VMAEv2-ViTB-16x4.yaml

DATA:
  PATH_TO_DATA_DIR: "/path/to/ava"
  NUM_FRAMES: 16
  SAMPLING_RATE: 4
  INPUT_CHANNEL_NUM: [3]
  DATASETS: ['ava_v2.2']
  TRAIN_MIN_SCALES: [256, 320]
  TEST_MIN_SCALES: [256]
AVA:
  ANNOTATION_DIR: "annotations_v2.2/"
MODEL:
  WEIGHT: "/path/to/weight.pth"
  BACKBONE:
    CONV_BODY: "MAE-ViT-B"
    PATHWAYS: 1
  STM:
    NUM_QUERIES: 100
    HIDDEN_DIM: 256
    NUM_STAGES: 6
    ACTION_CLASSES: 80
    OBJECT_CLASSES: 1
    NUM_HEADS: 8
    DROPOUT: 0.0
    DIM_FEEDFORWARD: 2048
    NUM_FCS: 2
    ACTIVATION: 'ReLU'
    SPATIAL_POINTS: 32
    TEMPORAL_POINTS: 8
    OUT_MULTIPLIER: 4
    N_GROUPS: 4
    NUM_CLS: 1
    NUM_ACT: 1
    NUM_REG: 1
    OBJECT_WEIGHT: 2.0
    ACTION_WEIGHT: 48.0
    GIOU_WEIGHT: 2.0
    L1_WEIGHT: 2.0
    BACKGROUND_WEIGHT: 0.1
    INTERMEDIATE_SUPERVISION: True
    PERSON_THRESHOLD: 0.6
    MEM_ACTIVE: False
ViT:
  EMBED_DIM: 768
  DEPTH: 12
  USE_CHECKPOINT: True
  LAYER_DECAY: 1.0
  WEIGHT_DECAY: 1e-4
  DROP_PATH_RATE: 0.2
SOLVER:
  ITER_PER_EPOCH: 23048
  MAX_EPOCH: 12
  BASE_LR: 0.00001
  WEIGHT_DECAY: 1e-4
  STEPS: (5, 8)
  WARMUP_FACTOR: 0.1
  WARMUP_EPOCH: 2
  CHECKPOINT_PERIOD: 1
  EVAL_PERIOD: 1
  EVAL_AFTER: 3
  VIDEOS_PER_BATCH: 8
  OPTIMIZING_METHOD: 'adamw'
TEST:
  VIDEOS_PER_BATCH: 16
OUTPUT_DIR: "data/output/vitbv2-train"