NetworkConfig.yml

common:
  data: /path/to/dataset/NTU-RGBD
  splits: /path/to/dataset/dataset/NTU-RGBD/dataset_splits/@CS # include: train.txt and test.txt

  #-------basic Hyparameter----------
  visdom:
    enable: True
    visname: NTU
  dataset: NTU #Database name e.g., NTU, THUREAD, NvGesture and IsoGD ...
  batch_size: 6
  test_batch_size: 6
  num_workers: 6
  learning_rate: 0.01
  learning_rate_min: 0.00001
  momentum: 0.9
  weight_decay: 0.0003
  init_epochs: 0
  epochs: 100 # if training on IsoGD dataset, set 300 is better.
  report_freq: 100
  optim: SGD
  dist: True
  vis_feature: True # Visualization?

  scheduler:
    name: cosin
    patience: 4
    warm_up_epochs: 3
  loss:
    name: CE
    labelsmooth: True
  MultiLoss: True
  loss_lamdb: [ 1, 0.5, 0.5, 0.5 ]
  distill: 1.

model:
  Network: I3DWTrans # e.g., I3DWTrans or FusionNet
  pretrained: '' # all of experiments are pre-trained on 20BN Jester V1 dataset except for NTU-RGBD.
  resume: ''
  resumelr: False
  sample_duration: 64
  sample_size: 224
  grad_clip: 5.
  SYNC_BN: 1
  w: 10
  temper: 0.5 # 0.5 for THUREAD and NTU-RGBD; 0.4 for NvGesture and IsoGD
  recoupling: True
  knn_attention: 0.7
  sharpness: True
  temp: [ 0.04, 0.07 ]
  frp: True
  SEHeads: 1
  N: 6 # Number of Transformer Blocks

  #-------Used for fusion network----------
  rgb_checkpoint: ''
  depth_checkpoint: ''

dataset:
  type: M # M: rgb, K: depth
  flip: 0.5 # set 0.0 for NvGesture and IsoGD
  rotated: 0.5 # THUREAD: 0.8, others: 0.5
  angle: (-10, 10) # Rotation angle. THUREAD: (-45, 45), others: (-10, 10)
  Blur: False
  resize: (320, 240) #NTU and THUREAD: (320, 240), others:(256, 256)
  crop_size: 224 # THUREAD: 200, others: 224
  low_frames: 16
  media_frames: 32
  high_frames: 48