segformer_b2_cityscapes_1024x512_160k.yml

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

batch_size: 1
iters: 160000

model:
  type: SegFormer
  backbone:
    type: MixVisionTransformer_B2
    pretrained: https://bj.bcebos.com/paddleseg/dygraph/backbone/mix_vision_transformer_b2.tar.gz
  embedding_dim: 768
  num_classes: 19

optimizer:
  _inherited_: False
  type: AdamW
  beta1: 0.9
  beta2: 0.999
  weight_decay: 0.01

lr_scheduler:
  type: PolynomialDecay
  learning_rate: 0.00006
  power: 1

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