test_single_depth.yaml

feeder: feeder.dataloader_joint.JointFeeder
phase: test
num_epoch: 40
work_dir: ./work_dir/test_single_depth
batch_size: 16
random_seed: 1234
test_batch_size: 16
num_worker: 4
device: 0
log_interval: 100
eval_interval: 1
save_interval: 5
load_weights: './weights/single_depth.pt'
optimizer_args:
  optimizer: Adam
  learning_rate:
    default: 0.00001
    head: 0.0001
  optim_args:
    betas: [0.9, 0.998]
    weight_decay: 0.00001
  scheduler_type: MultiStepLR
  scheduler_args:
    milestones: [20, 35]
  start_epoch: 0


feeder_args:
  train:
    input_list_file: ./data/train_self.json
    num_frames: 32
    data_type: ['d_features']
  val:
    input_list_file: ./data/val_self.json
    num_frames: 32
    data_type: ['d_features']
  # test only used for generate submission
  test:
    num_frames: 32
    data_type: ['d_features']
  kps_config: &kps_path ./configs/kps/best_kps_config.yaml

model: modules.islr_model.ISLRModel
model_args:
  visual_backbone_args:    
    depth:
      type: vit_base_patch16_224
      stride: 8
  
  head_args:
    classify:
      type: ClassifyHead
      in_dim: 768
      # specify the feat used for classification
      feat_type: depth
      norm_classifier: True
      norm_scale: 32
      class_num: 1000 # this will be overwrited by len(gloss_dict)
  temporal_arg:
    type: Normal
    use_loc_loss: True
    
  loss_weight:
    classify: 1.0
  
  # use which head for prediction
  pred_head: classify
  temporal_arg:
    type: Normal
    use_loc_loss: true
  
aug_poss:
  osxposs: 0.3