mask_rcnn_r50_fpn_1x.yml

architecture: MaskRCNN
use_gpu: true
max_iters: 180000
log_smooth_window: 20
save_dir: output
snapshot_iter: 10000
pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_cos_pretrained.tar
metric: COCO
weights: output/mask_rcnn_r50_fpn_1x/model_final
num_classes: 81
load_static_weights: True

# Model Achitecture
MaskRCNN:
  # model anchor info flow
  anchor: AnchorRPN
  proposal: Proposal
  mask: Mask
  # model feat info flow
  backbone: ResNet
  neck: FPN
  rpn_head: RPNHead
  bbox_head: BBoxHead
  mask_head: MaskHead

ResNet:
  # index 0 stands for res2
  depth: 50
  norm_type: bn
  freeze_at: 0
  return_idx: [0,1,2,3]
  num_stages: 4

FPN:
  in_channels: [256, 512, 1024, 2048]
  out_channel: 256
  min_level: 0
  max_level: 4
  spatial_scale: [0.25, 0.125, 0.0625, 0.03125]


RPNHead:
  rpn_feat:
    name: RPNFeat
    feat_in: 256
    feat_out: 256
  anchor_per_position: 3
  rpn_channel: 256

BBoxHead:
  bbox_feat:
    name: BBoxFeat
    roi_extractor:
      name: RoIExtractor
      resolution: 7
      sampling_ratio: 2
    head_feat:
      name: TwoFCHead
      in_dim: 256
      mlp_dim: 1024
  in_feat: 1024

MaskHead:
  mask_feat:
    name: MaskFeat
    num_convs: 4
    feat_in: 256
    feat_out: 256
    mask_roi_extractor:
      name: RoIExtractor
      resolution: 14
      sampling_ratio: 2
    share_bbox_feat: False
  feat_in: 256

AnchorRPN:
  anchor_generator:
    name: AnchorGeneratorRPN
    aspect_ratios: [0.5, 1.0, 2.0]
    anchor_start_size: 32
    stride: [4., 4.]
  anchor_target_generator:
    name: AnchorTargetGeneratorRPN
    batch_size_per_im: 256
    fg_fraction: 0.5
    negative_overlap: 0.3
    positive_overlap: 0.7
    straddle_thresh: 0.0

Proposal:
  proposal_generator:
    name: ProposalGenerator
    min_size: 0.0
    nms_thresh: 0.7
    train_pre_nms_top_n: 2000
    train_post_nms_top_n: 2000
    infer_pre_nms_top_n: 1000
    infer_post_nms_top_n: 1000
  proposal_target_generator:
    name: ProposalTargetGenerator
    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_thresh: [0.5,]
    fg_fraction: 0.25
  bbox_post_process: # used in infer
    name: BBoxPostProcess
    # decode -> clip -> nms
    decode_clip_nms:
      name: DecodeClipNms
      keep_top_k: 100
      score_threshold: 0.05
      nms_threshold: 0.5

Mask:
  mask_target_generator:
    name: MaskTargetGenerator
    mask_resolution: 28
  mask_post_process:
    name: MaskPostProcess
    mask_resolution: 28


# Train
LearningRate:
  base_lr: 0.01
  schedulers:
  - !PiecewiseDecay
    gamma: 0.1
    milestones: [120000, 160000]
  - !LinearWarmup
    start_factor: 0.3333
    steps: 500

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

_READER_: 'mask_reader.yml'