configs/coco/instance-segmentation/swin/opd_base.yaml

_BASE_: ../maskformer2_R50_bs16_50ep.yaml

INPUT:
  FORMAT: "RGB"
  IMAGE_SIZE: 256
  MAX_SIZE_TEST: 256
  MAX_SIZE_TRAIN: 256
  MIN_SIZE_TEST: 256
  MIN_SIZE_TRAIN:
  - 256
  # DATASET_MAPPER_NAME: "motion_instance"

DATALOADER:
  NUM_WORKERS: 4

DATASETS:
  TRAIN: ("MotionNet_train",)
  TEST: ("MotionNet_valid",)

MODEL:
  MOTIONNET:
    TYPE: BMOC_V0
  SEM_SEG_HEAD:
    NUM_CLASSES: 3
  MASK_ON: True # Useful for our MotionEvaluator, because it's from an older version detectron2
  MASK_FORMER:
    TRANSFORMER_DECODER_NAME: OPDMultiScaleMaskedTransformerDecoder
    CLASS_WEIGHT: 2.0
    MASK_WEIGHT: 5.0
    DICE_WEIGHT: 5.0
    MTYPE_WEIGHT: 2.0
    MORIGIN_WEIGHT: 16.0
    MAXIS_WEIGHT: 16.0
    MSTATE_WEIGHT: 16.0
    MSTATEMAX_WEIGHT: 16.0
    EXTRINSIC_WEIGHT: 30.0

SOLVER:
  IMS_PER_BATCH: 16
  BASE_LR: 0.0001
  STEPS: (36000, 48000)
  MAX_ITER: 60000
  CHECKPOINT_PERIOD: 10000

TEST:
  AUG:
    ENABLED: false
    FLIP: false
  EVAL_PERIOD: 10000
SEED: 42