import gradio as gr
import joblib
import numpy as np
from sklearn.feature_extraction.text import CountVectorizer

# Load models (change paths to where your .pkl files are stored)
logreg_model = joblib.load('best_lr_model.pkl')
knn_model = joblib.load('best_knn_model.pkl')
svc_model = joblib.load('best_svc_model.pkl')  # Ensure this model has 'probability=True' if needed
rf_model = joblib.load('best_rf_model.pkl')

# Preprocessing: Load the same vectorizer you used during training
vectorizer = joblib.load('vectorizer.pkl')

# Define the function to predict sentiment
def predict_sentiment(review_text, model_name, positive_threshold=0.6):
    # Preprocess the review text using the same vectorizer used during training
    transformed_text = vectorizer.transform([review_text])  # This is sparse matrix

    # Convert sparse matrix to dense (necessary for SVM and other models)
    transformed_text_dense = transformed_text.toarray()  # Convert sparse matrix to dense array

    # Choose model based on the dropdown selection
    if model_name == "Logistic Regression":
        model = logreg_model
    elif model_name == "K-Nearest Neighbors":
        model = knn_model
    elif model_name == "Random Forest":
        model = rf_model
    elif model_name == "SVM":
        model = svc_model

    # Make predictions and get probabilities
    if hasattr(model, 'predict_proba'):
        prob = model.predict_proba(transformed_text_dense)  # Use dense matrix here
        sentiment_prob = prob[0][1]  # Probability of positive sentiment
        sentiment = model.predict(transformed_text_dense)[0]

        # Apply threshold to determine sentiment class
        if sentiment_prob >= positive_threshold:
            sentiment_class = "Positive"
            probability = f"Positive Comment: {sentiment_prob*100:.2f}%, Negative Comment: {(1-sentiment_prob)*100:.2f}%"
        else:
            sentiment_class = "Negative"
            probability = f"Positive Comment: {sentiment_prob*100:.2f}%, Negative Comment: {(1-sentiment_prob)*100:.2f}%"
    else:
        sentiment = model.predict(transformed_text_dense)[0]
        probability = "Model does not support probability estimation."

    # Return the predicted sentiment class and probabilities
    sentiment_class = "Positive" if sentiment == 1 else "Negative"
    
    return sentiment_class, probability

# Step 6.4: Put all components together in Gradio's interface
with gr.Blocks() as demo:
    # Input components
    review_input = gr.Textbox(label="Review Comment", lines=2, placeholder="Enter your review comment here...")
    model_dropdown = gr.Dropdown(choices=["Logistic Regression", "K-Nearest Neighbors", "Random Forest", "SVM"], label="Select Model")
    
    # Output components
    output_sentiment = gr.Textbox(label="Predicted Sentiment Class", placeholder="Predicted sentiment will appear here...")
    output_probability = gr.Textbox(label="Predicted Probability", placeholder="Probabilities will appear here...")
    
    # Submit Button
    submit_button = gr.Button("Submit")
    
    # Define the action when the button is clicked
    submit_button.click(fn=predict_sentiment, inputs=[review_input, model_dropdown], outputs=[output_sentiment, output_probability])

# Launch the interface
demo.launch()