compressed_tokens': compressed_tokens,
            }

            return result
        else:
            result = {
                'compressed_prompt': "",
                'ratio': 0,
                'original_tokens': "",
                'compressed_tokens': "",
            }
            return result