vitpose_base_coco_256x196.yml

use_gpu: true
log_iter: 50
save_dir: output
snapshot_epoch: 10
weights: output/vitpose_base_simple_coco_256x192/model_final
epoch: 210
num_joints: &num_joints 17
pixel_std: &pixel_std 200
metric: KeyPointTopDownCOCOEval
num_classes: 1
train_height: &train_height 256
train_width: &train_width 192
trainsize: &trainsize [*train_width, *train_height]
hmsize: &hmsize [48, 64]
flip_perm: &flip_perm [[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12], [13, 14], [15, 16]]


#####model
architecture: VitPose_TopDown
pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/keypoint/mae_pretrain_vit_base.pdparams
VitPose_TopDown:
  backbone: ViT
  head: TopdownHeatmapSimpleHead
  post_process: VitPosePostProcess
  loss: KeyPointMSELoss
  flip_test: True

ViT:
  img_size: [256, 192]
  patch_size: 16
  embed_dim: 768
  depth: 12
  num_heads: 12
  ratio: 1
  mlp_ratio: 4
  qkv_bias: True
  drop_path_rate: 0.3
  epsilon: 0.000001


TopdownHeatmapSimpleHead:
  in_channels: 768
  num_deconv_layers: 2
  num_deconv_filters: [256,256]
  num_deconv_kernels: [4,4]
  out_channels: 17
  shift_heatmap: False
  flip_pairs: *flip_perm
  extra: {final_conv_kernel: 1}

VitPosePostProcess:
  use_dark: True

KeyPointMSELoss:
  use_target_weight: true
  loss_scale: 1.0

####optimizer
LearningRate:
  base_lr: 0.0005
  schedulers:
  - !PiecewiseDecay
    gamma: 0.1
    milestones: [170, 200]
  - !LinearWarmup
    start_factor: 0.001
    steps: 500

OptimizerBuilder:
  clip_grad_by_norm: 1.0
  optimizer:
    type: AdamWDL
    betas: [0.9, 0.999]
    weight_decay: 0.1
    num_layers: 12
    layer_decay: 0.75
    filter_bias_and_bn: True
    skip_decay_names: ['pos_embed','norm']
    set_param_lr_func: 'layerwise_lr_decay'




#####data
TrainDataset:
  !KeypointTopDownCocoDataset
    image_dir: train2017
    anno_path: annotations/person_keypoints_train2017.json
    dataset_dir: dataset/coco
    num_joints: *num_joints
    trainsize: *trainsize
    pixel_std: *pixel_std
    center_scale: 0.4




EvalDataset:
  !KeypointTopDownCocoDataset
    image_dir: val2017
    anno_path: annotations/person_keypoints_val2017.json
    dataset_dir: dataset/coco
    num_joints: *num_joints
    trainsize: *trainsize
    pixel_std: *pixel_std
    image_thre: 0.0
    use_gt_bbox: True

TestDataset:
  !ImageFolder
    anno_path: dataset/coco/keypoint_imagelist.txt

worker_num: 4
global_mean: &global_mean [0.485, 0.456, 0.406]
global_std: &global_std [0.229, 0.224, 0.225]
TrainReader:
  sample_transforms:
    - RandomFlipHalfBodyTransform:
        scale: 0.5
        rot: 40
        num_joints_half_body: 8
        prob_half_body: 0.3
        pixel_std: *pixel_std
        trainsize: *trainsize
        upper_body_ids: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
        flip_pairs: *flip_perm

    - TopDownAffine:
        trainsize: *trainsize
        use_udp: true
    - ToHeatmapsTopDown_UDP:
        hmsize: *hmsize
        sigma: 2

  batch_transforms:
    - NormalizeImage:
        mean: *global_mean
        std: *global_std
        is_scale: true
    - Permute: {}
  batch_size: 64
  shuffle: True
  drop_last: True

EvalReader:
  sample_transforms:
    - TopDownAffine:
        trainsize: *trainsize
        use_udp: true
  batch_transforms:
    - NormalizeImage:
        mean: *global_mean
        std: *global_std
        is_scale: true
    - Permute: {}
  batch_size: 64

TestReader:
  inputs_def:
    image_shape: [3, *train_height, *train_width]
  sample_transforms:
    - Decode: {}
    - TopDownEvalAffine:
        trainsize: *trainsize
    - NormalizeImage:
        mean: *global_mean
        std: *global_std
        is_scale: true
    - Permute: {}
  batch_size: 1
  fuse_normalize: false