pages/review_predictor.py

import json
import pickle

import pandas as pd
import streamlit as st
import torch
import torch.nn as nn
import transformers

from model.funcs import (create_model_and_tokenizer, load_model,
                         predict_sentiment)
from model.model import LSTMConcatAttentionEmbed
from preprocessing.preprocessing import data_preprocessing
from preprocessing.rnn_preprocessing import preprocess_single_string

# Load preprocessing steps
with open("vectorizer.pkl", "rb") as f:
    logreg_vectorizer = pickle.load(f)

# Load trained model
with open("logreg_model.pkl", "rb") as f:
    logreg_predictor = pickle.load(f)

model_concat_embed = LSTMConcatAttentionEmbed()
model_concat_embed.load_state_dict(torch.load("model/model_weights.pt"))

with open("model/vocab.json", "r") as f:
    vocab_to_int = json.load(f)

with open("model/int_vocab.json", "r") as f:
    int_to_vocab = json.load(f)

model_class = transformers.AutoModel
tokenizer_class = transformers.AutoTokenizer
pretrained_weights = "cointegrated/rubert-tiny2"
weights_path = "model/best_bert_weights.pth"
model = load_model(model_class, pretrained_weights, weights_path)
tokenizer = tokenizer_class.from_pretrained(pretrained_weights)


def plot_and_predict(review: str, SEQ_LEN: int, model: nn.Module):
    inp = preprocess_single_string(review, SEQ_LEN, vocab_to_int)
    model.eval()
    with torch.inference_mode():
        pred, _ = model(inp.long().unsqueeze(0))
    pred = pred.sigmoid().item()
    return 1 if pred > 0.75 else 0


def preprocess_text_logreg(text):
    # Apply preprocessing steps (cleaning, tokenization, vectorization)
    clean_text = data_preprocessing(
        text
    )  # Assuming data_preprocessing is your preprocessing function
    print("Clean text ", clean_text)
    vectorized_text = logreg_vectorizer.transform([" ".join(clean_text)])
    return vectorized_text


# Define function for making predictions
def predict_sentiment_logreg(text):
    # Preprocess input text
    processed_text = preprocess_text_logreg(text)
    # Make prediction
    prediction = logreg_predictor.predict(processed_text)
    return prediction


metrics = {
    "Models": ["Logistic Regression", "LSTM + attention", "ruBERTtiny2"],
    "f1-macro score": [0.94376, 1, 0.94070],
}


col1, col2 = st.columns([1, 3])
df = pd.DataFrame(metrics)
df.set_index("Models", inplace=True)
df.index.name = "Model"


st.sidebar.title("Model Selection")
model_type = st.sidebar.radio("Select Model Type", ["Classic ML", "LSTM", "BERT"])
st.title("Review Prediction")

# Streamlit app code
st.title("Sentiment Analysis with Logistic Regression")
text_input = st.text_input("Enter your review:")
if st.button("Predict"):
    if model_type == "Classic ML":
        prediction = predict_sentiment_logreg(text_input)
    elif model_type == "LSTM":
        prediction = plot_and_predict(
            review=text_input, SEQ_LEN=25, model=model_concat_embed
        )
    elif model_type == "BERT":
        prediction = predict_sentiment(text_input, model, tokenizer, "cpu")
        st.write(prediction)

    if prediction == 1:
        st.write("prediction")
        st.write("Отзыв положительный")
    elif prediction == 0:
        st.write("prediction")
        st.write("Отзыв отрицательный")

st.write(df)