next_word_predictor.py

import gradio as gr
import torch
from transformers import GPT2LMHeadModel, GPT2Tokenizer
import torch.nn.functional as F
import spaces

class NextWordPredictor:
    def __init__(self):
        # Load pre-trained GPT-2 model and tokenizer
        self.model_name = "gpt2"
        self.tokenizer = GPT2Tokenizer.from_pretrained(self.model_name)
        self.model = GPT2LMHeadModel.from_pretrained(self.model_name)
        
        # Set padding token
        self.tokenizer.pad_token = self.tokenizer.eos_token

        #make model smaller
        self.model.half()
        
        # Set model to evaluation mode
        self.model.eval()

    @spaces.GPU
    def predict_next_words(self, text, top_k=10):
        """
        Predict the next word given input text
        Returns top_k most likely words with their probabilities and suggested words
        """
        text = text.strip()
        if not text:
            return [], []
        
        # Tokenize input text
        inputs = self.tokenizer.encode(text, return_tensors='pt')
        
        # Get model predictions
        with torch.no_grad():
            outputs = self.model(inputs)
            predictions = outputs.logits[0, -1, :]  # Get last token predictions
        
        # Apply softmax to get probabilities
        probabilities = F.softmax(predictions, dim=-1)
        
        # Get top k predictions
        top_k_probs, top_k_indices = torch.topk(probabilities, top_k)
        
        # Convert to readable format with aligned progress bars
        results = []
        suggested_words = []
        
        # Find the longest word for alignment
        words_with_probs = []
        for prob, idx in zip(top_k_probs, top_k_indices):
            word = self.tokenizer.decode(idx.item()).strip()
            probability = prob.item()
            percentage = probability * 100
            words_with_probs.append((word, probability, percentage))
        
        # Find max word length for alignment
        max_word_length = max(len(word) for word, _, _ in words_with_probs)
        
        for word, probability, percentage in words_with_probs:
            # Create aligned progress bar with better blocks
            bar_length = 20
            filled_length = int(bar_length * probability)
            bar = '█' * filled_length + '▢' * (bar_length - filled_length)
            
            # Align everything properly
            word_padded = word.ljust(max_word_length)
            result = f"{word_padded} | {probability:.4f} ({percentage:5.2f}%) {bar}"
            
            results.append(result)
            suggested_words.append(word)
        
        return results, suggested_words

# Initialize the predictor
predictor = NextWordPredictor()

def update_predictions(text):
    """Update predictions based on current text"""
    predictions_list, suggested_words = predictor.predict_next_words(text)
    
    if not predictions_list:
        return [gr.update(visible=False, interactive=True)] * 10
    
    # Update buttons with predictions, hide unused ones, enable all buttons
    updates = []
    for i in range(10):
        if i < len(predictions_list):
            updates.append(gr.update(value=predictions_list[i], visible=True, interactive=True))
        else:
            updates.append(gr.update(visible=False, interactive=True))
    
    return updates

def disable_all_buttons():
    """Disable all prediction buttons"""
    return [gr.update(interactive=False)] * 10

def add_word_to_text(current_text, button_value):
    """Extract word from button and add to text"""
    if not button_value:
        return current_text
    
    # Extract the word (everything before the first "|")
    word = button_value.split(" | ")[0].strip()
    
    if not current_text.strip():
        return word
    # Add space if text doesn't end with space
    if current_text.endswith(' '):
        return current_text + word
    else:
        return current_text + ' ' + word

# Create Gradio interface
with gr.Blocks(title="Next Word Predictor", theme=gr.themes.Soft()) as demo:
    gr.Markdown("# Next Word Predictor")
    gr.Markdown("Type a sentence and see the top 10 most likely next words with their probabilities! **Click on any prediction to add that word to your text.**")
    
    with gr.Row():
        text_input = gr.Textbox(
            label="Enter your text",
            placeholder="Start typing a sentence...",
            lines=4,
            interactive=True
        )
    with gr.Row():
        # Examples
        gr.Examples(
            examples=[
                ["The weather today is"],
                ["I love to eat"],
                ["Machine learning is"],
                ["The quick brown fox"],
                # ["In the future, we will"]
            ],
            inputs=text_input
        )
        
    with gr.Row(): 
        gr.Markdown("### Top 10 Next Word Predictions")
        gr.Markdown("*Click any prediction below to add it to your text*")
        
    # Create 10 clickable buttons for predictions
    prediction_buttons = []
    for i in range(10):
        with gr.Row():
            btn = gr.Button(
                value="",
                visible=False,
                variant="secondary",
                size="sm",
                interactive=True
            )
            prediction_buttons.append(btn)

    # Update predictions as user types
    text_input.change(
        fn=update_predictions,
        inputs=text_input,
        outputs=prediction_buttons
    )

    # Add click handlers for each prediction button
    for btn in prediction_buttons:
        # First disable all buttons when any button is clicked
        btn.click(
            fn=disable_all_buttons,
            inputs=[],
            outputs=prediction_buttons
        ).then(
            # Then add the word to text
            fn=add_word_to_text,
            inputs=[text_input, btn],
            outputs=text_input
        ).then(
            # Finally update predictions (which will re-enable buttons)
            fn=update_predictions,
            inputs=text_input,
            outputs=prediction_buttons
        )
    

    
    # gr.Markdown("### How it works:")
    # gr.Markdown("""
    # - Uses GPT-2 language model to predict next words
    # - Applies softmax to convert logits to probabilities  
    # - Shows top 10 most likely words with percentages and aligned visual bars
    # - Updates predictions in real-time as you type
    # - **Click on any prediction button to add that word to your text automatically**
    # - **Buttons are disabled while processing to prevent multiple clicks**
    # - Progress bars show relative probability: █ = filled, ▢ = empty outline
    # - All bars are perfectly aligned for easy comparison
    # """)

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