app.py

from transformers import AutoTokenizer, AutoModelForMaskedLM
import torch
import gradio as gr
import numpy as np
import json

class BertEmbeddingsGenerator:
    def __init__(self, model_name="tahrirchi/tahrirchi-bert-base"):
        """Initialize the BERT model and tokenizer."""
        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
        self.model = AutoModelForMaskedLM.from_pretrained(model_name)
        self.model.eval()  # Set to evaluation mode

    def get_embeddings(self, text):
        """
        Generate embeddings for the input text.
        
        Args:
            text (str): Input text to embed
            
        Returns:
            np.ndarray: Text embeddings
        """
        # Tokenize input text
        inputs = self.tokenizer(
            text,
            return_tensors="pt",
            truncation=True,
            padding=True,
            max_length=512
        )
        
        # Generate embeddings
        with torch.no_grad():
            outputs = self.model(**inputs, output_hidden_states=True)
            
        # Get the hidden states from the last layer
        # The hidden states tuple contains embeddings from all layers, -1 gets the last layer
        last_hidden_state = outputs.hidden_states[-1]
            
        # Average token embeddings to get sentence embedding
        embeddings = last_hidden_state.mean(dim=1)
        
        # Convert to numpy and then to list
        return embeddings.squeeze().cpu().numpy()

def create_gradio_interface():
    """Create and configure the Gradio interface."""
    # Initialize the embeddings generator
    generator = BertEmbeddingsGenerator()
    
    def embed_text(input_text):
        """Gradio interface function."""
        try:
            if not input_text or not input_text.strip():
                return json.dumps({"error": "Matn kiritilmadi"})
                
            embeddings = generator.get_embeddings(input_text)
            
            # Convert numpy array to list and handle NaN/Infinity values
            embeddings_list = np.where(np.isfinite(embeddings), embeddings, None).tolist()
            
            # Create a structured output
            output = {
                "embeddings": embeddings_list,
                "dimensions": len(embeddings_list),
                "status": "success"
            }
            
            return json.dumps(output, ensure_ascii=False)
            
        except Exception as e:
            return json.dumps({
                "error": str(e),
                "status": "error"
            })
    
    # Create Gradio interface
    iface = gr.Interface(
        fn=embed_text,
        inputs=gr.Textbox(
            lines=2,
            placeholder="Matn kiriting...",
            label="Input Text"
        ),
        outputs=gr.JSON(label="Embeddings"),
        title="O'zbek tili uchun embedding",
        description="O'zbek tili uchun Tahrirchi BERT Base modeli orqali embedding generatsiya qilish",
        examples=[
            ["Assalomu alaykum, men o'zbek tili bilan ishlayman"],
            ["O'zbek tili uchun Tahrirchi BERT Base modeli orqali embedding generatsiya qilish uchun namuna matn."]
        ]
    )
    return iface

if __name__ == "__main__":
    # Create and launch the interface
    iface = create_gradio_interface()
    iface.launch()