# ============================================================
# Malignancy Classifier — Training Configuration
# ============================================================
# 3D ResNet classifier for nodule malignancy (demo feature only)
# NOT part of the research paper.

experiment_name: "malignancy_classifier"
description: "3D ResNet malignancy classifier trained on LIDC-IDRI annotations"

# ============================================================
# MODEL
# ============================================================
model:
  type: "malignancy_classifier"
  block1_channels: 64
  block2_channels: 128
  block3_channels: 256
  blocks_per_stage: 2
  hidden_dim: 128
  num_classes: 2          # Binary: benign vs malignant
  dropout: 0.5
  head_dropout: 0.3

# ============================================================
# TRAINING
# ============================================================
training:
  num_epochs: 50
  batch_size: 32
  learning_rate: 0.0001   # 1e-4
  weight_decay: 0.00001   # 1e-5
  optimizer: "AdamW"
  scheduler: "CosineAnnealingLR"
  scheduler_T_max: 50     # Same as num_epochs
  gradient_clip: 1.0
  use_amp: true
  early_stopping_patience: 15

# ============================================================
# DATA
# ============================================================
data:
  # Path to LIDC annotations CSV (must have 'nodule_id' and 'malignancy' columns)
  annotations_csv: "data/LIDC_annotations.csv"
  # Directory containing {nodule_id}.npy patch files (64x64x64)
  patches_dir: "data/nodule_patches"
  val_ratio: 0.2
  num_workers: 4

# ============================================================
# LOGGING & CHECKPOINTING
# ============================================================
logging:
  experiment_dir: "experiments/malignancy_classifier"
  checkpoint_dir: "experiments/malignancy_classifier/checkpoints"
  log_interval: 10
  save_best: true