CUDA_VISIBLE_DEVICES="0, 1, 2, 3" python3 -m torch.distributed.launch --nproc_per_node 4 ./src/train_swin-bart.py \
--input_file_dir <your dataset dir> \
--input_file <train file(pickle)> \
--input_val_file <valid file (pickle)> \
--input_test_file <test file (pickle )> \
--output {output_dir} \
--model_id swin-bart \
--ratio_train 0.9 \
--train_batch_size 4 \
--val_batch_size 8 \
--train_epoch 20 \
--train_accum_iter 12 \
--input_height 448 \
--input_width 896 \
--window_size 7 \
--encoder_depth 2 2 14 2 \
--encoder_num_heads 4 8 16 32 \
--store_names iter train_loss val_loss \
--valid_ratio 0.3 \
--train_clip_grad 1 \
--learning_rate 1e-4 \
--save_freq 5 \
--check_grad_norm \
--seed 40 \
--max_length 1024 \
--use_amp \
--use_ddp
CUDA_VISIBLE_DEVICES="0, 1, 2, 3" python3 -m torch.distributed.launch --nproc_per_node 4 ./src/train_layoutlmv3-bart.py \
--input_file_dir <your dataset dir> \
--input_file <train file(pickle)> \
--input_val_file <valid file (pickle)> \
--input_test_file <test file (pickle )> \
--output {output_dir} \
--ratio_train 0.9 \
--train_batch_size 4 \
--val_batch_size 8 \
--train_epoch 15 \
--train_accum_iter 14 \
--store_names iter train_loss val_loss \
--valid_ratio 0.3 \
--train_clip_grad 1 \
--learning_rate 1e-4 \
--save_freq 4 \
--check_grad_norm \
--seed ${seed} \
--num_encoder_layer 6 \
--encoder_max_length 512 \
--decoder_max_length 1024 \
--use_amp \
--use_ddp