prepare_data.py

import os
import numpy as np
import pickle
from imutils import paths
from sklearn.preprocessing import LabelEncoder
import pandas as pd
import logging
import sys
from PIL import Image

# Set up logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)

img_height, img_width = 28, 28

def get_data(imagePaths, csv_path, verbose=245):
    # Load the CSV mapping
    try:
        df = pd.read_csv(csv_path, encoding='utf-8')
        char_to_class = dict(zip(df['char'], df['class']))
        logger.info(f"Loaded CSV with {len(char_to_class)} character mappings")
        
    except Exception as e:
        logger.error(f"Error loading CSV: {e}")
        return None, None
    
    # initialize the list of features and labels
    data = []
    labels = []
    failed_images = 0
    successful_images = 0

    # loop over the input images
    for (i, imagePath) in enumerate(imagePaths):
        try:
            # Check if file exists and is readable
            if not os.path.exists(imagePath):
                failed_images += 1
                continue
            
            # Extract the folder name (character) from the path
            path_parts = imagePath.split(os.path.sep)
            if len(path_parts) < 2:
                failed_images += 1
                continue
                
            folder_name = path_parts[-2]  # Second to last part is folder name
            
            # Check if this character is in our mapping
            if folder_name not in char_to_class:
                failed_images += 1
                if failed_images <= 5:
                    logger.warning(f"Unmapped character: '{folder_name}' in path: {imagePath}")
                continue
            
            # Load and preprocess the image using PIL
            try:
                with Image.open(imagePath) as img:
                    # Convert to grayscale
                    if img.mode != 'L':
                        img = img.convert('L')
                    
                    # Resize to target dimensions
                    img = img.resize((img_width, img_height), Image.Resampling.LANCZOS)
                    
                    # Convert to numpy array and normalize
                    image_array = np.array(img, dtype=np.float32) / 255.0
                    
            except Exception as e:
                logger.debug(f"PIL failed to read {imagePath}: {e}")
                failed_images += 1
                continue
            
            # Get the class label from mapping
            label = char_to_class[folder_name]
            
            # Add to our dataset
            data.append(image_array)
            labels.append(label)
            successful_images += 1

            # show an update every 'verbose' images
            if verbose > 0 and successful_images > 0 and (successful_images) % verbose == 0:
                logger.info(f"Processed {successful_images} images successfully")

        except Exception as e:
            failed_images += 1
            if failed_images <= 5:  # Log first few errors
                logger.error(f"Error processing image {imagePath}: {e}")
            continue

    logger.info(f"Successfully processed {successful_images} images")
    logger.info(f"Failed to process {failed_images} images")
    
    if successful_images == 0:
        logger.error("No images were successfully processed!")
        return None, None

    # Convert to numpy arrays
    data = np.array(data)
    labels = np.array(labels)
    
    # Reshape data to add channel dimension (grayscale)
    data = data.reshape((data.shape[0], img_height, img_width, 1))
    
    # show some information on memory consumption of the images
    logger.info(f"Features matrix: {data.nbytes / (1024 * 1000.0):.1f}MB")
    logger.info(f"Number of classes: {len(np.unique(labels))}")
    logger.info(f"Total samples: {len(data)}")

    return data, labels

if __name__ == "__main__":
    # Get image paths
    dataset_path = './dataset_ka_kha'
    if not os.path.exists(dataset_path):
        logger.error(f"Dataset path {dataset_path} does not exist!")
        sys.exit(1)
        
    imagePaths = list(paths.list_images(dataset_path))
    logger.info(f"Found {len(imagePaths)} images in dataset")
    
    if len(imagePaths) == 0:
        logger.error("No images found in the dataset directory!")
        sys.exit(1)
    
    # Process the data using PIL instead of OpenCV
    data, labels = get_data(imagePaths, 'sample.csv', 20000)  # Reduced verbose for more frequent updates
    
    if data is None or labels is None:
        logger.error("Failed to prepare data!")
        sys.exit(1)
    
    # Save both data and labels
    os.makedirs('dataset_pickles', exist_ok=True)
    with open('dataset_pickles/tigrigna_dataset.pickle', 'wb') as f:
        pickle.dump((data, labels), f)
    
    logger.info("Data preparation complete. Saved to dataset_pickles/tigrigna_dataset.pickle")