app.py

import streamlit as st
import torch
from simplet5 import SimpleT5
from transformers import T5Tokenizer, T5ForConditionalGeneration
import chinese_converter

MODEL_PATH = "hululuzhu/chinese-poem-t5-mengzi-finetune"
class PoemModel(SimpleT5):
  def __init__(self) -> None:
    super().__init__()
    self.device = torch.device("cuda")

  def load_my_model(self):
    self.tokenizer = T5Tokenizer.from_pretrained(MODEL_PATH)
    self.model = T5ForConditionalGeneration.from_pretrained(MODEL_PATH)

# 有一些预先设定参数
AUTHOR_PROMPT = "模仿："
TITLE_PROMPT = "作诗："
EOS_TOKEN = '</s>'

poem_model = PoemModel()
poem_model.load_my_model()
poem_model.model = poem_model.model.to('cuda')

MAX_AUTHOR_CHAR = 4
MAX_TITLE_CHAR = 12
MIN_CONTENT_CHAR = 10
MAX_CONTENT_CHAR = 64

poem("百花")

def poem(title_str, opt_author=None, model=poem_model,
         is_input_traditional_chinese=False):
  model.model = model.model.to('cuda')
  if opt_author:
    in_request = TITLE_PROMPT + title_str[:MAX_TITLE_CHAR] + EOS_TOKEN + AUTHOR_PROMPT + opt_author[:MAX_AUTHOR_CHAR]
  else:
    in_request = TITLE_PROMPT + title_str[:MAX_TITLE_CHAR]
  if is_input_traditional_chinese:
    in_request = chinese_converter.to_simplified(in_request)
  out = model.predict(in_request,
                      max_length=MAX_CONTENT_CHAR)[0].replace(",", "，")
  if is_input_traditional_chinese:
    out = chinese_converter.to_traditional(out)
    st.text(f"標題： {in_request.replace('</s>', ' ')}\n詩歌： {out}")
  else:
    st.text(f"标题： {in_request.replace('</s>', ' ')}\n诗歌： {out}")