model_config.yaml

TRAIN:
  USE_FP16: true
  FP16_TYPE: bfloat16
MODEL:
  IMAGE_SIZE:
  - 512
  - 512
  IMAGE_MEAN:
  - 0.485
  - 0.456
  - 0.406
  IMAGE_STD:
  - 0.229
  - 0.224
  - 0.225
  BACKBONE:
    TYPE: dinov3_vith16plus
    PRETRAINED_WEIGHTS: ''
    FROZEN_STAGES: -1
    DROP_PATH_RATE: 0.1
  DECODER:
    TYPE: sam
    DIM: 1024
    DEPTH: 6
    HEADS: 8
    MLP_DIM: 1024
    DIM_HEAD: 64
    LAYER_SCALE_INIT: 0.0
    DROP_RATE: 0.0
    ATTN_DROP_RATE: 0.0
    DROP_PATH_RATE: 0.0
    FFN_TYPE: origin
    ENABLE_TWOWAY: false
    REPEAT_PE: true
    FROZEN: false
    CONDITION_TYPE: cliff
    USE_INTRIN_CENTER: true
    DO_INTERM_PREDS: true
    DO_INTERM_SUP: true
    DO_KEYPOINT_TOKENS: true
    DO_HAND_DETECT_TOKENS: true
    KEYPOINT_TOKEN_UPDATE: v2
    KEYPOINT_TOKEN_UPDATE_COORD_EMB_USE_MLP: true
    DO_KEYPOINT3D_TOKENS: true
  PROMPT_ENCODER:
    ENABLE: true
    MAX_NUM_CLICKS: 2
    PROMPT_KEYPOINTS: mhr70
    FROZEN: false
    KEYPOINT_SAMPLER:
      TYPE: v1
      WORST_RATIO: 0.8
      KEYBODY_RATIO: 0.8
      NEGATIVE_RATIO: 0.1
      DUMMY_RATIO: 0.1
      DISTANCE_THRESH: 0.0001
    MASK_EMBED_TYPE: v2
    MASK_PROMPT: v1
  PERSON_HEAD:
    POSE_TYPE: mhr
    CAMERA_ENABLE: true
    CAMERA_TYPE: perspective
    ZERO_POSE_INIT: true
    ZERO_POSE_INIT_BODY_FACTOR: 1
  MHR_HEAD:
    MLP_DEPTH: 2
    MLP_CHANNEL_DIV_FACTOR: 1
  CAMERA_HEAD:
    MLP_DEPTH: 2
    MLP_CHANNEL_DIV_FACTOR: 1
    DEFAULT_SCALE_FACTOR_HAND: 10
  ENABLE_BODY: true
  ENABLE_HAND: true