Table 1 Training parameters for FADiff.
From: Adaptive diffusion models for overcoming data scarcity in long-distance face recognition
Parameter | Value |
|---|---|
Optimizer | AdamW |
FCEM initial learning rate | 0.0001 |
FAIE initial learning rate | 0.0002 |
DM learning rate | 0.0001 |
Batch size | 64 |
Total training epochs | 250 (FCEM), 300 (FAIE), 400 (FADiff) |
Diffusion steps (T) | 1000 |
Weight decay | 1e-4 |
Gradient clipping norm | 1.0 |
Learning rate scheduler | Cosine Annealing |
Loss functions | ArcFace Loss (FCEM), L1 (FAIE), MSE (FADiff) |
Face image resolution | 128 × 128,128 \times 128,128 × 128 |
Mixed precision | Enabled (AMP) |
Hardware | 4 × NVIDIA A100 GPUs (40 GB) |
