custom_tools/liveness_detection_tool.py

from agentlego.tools import BaseTool
from PIL import Image
import torch
import tempfile
import cv2
import os

class LivenessDetectionTool(BaseTool):
    default_desc = 'Detects liveness in an image using a DinoV2 image classification model.'
    
    def __init__(self):
        super().__init__()
        # Move model loading inside the class initialization
        from transformers import AutoImageProcessor, AutoModelForImageClassification
        self.processor = AutoImageProcessor.from_pretrained("nguyenkhoa/dinov2_Liveness_detection_v2.2.3")
        self.model = AutoModelForImageClassification.from_pretrained("nguyenkhoa/dinov2_Liveness_detection_v2.2.3")
    
    def apply(self, image_path: str) -> str:
        try:
            # Load image
            image = Image.open(image_path).convert("RGB")

            # Preprocess and infer
            inputs = self.processor(images=image, return_tensors="pt")
            with torch.no_grad():
                outputs = self.model(**inputs)

            logits = outputs.logits
            probs = torch.nn.functional.softmax(logits, dim=-1)[0]

            # Get prediction
            predicted_class_idx = torch.argmax(probs).item()
            predicted_label = self.model.config.id2label[predicted_class_idx]
            confidence = round(probs[predicted_class_idx].item(), 4)

            # Format result
            result = f"Liveness: {predicted_label} (Confidence: {confidence})"
            return result

        except Exception as e:
            return f"Error during liveness detection: {str(e)}"