inference.py

import torch
from transformers import AutoModel, AutoImageProcessor
from model import DinoV3LinearMultiLinear

def load_model(weights_path, device="cuda"):
    """
    Load the pre-trained classifier.
    
    Args:
        weights_path: Path to the saved weights (.pt file)
        device: Device to load model on ('cuda' or 'cpu')
    
    Returns:
        model: Loaded DinoV3LinearMultiLinear model in eval mode
        processor: Image processor for preprocessing input images
    """

    # Load config
    import json
    with open("config.json", "r") as f:
        config = json.load(f)
    # Load backbone
    backbone = AutoModel.from_pretrained(config["model_name"])
    hidden_size = backbone.config.hidden_size
    # Instantiate classifier head
    model = DinoV3LinearMultiLinear(
        backbone=backbone,
        num_classes=config["num_classes"],
        hidden_size=hidden_size,
        freeze_backbone=True
    )
    
    # Load trained weights
    model.load_state_dict(torch.load(weights_path, map_location=device)["model_state_dict"])
    model.to(device)
    model.eval()
    
    # Load image processor
    processor = AutoImageProcessor.from_pretrained(config["model_name"])

    # Load labels
    with open("id2label.json", "r") as f:
        id2label = json.load(f)
    
    return model, processor, id2label
    

def probs_to_labels(probs, id2label):
    """
    Convert probability distribution to labels.
    """
    predicted_indices = probs.argmax(dim=1)
    predicted_labels = [id2label[str(idx.item())] for idx in predicted_indices]
    return predicted_labels