optimizer_base_1x.yml 434 B

12345678910111213141516171819202122
  1. epoch: 12
  2. LearningRate:
  3. base_lr: 0.0001
  4. schedulers:
  5. - !PiecewiseDecay
  6. gamma: 0.1
  7. milestones: [9, 11]
  8. - !LinearWarmup
  9. start_factor: 0.001
  10. steps: 1000
  11. OptimizerBuilder:
  12. optimizer:
  13. type: AdamWDL
  14. betas: [0.9, 0.999]
  15. layer_decay: 0.75
  16. weight_decay: 0.02
  17. num_layers: 12
  18. filter_bias_and_bn: True
  19. skip_decay_names: ['pos_embed', 'cls_token']
  20. set_param_lr_func: 'layerwise_lr_decay'