architecture: SOLOv2
use_gpu: true
max_iters: 135000
snapshot_iter: 20000
log_smooth_window: 20
save_dir: output
pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV3_large_x1_0_ssld_pretrained.tar
metric: COCO
weights: output/solov2/solov2_mobilenetv3_fpn_448_3x/model_final
num_classes: 81
use_ema: true
ema_decay: 0.9998

SOLOv2:
  backbone: MobileNetV3RCNN
  fpn: FPN
  bbox_head: SOLOv2Head
  mask_head: SOLOv2MaskHead

MobileNetV3RCNN:
  norm_type: bn
  freeze_norm: true
  norm_decay: 0.0
  feature_maps: [2, 3, 4, 6]
  conv_decay: 0.00001
  lr_mult_list: [0.25, 0.25, 0.5, 0.5, 0.75]
  scale: 1.0
  model_name: large

FPN:
  max_level: 6
  min_level: 2
  num_chan: 256
  spatial_scale: [0.03125, 0.0625, 0.125, 0.25]
  reverse_out: True

SOLOv2Head:
  seg_feat_channels: 256
  stacked_convs: 2
  num_grids: [40, 36, 24, 16, 12]
  kernel_out_channels: 128
  solov2_loss: SOLOv2Loss
  mask_nms: MaskMatrixNMS
  drop_block: True

SOLOv2MaskHead:
  in_channels: 128
  out_channels: 128
  start_level: 0
  end_level: 3

SOLOv2Loss:
  ins_loss_weight: 3.0
  focal_loss_gamma: 2.0
  focal_loss_alpha: 0.25

MaskMatrixNMS:
  pre_nms_top_n: 500
  post_nms_top_n: 100

LearningRate:
  base_lr: 0.02
  schedulers:
  - !PiecewiseDecay
    gamma: 0.1
    milestones: [90000, 120000]
  - !LinearWarmup
    start_factor: 0.
    steps: 1000

OptimizerBuilder:
  optimizer:
    momentum: 0.9
    type: Momentum
  regularizer:
    factor: 0.0001
    type: L2

_READER_: 'solov2_light_448_reader.yml'