FADE-Net

FADE-Net is a lightweight facial age-estimation model for AFAD. The active training path is MobileNetV4-Conv-Small with Distribution-Conditioned Scale Routing (DCSR) and Correction-Need Guided Bounded Residual Refinement (CGBR).

Active code path

The repository keeps one active source path:

src/
├── train_fade_net.py   # training, validation, checkpointing, EMA
├── fade_net.py         # FADE-Net model
├── dcsr_cgbr.py        # DCSR, CGBR, adapters, FADELoss
├── config.py           # model and training configuration
├── backbones.py        # timm backbone adapter
└── experiment.py       # shared configuration metadata helpers

The previous train.py/model.py/dataset.py/evaluation.py/utils.py stack has been removed from the active source tree.

Model and protocol

• Backbone: timm mobilenetv4_conv_small, pretrained
• Age range: 15–40, 26 classes
• Input: RGB 256×256
• Multi-scale features: 32/96/960 channels from shallow, middle and deep stages
• Head: coarse age distribution → DCSR → main age distribution → CGBR refinement
• Dataset: AFAD, identity-disjoint folds
• Optimizer: AdamW with backbone/head differential learning rates
• Scheduler: cosine annealing
• EMA: updated after every optimizer step; model buffers are synchronized

Training

Install dependencies first:

pip install -r requirements.txt

For one fold:

python src/train_fade_net.py \
  --afad_dir datasets/AFAD \
  --split_dir . \
  --split_id 0 \
  --output_dir outputs/fade_net_ema_fix

The split directory must contain files named dataset_split_AFAD_15_40_iddisjoint_fold0.json through fold4.json. The server five-fold launcher is scripts/train_fade_net.sh.

Outputs

Checkpoints and runtime logs are experiment artifacts and should not be committed to Git. Keep paper result summaries and split metadata under version control when they are part of the evidence chain.

Documentation

• docs/architecture_review.md: architecture and implementation review
• docs/paper_result_summary.md: historical result summary
• docs/dataset_setup.md: AFAD setup notes