config files about vit-small and vit-base

[stonewjf]

can you tell me vit-small and vit-base training configuration

[paritosh-101]

My config file for vit_small. Adjust the patch size according to the image dimensions:

data:
batch_size: 128
color_jitter_strength: 0.0
crop_scale:

• 0.3
• 1.0
  crop_size: 32
  image_folder:
  num_workers: 10
  pin_mem: true
  root_path:
  use_color_distortion: false
  use_gaussian_blur: false
  use_horizontal_flip: false
  logging:
  folder:
  write_tag: jepa
  mask:
  allow_overlap: false
  aspect_ratio:
• 0.75
• 1.5
  enc_mask_scale:
• 0.85
• 1.0
  min_keep: 4
  num_enc_masks: 1
  num_pred_masks: 2
  patch_size: 4
  pred_mask_scale:
• 0.15
• 0.2
  meta:
  copy_data: false
  load_checkpoint: false
  model_name: vit_small
  pred_depth: 8
  pred_emb_dim: 192
  read_checkpoint: null
  use_bfloat16: false
  optimization:
  ema:
• 0.996
• 1.0
  epochs: 100
  final_lr: 1.0e-06
  final_weight_decay: 0.4
  ipe_scale: 1.0
  lr: 0.001
  start_lr: 0.0002
  warmup: 40
  weight_decay: 0.04

[Oguzhanercan]

when I change the crop_size, an error occurs about masks, can you share your mask code?

[paritosh-101]

I did not make any changes to the mask code, just adjusted the patch size according to the crop size (which is equal to the img size in my case). You can see this file (scroll down to the bottom): "src/models/vision_transformer.py", to get a better idea of adjustments.