Cauchy-Schwarz Regularizers

This is the code for the numerical results in the paper "Cauchy-Schwarz Regularizers", S. Taner, Z. Wang, and C. Studer, ICLR, 2025. (c) 2025 Sueda Taner, Ziyi Wang

email: taners@ethz.ch

Important Information

If you are using this simulator (or parts of it) for a publication, then you must cite the following paper:

"Cauchy-Schwarz Regularizers", S. Taner, Z. Wang, and C. Studer, ICLR, 2025.

How to use this code...

...for the solution recovery results from Sections 3.1-3.3 and Appendix C

• Use folder solution_recovery:
• For recovering discrete-valued vectors:
  • For binary recovery success rate results with respect to the system dimensions as in Section 3.1, run binary_recovery.m.
  • For one-sided binary and ternary recovery success rate results with respect to the linear system dimensions as in Section 3.1, run osb_ternary_recovery.m.
  • For one-sided binary and ternary recovery success rate results with respect to the density as in Appendix C.4, run osb_ternary_vary_density.m.
  • For recovering two-bit valued vectors as in Appendix C.5, run bbit_recovery.m.
• For recovering eigenvectors as in Section 3.2, run eigvec_recovery.m.
• For recovering orthogonal-column matrices as in Section 3.3, run orth_matrix_recovery.m.

...for the neural network quantization results from Section 3.4

• Use folder quantized_nns.
• Create your environment using requirements.txt.
• Set the directory paths for ImageNet and CIFAR10 datasets in train_from_cfg.py.
• Create a folder named pretrained_models and download the pretrained ResNet-20 model from here into it.
• Set your training configuration in main_train.py, e.g., for training ResNet-18 on ImageNet (or ResNet-20 on CIFAR-10), for binary (or ternary) regularization, and run.

Version history

Version 0.1: taners@ethz.ch - initial version for GitHub release.

Acknowledgments

This project makes use of the following external code:

• fasta-matlab by T. Goldstein, accessed on 1/3/2024: used in the solution recovery results in Sections 3.1-3.3 and Appendix C.
• pytorch_resnet_cifar10 by I. Idelbayev, accessed on 1/12/2023: used for the pretrained ResNet-20 model in Section 3.4.