app.py

import gradio as gr
from transformers import pipeline
import time

# Load the sentiment analysis model
print("Loading model...")
# Force reload without cache
classifier = pipeline(
    "sentiment-analysis", 
    model="./istrenirani_model",
    device=-1  # Force CPU (same as Streamlit)
)
print("Model loaded successfully!")

# Translation function
def prevedi_labelu(label, score):
    """Translate model labels to Serbian"""
    # Note: Your model's label mapping from training:
    # LABEL_0 = Negativan (0)
    # LABEL_1 = Neutralan (1)  
    # LABEL_2 = Pozitivan (2)
    if label == 'LABEL_0':
        return "🔴 Negativan", score, "#ff4444"
    elif label == 'LABEL_1':
        return "⚪ Neutralan", score, "#888888"
    elif label == 'LABEL_2':
        return "🟢 Pozitivan", score, "#44ff44"
    return "❓ Nepoznat", score, "#888888"

# Main prediction function
def analyze_sentiment(text):
    """Analyze sentiment of Serbian text"""
    if not text or not text.strip():
        return "⚠️ Molim vas unesite tekst!"
    
    try:
        # Get prediction
        result = classifier(text)[0]
        
        # Translate result
        sentiment, confidence, color = prevedi_labelu(result['label'], result['score'])
        
        # Create colored HTML output
        sentiment_html = f"""
        <div style='text-align: center; padding: 40px; border-radius: 15px; background-color: {color}22; border: 3px solid {color};'>
            <h1 style='color: {color}; margin: 0; font-size: 56px;'>{sentiment}</h1>
        </div>
        """
        
        return sentiment_html
        
    except Exception as e:
        return f"❌ Greška: {str(e)}"


# Custom CSS for better styling
custom_css = """
#sentiment-output {
    font-size: 24px;
    font-weight: bold;
}
.gradio-container {
    max-width: 800px;
    margin: auto;
}
"""

# Create Gradio interface
with gr.Blocks(css=custom_css, title="Serbian Sentiment Analyzer", theme=gr.themes.Soft()) as demo:
    
    # Header
    gr.Markdown(
        """
        # 🎭 Analiza Sentimenta - Srpski Jezik
        
        Unesite tekst na srpskom jeziku da biste analizirali sentiment.
        
        Model prepoznaje: **Pozitivan** 🟢, **Neutralan** ⚪, **Negativan** 🔴
        
        ---
        """
    )
    
    # Input and output
    with gr.Row():
        with gr.Column():
            text_input = gr.Textbox(
                label="Unesite tekst za analizu",
                placeholder="Npr: Ovaj film je bio fenomenalan!",
                lines=5
            )
            
            analyze_btn = gr.Button("🔍 Analiziraj Sentiment", variant="primary", size="lg")
    
    with gr.Row():
        sentiment_output = gr.HTML(label="Rezultat", elem_id="sentiment-output")
    
    # Connect button to function
    analyze_btn.click(
        fn=analyze_sentiment,
        inputs=text_input,
        outputs=sentiment_output
    )
    
    # Footer
    gr.Markdown(
        """
        ---
        
        ### ℹ️ O Modelu
        
        - **Baziran na**: [dejanseo/BERTic-sentiment](https://huggingface.co/dejanseo/BERTic-sentiment)
        - **Istreniran na**: 4500 srpskih rečenica (1500 po kategoriji)
        - **Klase**: Pozitivan, Neutralan, Negativan
        
        🤖 Napravljeno sa ❤️ koristeći Transformers i Gradio (Iako mi je Streamlit bolji)
        
        ---
        
        ### 🔌 Korišćenje API-ja

        Ovaj Space podržava API koji možete pozivati programski...

        Primer (JavaScript):
        
        ```javascript
        const response = await fetch("https://huggingface.co/spaces/codepanov/serbian-sentiment-analyzer/api/predict", {
            method: "POST",
            headers: { "Content-Type": "application/json" },
            body: JSON.stringify({ data: ["Vaš tekst ovde"] })
        });
        const result = await response.json();
        ```
        """
    )

# Launch the app
if __name__ == "__main__":
    demo.launch()