SIRA Checkpoint

This repository hosts the DeepSpeed checkpoint for SIRA: Reasoning-Aware Surgical Instrument Segmentation via Query-Anchored Alignment.

Code and usage instructions are available in the GitHub repository: linxir226/SIRA.

The SurgRS dataset is available at: linxir226/SurgRS.

Checkpoint Contents

ckpt_model/
  latest
  zero_to_fp32.py
  global_step65000/
    mp_rank_00_model_states.pt
    bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
    bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

This is a complete DeepSpeed checkpoint directory. Use the ckpt_model/ directory as CHECKPOINT_PATH; do not point to global_step65000/, an individual .pt file, or a meta_log_epoch*.pth file.

Inference

Clone the code repository and follow its installation instructions:

git clone https://github.com/linxir226/SIRA.git
cd SIRA

After downloading this checkpoint repository, run validation with:

DATA_ROOT=/path/to/datasets \
CHECKPOINT_PATH=/path/to/this/repo/ckpt_model \
bash scripts/valid_inference.sh

To save visualizations:

DATA_ROOT=/path/to/datasets \
CHECKPOINT_PATH=/path/to/this/repo/ckpt_model \
bash scripts/valid_inference.sh --vis_enable

Citation

@misc{zhang2026sira,
  title={SIRA: Reasoning-Aware Surgical Instrument Segmentation via Query-Anchored Alignment},
  author={Zhang, Zhibo and Wang, Qijie and Yan, Zengqiang},
  year={2026},
  url={https://github.com/linxir226/SIRA}
}