ml/configs/vit.yaml

# Vision Transformer (ViT) Transfer Learning Configuration
# Inherits from base.yaml

model:
  name: "vit"
  architecture:
    backbone: "vit_base_patch16_224"
    pretrained: true
    freeze_backbone: true
    unfreeze_after_epochs: 8
    unfreeze_layers: ["blocks.10", "blocks.11", "head", "norm"]

training:
  batch_size: 16               # Smaller batch due to memory
  num_epochs: 25
  learning_rate: 0.0005
  fine_tune_lr: 0.00005
  weight_decay: 0.01
  optimizer: "adamw"
  scheduler: "cosine"
  warmup_epochs: 3
  early_stopping_patience: 7
  label_smoothing: 0.1

image:
  size: 224                     # ViT-B/16 uses 224x224