inference.py

import torch
from transformers import SwinForImageClassification, AutoImageProcessor
from PIL import Image
import joblib
import numpy as np
import matplotlib.pyplot as plt
import cv2
import os
from pathlib import Path

class CoinPredictor:
    def __init__(self, model_dir='model_checkpoints', top_n=10):
        """
        Initialize the predictor with trained model and necessary components.
        
        Args:
            model_dir (str): Directory containing the saved model and label encoder
            top_n (int): Number of top predictions to return
        """
        self.top_n = top_n
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.crop_percentage = 0.15  # 15% crop from each side
        
        # Load the image processor
        self.image_processor = AutoImageProcessor.from_pretrained("microsoft/swin-tiny-patch4-window7-224")
        
        # Load the trained model
        model_path = os.path.join(model_dir, 'best_model')
        self.model = SwinForImageClassification.from_pretrained(model_path)
        self.model.to(self.device)
        self.model.eval()
        
        # Load the label encoder
        encoder_path = os.path.join(model_dir, 'label_encoder.joblib')
        self.label_encoder = joblib.load(encoder_path)
        
        print(f"Model loaded and running on {self.device}")
    
    def crop_center(self, image):
        """
        Crop the center portion of the image.
        """
        h, w = image.shape[:2]
        crop_h = int(h * self.crop_percentage)
        crop_w = int(w * self.crop_percentage)
        
        return image[crop_h:h-crop_h, crop_w:w-crop_w]
    
    def preprocess_image(self, image_path):
        """
        Preprocess a single image for prediction.
        """
        # Read image
        image = cv2.imread(image_path)
        if image is None:
            raise ValueError(f"Could not load image: {image_path}")
        
        # Crop center
        image = self.crop_center(image)
        
        # Convert to RGB (from BGR)
        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
        
        # Convert to PIL Image
        image = Image.fromarray(image)
        
        return image
    
    def predict(self, image_path):
        """
        Make prediction for a single image.
        
        Args:
            image_path (str): Path to the image file
            
        Returns:
            list of tuples: (label, probability) for top N predictions
        """
        # Preprocess image
        image = self.preprocess_image(image_path)
        
        # Prepare image for model
        inputs = self.image_processor(image, return_tensors="pt")
        inputs = {k: v.to(self.device) for k, v in inputs.items()}
        
        # Get predictions
        with torch.no_grad():
            outputs = self.model(**inputs)
            probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)
        
        # Get top N predictions
        top_probs, top_indices = torch.topk(probabilities[0], self.top_n)
        
        # Convert to labels and probabilities
        predictions = []
        for prob, idx in zip(top_probs.cpu().numpy(), top_indices.cpu().numpy()):
            label = self.label_encoder.inverse_transform([idx])[0]
            predictions.append((label, float(prob)))
        
        return predictions, image
    
    def visualize_prediction(self, image_path, predictions, reference_dir="all_coins_cropped"):
        """
        Visualize the input image and top N matching reference images with probabilities.
        
        Args:
            image_path (str): Path to the query image
            predictions (tuple): (predictions, preprocessed_image)
            reference_dir (str): Directory containing reference images
        """
        predictions, processed_image = predictions
        
        # Calculate grid size
        n_cols = 4  # 4 images per row
        n_rows = (self.top_n + 3) // n_cols  # +3 for query images and ceiling division
        
        # Create figure
        fig = plt.figure(figsize=(15, 4 * n_rows))
        
        # Plot original query image
        plt.subplot(n_rows, n_cols, 1)
        original_img = cv2.imread(image_path)
        original_img = cv2.cvtColor(original_img, cv2.COLOR_BGR2RGB)
        plt.imshow(original_img)
        plt.title("Original Query")
        plt.axis('off')
        
        # Plot processed query image
        plt.subplot(n_rows, n_cols, 2)
        plt.imshow(processed_image)
        plt.title("Processed Query")
        plt.axis('off')
        
        # Plot top N predictions with their reference images
        for i, (label, prob) in enumerate(predictions, 3):
            # Find reference image
            ref_path = None
            for ext in ['.jpg', '.jpeg', '.png']:
                test_path = os.path.join(reference_dir, label + ext)
                if os.path.exists(test_path):
                    ref_path = test_path
                    break
            
            if ref_path:
                plt.subplot(n_rows, n_cols, i)
                ref_img = cv2.imread(ref_path)
                ref_img = cv2.cvtColor(ref_img, cv2.COLOR_BGR2RGB)
                plt.imshow(ref_img)
                plt.title(f"{label}\n{prob:.1%}")
                plt.axis('off')
        
        plt.tight_layout()
        plt.show()

def main():
    # Initialize predictor
    predictor = CoinPredictor()
    
    # Get image path from user
    image_path = input("Enter the path to the coin image: ")
    
    try:
        # Make prediction
        predictions = predictor.predict(image_path)
        
        # Print results
        print("\nPredictions:")
        for i, (label, prob) in enumerate(predictions[0], 1):
            print(f"{i}. {label}: {prob:.1%}")
        
        # Visualize results
        predictor.visualize_prediction(image_path, predictions)
        
    except Exception as e:
        print(f"Error: {e}")

if __name__ == "__main__":
    main()