# Architecture of detection, which is also the prefix of data feed module
architecture: MaskRCNN

# Data feed module 
train_feed: MaskRCNNTrainFeed
eval_feed: MaskRCNNEvalFeed
test_feed: MaskRCNNTestFeed

# Use GPU or CPU, true by default
use_gpu: true

# Maximum number of iteration. 
# In rcnn models, max_iters is 180000 if lr schedule is 1x and batch_size is 1.
max_iters: 180000

# Snapshot period. If training and test at same time, evaluate model at each snapshot_iter. 10000 by default.
snapshot_iter: 10000

# Smooth the log output in specified iterations, 20 by default.
log_smooth_window: 20

# The number of iteration interval to display in training log.
log_iter: 20

# The directory to save models.
save_dir: output

# The path of oretrained wegiths. If url is provided, it will download the pretrain_weights and decompress automatically.
pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_cos_pretrained.tar

# Evalution method, COCO and VOC are available. 
metric: COCO

# The path of final model for evaluation and test.
weights: output/mask_rcnn_r50_fpn_1x/model_final/

# Number of classes, 81 for COCO and 21 for VOC
num_classes: 81

# Mask RCNN architecture, see https://arxiv.org/abs/1703.06870
MaskRCNN:
  backbone: ResNet
  fpn: FPN
  roi_extractor: FPNRoIAlign
  rpn_head: FPNRPNHead
  bbox_assigner: BBoxAssigner
  bbox_head: BBoxHead
  mask_assigner: MaskAssigner
  mask_head: MaskHead
  rpn_only: false

# Backbone module
ResNet:
  # Index of stages using deformable conv v2, [] by default
  dcn_v2_stages: []
  # ResNet depth, 50 by default
  depth: 50
  # Stage index of returned feature map, [2,3,4,5] by default
  feature_maps:
  - 2
  - 3
  - 4
  - 5
  # Stage Index of backbone to freeze, 2 by default
  freeze_at: 2
  # Whether freeze normalization layers, true by default
  freeze_norm: true
  # Weight decay for normalization layer weights, 0. by default
  norm_decay: 0.0
  # Normalization type, bn/sync_bn/affine_channel, affine_channel by default
  norm_type: affine_channel
  # ResNet variant, supports 'a', 'b', 'c', 'd' currently, b by default
  variant: b

# FPN module
FPN:
  # Whether has extra conv in higher levels, false by default
  has_extra_convs: false
  # Highest level of the backbone feature map to use, 6 by default
  max_level: 6
  # Lowest level of the backbone feature map to use, 6 by default
  min_level: 2
  # FPN normalization type, bn/sync_bn/affine_channel, null by default
  norm_type: null
  # Number of feature channels, 256 by default
  num_chan: 256
  # Feature map scaling factors, [0.03125, 0.0625, 0.125, 0.25] by default
  spatial_scale:
  - 0.03125
  - 0.0625
  - 0.125
  - 0.25

# RPN module, if use non-FPN architecture, use RPNHead instead
# Extract proposals according to anchors and assign box targets and
# score targets to selected proposals to compute RPN loss. For FPN
# architecture, RPN is computed from each levels and collect proposals
# together.
FPNRPNHead:
  # fluid.layers.anchor_generator
  # Generate anchors for RCNN models. Each position of input produces
  # N anchors. N = anchor_sizes * aspect_ratios. In FPNRPNHead, aspect_ratios
  # is provided and anchor_sizes depends on FPN levels and anchor_start_size.
  anchor_generator:
    aspect_ratios:
    - 0.5
    - 1.0
    - 2.0
    variance:
    - 1.0
    - 1.0
    - 1.0
    - 1.0
  # fluid.layers.rpn_target_assign
  # Assign classification and regression targets to each anchor according
  # to Intersection-over-Union(IoU) overlap between anchors and ground
  # truth boxes. The classification targets is binary class labels. the 
  # positive labels are two kinds of anchors: the anchors with the highest 
  # IoU overlap with a ground-truth box, or an anchor that has an IoU overlap 
  # higher than rpn_positive_overlap with any ground-truth box.  
  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
  # fluid.layers.generate_proposals in training
  # Generate RoIs according to each box with probability to be a foreground 
  # object. The operation performs following steps: Transposes and resizes 
  # scores and bbox_deltas; Calculate box locations as proposal candidates;
  # Clip boxes to image; Remove predicted boxes with small area; Apply NMS to 
  # get final proposals as output.
  train_proposal:
    min_size: 0.0
    nms_thresh: 0.7
    post_nms_top_n: 2000
    pre_nms_top_n: 2000
  # fluid.layers.generate_proposals in test
  test_proposal:
    min_size: 0.0
    nms_thresh: 0.7
    post_nms_top_n: 1000
    pre_nms_top_n: 1000
  # Size of anchor at the first scale, 32 by default
  anchor_start_size: 32
  # highest level of FPN output, 6 by default
  max_level: 6
  # Lowest level of FPN output, 2 by default
  min_level: 2
  # Number of FPN output channels, 256 by default
  num_chan: 256
  # Number of classes in RPN output, 1 by default
  num_classes: 1

# RoI extractor module, if use non-FPN architecture, use RoIAlign instead
# For FPN architecture, proposals are distributed to different levels and
# apply roi align at each level. Then concat the outputs.
FPNRoIAlign:
  # The canconical FPN feature map level, 4 by default
  canconical_level: 4
  # The canconical FPN feature map size, 224 by default
  canonical_size: 224
  # The highest level of FPN layer, 5 by default
  max_level: 5
  # The lowest level of FPN layer, 2 by default
  min_level: 2
  # Number of sampling points, 0 by default
  sampling_ratio: 2
  # Box resolution, 7 by default
  box_resolution: 7
  # Mask RoI resolution, 14 by default
  mask_resolution: 14

# Mask head module  
# Generate mask output and compute loss mask.
MaskHead:
  # Number of convolutions, 4 for FPN, 0 otherwise. 0 by default
  num_convs: 4
  # size of the output mask, 14 by default
  resolution: 28
  # Dilation rate, 1 by default
  dilation: 1
  # Number of channels after first conv, 256 by default
  num_chan_reduced: 256
  # Number of output classes, 81 by default
  num_classes: 81

# fluid.layers.generate_proposal_labels
# Combine boxes and gt_boxes, and sample foreground proposals and background 
# prosals.Then assign classification and regression targets to selected RoIs. 
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
  num_classes: 81
  shuffle_before_sample: true

# fluid.layers.generate_mask_labels
# For given the RoIs and corresponding labels, sample foreground RoIs.
# Assign mask targets to selected RoIs which are encoded to K binary masks 
# of resolution M x M. 
MaskAssigner:
  resolution: 28
  num_classes: 81

# BBox head module
# Faster bbox head following the RoI extractor, and apply post process, such as
# NMS and box coder..
BBoxHead:
  # Head after RoI extractor, ResNetC5/TwoFCHead
  head: TwoFCHead
  # fluid.layers.multiclass_nms
  # Select a subset of detection bounding boxes that have high scores larger 
  # than score_threshold. Then  prune away boxes that have high IoU overlap 
  # with already selected boxes by nms_threshold.
    keep_top_k: 100
    nms_threshold: 0.5
    score_threshold: 0.05
  # fluid.layers.box_coder
  box_coder:
    axis: 1
    box_normalized: false
    code_type: decode_center_size
    prior_box_var:
    - 0.1
    - 0.1
    - 0.2
    - 0.2
  num_classes: 81

# RCNN head with two Fully Connected layers
TwoFCHead:
  # The number of output channels, 1024 by default
  num_chan: 1024

# Learning rate configuration
LearningRate:
  # Base learning rate, 0.01 by default
  base_lr: 0.01
  # Learning rate schedulers, PiecewiseDecay and LinearWarmup by default
  schedulers:
  # fluid.layers.piecewise_decay
  # Values has higher priority and if values is null, learning rate is multipled by gamma at each stage
  - !PiecewiseDecay
    gamma: 0.1
    milestones:
    - 120000
    - 160000
    values: null
  # fluid.layers.linear_lr_warmup
  # Start learning rate equals to base_lr * start_factor
  - !LinearWarmup
    start_factor: 0.3333333333333333
    steps: 500

# Optimizer module
OptimizerBuilder:
  # fluid.optimizer
  optimizer:
    momentum: 0.9
    type: Momentum
  # fluid.regularizer
  regularizer:
    factor: 0.0001
    type: L2

# Data feed module for training
MaskRCNNTrainFeed:
  # Batch size per device, 1 by default
  batch_size: 1
  # Dataset module
  dataset:
    # Annotation file path
    annotation: annotations/instances_train2017.json
    # Dataset directory
    dataset_dir: dataset/coco
    # Directory where image files are stored
    image_dir: train2017
  # List of data fields needed
  fields:
  - image
  - im_info
  - im_id
  - gt_box
  - gt_label
  - is_crowd
  - gt_mask
  # list of image dims
  image_shape:
  - 3
  - 800
  - 1333
  # List of sample transformations to use
  sample_transforms:
  # Transform the image data to numpy format.
  - !DecodeImage
    to_rgb: true  # default: true
    with_mixup: false  # default: false
  # Flip images randomly
  # Transform the x coordinates of bboxes and segmentations
  - !RandomFlipImage
    is_mask_flip: true  # default: false
    # Whether bbox is normalized
    is_normalized: false  # default: false
    prob: 0.5  # default: 0.5
  # Normalize the image
  - !NormalizeImage
    # The format of image, [H, W, C]/[C, H, W], true by default
    is_channel_first: false
    # Whether divide by 255, true by default
    is_scale: true
    # default: [0.485, 0.456, 0.406]
    mean:
    - 0.485
    - 0.456
    - 0.406
    # default: [1, 1, 1]
    std:
    - 0.229
    - 0.224
    - 0.225
  # Rescale image to the specified target size, and capped at max_size
  - !ResizeImage
    # Resize method, cv2.INTER_LINEAR(1) by default
    interp: 1
    max_size: 1333
    target_size: 800
    use_cv2: true  # default: true
  # Change the channel
  - !Permute
    # The format of image, [H, W, C]/[C, H, W], true by default
    channel_first: true
    to_bgr: false  # default: true
  # List of batch transformations to use
  batch_transforms:
  # Pad a batch of samples to same dimensions
  - !PadBatch
    pad_to_stride: 32  # default: 32
  # Drop last batch if size is uneven, false by default
  drop_last: false
  # Number of workers processes(or threads), 2 by default
  num_workers: 2
  # Number of samples, -1 represents all samples. -1 by default
  samples: -1
  # If samples should be shuffled, true by default
  shuffle: true
  # If update im_info after padding, false by default
  use_padded_im_info: false
  # If use multi-process, false by default
  use_process: false

# Data feed module for test
MaskRCNNEvalFeed:
  #  Batch size per device, 1 by default
  batch_size: 1
  # Dataset module
  dataset:
    # Annotation file path
    annotation: annotations/instances_val2017.json
    # Dataset directory
    dataset_dir: dataset/coco
    # Directory where image files are stored
    image_dir: val2017
  # List of data fields needed
  fields:
  - image
  - im_info
  - im_id
  - im_shape
  # list of image dims
  image_shape:
  - 3
  - 800
  - 1333
  # List of sample transformations to use
  sample_transforms:
  # Transform the image data to numpy format.
  - !DecodeImage
    to_rgb: true  # default: true
    with_mixup: false  # default: false
  # Normalize the image
  - !NormalizeImage
    # The format of image, [H, W, C]/[C, H, W], true by default
    is_channel_first: false
    # Whether divide by 255, true by default
    is_scale: true
    # default: [0.485, 0.456, 0.406]
    mean:
    - 0.485
    - 0.456
    - 0.406
    # default: [1, 1, 1]
    std:
    - 0.229
    - 0.224
    - 0.225
  # Rescale image to the specified target size, and capped at max_size
  - !ResizeImage
    # Resize method, cv2.INTER_LINEAR(1) by default
    interp: 1
    max_size: 1333
    target_size: 800
    use_cv2: true  # default: true
  # Change the channel
  - !Permute
    # The format of image, [H, W, C]/[C, H, W], true by default
    channel_first: true
    to_bgr: false  # default: true
  # List of batch transformations to use
  batch_transforms:
  # Pad a batch of samples to same dimensions
  - !PadBatch
    pad_to_stride: 32  # default: 32
  # Drop last batch if size is uneven, false by default
  drop_last: false
  # Number of workers processes(or threads), 2 by default
  num_workers: 2
  # Number of samples, -1 represents all samples. -1 by default
  samples: -1
  # If samples should be shuffled, true by default
  shuffle: false
  # If update im_info after padding, false by default
  use_padded_im_info: true
  # If use multi-process, false by default
  use_process: false

# Data feed module for test
MaskRCNNTestFeed:
  # Batch size per device, 1 by default  
  batch_size: 1
  # Dataset module
  dataset:
    # Annotation file path
    annotation: dataset/coco/annotations/instances_val2017.json
  # List of data fields needed
  fields:
  - image
  - im_info
  - im_id
  - im_shape
  # list of image dims
  image_shape:
  - 3
  - 800
  - 1333
  # List of sample transformations to use
  sample_transforms:
  # Transform the image data to numpy format.
  - !DecodeImage
    to_rgb: true  # default: true
    with_mixup: false  # default: false
  # Normalize the image
  - !NormalizeImage
    # The format of image, [H, W, C]/[C, H, W], true by default
    is_channel_first: false
    # Whether divide by 255, true by default
    is_scale: true
    # default: [0.485, 0.456, 0.406]
    mean:
    - 0.485
    - 0.456
    - 0.406
    # default: [1, 1, 1]
    std:
    - 0.229
    - 0.224
    - 0.225
  # Change the channel
  - !Permute
    # The format of image, [H, W, C]/[C, H, W], true by default
    channel_first: true
    to_bgr: false  # default: true
  # List of batch transformations to use
  batch_transforms:
  # Pad a batch of samples to same dimensions
  - !PadBatch
    pad_to_stride: 32  # default: 32
  # Drop last batch if size is uneven, false by default 
  drop_last: false
  # Number of workers processes(or threads), 2 by default
  num_workers: 2
  # Number of samples, -1 represents all samples. -1 by default
  samples: -1
  # If samples should be shuffled, true by default
  shuffle: false
  # If update im_info after padding, false by default
  use_padded_im_info: true
  # If use multi-process, false by default
  use_process: false