app.py

import transformers
import torch
import tokenizers
import streamlit as st
import re

from PIL import Image


@st.cache(hash_funcs={tokenizers.Tokenizer: lambda _: None, tokenizers.AddedToken: lambda _: None, re.Pattern: lambda _: None}, allow_output_mutation=True, suppress_st_warning=True)
def get_model(model_name, model_path):
    tokenizer = transformers.GPT2Tokenizer.from_pretrained(model_name)
    tokenizer.add_special_tokens({
        'eos_token': '[EOS]'
    })
    model = transformers.GPT2LMHeadModel.from_pretrained(model_name)
    model.resize_token_embeddings(len(tokenizer))
    model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))
    model.eval()
    return model, tokenizer


def predict(text, model, tokenizer, n_beams=5, temperature=2.5, top_p=0.8, length_of_generated=300):
    text += '\n'
    input_ids = tokenizer.encode(text, return_tensors="pt")
    length_of_prompt = len(input_ids[0])
    with torch.no_grad():
        out = model.generate(input_ids,
                             do_sample=True,
                             num_beams=n_beams,
                             temperature=temperature,
                             top_p=top_p,
                             max_length=length_of_prompt + length_of_generated,
                             eos_token_id=tokenizer.eos_token_id
                             )

    return list(map(tokenizer.decode, out))[0]


medium_model, medium_tokenizer = get_model('sberbank-ai/rugpt3medium_based_on_gpt2', 'korzh-medium_best_eval_loss.bin')
large_model, large_tokenizer = get_model('sberbank-ai/rugpt3large_based_on_gpt2', 'korzh-large_best_eval_loss.bin')

# st.title("NeuroKorzh")

image = Image.open('korzh.jpg')
st.image(image, caption='НейроКорж')

option = st.selectbox('Выберите своего Коржа', ('Быстрый', 'Глубокий'))
craziness = st.slider(label='Абсурдность', min_value=0.5, max_value=4., value=2.5, step=0.1)

st.markdown("\n")

text = st.text_area(label='Напишите начало песни', value='Что делать, Макс?', height=100)
button = st.button('Старт')

if button:
    try:
        with st.spinner("Пушечка пишется"):
            if option == 'Быстрый':
                result = predict(text, medium_model, medium_tokenizer, temperature=craziness)
            elif option == 'Глубокий':
                result = predict(text, large_model, large_tokenizer, temperature=craziness)
            else:
                raise st.error('Error in selectbox')
        
        st.text_area(label='', value=result, height=1100)
    
    except Exception:
        st.error("Ooooops, something went wrong. Please try again and report to me, tg: @vladyur")