FastChannelVimS_maxpool_ps8.yaml

batch_size: 32
num_workers: 12
num_nodes: 1

pl_seed: 0

patch_size: 8 # for now only implemented when both stride and patch size is same

sort_channels: True
scanpath_type: rowwise  # rowwise, colwise
use_norm_after_ssm: True
rotate_every_block: True
collapse_method: max # mean, max

training_epochs: 100
warmup_epochs: 10
lr: 0.001
warmup_initial_lr: 0
min_lr: 0.000001
scaling_rule: linear
weight_decay: 0.04
weight_decay_end: 0.4
num_classes: 161

drop_path_rate: 0.05 

channelvim_config:  
  _target_: models.channel_wise_tokenization.models_channel_mamba_faster.channelvim_small_patch16_224_final_pool_mean_abs_pos_embed_with_noclstok_div2
  pretrained: false
  num_classes: ${num_classes}
  drop_path_rate: ${drop_path_rate}
  scan_order: 'Channel-First'   # Spatial-First, Channel-First  
  hcs: True
  channels: 8
  sort_channels: ${sort_channels}
  patch_size: ${patch_size}
  stride: ${patch_size}
  scanpath_type: ${scanpath_type}
  rotate_every_block: ${rotate_every_block}
  use_norm_after_ssm: ${use_norm_after_ssm}
  collapse_method: ${collapse_method}

# Configuration for the dataloader
data_config:
  _target_: cell_imaging.datasets_supervised.load_DataModule
  batch_size: ${batch_size}
  num_workers: ${num_workers}
  #augmentation
  perturbation_list: ['compound']
  channels: [0,1,2,3,4,5,6,7]
  normalization_mean:
    - 4.031743599139058
    - 1.565935237087539
    - 3.77367898215863
    - 3.4605251427133257
    - 4.1723172504050225
    - 6.780529773318951
    - 6.787385700135139
    - 6.778120829362721
  normalization_std:
    - 17.318438884455695
    - 12.015918256263747
    - 16.966058078452495
    - 15.064776266287147
    - 17.964118200870608
    - 21.638766346725316
    - 21.670565699654457
    - 21.639488585095584

# Configuration for the learning rate schedule
lr_config:
  _target_: cell_imaging.utils.get_lr_scheduler
  dataloader: ${data_config}
  num_nodes: ${num_nodes}
  batch_size: ${batch_size}
  scheduling_epochs: ${training_epochs}
  warmup_epochs: ${warmup_epochs}
  lr: ${lr}
  warmup_initial_lr: ${warmup_initial_lr}
  min_lr: ${min_lr}
  scaling_rule: ${scaling_rule}

# Configuration for the weight decay schedule
wd_config:
  _target_: cell_imaging.utils.get_wd_scheduler
  num_nodes: ${num_nodes}
  scheduling_epochs: ${training_epochs}
  dataloader: ${data_config}
  weight_decay: ${weight_decay}
  weight_decay_end: ${weight_decay_end}

model_config:
  _target_: cell_imaging.supervised.SupervisedModule
  backbone: ${channelvim_config}
  lr_schedule: ${lr_config}
  wd_schedule: ${wd_config}
  num_classes: ${num_classes}
  label_smoothing: 0.0
  mixup: 0.0
  cutmix: 0.0
  mixup_mode: 'batch'
  mixup_prob: 1.0
  mixup_switch_prob: 0.5