batch_size: 4
iters: 160000

train_dataset:
  type: Dataset
  dataset_root: data/Cityscapes/
  train_path: data/Cityscapes/train.txt
  num_classes: 150
  transforms: 
    - type: ResizeByShort
      short_size: [256, 307, 358, 409, 460, 512, 563, 614, 665, 716, 768, 819, 870, 921, 972, 1024]
      max_size: 2048
    - type: RandomPaddingCrop
      crop_size: [512, 512] 
    - type: RandomDistort
      brightness_range: 0.125
      brightness_prob: 1.0
      contrast_range: 0.5
      contrast_prob: 1.0
      saturation_range: 0.5
      saturation_prob: 1.0
      hue_range: 18
      hue_prob: 1.0
    - type: RandomHorizontalFlip
    - type: GenerateInstanceTargets
      num_classes: 150
      ignore_index: 255
    - type: Normalize
      mean: [0.485, 0.456, 0.406]
      std: [0.229, 0.224, 0.225]
  mode: train

val_dataset:
  type: Dataset
  dataset_root: datasets/Cityscapes
  val_path: datasets/Cityscapes/val.txt
  num_classes: 150
  transforms:
    - type: Resize
      target_size: [512, 512]
      keep_ratio: False
      interp: LINEAR
    - type: Normalize
      mean: [0.485, 0.456, 0.406]
      std: [0.229, 0.224, 0.225]
  mode: val

model:
  type: MaskFormer
  num_classes: 150
  backbone:
    type: SwinTransformer_tiny_patch4_window7_224_maskformer
  pretrained: https://bj.bcebos.com/paddleseg/paddleseg/dygraph/ade20k/maskformer_ade20k_swin_tiny/pretrain/model.pdparams

optimizer:
  type: AdamW
  weight_decay: 0.01
  custom_cfg:
  - name: backbone
    lr_mult: 1.0
  - name: norm
    weight_decay_mult: 0.0
  - name: relative_position_bias_table
    weight_decay_mult: 0.0
  grad_clip_cfg: 
    name: ClipGradByNorm
    clip_norm: 0.01

lr_scheduler:
  type: PolynomialDecay
  warmup_iters: 1500
  warmup_start_lr: 6.0e-11
  learning_rate: 6.0e-05
  end_lr: 0
  power: 0.9

loss:
  types:
    - type: MaskFormerLoss
      num_classes: 150
      eos_coef: 0.1
  coef: [1]