# global configs
Global:
  checkpoints: null
  pretrained_model: null
  output_dir: "./output/"
  device: "gpu"
  iter_per_epoch: &iter_per_epoch 50
  warmup_iter: 10
  save_interval: 8
  eval_during_train: True
  eval_interval: 8
  epochs: &epochs 348    # 348*50 = 120*145 = 17400 iters
  print_batch_step: 25
  use_visualdl: False
  eval_mode: "retrieval"
  retrieval_feature_from: "features" # 'backbone' or 'features'
  re_ranking: False
  # used for static mode and model export
  image_shape: [3, 256, 128]
  save_inference_dir: "./inference"
  train_mode: 'metabin'

AMP:
  scale_loss: 65536
  use_dynamic_loss_scaling: True

# model architecture
Arch:
  name: "RecModel"
  Backbone:
    name: "ResNet50_metabin"
    pretrained: https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/others/resnet50-19c8e357_torch2paddle.pdparams
    bias_lr_factor: 2.0
  BackboneStopLayer:
    name: "flatten"
  Neck:
    name: MetaBNNeck
    num_features: &feat_dim 2048
    use_global_stats: True
  Head:
    name: "FC"
    embedding_size: *feat_dim
    class_num: &class_num 751
    weight_attr:
      initializer:
        name: KaimingUniform
        negative_slope: 2.23606   # math.sqrt(5)
        nonlinearity: "leaky_relu"
    bias_attr: False

# data loader for train and eval
DataLoader:
  Train:
    dataset:
      name: "Market1501"
      image_root: "./dataset/"
      cls_label_path: "bounding_box_train"
      backend: "pil"
      transform_ops:
        - ResizeImage:
            size: [128, 256]
            return_numpy: False
            interpolation: "bicubic"
            backend: "pil"
        - RandFlipImage:
            flip_code: 1
        - Pad:
            padding: 10
        - RandCropImageV2:
            size: [128, 256]
        - ColorJitter:
            brightness: 0.15
            contrast: 0.15
            saturation: 0.1
            hue: 0.1
        - ToTensor:
        - Normalize:
            mean: [0.485, 0.456, 0.406]
            std: [0.229, 0.224, 0.225]
    sampler:
      name: NaiveIdentityBatchSampler
      batch_size: 96
      num_instances: 4
      drop_last: True
    loader:
      num_workers: 4
      use_shared_memory: True
      
  Metalearning:
    Train:
      dataset:
        name: "Market1501"
        image_root: "./dataset/"
        cls_label_path: "bounding_box_train"
        backend: "pil"
        transform_ops:
          - ResizeImage:
              size: [128, 256]
              return_numpy: False
              interpolation: "bicubic"
              backend: "pil"
          - RandFlipImage:
              flip_code: 1
          - Pad:
              padding: 10
          - RandCropImageV2:
              size: [128, 256]
          - ColorJitter:
              brightness: 0.15
              contrast: 0.15
              saturation: 0.1
              hue: 0.1
          - ToTensor:
          - Normalize:
              mean: [0.485, 0.456, 0.406]
              std: [0.229, 0.224, 0.225]
      sampler:
        name: DomainShuffleBatchSampler
        batch_size: 96
        num_instances: 4
        drop_last: True
        camera_to_domain: True
      loader:
        num_workers: 4
        use_shared_memory: True

  Eval:
    Query:
      dataset:
        name: "DukeMTMC"
        image_root: "./dataset/"
        cls_label_path: "query"
        backend: "pil"
        transform_ops:
          - ResizeImage:
              size: [128, 256]
              return_numpy: False
              interpolation: "bicubic"
              backend: "pil"
          - ToTensor:
          - Normalize:
              mean: [0.485, 0.456, 0.406]
              std: [0.229, 0.224, 0.225]
      sampler:
        name: DistributedBatchSampler
        batch_size: 128
        drop_last: False
        shuffle: False
      loader:
        num_workers: 4
        use_shared_memory: True

    Gallery:
      dataset:
        name: "DukeMTMC"
        image_root: "./dataset/"
        cls_label_path: "bounding_box_test"
        backend: "pil"
        transform_ops:
          - ResizeImage:
              size: [128, 256]
              return_numpy: False
              interpolation: "bicubic"
              backend: "pil"
          - ToTensor:
          - Normalize:
              mean: [0.485, 0.456, 0.406]
              std: [0.229, 0.224, 0.225]
      sampler:
        name: DistributedBatchSampler
        batch_size: 128
        drop_last: False
        shuffle: False
      loader:
        num_workers: 4
        use_shared_memory: True

# loss function config for traing/eval process
Loss:
  Train:
    - CELossForMetaBIN:
        weight: 1.0
        epsilon: 0.1
    - TripletLossForMetaBIN:
        weight: 1.0
        margin: 0.3
        feature_from: "backbone"
    - IntraDomainScatterLoss:
        weight: 1.0
        normalize_feature: True
        feature_from: "backbone"
    - InterDomainShuffleLoss:
        weight: 1.0
        normalize_feature: False
        feature_from: "backbone"
  Basic:
    - CELossForMetaBIN:
        weight: 1.0
        epsilon: 0.1
    - TripletLossForMetaBIN:
        weight: 1.0
        margin: 0.3
        feature_from: "backbone"
  MetaTrain:
    - CELossForMetaBIN:
        weight: 1.0
        epsilon: 0.1
    - TripletLossForMetaBIN:
        weight: 1.0
        margin: 0.3
        feature_from: "backbone"
    - IntraDomainScatterLoss:
        weight: 1.0
        normalize_feature: True
        feature_from: "backbone"
    - InterDomainShuffleLoss:
        weight: 1.0
        normalize_feature: False
        feature_from: "backbone"
  MetaTest:
    - CELossForMetaBIN:
        weight: 1.0
        epsilon: 0.1
    - TripletLossForMetaBIN:
        weight: 1.0
        margin: 0.3
        feature_from: "backbone"
  Eval:
    - TripletLossForMetaBIN:
        weight: 1.0
        margin: 0.3
        feature_from: "backbone"

Optimizer:
  - Momentum:
      scope: ".*(conv|batch_norm|instance_norm|feat_bn|fc)"
      lr:
        name: MultiStepDecay
        epochs: *epochs
        learning_rate: 0.01
        step_each_epoch: *iter_per_epoch
        milestones: [145, 261]
        gamma: 0.1
        warmup_epoch: 29
        warmup_start_lr: 0.0001
        by_epoch: False
        last_epoch: -1
      momentum: 0.9
      regularizer:
        name: "L2"
        coeff: 0.0005
  - SGD:
      scope: "backbone.*gate"
      lr:
        name: Constant
        learning_rate: 0.2
        last_epoch: -1
  - SGD:
      scope: "RecModel"
      lr:
        name: Cyclic
        epochs: *epochs
        step_each_epoch: *iter_per_epoch
        base_learning_rate: 0.001
        max_learning_rate: 0.1
        warmup_epoch: 0
        warmup_start_lr: 1
        step_size_up: 1095
        step_size_down: 1095
        by_epoch: False
        last_epoch: 0

Metric:
  Eval:
    - Recallk:
        topk: [1, 5, 10]
    - mAP: {}