config.json

{
  "batch_size": 2,
  "epochs": 30,
  "lr": 5e-05,
  "weight_decay": 0.0001,
  "seed": 42,
  "num_workers": 8,
  "frozen_backbone": false,
  "unfreeze_blocks": 4,
  "backbone_lr_mult": 0.1,
  "use_focal_loss": false,
  "focal_gamma": 2.0,
  "label_smoothing": 0.1,
  "use_class_weights": true,
  "no_weighted_sampler": true,
  "use_aux_severity": true,
  "aux_severity_weight": 0.5,
  "aux_severity_targets": "0.0,0.3,0.7,0.5,1.0",
  "gate_class_weights": null,
  "la_tau": 1.0,
  "use_ema": true,
  "ema_decay": 0.9995,
  "drop_path": 0.0,
  "no_augment": false,
  "use_copy_paste": false,
  "copy_paste_max_prob": 0.3,
  "copy_paste_warmup_epochs": 2,
  "copy_paste_max_pastes": 2,
  "copy_paste_feather": 3,
  "use_balanced_sampling": false,
  "samples_per_class": 3000,
  "resume": "runs/ddp_v4_best_recipe/best_model.pth",
  "name": "ddp_v11_2gpu_30ep",
  "stratified_splits": "LAFIRE/tiles_1024/stratified_splits.json",
  "val_split_key": "val",
  "no_multiscale": true,
  "no_attention_pool": true,
  "use_sfp": true,
  "sfp_stride": 8,
  "use_hierarchical": true,
  "use_gated_head": true,
  "gate_weight": 1.0,
  "leaf_weight": 2.0,
  "use_tiles": true,
  "type_weight": 5.0,
  "extent_weight": 5.0,
  "disable_hard_class_boost": false,
  "accumulation_steps": 8
}