configs/vitb_r101_384.yaml

_BASE_: config.yaml
MODEL:
  META_ARCHITECTURE: "CATSeg"
  BACKBONE:
    FREEZE_AT: 0
    NAME: "build_resnet_backbone"
  WEIGHTS: "R-101.pkl"
  RESNETS:
    DEPTH: 101
    STEM_TYPE: "basic" 
    STEM_OUT_CHANNELS: 64
    STRIDE_IN_1X1: False
    OUT_FEATURES: ["res2", "res3", "res4"]
  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-B/16"
    PROMPT_DEPTH: 0
    PROMPT_LENGTH: 0
    TEXT_AFFINITY_DIM: 512
    TEXT_AFFINITY_PROJ_DIM: 128
    APPEARANCE_AFFINITY_DIM: 1024
    APPEARANCE_AFFINITY_PROJ_DIM: 128
    DECODER_DIMS: [64, 32]
    DECODER_AFFINITY_DIMS: [512, 256]
    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"
SOLVER:
  BACKBONE_MULTIPLIER: 0.01