app.py

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"

from huggingface_hub import hf_hub_download
import gradio as gr
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import torch.nn as nn
import pickle
import numpy as np
import re
import fasttext

svm_repo_id = "HighFive-OPJ/svm-sentiment-model"
svm_model_path = hf_hub_download(repo_id=svm_repo_id, filename="svm_model.pkl")
with open(svm_model_path, "rb") as f:
    svm_model = pickle.load(f)
vectorizer_path = hf_hub_download(repo_id=svm_repo_id, filename="vectorizer.pkl")
with open(vectorizer_path, "rb") as f:
    vectorizer = pickle.load(f)

fasttext_path = hf_hub_download(
    repo_id="HighFive-OPJ/Deep_Learning",
    filename="FastText.bin",
    repo_type="dataset"
)
ft_model = fasttext.load_model(fasttext_path)

class LSTMClassifier(nn.Module):
    def __init__(self, input_dim=300, hidden_dim=256, num_classes=3):
        super().__init__()
        self.lstm = nn.LSTM(input_dim, hidden_dim, batch_first=True, bidirectional=True)
        self.fc = nn.Linear(hidden_dim * 2, num_classes)
    
    def forward(self, x):
        _, (hn, _) = self.lstm(x)
        hn = torch.cat((hn[-2], hn[-1]), dim=1)
        out = self.fc(hn)
        return out

lstm_repo_id = "HighFive-OPJ/lstm-sentiment-model"
lstm_model_path = hf_hub_download(repo_id=lstm_repo_id, filename="fasttext_lstm.pt")

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
lstm_model = LSTMClassifier()
lstm_model.load_state_dict(torch.load(lstm_model_path, map_location=device))
lstm_model.to(device)
lstm_model.eval()

bert_repo_id = "HighFive-OPJ/bertic_sentiment"
bert_tokenizer = AutoTokenizer.from_pretrained(bert_repo_id)
bert_model = AutoModelForSequenceClassification.from_pretrained(bert_repo_id)
bert_model.to(device)
bert_model.eval()

def preprocess_text(text):
    text = text.lower()
    text = re.sub(r"[^a-zA-Z\s]", "", text).strip()
    return text

def text_to_fasttext_tensor(text, max_len=200):
    tokens = preprocess_text(text).split()
    vectors = []
    for t in tokens[:max_len]:
        vec = ft_model.get_word_vector(t)
        vectors.append(vec)
    while len(vectors) < max_len:
        vectors.append(np.zeros(300))
    return torch.tensor([vectors], dtype=torch.float32).to(device)

def predict_with_svm(text):
    transformed = vectorizer.transform([text])
    prediction = svm_model.predict(transformed)
    return int(prediction[0])

def predict_with_lstm(text):
    input_tensor = text_to_fasttext_tensor(text)
    with torch.no_grad():
        outputs = lstm_model(input_tensor)
        pred = torch.argmax(outputs, dim=1).item()
    return pred

def predict_with_bert(text):
    inputs = bert_tokenizer([text], padding=True, truncation=True, max_length=512, return_tensors="pt").to(device)
    with torch.no_grad():
        outputs = bert_model(**inputs)
        logits = outputs.logits
        predictions = logits.argmax(axis=-1).cpu().numpy()
    bert_score = int(predictions[0])
    if bert_score <= 2:
        return 0
    elif bert_score == 3:
        return 1
    else:
        return 2

def analyze_sentiment(text):
    try:
        svm_result = predict_with_svm(text)
    except Exception as e:
        svm_result = f"Error: {str(e)}"
    
    try:
        lstm_result = predict_with_lstm(text)
    except Exception as e:
        lstm_result = f"Error: {str(e)}"
    
    try:
        bert_result = predict_with_bert(text)
    except Exception as e:
        bert_result = f"Error: {str(e)}"

    try:
        scores = []
        for r in [svm_result, lstm_result, bert_result]:
            if isinstance(r, int):
                scores.append(r)
        average = np.mean(scores) if scores else float("nan")
        stats = f"Average Score (0=Pos,1=Neg,2=Neu): {average:.2f}\n"
    except Exception as e:
        stats = f"Error calculating stats: {str(e)}"

    def format_output(result):
        return convert_to_stars(result) if isinstance(result, int) else result

    return (
        format_output(svm_result),
        format_output(lstm_result),
        format_output(bert_result),
        stats
    )
def convert_to_stars(score):
    star_map = {0: 5, 1: 1, 2: 3}
    stars = star_map.get(score, 3)  
    return "★" * stars + "☆" * (5 - stars)

def process_input(text):
    if not text.strip():
        return ("", "", "", "Please enter valid text.")
    try:
        return analyze_sentiment(text)
    except Exception as e:
        error_message = f"Error during sentiment analysis:\n{str(e)}"
        return ("error", "error", "error", error_message)

with gr.Blocks() as demo:
    gr.Markdown("# Sentiment Analysis Demo")
    gr.Markdown("""
Enter a review and see how different models evaluate its sentiment! This app uses:
- SVM for classic machine learning  
- LSTM for deep learning (using FastText)  
- BERTić for transformer-based analysis  

Rating guide:  
5 ★ → positive sentiment  
3 ★ → neutral sentiment  
1 ★ → negative sentiment
""")

    with gr.Row():
        with gr.Column():
            input_text = gr.Textbox(label="Enter your review:", lines=3)
            analyze_button = gr.Button("Analyze Sentiment")

        with gr.Column():
            svm_output = gr.Textbox(label="SVM", interactive=False)
            lstm_output = gr.Textbox(label="LSTM", interactive=False)
            bert_output = gr.Textbox(label="BERTić", interactive=False)
            stats_output = gr.Textbox(label="Statistics", interactive=False)

    analyze_button.click(
        process_input,
        inputs=[input_text],
        outputs=[svm_output, lstm_output, bert_output, stats_output]
    )

demo.launch()