LVIS_OVD_Base_PIS.yaml

_BASE_: "mask_rcnn_R50FPN_CLIP_sigmoid.yaml"
MODEL:
  META_ARCHITECTURE: "CustomRCNNMViT"
  WEIGHTS: saved_models/lvis_ovd_base.pth
  PIS_PROP_PATH: '' # datasets/MAVL_proposals/lvis_props/class_specific/imagenet_lvis_props (if pseudo props must be loaded from files)
  ROI_BOX_HEAD:
    WEIGHT_TRANSFER: False
    ZEROSHOT_WEIGHT_PATH: 'datasets/zeroshot_weights/lvis_v1_clip_a+photo+cname.npy'
    CAT_FREQ_PATH: 'datasets/lvis/lvis_v1_train_norare_cat_info.json'
    IMAGE_LABEL_LOSS: 'pseudo_max_score'
    USE_BIAS: -4.6
INPUT:
  CUSTOM_AUG: "ResizeShortestEdge"
DATASETS:
  TRAIN: ("lvis_v1_train_norare", "imagenet_lvis_v1_pis")
  TEST: ("lvis_v1_val",)
TEST:
  DETECTIONS_PER_IMAGE: 300
SOLVER:
  BASE_LR: 0.02
  STEPS: (60000, 80000)
  MAX_ITER: 90000
  WARMUP_FACTOR: 0.001
  WARMUP_ITERS: 1000
DATALOADER:
  SAMPLER_TRAIN: "MultiDatasetSampler"
  DATASET_RATIO: [ 1, 4 ]
  USE_DIFF_BS_SIZE: True
  DATASET_BS: [2, 8] # Batch-size per GPU
  USE_RFS: [True, False]
  DATASET_MIN_SIZES: [[640, 672, 704, 736, 768, 800], [400, 400]]
  DATASET_MAX_SIZES: [1333, 667]
  FILTER_EMPTY_ANNOTATIONS: False
  MULTI_DATASET_GROUPING: True
  DATASET_ANN: ['box', 'image']
  NUM_WORKERS: 8
WITH_IMAGE_LABELS: True
OUTPUT_DIR: output/lvis_ovd_PIS