dd3d_kitti_omninets.yaml

# @package _global_
defaults:
  - /evaluators/kitti_3d@EVALUATORS.KITTI3D
  - override /meta_arch: dd3d
  - override /train_datasets@DATASETS.TRAIN: kitti_3d
  - override /test_datasets@DATASETS.TEST: kitti_3d
  - override /feature_extractors@FE: dla34_fpn

MODEL:
  # from-coco, IODA-pretrained.
  backbone_with_fpn: 
  width_mult: 1.0
  depth_mult: 1.0

CKPT: https://tri-ml-public.s3.amazonaws.com/github/dd3d/pretrained/depth_pretrained_omninet-small-3nxjur71.pth

FE:
  BACKBONE:
    NORM: FrozenBN
  FPN:
    NORM: FrozenBN
  OUT_FEATURES: ${.FPN.OUT_FEATURES}

DD3D:
  FCOS2D:
    NORM: BN
    INFERENCE:
      NMS_THRESH: 0.75

  FCOS3D:
    NORM: FrozenBN

INPUT:
  RESIZE:
    # KITTI images are (370, 1224)
    MIN_SIZE_TRAIN: [288, 304, 320, 336, 352, 368, 384, 400, 416, 448, 480, 512, 544, 576]
    MAX_SIZE_TRAIN: 10000
    MIN_SIZE_TEST: 384
    MAX_SIZE_TEST: 100000

SOLVER:
  IMS_PER_BATCH: 64 # need at least 128 GPU mem (with fp16).
  BASE_LR: 0.002
  MAX_ITER: 25000
  STEPS: [21500, 24000]
  WARMUP_ITERS: 2000
  MIXED_PRECISION_ENABLED: True
  CHECKPOINT_PERIOD: 2000

TEST:
  IMS_PER_BATCH: 80
  EVAL_PERIOD: 2000
  AUG:
    ENABLED: True
    MIN_SIZES: [320, 384, 448, 512, 576]
    MAX_SIZE: 100000
    FLIP: True

DATALOADER:
  TRAIN:
    NUM_WORKERS: 12
    SAMPLER: RepeatFactorTrainingSampler
    REPEAT_THRESHOLD: 0.4

WANDB:
  TAGS: [kitti-val, dla34, bn]