Refactor/data.py

[agentksimha]

Summary

Three targeted improvements to ArtExtract_Soyoung/utils/data.py:

1. File path validation

• Validates all directories (images, masks, sub-dirs) at init time with clear, actionable error messages
• Detects unpaired images (no matching masks) before training starts
• Runtime checks in __getitem__ catch deleted files or broken symlinks mid-run
• Replaces brittle string concatenation with os.path.join

2. Fix double-normalisation on masks

• ToTensor() already scales uint8 PIL images to [0, 1]; the previous / 255.0 further compressed values to [0, ~0.004], effectively zeroing all masks
• Removed the redundant division — masks now normalise exactly once

3. Reproducible DataLoader with num_workers

• num_workers = min(4, os.cpu_count()) overlaps I/O with GPU compute
• persistent_workers=True avoids re-spawning workers each epoch
• Seeded torch.Generator makes shuffle order deterministic across runs

No changes to model architecture, training loop, or public API.