# Stage 1 Binary Detection — accuracy + recall oriented
# Target: 90–95% overall accuracy with strong cancer sensitivity.

model_name: "efficientnet_b0"
pretrained: true
in_channels: 3
num_classes: 1
img_size: 224
seed: 42

batch_size: 16
learning_rate: 0.0003
weight_decay: 0.0001
epochs: 14
# Train classifier head only for N epochs, then fine-tune full net
freeze_epochs: 2
amp_enabled: true
# Print progress every N train batches (CPU runs are long)
log_every: 25

threshold: 0.35
# FN penalty to push recall higher on ambiguous cancer patches
pos_weight_boost: 1.5
# Clip exploding gradients during full fine-tune
grad_clip: 1.0
# Stop if val accuracy stalls (0 = disabled)
early_stopping_patience: 5
device: "cpu"

data:
  dataset_ids:
    - lc25000
    - patch_camelyon
    - breast_histopathology_patches
    - skin_lesion_hm10000_binary_2k
    - camelyon17_jxie
    - camelyon17_djghosh
  # Increased sample count for multi-organ histology coverage
  max_per_dataset: 4000
  balanced_sampling: true
  apply_stain_norm: false
  num_workers: 0

output:
  artifacts_dir: "models/stage1/artifacts"