unilm/layoutlmv3/examples/object_detection/cascade_layoutlmv3.yaml

MODEL:
  MASK_ON: True
  IMAGE_ONLY: True
  META_ARCHITECTURE: "VLGeneralizedRCNN"
  PIXEL_MEAN: [ 127.5, 127.5, 127.5 ]
  PIXEL_STD: [ 127.5, 127.5, 127.5 ]
  WEIGHTS: "/path/to/models/layoutlmv3/pts/layoutlmv3-base/pytorch_model.bin"
  BACKBONE:
    NAME: "build_vit_fpn_backbone"
  VIT:
    NAME: "layoutlmv3_base"
    OUT_FEATURES: [ "layer3", "layer5", "layer7", "layer11" ]
    DROP_PATH: 0.1
    IMG_SIZE: [ 224,224 ]
    POS_TYPE: "abs"
  ROI_HEADS:
    NAME: CascadeROIHeads
    IN_FEATURES: [ "p2", "p3", "p4", "p5" ]
    NUM_CLASSES: 5
  ROI_BOX_HEAD:
    CLS_AGNOSTIC_BBOX_REG: True
    NAME: "FastRCNNConvFCHead"
    NUM_FC: 2
    POOLER_RESOLUTION: 7
  ROI_MASK_HEAD:
    NAME: "MaskRCNNConvUpsampleHead"
    NUM_CONV: 4
    POOLER_RESOLUTION: 14
  FPN:
    IN_FEATURES: [ "layer3", "layer5", "layer7", "layer11" ]
  ANCHOR_GENERATOR:
    SIZES: [ [ 32 ], [ 64 ], [ 128 ], [ 256 ], [ 512 ] ]  # One size for each in feature map
    ASPECT_RATIOS: [ [ 0.5, 1.0, 2.0 ] ]  # Three aspect ratios (same for all in feature maps)
  RPN:
    IN_FEATURES: [ "p2", "p3", "p4", "p5", "p6" ]
    PRE_NMS_TOPK_TRAIN: 2000  # Per FPN level
    PRE_NMS_TOPK_TEST: 1000  # Per FPN level
    # Detectron1 uses 2000 proposals per-batch,
    # (See "modeling/rpn/rpn_outputs.py" for details of this legacy issue)
    # which is approximately 1000 proposals per-image since the default batch size for FPN is 2.
    POST_NMS_TOPK_TRAIN: 2000
    POST_NMS_TOPK_TEST: 1000
DATASETS:
  TRAIN: ("publaynet_train",)
  TEST: ("publaynet_val",)
SOLVER:
  GRADIENT_ACCUMULATION_STEPS: 1
  BASE_LR: 0.0002
  WARMUP_ITERS: 1000
  IMS_PER_BATCH: 32
  MAX_ITER: 60000
  CHECKPOINT_PERIOD: 2000
  LR_SCHEDULER_NAME: "WarmupCosineLR"
  AMP:
    ENABLED: True
  OPTIMIZER: "ADAMW"
  BACKBONE_MULTIPLIER: 1.0
  CLIP_GRADIENTS:
    ENABLED: True
    CLIP_TYPE: "full_model"
    CLIP_VALUE: 1.0
    NORM_TYPE: 2.0
  WARMUP_FACTOR: 0.01
  WEIGHT_DECAY: 0.05
TEST:
  EVAL_PERIOD: 2000
INPUT:
  CROP:
    ENABLED: True
    TYPE: "absolute_range"
    SIZE: (384, 600)
  MIN_SIZE_TRAIN: (480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800)
  FORMAT: "RGB"
DATALOADER:
  FILTER_EMPTY_ANNOTATIONS: False
VERSION: 2
AUG:
  DETR: True
SEED: 42
OUTPUT_DIR: "/path/to/models/layoutlmv3/fts/publaynet-base/"
PUBLAYNET_DATA_DIR_TRAIN: "/path/to/data/PubLayNet/publaynet/train"
PUBLAYNET_DATA_DIR_TEST: "/path/to/data/PubLayNet/publaynet/val"
CACHE_DIR: "/path/to/cache/huggingface"