app.py

import streamlit as st
from kdnv_preprocess import data_preprocessing
from sentence_transformers import SentenceTransformer
import faiss
import numpy as np
import json
from collections import Counter

@st.cache_resource
def load_model():
    return SentenceTransformer("sentence-transformers/paraphrase-multilingual-mpnet-base-v2")

@st.cache_resource
def load_index():
    indices = {
        'L2': faiss.read_index("models/index_l2.faiss"),
        'Dot': faiss.read_index("models/index_dot.faiss"),
        'Cos': faiss.read_index("models/index_cosine.faiss")
    }
    return indices

model = load_model()
indices = load_index()

with open('class_dict.json', 'r') as file:
    class_dict = json.load(file)

st.header('Кальянный угадыватель')
st.markdown('[для Кобза](https://www.youtube.com/watch?v=dQw4w9WgXcQ)')
st.divider()

with st.form(key='pred'):
    text = st.text_area(label='Введи сюда описание табака')
    button = st.form_submit_button('Узнать предсказание')

if button:
    text = data_preprocessing(text)
    prompt_embedding = model.encode(text).astype('float32')
    prompt_embedding = prompt_embedding[np.newaxis, :]

    _, indices_result_l2 = indices['L2'].search(prompt_embedding, 1)
    _, indices_result_dot = indices['Dot'].search(prompt_embedding, 1)
    _, indices_result_cosine = indices['Cos'].search(prompt_embedding, 1)

    pred_l2 = class_dict[str(indices_result_l2[0][0])]
    pred_dot = class_dict[str(indices_result_dot[0][0])]
    pred_cosine = class_dict[str(indices_result_cosine[0][0])]

    predictions = [pred_l2, pred_dot, pred_cosine]

    prediction_counts = Counter(predictions)

    final_prediction = prediction_counts.most_common(1)[0][0]

    if len(prediction_counts) == len(predictions):
        final_prediction = pred_l2

    st.subheader(f'Я считаю, что это: {final_prediction}')