demo_chatbot_jddc.py

# coding=utf-8
# author: xusong <[email protected]>
# time: 2022/9/05 14:12

"""
TODO: 还要能判断是否需要回复。
"""

import torch
import gradio as gr
from info import article
from kplug import modeling_kplug_s2s_patch
from transformers import BertTokenizer, BartForConditionalGeneration

model = BartForConditionalGeneration.from_pretrained("eson/kplug-base-jddc")
tokenizer = BertTokenizer.from_pretrained("eson/kplug-base-jddc")


def predict(input, history=[]):
    """
    拼接方案：直接拼接history作为输入，不区分角色。虽然简单粗糙，但是encoder-decoder架构不会混淆输入和输出（如果是gpt架构就需要区分角色了）。
    """
    # append the new user input tokens to the chat history
    history = history + [input]  # history如果包含错误的response，可能会造成误差传递

    # tokenize the new input sentence
    bot_input_ids = tokenizer.encode("".join(history)[-500:], return_tensors='pt')

    # bot_input_ids = torch.cat([torch.LongTensor(history), new_user_input_ids], dim=-1)

    # generate a response
    response = model.generate(bot_input_ids, max_length=1000, pad_token_id=tokenizer.eos_token_id).tolist()

    # convert the tokens to text, and then split the responses into lines
    response = "".join(tokenizer.decode(response[0], skip_special_tokens=True).split())
    history = history + [response]
    response = [(history[i], history[i + 1]) for i in range(0, len(history) - 1, 2)]  # convert to tuples of list
    return response, history


jddc_examples = [
    # 价保
    "昨天刚买的怎么就降了几十块，应该补给我差价吧",

    "请问这个猕猴桃是有货的吗?",
    # 到货时间
    "我下的这个单怎么还没到",
    # 快递
    "发什么快递",
    "能发邮政吗",
]

jddc_iface = gr.Interface(
    fn=predict,
    # inputs=["text", "state"],
    inputs=[
        gr.Textbox(
            label="输入文本",
            value="发什么快递"),  # gr.State() 报错
        "state"
    ],
    outputs=["chatbot", "state"],
    examples=jddc_examples,
    title="电商客服-生成式对话（Response Generation）",
    article=article,
)

if __name__ == "__main__":
    jddc_iface.launch()