super-gradients/src/super_gradients/recipes/imagenet_vit_base.yaml

#  ViT Imagenet1K fine tuning from Imagenet21K classification training:
#  This example trains with batch_size = 64 * 8 GPUs, total 512.
#  Training time on 8 x GeForce RTX A5000 is 15min / epoch.
#  ViT base : 84.15
#
#  Log and tensorboard at s3://deci-pretrained-models/vit_base_imagenet1k/

# Instructions:
# running from the command line, set the PYTHONPATH environment variable: (Replace "YOUR_LOCAL_PATH" with the path to the downloaded repo):
#   export PYTHONPATH="YOUR_LOCAL_PATH"/super_gradients/
# Then:
# for vit_base:
#   python -m torch.distributed.launch --nproc_per_node=8 train_from_recipe_example/train_from_recipe.py --config-name=imagenet_vit_base


defaults:
  - training_hyperparams: imagenet_vit_train_params
  - dataset_params: imagenet_dataset_params
  - arch_params: vit_base_arch_params
  - checkpoint_params: vit_base_imagenet_checkpoint_params


dataset_params:
  resize_size: 249
  batch_size: 64
  random_erase_prob: 0
  random_erase_value: random
  train_interpolation: random
  rand_augment_config_string: rand-m7-mstd0.5
  cutmix: True
  cutmix_params:
    mixup_alpha: 0.2
    cutmix_alpha: 1.0
    label_smoothing: 0.1
  img_mean: [0.5, 0.5, 0.5]
  img_std: [0.5, 0.5, 0.5]

dataset_interface:
  imagenet:
    dataset_params: ${dataset_params}

data_loader_num_workers: 8

model_checkpoints_location: local
load_checkpoint: True
load_weights_only: True

experiment_name: vit_base_imagenet1k

sg_model:
  _target_: super_gradients.SgModel
  experiment_name: ${experiment_name}
  model_checkpoints_location: ${model_checkpoints_location}
  multi_gpu: AUTO

architecture: vit_base