gpt_v1.py

from transformers import GPT2LMHeadModel, GPT2Tokenizer
import torch
import streamlit as st

model = GPT2LMHeadModel.from_pretrained(
    'sberbank-ai/rugpt3small_based_on_gpt2',
    output_attentions = False,
    output_hidden_states = False,
)

tokenizer = GPT2Tokenizer.from_pretrained('sberbank-ai/rugpt3small_based_on_gpt2')


# Вешаем сохраненные веса на нашу модель
model.load_state_dict(torch.load('models/model.pt', map_location=torch.device('cpu')))

prompt = st.text_input('Введите текст prompt:')
length = st.slider('Длина генерируемой последовательности:', 10, 1000, 50)
num_samples = st.slider('Число генераций:', 1, 10, 1)
temperature = st.slider('Температура:', 0.1, 1.0, 0.5)


def generate_text(model, tokenizer, prompt, length, num_samples, temperature):
    input_ids = tokenizer.encode(prompt, return_tensors='pt')
    output_sequences = model.generate(
        input_ids=input_ids,
        max_length=length,
        num_return_sequences=num_samples,
        temperature=temperature
    )

    generated_texts = []
    for output_sequence in output_sequences:
        generated_text = tokenizer.decode(output_sequence, clean_up_tokenization_spaces=True)
        generated_texts.append(generated_text)

    return generated_texts


if st.button('Сгенерировать текст'):
    generated_texts = generate_text(model, tokenizer, prompt, length, num_samples, temperature)
    for i, text in enumerate(generated_texts):
        st.write(f'Текст {i+1}:')
        st.write(text)