panoptic_deeplab_X_65_os16_mg124_poly_200k_bs64_crop_640_640_coco_dsconv.yaml

_BASE_: ../Cityscapes-PanopticSegmentation/Base-PanopticDeepLab-OS16.yaml
MODEL:
  WEIGHTS: "x65.pkl"
  PIXEL_MEAN: [128, 128, 128]
  PIXEL_STD: [128, 128, 128]
  BACKBONE:
    NAME: "d2_xception_65"
  RESNETS:
    NORM: "SyncBN"
    RES4_DILATION: 1
    RES5_DILATION: 2
  PANOPTIC_DEEPLAB:
    STUFF_AREA: 4096
    NMS_KERNEL: 41
    SIZE_DIVISIBILITY: 640
    USE_DEPTHWISE_SEPARABLE_CONV: True
  SEM_SEG_HEAD:
    NUM_CLASSES: 133
    LOSS_TOP_K: 1.0
    USE_DEPTHWISE_SEPARABLE_CONV: True
DATASETS:
  TRAIN: ("coco_2017_train_panoptic",)
  TEST: ("coco_2017_val_panoptic",)
SOLVER:
  BASE_LR: 0.0005
  MAX_ITER: 200000
  IMS_PER_BATCH: 64
INPUT:
  FORMAT: "RGB"
  GAUSSIAN_SIGMA: 8
  MIN_SIZE_TRAIN: !!python/object/apply:eval ["[int(x * 0.1 * 640) for x in range(5, 16)]"]
  MIN_SIZE_TRAIN_SAMPLING: "choice"
  MIN_SIZE_TEST: 640
  MAX_SIZE_TRAIN: 960
  MAX_SIZE_TEST: 640
  CROP:
    ENABLED: True
    TYPE: "absolute"
    SIZE: (640, 640)