configs/CLIP_fast_rcnn_swin_base_C4.yaml

_BASE_: "./Base-RCNN-C4.yaml"
MODEL:
  META_ARCHITECTURE: "CLIPFastRCNN" # "CLIPRCNN" # "GeneralizedRCNN"
  BACKBONE:
    NAME: "build_clip_swin" # "build_resnet_fpn_backbone"
    FREEZE_AT: 2
  TEXT_BACKBONE:
    NAME: "build_clip_swin_text_backbone"
  SPEC:
    EMBED_DIM: 512
    VISION:
      PATCH_SIZE: 4
      IN_CHANS: 3
      EMBED_DIM: 128
      DEPTHS: [ 2, 2, 18, 2 ]
      NUM_HEADS: [ 4, 8, 16, 32 ]
      WINDOW_SIZE: 7
      MLP_RATIO: 4.
      QKV_BIAS: True
      APE: False
      PATCH_NORM: True
      DROP_RATE: 0.0
      DROP_PATH_RATE: 0.2  
      OUT_FEATURES: ["stage2", "stage3", "stage4", "stage5"]   
    TEXT:
      NAME: 'transformer'
      TOKENIZER: clip
      CONTEXT_LENGTH: 77
      WIDTH: 512
      HEADS: 8
      LAYERS: 12
  WEIGHTS: "" # "detectron2://ImageNetPretrained/MSRA/R-50.pkl"
  MASK_ON: True
  RPN:
    HEAD_NAME: StandardRPNHead
    IN_FEATURES: ["stage4"]
  ROI_HEADS:
    NAME: "CLIPSwinROIHeads" # "Res5ROIHeads" # "StandardROIHeads"
    IN_FEATURES: ["stage4"]
    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: [0.485, 0.456, 0.406]
  PIXEL_STD: [0.229, 0.224, 0.225]
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)
  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"