ijepa_2/config.yaml

mode: ijepa

data:
  datasets_path: /home/lmk22/datasets
  train_dataset: imagenet
  batch_size: 128 # 2 gpu
  crop_size: 224
  num_workers: 32 # 2 gpu
  prefetch_factor: 1
  pin_memory: false
  drop_last: true
  crop_scale:
    - 0.3
    - 1.0
  crop_ratio:
    - 0.75
    - 1.3333333333333333
  normalize:
    mean:
      - 0.485
      - 0.456
      - 0.406
    std:
      - 0.229
      - 0.224
      - 0.225
  separate_val_subset:
    use: false
    size: 0.1
  
mask:
  target_aspect_ratio:
  - 0.75
  - 1.5
  context_mask_scale:
  - 0.85
  - 1.0
  min_context_patches: 10
  num_target_masks: 4
  patch_size: 16
  target_mask_scale:
  - 0.15
  - 0.2

meta:
  model_name: vit_base
  checkpoint: false
  predictor_depth: 6
  predictor_emb_dim: 384
  predictor_num_heads: 12
  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
  ema:
  - 0.996
  - 1.0
  lr:
  - 0.0001
  - 0.001
  - 0.000001
  weight_decay:
  - 0.04
  - 0.4
  epochs: 600
  warmup_epochs: 15
  optimizer: adamw
  criterion: l1_smooth_loss
  decay_bias: false
  decay_norm: false