architecture: MaskRCNN
use_gpu: false
use_xpu: true
max_iters: 1200
snapshot_iter: 100
log_iter: 20
save_dir: output
pretrain_weights: https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_r50_2x.tar
metric: COCO
weights: output/mask_rcnn_r50_1x_cocome_kunlun/model_final
num_classes: 2
finetune_exclude_pretrained_params: ['cls_score']

MaskRCNN:
  backbone: ResNet
  rpn_head: RPNHead
  roi_extractor: RoIAlign
  bbox_assigner: BBoxAssigner
  bbox_head: BBoxHead
  mask_assigner: MaskAssigner
  mask_head: MaskHead

ResNet:
  norm_type: affine_channel
  norm_decay: 0.
  depth: 50
  feature_maps: 4
  freeze_at: 2

ResNetC5:
  depth: 50
  norm_type: affine_channel

RPNHead:
  anchor_generator:
    anchor_sizes: [32, 64, 128, 256, 512]
    aspect_ratios: [0.5, 1.0, 2.0]
    stride: [16.0, 16.0]
    variance: [1.0, 1.0, 1.0, 1.0]
  rpn_target_assign:
    rpn_batch_size_per_im: 256
    rpn_fg_fraction: 0.5
    rpn_negative_overlap: 0.3
    rpn_positive_overlap: 0.7
    rpn_straddle_thresh: 0.0
  train_proposal:
    min_size: 0.0
    nms_thresh: 0.7
    pre_nms_top_n: 12000
    post_nms_top_n: 2000
  test_proposal:
    min_size: 0.0
    nms_thresh: 0.7
    pre_nms_top_n: 6000
    post_nms_top_n: 1000

RoIAlign:
  resolution: 14
  spatial_scale: 0.0625
  sampling_ratio: 0

BBoxHead:
  head: ResNetC5
  nms:
    keep_top_k: 100
    nms_threshold: 0.5
    normalized: false
    score_threshold: 0.05

MaskHead:
  dilation: 1
  conv_dim: 256
  resolution: 14

BBoxAssigner:
  batch_size_per_im: 512
  bbox_reg_weights: [0.1, 0.1, 0.2, 0.2]
  bg_thresh_hi: 0.5
  bg_thresh_lo: 0.0
  fg_fraction: 0.25
  fg_thresh: 0.5

MaskAssigner:
  resolution: 14

LearningRate:
  base_lr: 0.001
  schedulers:
  - !PiecewiseDecay
    gamma: 0.1
    milestones: [900, 1100]
  - !LinearWarmup
    start_factor: 0.1
    steps: 300

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

_READER_: 'mask_reader_cocome.yml'