Rotary Position Embedding for Vision Transformer -- Swin code

This folder is RoPE Swin training code based on Swin-Transformer codebase.

RoPE Swin is implemented in models/swin_transformer_rope.py. You can find configs from configs/swinrope/swin_rope_*.yaml

Training

OMP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 main.py \
 --cfg configs/swinrope/${config_file} --data-path ${data_path} --output ${save_path} --batch-size 128

Evaluation

• With pre-trained model (huggingface hub)

  OMP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=1 --nnodes=1 main.py \
  --eval --cfg configs/swinrope/${config_file} --resume huggingface --data-path ${data_path} --output ${save_path} --batch-size 128

• With custom ${checkpoint_file}

  OMP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=1 --nnodes=1 main.py \
  --eval --cfg configs/swinrope/${config_file} --resume ${checkpoint_file} --data-path ${data_path} --output ${save_path} --batch-size 128