dd3d_nusc_v99.yaml

# @package _global_
defaults:
  - /evaluators/nuscenes@EVALUATORS.NUSCENES
  - override /meta_arch: dd3d
  - override /train_datasets@DATASETS.TRAIN: nuscenes
  - override /test_datasets@DATASETS.TEST: nuscenes
  - override /feature_extractors@FE: v2_99_fpn

MODEL:
  META_ARCHITECTURE: NuscenesDD3D
  # from-coco, IODA-pretrained.
  CKPT: https://tri-ml-public.s3.amazonaws.com/github/dd3d/pretrained/depth_pretrained_v99-3jlw0p36-20210423_010520-model_final-remapped.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

  NUSC:
    LOSS:
      WEIGHT_ATTR: 0.2
      WEIGHT_SPEED: 0.2
    INFERENCE:
      NUM_IMAGES_PER_SAMPLE: ${DATASETS.TEST.NUM_IMAGES_PER_SAMPLE}
      MAX_NUM_DETS_PER_SAMPLE: 500

INPUT:
  RESIZE:
    # Nuscens images are (900, 1600)
    MIN_SIZE_TRAIN: [640, 672, 704, 736, 768, 800, 832, 864, 896, 928, 960, 992, 1024, 1056, 1088, 1120, 1152]
    MAX_SIZE_TRAIN: 10000
    MIN_SIZE_TEST: 896
    MAX_SIZE_TEST: 100000

SOLVER:
  IMS_PER_BATCH: 64 # need at least 400 GPU mem (with fp16).
  BASE_LR: 0.002
  MAX_ITER: 120000
  STEPS: [100000, 115000]
  WARMUP_ITERS: 2000
  MIXED_PRECISION_ENABLED: True
  CHECKPOINT_PERIOD: 2000

TEST:
  IMS_PER_BATCH: 192 # 6 * 32 (must be multiple of 6 x #GPUs.)
  EVAL_PERIOD: 2000
  AUG:
    ENABLED: True
    MIN_SIZES: [640, 768, 896, 1024, 1152]
    MAX_SIZE: 100000
    FLIP: True

DATALOADER:
  TRAIN:
    NUM_WORKERS: 12
    SAMPLER: RepeatFactorTrainingSampler
    REPEAT_THRESHOLD: 0.8
  TEST:
    SAMPLER: InferenceGroupSampler
    NUM_IMAGES_PER_GROUP: ${DATASETS.TEST.NUM_IMAGES_PER_SAMPLE}

WANDB:
  TAGS: [nusc-val, v2-99, bn]