correct.py

# -*- coding: utf-8 -*-
import torch
from transformers import AutoTokenizer, GPT2LMHeadModel

O_TKN = '<origin>'
C_TKN = '<correct>'
BOS = "</s>"
EOS = "</s>"
PAD = "<pad>"
MASK = '<unused0>'
SENT = '<unused1>'


def chat():
    tokenizer = AutoTokenizer.from_pretrained('skt/kogpt2-base-v2',
                                               eos_token=EOS, unk_token='<unk>',
                                               pad_token=PAD, mask_token=MASK)
    model = GPT2LMHeadModel.from_pretrained('Moo/kogpt2-proofreader')
    with torch.no_grad():
        while True:
            q = input('원래문장: ').strip()
            if q == 'quit':
                break
            a = ''
            while True:
                input_ids = torch.LongTensor(tokenizer.encode(O_TKN + q + C_TKN + a)).unsqueeze(dim=0)
                pred = model(input_ids)
                gen = tokenizer.convert_ids_to_tokens(
                    torch.argmax(
                        pred[0],
                        dim=-1).squeeze().numpy().tolist())[-1]
                if gen == EOS:
                    break
                a += gen.replace('▁', ' ')
            print(f"교정: {a.strip()}")


if __name__ == "__main__":
    chat()