boxsup_spf_scannet.yaml

train:
  seed: 1999
  epochs: &epoch 512
  interval: 16
  pretrain: pretrains/scannetv2/best_pretrain_softgroup_spp_pool.pth

test:
  seed: 1999
  test_topk_per_scene: 100
  test_score_thresh: 0.0
  test_npoint_thresh: 100

dataloader:
  train:
    batch_size: 10
    num_workers: 10
    persistent_workers: True
  val:
    batch_size: 1
    num_workers: 4
    persistent_workers: True
  test:
    batch_size: 1
    num_workers: 4
    persistent_workers: True

data:
  train:
    type: scannetv2
    data_root: data/scannetv2
    prefix: train
    suffix: _inst_nostuff.pth
    label_type: &label_type_ gaussian_process_kl_pseudo_labels
    training: True
    mode: 4
    with_elastic: True
    use_xyz: True
    repeat: 1
    voxel_cfg:
      scale: 50
      spatial_shape: [128, 512]
      max_npoint: 250000
  val:
    type: scannetv2
    data_root: data/scannetv2
    prefix: val
    suffix: _inst_nostuff.pth
    label_type: *label_type_
    training: False
    mode: 4
    with_elastic: False
    use_xyz: True
    voxel_cfg:
      scale: 50
      spatial_shape: [128, 512]
      max_npoint: 250000
  test:
    type: scannetv2
    data_root: data/scannetv2
    prefix: val
    suffix: _inst_nostuff.pth
    label_type: *label_type_
    training: False
    mode: 4
    with_elastic: False
    use_xyz: True
    voxel_cfg:
      scale: 50
      spatial_shape: [128, 512]
      max_npoint: 250000

model:
  input_channel: 6
  blocks: 5
  block_reps: 2
  media: 32
  normalize_before: True
  return_blocks: True
  pool: mean
  num_class: 18
  decoder:
    num_layer: 6
    num_query: 400
    d_model: 256
    nhead: 8
    hidden_dim: 1024
    dropout: 0.0
    activation_fn: gelu
    iter_pred: True
    attn_mask: True
    pe: False
  criterion:
    loss_weight: [0.5, 1.0, 1.0, 0.5, 0.2]
    cost_weight: [0.5, 1.0, 1.0]
    non_object_weight: 0.1
  test_cfg:
    topk_insts: 100
    score_thr: 0.0
    npoint_thr: 100
  norm_eval: False
  fix_module: []

optimizer:
  type: AdamW
  lr: 0.0002
  weight_decay: 0.05

lr_scheduler:
  type: PolyLR
  max_iters: *epoch
  power: 0.9
  constant_ending: 0.0


fp16: True