_BASE_: "./Base-RCNN-C4.yaml"
MODEL:
  META_ARCHITECTURE: "GeneralizedRCNN"
  BACKBONE:
    NAME: "build_clip_resnet_backbone" #"build_clip_resnet_fpn_backbone" # "build_resnet_fpn_backbone"
    FREEZE_AT: 2
  WEIGHTS: "" # "detectron2://ImageNetPretrained/MSRA/R-50.pkl"
  MASK_ON: True
  RESNETS:
    DEPTH: 50
    OUT_FEATURES: ["res4"]
    NORM: FrozenBN
    STEM_OUT_CHANNELS: 64
    RES2_OUT_CHANNELS: 256
  RPN:
    HEAD_NAME: StandardRPNHead
    IN_FEATURES: ["res4"]
  ROI_HEADS:
    NAME: "CLIPRes5ROIHeads" # "Res5ROIHeads" # "StandardROIHeads"
    IN_FEATURES: ["res4"]
    NUM_CLASSES: 1203
    SCORE_THRESH_TEST: 0.0001
  ROI_BOX_HEAD:
    NAME: ""
    NUM_FC: 0
    POOLER_RESOLUTION: 14
  ROI_MASK_HEAD:
    NAME: "MaskRCNNConvUpsampleHead"
    NUM_CONV: 0
    POOLER_RESOLUTION: 14
  PIXEL_MEAN: [0.48145466, 0.4578275, 0.40821073] # [103.530, 116.280, 123.675] # 
  PIXEL_STD: [0.26862954, 0.26130258, 0.27577711] # [1.0, 1.0, 1.0] # 
INPUT:
  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)
DATASETS:
  TRAIN: ("lvis_v1_train",)
  TEST: ("lvis_v1_val",)
TEST:
  DETECTIONS_PER_IMAGE: 300  # LVIS allows up to 300
  EVAL_PERIOD: 25000
SOLVER:
  IMS_PER_BATCH: 16
  BASE_LR: 0.02
  STEPS: (120000, 160000) # (140000,)  #
  MAX_ITER: 180000  # 180000 * 16 / 100000 ~ 28.8 epochs
DATALOADER:
  SAMPLER_TRAIN: "RepeatFactorTrainingSampler"
  REPEAT_THRESHOLD: 0.001
INPUT:
  MIN_SIZE_TRAIN_SAMPLING: choice
  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)
  MAX_SIZE_TRAIN: 1333
  MIN_SIZE_TEST: 800
  MAX_SIZE_TEST: 1333
  FORMAT: "RGB" # "BGR"