RT-K-Net-Cityscapes.yaml

MODEL:
  META_ARCHITECTURE: "RTKNet"
  # ImageNet pretrained RTFormer backbone weights
  WEIGHTS: "https://drive.google.com/file/d/1W_axRYN1UP7uLshDmQVBQvAcOIPzDyIV/view?usp=sharing"
  PIXEL_MEAN: [ 123.675, 116.280, 103.530 ]
  PIXEL_STD: [ 58.395, 57.120, 57.375 ]
  BACKBONE:
    FREEZE_AT: 0
    NAME: "build_rt_former_backbone"
  RT_FORMER_BACKBONE:
    VARIANT: "base"
    NORM: "SyncBN"
  FEATURE_MAP_GENERATOR:
    NAME: "RTFormerHead"
    INIT_FUNC: "rtknet.layers.weight_init.kaiming_init"
  RT_FORMER_HEAD:
    NORM: "SyncBN"
  SEM_SEG_HEAD:
    NAME: "RTKNetHead"
    NUM_CLASSES: 19
    NORM: "SyncBN"
    NUM_KERNEL_UPDATE_HEADS: 4
  TEST:
    OVERLAP_THRESHOLD: 0.6
DATASETS:
  TRAIN: ("cityscapes_fine_panoptic_train",)
  TEST: ("cityscapes_fine_panoptic_val",)
SOLVER:
  IMS_PER_BATCH: 32
  OPTIMIZER: "ADAMW"
  BASE_LR: 0.0002
  WEIGHT_DECAY: 0.05
  BACKBONE_MULTIPLIER: 1.0
  CLIP_GRADIENTS:
    ENABLED: True
    CLIP_TYPE: "full_model"
    CLIP_VALUE: 1.0
    NORM_TYPE: 2.0
  LR_SCHEDULER_NAME: "WarmupPolyLR"
  WARMUP_ITERS: 1000
  WARMUP_FACTOR: 0.001
  MAX_ITER: 90000
  AMP:
    ENABLED: True
INPUT:
  TRAIN_DATASET_MAPPER: "rtknet.data.PanopticDatasetMapper"
  TEST_DATASET_MAPPER: "detectron2.data.DatasetMapper"
  MIN_SIZE_TRAIN: !!python/object/apply:eval ["[int(x * 0.1 * 1024) for x in range(5, 21)]"]
  MIN_SIZE_TRAIN_SAMPLING: "choice"
  MAX_SIZE_TRAIN: 4096
  MIN_SIZE_TEST: 1024
  MAX_SIZE_TEST: 2048
  CROP:
    ENABLED: True
    TYPE: "absolute"
    SIZE: (512, 1024)
  FORMAT: "RGB"
TEST:
  AMP:
    ENABLED: True
  EVAL_PERIOD: 5000
DATALOADER:
  FILTER_EMPTY_ANNOTATIONS: True
  NUM_WORKERS: 8
VERSION: 2