Code-DFQ-SAM (for review)

Here is the source code to reproduce DFQ-SAM.

Please note that the code is only used for review purposes and has not been publicly released.

Environment

• The GPU is recommended to be a single NVIDIA A6000.
• Representative package: python=3.10, torch=2.0.1 (In our experiments, DFQ-SAM has good compatibility with versions of other packages.)

Data-free Quantization for SAM

First, download the pre-trained model checkpoint from Model.

Then, you can run the following command to quantize SAM via the proposed DFQ-SAM.

export CUDA_VISIBLE_DEVICES=0
cd ./DFQ-SAM
python test_quant.py

• data_path: Path to test dataset.
• sam_checkpoint: Path to SAM checkpoint.
• batch_size: 1.
• image_size: Default value is 256.
• boxes_prompt: Use Bbox prompt to get segmentation results.
• point_num: Specifies the number of points. Default value is 1.
• iter_point: Specifies the number of iterations for point prompts.
• encoder_adapter: Set to True if using SAM-Med2D's pretrained weights.
• save_pred: Whether to save the prediction results.

Acknowledgments

The code of DFQ-SAM is based on SAM-Med2D. We thank for their open-sourced code.