model.py

import torch
import torchvision.transforms as T
from PIL import Image
import joblib
import json

# Define image transformation
transform_image = T.Compose([
    T.Resize(224),
    T.CenterCrop(224),
    T.ToTensor(),
    T.Normalize([0.5], [0.5])
])

def load_image(img_path: str) -> torch.Tensor:
    """
    Load an image and return a tensor that can be used as an input to the model.
    """
    img = Image.open(img_path).convert("RGB")
    transformed_img = transform_image(img).unsqueeze(0)
    return transformed_img

# Load DINOv2 model for feature extraction
dinov2_model = torch.hub.load('facebookresearch/dino:main', 'dino_vits16')
device = torch.device('cuda' if torch.cuda.is_available() else "cpu")
dinov2_model.to(device)
dinov2_model.eval()  # Set the model to evaluation mode

# Load the pre-trained SVM classifier
clf = joblib.load('svm_model.joblib')

# Function to predict the class for a new image
def predict(image_path: str):
    # Load and transform the image
    new_image = load_image(image_path).to(device)
    
    # Extract features using DINOv2
    with torch.no_grad():
        features = dinov2_model(new_image)
    
    # Flatten features to 2D for SVM input
    embedding = features.cpu().numpy().reshape(1, -1)
    
    # Predict the class using the SVM classifier
    prediction = clf.predict(embedding)
    
    return prediction[0]

# If running as a script
if __name__ == "__main__":
    import sys
    image_path = sys.argv[1]  # Get image path from command line arguments
    predicted_class = predict(image_path)
    print("Predicted class:", predicted_class)