swinl_coco.yaml

MODEL:
  META_ARCHITECTURE: "MDQE"
  WEIGHTS: "pretrained/imagenet/swinv2_large_patch4_window12_192_22k_d2.pth"
  PIXEL_MEAN: [ 123.675, 116.280, 103.530 ]
  PIXEL_STD: [ 58.395, 57.120, 57.375 ]
  MASK_ON: True
  BACKBONE:
    NAME: "build_swinv2_backbone"
  SWIN:
    EMBED_DIM: 192
    DEPTHS: [ 2, 2, 18, 2 ]
    NUM_HEADS: [ 6, 12, 24, 48 ]
    WINDOW_SIZE: 12
    MLP_RATIO: 4
    DROP_PATH_RATE: 0.2
    APE: False
  MDQE:
    NUM_OBJECT_QUERIES: 200
    MLP_RATIO: 4
    ENC_LAYERS: 6
    DEC_LAYERS: 6
    NUM_FEATURE_LEVELS: 4
    DEC_TEMPORAL: True
    HIDDEN_DIM: 192
    NUM_CLASSES: 80
    QUERY_EMBED_DIM: 64
DATASETS:
  TRAIN: ("coco_2017_train",)
  TEST: ("coco_2017_val",)
SOLVER:
  IMS_PER_BATCH: 8
  BASE_LR: 0.00005
  STEPS: (472000,)  # 3x on 118k images in CoCo training data
  MAX_ITER: 534000
  WARMUP_FACTOR: 1.0
  WARMUP_ITERS: 10
  WEIGHT_DECAY: 0.05
  OPTIMIZER: "ADAMW"
  BACKBONE_MULTIPLIER: 0.1
  CLIP_GRADIENTS:
    ENABLED: True
    CLIP_TYPE: "full_model"
    CLIP_VALUE: 0.01
    NORM_TYPE: 2.0
INPUT:
  FORMAT: "RGB"
  SAMPLING_FRAME_NUM: 1
  AUGMENTATIONS: []
  RANDOM_FLIP: "flip_by_clip"
  MIN_SIZE_TRAIN_SAMPLING: "choice_by_clip"
  MIN_SIZE_TRAIN: (352, 384, 416, 448, 480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800)
  MAX_SIZE_TRAIN: 1333
  MIN_SIZE_TEST: 800
  CROP:
    ENABLED: True
    TYPE: "absolute_range"
    SIZE: (384, 600)
TEST:
  EVAL_PERIOD: 5000
  DETECTIONS_PER_IMAGE: 100
DATALOADER:
  FILTER_EMPTY_ANNOTATIONS: False
  NUM_WORKERS: 4
VERSION: 2
OUTPUT_DIR: output/coco/mdqe_swinl_patch4_window24_384_22k_coco_bs8_3x_f1/