app.py

import gradio as gr
from typing import List, Optional
from transformers import BertTokenizer, BartForConditionalGeneration

title = "HIT-TMG/dialogue-bart-large-chinese"
description = """
This is a seq2seq model pre-trained on several Chinese dialogue datasets, from bart-large-chinese. 
However it is just a simple demo for this pre-trained model. It's better to fine-tune it on downstream tasks for better performance \n
See some details of model card at https://huggingface.co/HIT-TMG/dialogue-bart-large-chinese . \n\n
Besides starting the conversation from scratch, you can also input the whole dialogue history utterance by utterance seperated by '[SEP]'. \n
"""


tokenizer = BertTokenizer.from_pretrained("HIT-TMG/dialogue-bart-large-chinese")
model = BartForConditionalGeneration.from_pretrained("HIT-TMG/dialogue-bart-large-chinese")

tokenizer.truncation_side = 'left'
max_length = 512

examples = [
    ["你有什么爱好吗"],
    ["你好。[SEP]嘿嘿你好，请问你最近在忙什么呢？[SEP]我最近养了一只狗狗，我在训练它呢。"]
]


def chat_func(input_utterance: str, history: Optional[List[str]] = None):
    if history is not None:
        history.extend(input_utterance.split(tokenizer.sep_token))
    else:
        history = input_utterance.split(tokenizer.sep_token)

    history_str = "对话历史：" + tokenizer.sep_token.join(history)

    input_ids = tokenizer(history_str,
                          return_tensors='pt',
                          truncation=True,
                          max_length=max_length,
                          ).input_ids

    output_ids = model.generate(input_ids,
                                max_new_tokens=30,
                                top_k=32,
                                num_beams=4,
                                repetition_penalty=1.2,
                                no_repeat_ngram_size=4)[0]

    response = tokenizer.decode(output_ids, skip_special_tokens=True)

    history.append(response)


    if len(history) % 2 == 0:
        display_utterances = [(history[i], history[i + 1]) for i in range(0, len(history) - 1, 2)]
    else:
        display_utterances = [("", history[0])] + [(history[i], history[i + 1]) for i in range(1, len(history) - 1, 2)]

    return display_utterances, history


demo = gr.Interface(fn=chat_func,
                    title=title,
                    description=description,
                    inputs=[gr.Textbox(lines=1, placeholder="Input current utterance"), "state"],
                    examples=examples,
                    outputs=["chatbot", "state"])


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


# def chat(history):
#     history_prefix = "对话历史："
#     history = history_prefix + history
#
#     outputs = tokenizer(history,
#                         return_tensors='pt',
#                         padding=True,
#                         truncation=True,
#                         max_length=512)
#
#     input_ids = outputs.input_ids
#     output_ids = model.generate(input_ids)[0]
#
#     return tokenizer.decode(output_ids, skip_special_tokens=True)
#
#
# chatbot = gr.Chatbot().style(color_map=("green", "pink"))
# demo = gr.Interface(
#     chat,
#     inputs=gr.Textbox(lines=8, placeholder="输入你的对话历史（请以'[SEP]'作为每段对话的间隔）\nInput the dialogue history (Please split utterances by '[SEP]')"),
#     title=title,
#     description=description,
#     outputs =["text"]
# )
#
#
# if __name__ == "__main__":
#     demo.launch()