generate_splits.py

"""
Generate train/val/test splits for the SoloFace2 dataset.

Uses the same stratified split logic as the training pipeline:
  - 80% train, 10% val, 10% test
  - Stratified by column 'p'
  - random_state = 42

Output: splits.csv (image, split)
"""
import pandas as pd
from sklearn.model_selection import train_test_split
from pathlib import Path

LABELS_CSV = Path(r"C:\Users\Bidyut\Desktop\codebase\face-reg\fd-dataset\fd-dataset\labels.csv")
OUT_DIR = Path(r"C:\Users\Bidyut\Desktop\codebase\soloface2")
SEED = 42
TEST_SPLIT = 0.1
VAL_SPLIT = 0.1

print(f"Loading {LABELS_CSV}...")
df = pd.read_csv(LABELS_CSV)
print(f"  Total: {len(df):,} images")
print(f"  Face (p=1): {df['p'].sum():,}")
print(f"  No-face (p=0): {(df['p'] == 0).sum():,}")

# Stratified split: 80/10/10
train_val, test = train_test_split(
    df, test_size=TEST_SPLIT, random_state=SEED, stratify=df['p'],
)
val_frac = VAL_SPLIT / (1.0 - TEST_SPLIT)
train, val = train_test_split(
    train_val, test_size=val_frac, random_state=SEED, stratify=train_val['p'],
)

# Assign split labels
df['split'] = 'train'
df.loc[val.index, 'split'] = 'val'
df.loc[test.index, 'split'] = 'test'

# Save splits.csv
splits_path = OUT_DIR / "splits.csv"
df[['image', 'split']].to_csv(splits_path, index=False)

# Print stats
print(f"\n{'='*50}")
print(f"SPLIT DISTRIBUTION")
print(f"{'='*50}")
for split in ['train', 'val', 'test']:
    sdf = df[df['split'] == split]
    face = (sdf['p'] == 1).sum()
    noface = (sdf['p'] == 0).sum()
    print(f"  {split:5s}: {len(sdf):>8,}  face={face:>8,}  no-face={noface:>8,}  ({len(sdf)/len(df)*100:.1f}%)")

print(f"\n  Total:   {len(df):>8,}")

# Also by source
print(f"\n{'='*50}")
print(f"SOURCE DISTRIBUTION")
print(f"{'='*50}")
for src_prefix, src_name in [
    ('vgg_', 'VGGFace2'),
    ('sf_train_', 'SoloFace train'),
    ('sf_test_', 'SoloFace test'),
    ('sf_val_', 'SoloFace val'),
    ('wf_train_', 'WIDER FACE train'),
    ('wf_val_', 'WIDER FACE val'),
    ('coco_', 'COCO'),
]:
    sdf = df[df['image'].str.startswith(src_prefix)]
    if len(sdf):
        face = (sdf['p'] == 1).sum()
        noface = (sdf['p'] == 0).sum()
        print(f"  {src_name:<20s}: {len(sdf):>8,}  face={face:>8,}  no-face={noface:>8,}")

print(f"\nSplits saved to: {splits_path}")