handler.py

from transformers import BertForTokenClassification, BertTokenizer, AutoConfig
import torch
from typing import Dict, List, Any

class EndpointHandler:
    def __init__(self, path: str = "dejanseo/LinkBERT"):
        # Load the configuration from the saved model
        self.config = AutoConfig.from_pretrained(path)

        # Make sure to specify the correct model name for bert-large-cased
        # Adjust num_labels according to your model's configuration
        self.model = BertForTokenClassification.from_pretrained(
            path,
            config=self.config
        )
        self.model.eval()  # Set model to evaluation mode

        # Load the tokenizer for bert-large-cased
        self.tokenizer = BertTokenizer.from_pretrained("bert-large-cased")

    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
        # Extract input text from the request
        inputs = data.get("inputs", "")

        # Tokenize the inputs
        inputs_tensor = self.tokenizer(inputs, return_tensors="pt", add_special_tokens=True)
        input_ids = inputs_tensor["input_ids"]

        # Run the model
        with torch.no_grad():
            outputs = self.model(input_ids)
            predictions = torch.argmax(outputs.logits, dim=-1)

        # Process the predictions to generate readable output
        tokens = self.tokenizer.convert_ids_to_tokens(input_ids[0])[1:-1]  # Exclude CLS and SEP tokens
        predictions = predictions[0][1:-1].tolist()

        # Reconstruct the text with annotations for token classification
        result = []
        for token, pred in zip(tokens, predictions):
            if pred == 1:  # Adjust this based on your classification needs
                result.append(f"<u>{token}</u>")
            else:
                result.append(token)

        reconstructed_text = " ".join(result).replace(" ##", "")
        
        # Return the processed text in a structured format
        return [{"text": reconstructed_text}]

# Note: Ensure the path "dejanseo/LinkBERT" is correctly pointing to your model's location
# If the model is locally saved, adjust the path accordingly