_BASE_: config.yaml
MODEL:
  META_ARCHITECTURE: "CATSeg"
  BACKBONE:
    FREEZE_AT: 0
    NAME: "D2SwinTransformer"
  SWIN:
    EMBED_DIM: 128
    DEPTHS: [2, 2, 18]
    NUM_HEADS: [4, 8, 16]
    WINDOW_SIZE: 12
    APE: False
    DROP_PATH_RATE: 0.3
    PATCH_NORM: True
    PRETRAIN_IMG_SIZE: 384
    OUT_FEATURES: ["res2", "res3", "res4"]
  WEIGHTS: "swin_base_patch4_window12_384_22k.pkl"
  PIXEL_MEAN: [123.675, 116.280, 103.530]
  PIXEL_STD: [58.395, 57.120, 57.375]
  SEM_SEG_HEAD:
    NAME: "CATSegHead"
    IN_FEATURES: ["res2", "res3", "res4"]
    IGNORE_VALUE: 255
    NUM_CLASSES: 171
    TRAIN_CLASS_JSON: "datasets/coco.json"
    TEST_CLASS_JSON: "datasets/coco.json"
    CLIP_PRETRAINED: "ViT-L/14@336px"
    PROMPT_DEPTH: 0
    PROMPT_LENGTH: 0
    TEXT_AFFINITY_DIM: 768
    TEXT_AFFINITY_PROJ_DIM: 128
    APPEARANCE_AFFINITY_DIM: 512
    APPEARANCE_AFFINITY_PROJ_DIM: 128
    DECODER_DIMS: [64, 32]
    DECODER_AFFINITY_DIMS: [256, 128]
    DECODER_AFFINITY_PROJ_DIMS: [32, 16]
    NUM_LAYERS: 2
    NUM_HEADS: 4
    HIDDEN_DIMS: 128
    POOLING_SIZES: [2, 2]
    FEATURE_RESOLUTION: [24, 24]
    WINDOW_SIZES: 12
    ATTENTION_TYPE: "linear"
    CLIP_FINETUNE: "attention"
  PROMPT_ENSEMBLE_TYPE: "imagenet"
INPUT:
  MIN_SIZE_TRAIN: (384, )
  MIN_SIZE_TRAIN_SAMPLING: "choice"
  MIN_SIZE_TEST: 640
  CROP:
    ENABLED: True
    TYPE: "absolute"
    SIZE: (384, 384)
  SIZE_DIVISIBILITY: 384 
  FORMAT: "RGB"
  DATASET_MAPPER_NAME: "mask_former_semantic"
SOLVER:
  IMS_PER_BATCH: 4 
  LR_SCHEDULER_NAME: WarmupCosineLR
  BASE_LR: 0.0002
  MAX_ITER: 80000
  BACKBONE_MULTIPLIER: 0.0
  CLIP_MULTIPLIER: 0.01
TEST:
  EVAL_PERIOD: 5000