language: zh tags:

• gpt2
• 对话系统 license: cc-by-nc-sa-4.0

韩语对话GPT模型

使用方法

from transformers import PreTrainedTokenizerFast, GPT2LMHeadModel
import torch


device = 'cuda' if torch.cuda.is_available() else 'cpu'

tokenizer = PreTrainedTokenizerFast.from_pretrained('byeongal/Ko-DialoGPT')
model = GPT2LMHeadModel.from_pretrained('byeongal/Ko-DialoGPT').to(device)

past_user_inputs = []
generated_responses = []

while True:
    user_input = input(">> 用户输入:")
    if user_input == 'bye':
        break
    text_idx = tokenizer.encode(user_input + tokenizer.eos_token, return_tensors='pt')
    for i in range(len(generated_responses)-1, len(generated_responses)-3, -1):
        if i < 0:
            break
        encoded_vector = tokenizer.encode(generated_responses[i] + tokenizer.eos_token, return_tensors='pt')
        if text_idx.shape[-1] + encoded_vector.shape[-1] < 1000:
            text_idx = torch.cat([encoded_vector, text_idx], dim=-1)
        else:
            break
        encoded_vector = tokenizer.encode(past_user_inputs[i] + tokenizer.eos_token, return_tensors='pt')
        if text_idx.shape[-1] + encoded_vector.shape[-1] < 1000:
            text_idx = torch.cat([encoded_vector, text_idx], dim=-1)
        else:
            break
    text_idx = text_idx.to(device)
    inference_output = model.generate(
            text_idx,
            max_length=1000,
            num_beams=5,
            top_k=20,
            no_repeat_ngram_size=4,
            length_penalty=0.65,
            repetition_penalty=2.0,
        )
    inference_output = inference_output.tolist()
    bot_response = tokenizer.decode(inference_output[0][text_idx.shape[-1]:], skip_special_tokens=True)
    print(f"机器人回复: {bot_response}")
    past_user_inputs.append(user_input)
    generated_responses.append(bot_response)

参考文献

• SKT韩语GPT2模型
• 韩国电子技术研究院研发数据
• 韩语对话摘要数据集