app.py

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import shap
import torch
import gradio as gr
import matplotlib.pyplot as plt
import numpy as np

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("nlptown/bert-base-multilingual-uncased-sentiment")
model = AutoModelForSequenceClassification.from_pretrained("nlptown/bert-base-multilingual-uncased-sentiment")

# Define prediction function
def predict(texts):
    processed_texts = []
    for text in texts:
        if isinstance(text, list):
            processed_text = tokenizer.convert_tokens_to_string(text)
        else:
            processed_text = text
        processed_texts.append(processed_text)
    
    inputs = tokenizer(
        processed_texts, 
        return_tensors="pt", 
        padding=True, 
        truncation=True, 
        max_length=512,
        add_special_tokens=True
    )
    
    with torch.no_grad():
        outputs = model(**inputs)
    
    probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)
    return probabilities.numpy()

# Initialize SHAP components
output_names_list = [model.config.id2label[i] for i in range(len(model.config.id2label))]
masker = shap.maskers.Text(tokenizer=tokenizer, mask_token=tokenizer.mask_token, collapse_mask_token=True)
explainer = shap.Explainer(model=predict, masker=masker, output_names=output_names_list)

def analyze_text(text):
    # Get predictions
    probabilities = predict([text])[0]
    predicted_class = np.argmax(probabilities)
    predicted_label = model.config.id2label[predicted_class]
    
    # Generate SHAP explanations
    shap_values = explainer([text])
    
    # Create HTML visualizations for all classes
    html_plots = []
    for i in range(shap_values.shape[-1]):
        # Create SHAP text plot and convert to HTML
        plot_html = shap.plots.text(shap_values[0, :, i], display=False)
        html_plots.append(plot_html)
    
    # Format confidence scores
    confidence_scores = {model.config.id2label[i]: float(probabilities[i]) 
                        for i in range(len(probabilities))}
    
    return (predicted_label, 
            confidence_scores, 
            *html_plots)

# Create Gradio interface with HTML components
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("## 🔍 BERT Sentiment Analysis with SHAP Explanations")
    
    with gr.Row():
        input_text = gr.Textbox(label="Input Text", placeholder="Enter text to analyze...")
        
    with gr.Row():
        predict_btn = gr.Button("Analyze Sentiment")
        
    with gr.Row():
        label_output = gr.Label(label="Predicted Sentiment")
        prob_output = gr.Label(label="Confidence Scores")
        
    with gr.Row():
        gr.Markdown("""
        ### SHAP Explanations
        Below you can see how each word contributes to different sentiment scores (1-5 stars).  
        Red text increases the score, blue decreases it.
        """)
    
    # Individual Explanation Rows
    plot_components = []
    for i in range(5):
        with gr.Row():
            plot_components.append(
                gr.HTML(
                    label=f"Explanation for {model.config.id2label[i]}",
                    elem_classes=f"shap-plot-{i+1}"
                )
            )

    predict_btn.click(
        fn=analyze_text,
        inputs=input_text,
        outputs=[label_output, prob_output] + plot_components
    )
    
    examples = gr.Examples(
        examples=[
            ["This product exceeded all my expectations!"],
            ["Terrible customer service experience."],
            ["The movie was okay, nothing special."],
            ["You are kinda cool"],
        ],
        inputs=input_text
    )

if __name__ == "__main__":
    demo.launch(debug = True)