configs/CLIP_fast_rcnn_R_50_C4.yaml

_BASE_: "./Base-RCNN-C4.yaml"
MODEL:
  META_ARCHITECTURE: "CLIPFastRCNN" # "CLIPRCNN" # "GeneralizedRCNN"
  MASK_ON: False
  WEIGHTS: "./regionclip_r50x4.pth"
  BACKBONE:
    NAME: "build_clip_resnet_backbone" # "build_resnet_fpn_backbone"
    FREEZE_AT: 2
  TEXT_BACKBONE:
    NAME: "build_clip_language_encoder"    
  CLIP:
    CROP_REGION_TYPE: "RPN"
    OFFLINE_RPN_CONFIG: "./configs/mask_rcnn_R_50_FPN_1x.yaml"
    USE_TEXT_EMB_CLASSIFIER: True
    TEXT_EMB_PATH: "./lvis_1203_cls_emb_notnorm_rn50x4.pth"
    NO_BOX_DELTA: True
    OFFLINE_RPN_NMS_THRESH: 0.7
    CLSS_TEMP: 0.01
    MULTIPLY_RPN_SCORE: True
    TEXT_EMB_DIM: 640
  RESNETS:
    DEPTH: 200
    OUT_FEATURES: ["res4"]
    NORM: FrozenBN
    STEM_OUT_CHANNELS: 64
    RES2_OUT_CHANNELS: 256
  RPN:
    HEAD_NAME: StandardRPNHead
    IN_FEATURES: ["res4"]
    POST_NMS_TOPK_TEST: 500
    NMS_THRESH: 
  ROI_HEADS:
    NAME: "CLIPRes5ROIHeads" # "Res5ROIHeads" # "StandardROIHeads"
    IN_FEATURES: ["res4"]
    NUM_CLASSES: 1203
    NMS_THRESH_TEST: 0.3
    SCORE_THRESH_TEST: 0.0
  ROI_BOX_HEAD:
    NAME: ""
    NUM_FC: 0
    CLS_AGNOSTIC_BBOX_REG: True
    POOLER_RESOLUTION: 18
  ROI_MASK_HEAD:
    NAME: "MaskRCNNConvUpsampleHead"
    NUM_CONV: 0
    POOLER_RESOLUTION: 14
  PIXEL_MEAN: [0.48145466, 0.4578275, 0.40821073]
  PIXEL_STD: [0.26862954, 0.26130258, 0.27577711]
INPUT:
  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)
DATASETS:
  TRAIN: ("lvis_v1_train",)
  TEST: ("lvis_v1_val",)
TEST:
  DETECTIONS_PER_IMAGE: 10  # 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"