GTR_TAO_DR2101.yaml

_BASE_: "Base-CenterNet2.yaml"
MODEL:
  FREEZE_TYPE: BackboneBottomup
  META_ARCHITECTURE: "GTRRCNN"
  BACKBONE:
    NAME: "build_p67_res2net_fpn_backbone"
  WEIGHTS: "output/C2_LVISCOCO_DR2101_4x.pth"
  RESNETS:
    DEPTH: 101
    WIDTH_PER_GROUP: 26
    DEFORM_ON_PER_STAGE: [False, False, True, True] # on Res4, Res5
    DEFORM_MODULATED: True
  PIXEL_MEAN: [123.675, 116.280, 103.530]
  PIXEL_STD: [58.395, 57.12, 57.375]
  CENTERNET:
    USE_DEFORMABLE: True
  ROI_BOX_HEAD:
    USE_SIGMOID_CE: True
    USE_FED_LOSS: True
    DELAY_CLS: True
  ROI_HEADS:
    IN_FEATURES: ["p3"]
    NAME: GTRROIHeads
    PROPOSAL_APPEND_GT: False
  ASSO_ON: True
  ASSO_HEAD:
    ASSO_THRESH: 0.1
    ASSO_THRESH_TEST: 0.4
    ASSO_WEIGHT: 1.0
    NO_POS_EMB: True
SOLVER:
  USE_CUSTOM_SOLVER: True
  MAX_ITER: 45000
  LR_SCHEDULER_NAME: "WarmupCosineLR"
  BASE_LR: 0.00005
  OPTIMIZER: ADAMW
  IMS_PER_BATCH: 4
  CLIP_GRADIENTS:
    ENABLED: True
    CLIP_TYPE: "full_model"
    CLIP_VALUE: 0.1
    NORM_TYPE: 2.0
  BACKBONE_MULTIPLIER: 0.1
INPUT:
  FORMAT: RGB
  CUSTOM_AUG: EfficientDetResizeCrop
  TRAIN_SIZE: 896
  VIDEO:
    TRAIN_LEN: 8
    TEST_LEN: 16
VIDEO_INPUT: True
DATASETS:
  TRAIN: ("lvis_v1_train+coco_box",)
  TEST: ('tao_val_v1',)
DATALOADER:
  SAMPLER_TRAIN: "TrainingSampler"
OUTPUT_DIR: "./output/GTR-TAO/auto"
FIND_UNUSED_PARAM: True