pretrain_FastVimL.yaml

accum_iter: 1  #modify to keep overall batch size 4096
batch_size: 128
num_workers: 12
num_nodes: 1

pl_seed: 0
img_size: 224
patch_size: 16
norm_pix_loss: true
mask_ratio: 0.75
in_chans: 3

scanpath_type: "rowwise"
rotate_every_block: True
collapse_method: 'mean'
use_norm_after_ssm: True

training_epochs: 1600
warmup_epochs: 40
blr: 0.00015
min_lr: 0
weight_decay: 0.05

vim_config:
  _target_: models.mae.fastvim_mae_vimdecoder.mae_FastVim_large_dec512d2b
  # _target_: models.mae.fastvim_mae_vimdecoder_v2.mae_FastVim_large_dec512d2b
  img_size: ${img_size}
  patch_size: ${patch_size}
  stride: ${patch_size}
  in_chans: ${in_chans}
  norm_pix_loss: ${norm_pix_loss}
  scanpath_type: ${scanpath_type}
  rotate_every_block: ${rotate_every_block}
  collapse_method: ${collapse_method}
  use_norm_after_ssm: ${use_norm_after_ssm}

# Configuration for the dataloader
data_config:
  _target_: mae.datasets_mae.load_DataModule
  batch_size: ${batch_size}
  num_workers: ${num_workers}
  #augmentation
  img_size: 224
  hflip: 0.5
  interpolation: 'bicubic'

model_config:
  _target_: mae.mae_imagenet.SSLModule
  backbone: ${vim_config}
  weight_decay: ${weight_decay}
  blr: ${blr}
  batch_size: ${batch_size}
  warmup_epochs: ${warmup_epochs}
  scheduling_epochs: ${training_epochs}
  accum_iter: ${accum_iter}
  min_lr: ${min_lr}
  mask_ratio: ${mask_ratio}