paraphraser.py

from transformers import (
    AutoTokenizer,
    AutoModelForSeq2SeqLM,
)

class Paraphraser:
    def __init__(self, model_name='humarin/chatgpt_paraphraser_on_T5_base'):
        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
        self.model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
        self.model.eval()

    def paraphrase(self, text, num_return_sequences=5, num_beams=5, 
                  num_beam_groups=1, diversity_penalty=0.0, device=None):
        try:
            input_text = "paraphrase: " + text + " </s>"
            encoding = self.tokenizer.encode_plus(input_text, return_tensors="pt")
            
            # Move input tensors to specified device if provided
            if device is not None:
                input_ids = encoding["input_ids"].to(device)
                self.model = self.model.to(device)
            else:
                input_ids = encoding["input_ids"]

            outputs = self.model.generate(
                input_ids=input_ids,
                max_length=256,
                num_beams=num_beams,
                num_beam_groups=num_beam_groups,
                num_return_sequences=num_return_sequences,
                diversity_penalty=diversity_penalty,
                early_stopping=True
            )
            
            # Move outputs back to CPU for tokenizer decoding
            outputs = outputs.cpu() if device is not None else outputs
            paraphrases = [self.tokenizer.decode(output, skip_special_tokens=True) 
                          for output in outputs]
            return paraphrases
        except Exception as e:
            print(f"Error in paraphrasing: {e}")
            return []