v2.0.0 - EfficientNetV2-RW-S (82K images)

SkinAge Model v2.0.0

Architecture

• EfficientNetV2-RW-S backbone (25.8M params)
• UNet decoder (4-channel heatmaps)
• Quality head (1792→512→28) + Age head (1792→256→1)

Training

• 82,394 images (3.7K original + 33K UTKFace + 46K FairFace)
• Phase 1: 2 epochs frozen backbone (LR=1e-3)
• Phase 2: 2 epochs full fine-tune (LR=3e-5, cosine annealing)
• Mixed precision (AMP) on RTX 4060

Changes from v1

• Upgraded from EfficientNet-B2 to EfficientNetV2-RW-S
• 22x more training data (3.7K → 82K images)
• On-the-fly heatmap generation (texture-map pseudo-labels)
• Improved quality score differentiation across zones

v1.0.0 - Trained Model

EfficientNet-B2 SkinAge model checkpoint (13 epochs, val_loss 3.09, age MAE ~2 years)