mae/config.yaml

mode: mae

data:
  datasets_path: /home/lmk22/datasets
  train_dataset: imagenet
  batch_size: 256 # For each device, so the effective batch size is batch_size * num_devices
  crop_size: 224
  num_workers: 32 # For each device, so the effective number of workers is num_workers * num_devices
  prefetch_factor: 1
  pin_memory: false
  drop_last: true
  normalize:
    mean:
      - 0.485
      - 0.456
      - 0.406
    std:
      - 0.229
      - 0.224
      - 0.225
  random_resized_crop:
    use: true
    scale:
      - 0.2
      - 1.0
    ratio:
      - 0.75
      - 1.3333
  random_horizontal_flip:
    use: true
    p: 0.5
  separate_val_subset:
    use: false
    size: 0.1
  
mask:
  mask_ratio: 0.75 # Percentage of removed masks  

meta:
  model_name: vit_base
  checkpoint: false
  pretrained_weights: null # Path to the pretrained weights file. Set to null if you don't want to load pretrained weights.
  save_every: 100 # Set to 0 if you don't want to save intermediate checkpoints during training

optimization:
  ipe_scale: 1.0
  lr:
  - 0.00001
  - 0.00003
  - 0.0
  weight_decay:
  - 0.05
  - 0.05
  epochs: 800
  warmup_epochs: 40
  optimizer: adamw