split_dataset.py

"""
Script to split the lines_dataset into train/validation/test splits.
- 80% train
- 10% validation  
- 10% test

The split is done by source image to prevent data leakage - all crops from the
same source image go into the same split.
"""

import json
import shutil
import random
from pathlib import Path
from collections import defaultdict

# Configuration
RANDOM_SEED = 42
TRAIN_RATIO = 0.8
VAL_RATIO = 0.1
TEST_RATIO = 0.1

BASE_DIR = Path(__file__).parent
CURRENT_TRAIN_DIR = BASE_DIR / "train"
METADATA_FILE = CURRENT_TRAIN_DIR / "metadata.jsonl"


def load_metadata():
    """Load all metadata from the jsonl file."""
    if not METADATA_FILE.exists():
        print(f"Metadata file not found: {METADATA_FILE}")
        return []
    
    metadata = []
    with open(METADATA_FILE, 'r') as f:
        for line in f:
            if line.strip():
                metadata.append(json.loads(line))
    return metadata


def group_by_source(metadata):
    """Group samples by their source image index."""
    groups = defaultdict(list)
    for item in metadata:
        source_idx = item.get("source_image_idx", 0)
        groups[source_idx].append(item)
    return groups


def split_sources(source_indices, train_ratio, val_ratio, test_ratio):
    """Split source indices into train/val/test sets."""
    random.shuffle(source_indices)
    
    n = len(source_indices)
    n_train = int(n * train_ratio)
    n_val = int(n * val_ratio)
    
    train_sources = source_indices[:n_train]
    val_sources = source_indices[n_train:n_train + n_val]
    test_sources = source_indices[n_train + n_val:]
    
    return train_sources, val_sources, test_sources


def create_split_directory(split_name, samples, base_dir, source_dir):
    """Create a split directory with images and metadata."""
    split_dir = base_dir / split_name
    split_dir.mkdir(parents=True, exist_ok=True)
    
    # Copy images and prepare metadata
    metadata_entries = []
    
    for sample in samples:
        file_name = sample["file_name"]
        src_path = source_dir / file_name
        dst_path = split_dir / file_name
        
        if src_path.exists():
            shutil.copy2(src_path, dst_path)
            metadata_entries.append(sample)
        else:
            print(f"Warning: Image not found: {src_path}")
    
    # Write metadata
    metadata_path = split_dir / "metadata.jsonl"
    with open(metadata_path, 'w') as f:
        for entry in metadata_entries:
            f.write(json.dumps(entry) + '\n')
    
    return len(metadata_entries)


def main():
    random.seed(RANDOM_SEED)
    
    print("Loading metadata...")
    metadata = load_metadata()
    print(f"Total samples: {len(metadata)}")
    
    # Group by source image
    print("Grouping by source image...")
    groups = group_by_source(metadata)
    source_indices = list(groups.keys())
    print(f"Total source images: {len(source_indices)}")
    
    # Split source indices
    print(f"\nSplitting sources ({TRAIN_RATIO:.0%} train, {VAL_RATIO:.0%} val, {TEST_RATIO:.0%} test)...")
    train_sources, val_sources, test_sources = split_sources(
        source_indices, TRAIN_RATIO, VAL_RATIO, TEST_RATIO
    )
    
    print(f"  Train sources: {len(train_sources)}")
    print(f"  Validation sources: {len(val_sources)}")
    print(f"  Test sources: {len(test_sources)}")
    
    # Gather samples for each split
    train_samples = []
    val_samples = []
    test_samples = []
    
    for src_idx in train_sources:
        train_samples.extend(groups[src_idx])
    for src_idx in val_sources:
        val_samples.extend(groups[src_idx])
    for src_idx in test_sources:
        test_samples.extend(groups[src_idx])
    
    print(f"\nSamples per split:")
    print(f"  Train: {len(train_samples)} ({len(train_samples)/len(metadata)*100:.1f}%)")
    print(f"  Validation: {len(val_samples)} ({len(val_samples)/len(metadata)*100:.1f}%)")
    print(f"  Test: {len(test_samples)} ({len(test_samples)/len(metadata)*100:.1f}%)")
    
    # Create temporary directory for new structure
    temp_dir = BASE_DIR / "_temp_splits"
    temp_dir.mkdir(exist_ok=True)
    
    print("\nCreating split directories...")
    
    # Create each split
    n_train = create_split_directory("train", train_samples, temp_dir, CURRENT_TRAIN_DIR)
    print(f"  Created train split: {n_train} samples")
    
    n_val = create_split_directory("validation", val_samples, temp_dir, CURRENT_TRAIN_DIR)
    print(f"  Created validation split: {n_val} samples")
    
    n_test = create_split_directory("test", test_samples, temp_dir, CURRENT_TRAIN_DIR)
    print(f"  Created test split: {n_test} samples")
    
    # Remove old train directory and move new splits
    print("\nReorganizing directory structure...")
    
    # Remove old train directory
    shutil.rmtree(CURRENT_TRAIN_DIR)
    
    # Move new splits from temp to base
    for split_name in ["train", "validation", "test"]:
        src = temp_dir / split_name
        dst = BASE_DIR / split_name
        shutil.move(str(src), str(dst))
    
    # Remove temp directory
    temp_dir.rmdir()
    
    print("\nDone! New directory structure:")
    print(f"  {BASE_DIR}/train/ ({n_train} samples)")
    print(f"  {BASE_DIR}/validation/ ({n_val} samples)")
    print(f"  {BASE_DIR}/test/ ({n_test} samples)")
    
    # Count total lines per split
    def count_lines(samples):
        return sum(len(s.get("lines", {}).get("segments", [])) for s in samples)
    
    print(f"\nTotal lines per split:")
    print(f"  Train: {count_lines(train_samples)}")
    print(f"  Validation: {count_lines(val_samples)}")
    print(f"  Test: {count_lines(test_samples)}")


if __name__ == "__main__":
    main()