convert_best2019.py

import os
import pandas as pd
from PIL import Image
import codecs
import numpy as np
import glob
import io

def create_dataset():
    print("Starting dataset creation...")
    
    # Create output directory for images
    os.makedirs("processed_images", exist_ok=True)
    
    # Find all label files matching pattern
    print("Finding label files...")
    label_files = glob.glob("./original/test/round3?_best2019.label")
    print(f"Found {len(label_files)} label files")
    
    # Parse image filenames and captions
    data = []
    for i, label_path in enumerate(label_files):
        print(f"\nProcessing label file {i+1}/{len(label_files)}: {label_path}")
        
        # Read label file with Windows-874 encoding
        print("Reading label file...")
        with codecs.open(label_path, 'r', encoding='cp874') as f:
            lines = f.readlines()
        print(f"Found {len(lines)} entries")
        
        for j, line in enumerate(lines):
            if j % 100 == 0:
                print(f"Processing entry {j}/{len(lines)}")
                
            filename, caption = line.strip().split(' ', 1)
            
            # Get folder name from first 3 chars of filename
            folder = filename[:3]
            image_path = f"./original/test/{folder}/{filename}"
            # Load and verify image exists
            if os.path.exists(image_path):
                try:
                    # Load image and convert to bytes
                    img = Image.open(image_path)
                    img_byte_arr = io.BytesIO()
                    img.save(img_byte_arr, format='PNG')  # Force PNG format
                    img_bytes = {"bytes":bytearray(img_byte_arr.getvalue())}

                    data.append({
                        'image': img_bytes,  # Store as numpy array of bytes
                        'text': caption,
                        'label_file': os.path.basename(label_path)
                    })
                except Exception as e:
                    print(f"Error processing image {image_path}: {e}")
    print(f"\nProcessed {len(data)} total images successfully")
    
    # Convert to dataframe and save as parquet
    print("Converting to dataframe...")
    df = pd.DataFrame(data)
    print("Saving to parquet file...")
    df.to_parquet("train-0000.parquet", index=False)
    print("Dataset creation complete!")

if __name__ == "__main__":
    create_dataset()