config/train_rvlcdip.yaml

resume_from_checkpoint_path: null 
result_path: "./result"
pretrained_model_name_or_path: "naver-clova-ix/donut-base"
dataset_name_or_paths: ["./dataset/rvlcdip"] # should be prepared from https://www.cs.cmu.edu/~aharley/rvl-cdip/
sort_json_key: True
train_batch_sizes: [2]
val_batch_sizes: [4]
input_size: [2560, 1920]
max_length: 8
align_long_axis: False
# num_nodes: 8 # memo: donut-base-finetuned-rvlcdip was trained with 8 nodes
num_nodes: 1
seed: 2022
lr: 2e-5
warmup_steps: 10000
num_training_samples_per_epoch: 320000
max_epochs: 100
max_steps: -1
num_workers: 8
val_check_interval: 1.0
check_val_every_n_epoch: 1
gradient_clip_val: 1.0
verbose: True