app.py

import re
import time
import streamlit as st
import pandas as pd
from wordcloud import WordCloud
from googletrans import Translator
from transformers import AutoTokenizer, AutoModelWithLMHead, pipeline

from enums import MODEL_NAME, MESSAGES, DESCRIPTION

def iso2lang(iso):
    return MESSAGES["iso"][iso]

def create_df_from_io(input, output):
    return pd.DataFrame([[input, output, time.time()]], columns=["input", "output", "timestamp"])

def simple_clean(text):
    return re.sub('[!@#$.,\n-?]', ' ', text.lower())

def load_tokenizer():
    return AutoTokenizer.from_pretrained(MODEL_NAME)

@st.cache(allow_output_mutation=True)
def load_model():
    return AutoModelWithLMHead.from_pretrained(MODEL_NAME)

def load_pipe():
    model = load_model()
    tokenizer = load_tokenizer()
    return pipeline("text-generation", model=model, tokenizer=tokenizer)
# ---------------------------------------------------------------------- #
st.write(DESCRIPTION)

lang = st.radio('Хэл / Language', ('mn', 'en'), format_func=iso2lang)
translator = Translator()
if "df" not in st.session_state:
    st.session_state.df = pd.DataFrame(columns=["input", "output", "timestamp"])

with st.spinner(MESSAGES["loading_text"][lang]):
    pipe = load_pipe()
st.success(MESSAGES["success_model_load"][lang])

text = st.text_input(
    MESSAGES["input_description"][lang], MESSAGES["input_default"][lang])

with st.spinner(MESSAGES["loading_text"][lang]):
    if lang == "mn":
        result = pipe(text)[0]['generated_text']
        st.write(result)
    elif lang == "en":
        text = translator.translate(text, src='en', dest='mn').text
        result = pipe(text)[0]['generated_text']
        result_en = translator.translate(result, src='mn', dest='en').text
        st.write(f"*Translated:* {result_en}")
        st.write(f"> *Original:* {result}")
        st.warning('Translation is done by [`googletrans`](https://github.com/ssut/py-googletrans). Please check out the usage. https://github.com/ssut/py-googletrans#note-on-library-usage')

    st.session_state.df = st.session_state.df.append(create_df_from_io(text, result))


st.write("### WordCloud based on previous outputs")
with st.spinner(MESSAGES["loading_text"][lang]):
    wordcloud_input = ""
    for text in st.session_state.df.output.tolist():
        wordcloud_input += simple_clean(text)

    wordcloud = WordCloud(width = 800, height = 800,
                background_color ='white',
                min_font_size = 10).generate(wordcloud_input)

    st.image(wordcloud.to_array())

st.write("### Түүх / History")
with st.spinner(MESSAGES["loading_text"][lang]):
    st.table(st.session_state.df.sort_values(by="timestamp", ascending=False))