config.yaml

TASK: classification
DATA:
  dataset_name: mvfouls
  data_dir: /home/vorajv/opensportslib/SoccerNet/mvfouls
  data_modality: video
  view_type: multi  # multi or single
  num_classes: 8 # mvfoul
  train: 
    type: annotations_train.json
    video_path: ${DATA.data_dir}/train
    path: ${DATA.train.video_path}/annotations-train.json
    dataloader:
      batch_size: 8
      shuffle: true
      num_workers: 4
      pin_memory: true
  valid:
    type: annotations_valid.json
    video_path: ${DATA.data_dir}/valid
    path: ${DATA.valid.video_path}/annotations-valid.json
    dataloader:
      batch_size: 1
      num_workers: 1
      shuffle: false
  test:
    type: annotations_test.json
    video_path: ${DATA.data_dir}/test
    path: ${DATA.test.video_path}/annotations-test.json
    dataloader:
      batch_size: 1
      num_workers: 1
      shuffle: false
  num_frames: 16               # 8 before + 8 after the foul
  input_fps: 25                # Original FPS of video
  target_fps: 17               # Temporal downsampling to 1s clip (approx)
  start_frame: 63            # Start frame of clip relative to foul frame
  end_frame: 87              # End frame of clip relative to foul frame
  frame_size: [224, 224]       # Spatial resolution (HxW)
  augmentations:
    random_affine: true
    translate: [0.1, 0.1]        
    affine_scale: [0.9, 1.0]     
    random_perspective: true
    distortion_scale: 0.3        
    perspective_prob: 0.5
    random_rotation: true
    rotation_degrees: 5          
    color_jitter: true
    jitter_params: [0.2, 0.2, 0.2, 0.1]   # brightness, contrast, saturation, hue
    random_horizontal_flip: true
    flip_prob: 0.5
    random_crop: false

MODEL:
  type: custom # huggingface, custom 
  backbone: 
    type: mvit_v2_s # video_mae, r3d_18, mc3_18, r2plus1d_18, s3d, mvit_v2_s
  neck:
    type: MV_Aggregate
    agr_type: max   # max, mean, attention
  head: 
    type: MV_LinearLayer
  pretrained_model: mvit_v2_s # MCG-NJU/videomae-base, OpenGVLab/VideoMAEv2-Base, r3d_18, mc3_18, r2plus1d_18, s3d, mvit_v2_s
  unfreeze_head: true  # for videomae backbone
  unfreeze_last_n_layers: 3 # for videomae backbone
    

TRAIN:
  monitor: balanced_accuracy # balanced_accuracy, loss
  mode: max # max or min
  enabled: true
  use_weighted_sampler: false
  use_weighted_loss: true
  epochs: 20 #20
  log_interval: 10
  save_every: 2 #5

  criterion:
    type: CrossEntropyLoss

  optimizer:
    type: AdamW
    lr: 0.0001  #0.001
    backbone_lr: 0.00005
    head_lr: 0.001
    betas: [0.9, 0.999]
    eps: 0.0000001
    weight_decay: 0.001 #0.01 - videomae, 0.001 - others
    amsgrad: false
  
  scheduler:
    type: StepLR
    step_size: 3
    gamma: 0.1

SYSTEM:
  log_dir: ./logs
  save_dir: ./checkpoints
  use_seed: false
  seed: 42
  GPU: 4
  device: cuda   # auto | cuda | cpu
  gpu_id: 0