portraitnet_eg1800_224x224_46k.yml

batch_size: 64
iters: 46000

train_dataset:
  type: EG1800
  dataset_root: data/EG1800
  common_transforms:
    - type: RandomAffine
      max_rotation: 45
      min_scale_factor: 0.5
      max_scale_factor: 1.5
      size: [ 224, 224 ]
      translation_offset: 56
    - type: RandomHorizontalFlip
  transforms1:
    - type: Normalize
      mean: [0.485, 0.458, 0.408]
      std: [0.23, 0.23, 0.23]
  transforms2:
    - type: RandomDistort
      brightness_range: 0.6
      contrast_range: 0.4
      saturation_range: 0.6
      hue_prob: 0.0
      sharpness_range: 0.2
      sharpness_prob: 0.5
    - type: RandomBlur
      prob: 0.5
      blur_type: random
    - type: RandomNoise
    - type: Normalize
      mean: [ 0.485, 0.458, 0.408 ]
      std: [ 0.23, 0.23, 0.23 ]
  mode: train

val_dataset:
  type: EG1800
  dataset_root: data/EG1800
  common_transforms:
    - type: ScalePadding
      target_size: [ 224, 224 ]
      im_padding_value: [127.5, 127.5, 127.5]
      label_padding_value: 0
    - type: Normalize
      mean: [0.485, 0.458, 0.408]
      std: [0.23, 0.23, 0.23]
  transforms1: null
  transforms2: null
  mode: val

optimizer:
  type: adam
  weight_decay: 5.0e-4

lr_scheduler:
  type: StepDecay
  learning_rate: 0.001
  step_size: 460
  gamma: 0.95

loss:
  types:
    - type: CrossEntropyLoss
    - type: CrossEntropyLoss
    - type: FocalLoss
    - type: KLLoss
  coef: [1, 1, 0.3, 2]

model:
  type: PortraitNet
  backbone:
    type: MobileNetV2_x1_0
    channel_ratio: 1.0
    min_channel: 16
    pretrained: https://paddleseg.bj.bcebos.com/dygraph/backbone/mobilenetv2_x1_0_ssld.tar.gz
  add_edge: True
  num_classes: 2