pages/GPT.py

import streamlit as st
import torch
import numpy as np
import transformers 
import random
import textwrap

@st.cache_data
def load_model():
    model_finetuned = transformers.AutoModelWithLMHead.from_pretrained(
        'tinkoff-ai/ruDialoGPT-small',
        output_attentions = False,
        output_hidden_states = False
    )
    model_finetuned.load_state_dict(torch.load('GPT_sonnik_only.pt', map_location=torch.device('cpu')))
    tokenizer = transformers.AutoTokenizer.from_pretrained('tinkoff-ai/ruDialoGPT-small')
    return model_finetuned, tokenizer

def preprocess_text(text_input, tokenizer):
    prompt = tokenizer.encode(text_input, return_tensors='pt')
    return prompt

def predict_sentiment(model, prompt, temp, num_generate):
    print('1')
    with torch.inference_mode():
        print('2')
        result = model.generate(
            input_ids=prompt,
            max_length=100,
            num_beams=5,
            do_sample=True,
            temperature=float(temp),
            top_k=50,
            top_p=0.6,
            no_repeat_ngram_size=3,
            num_return_sequences=num_generate,
            ).cpu().numpy()
        print(result)
    return result

st.title('Text generation with dreambook')

model, tokenizer = load_model()

text_input = st.text_input("Enter some text about movie")
max_len = st.slider('Length of sequence', 0, 100, 50)
temp = st.slider('Temperature', 1, 30, 1)
num_generate = st.text_input("Enter number of sequences")
    
if st.button('Generate'):
    print('uirhf')
    prompt = preprocess_text(text_input, tokenizer)
    print('uirhf')
    result = predict_sentiment(model, prompt, temp, int(num_generate))
    print('uirhf')
    for i in result:
        st.write(textwrap.fill(tokenizer.decode(i), max_len))