emanet_resnet101_os8_cityscapes_1024x512_80k.yml

_base_: '../_base_/cityscapes.yml'

batch_size: 2
iters: 80000

model:
  type: EMANet
  backbone:
    type: ResNet101_vd
    output_stride: 8
    pretrained: https://bj.bcebos.com/paddleseg/dygraph/resnet101_vd_ssld.tar.gz
  num_classes: 19
  ema_channels: 512
  gc_channels: 256
  num_bases: 64
  stage_num: 3
  momentum: 0.1
  concat_input: True
  enable_auxiliary_loss: True
  align_corners: False

optimizer:
  type: sgd
  momentum: 0.9
  weight_decay: 0.0005

loss:
  types:
    - type: CrossEntropyLoss
    - type: CrossEntropyLoss
  coef: [1, 0.4]