[AAAI'2025] Enhancing Implicit Neural Representations via Symmetric Power Transformation

Project Page | Paper | Code | Checkpoints

Weixiang Zhang, Shuzhao Xie, Chengwei Ren, Shijia Ge, Mingzi Wang, Zhi Wang*
Tsinghua University
*: Corresponding Author

This is the official PyTorch implementation of Enhancing Implicit Neural Representations via Symmetric Power Transformation" (accepted by AAAI 2025).

Overview

Abstract. We propose symmetric power transformation to enhance the capacity of Implicit Neural Representation (INR) from the perspective of data transformation. Unlike prior work utilizing random permutation or index rearrangement, our method features a reversible operation that does not require additional storage consumption. Specifically, we first investigate the characteristics of data that can benefit the training of INR, proposing the Range-Defined Symmetric Hypothesis, which posits that specific range and symmetry can improve the expressive ability of INR. Based on this hypothesis, we propose a nonlinear symmetric power transformation to achieve both range-defined and symmetric properties simultaneously. We use the power coefficient to redistribute data to approximate symmetry within the target range. To improve the robustness of the transformation, we further design deviation-aware calibration and adaptive soft boundary to address issues of extreme deviation boosting and continuity breaking. Extensive experiments are conducted to verify the performance of the proposed method, demonstrating that our transformation can reliably improve INR compared with other data transformations. We also conduct 1D audio, 2D image and 3D video fitting tasks to demonstrate the effectiveness and applicability of our method.

Quick Start

Clone Repository

git clone https://github.com/zwx-open/Symmetric-Power-Transformation-INR.git
cd Symmetric-Power-Transformation-INR

Enviroment Setup

Tested Enviroments:
- Ubuntu 20.04 with PyTorch 1.12.1 & CUDA 11.3.1 on RTX 3090

Run Demo

Demo: Fit DIV2K/test/00.png with SIREN + Sym-power-trans (5k epochs; ~5minutes)

python run.py 

High Level Structure

├── assets
│   └── intro.png
├── components
│   ├── lpips.py
│   ├── ssim.py
│   ├── ssim_.py
│   └── transform.py ## The implementaion of Symm-Power-Trans
├── config.py ## The config of different experimental settings 
├── data
│   ├── div2k
│   ├── Kodak
│   ├── libri_test_clean_121726
│   └── text
├── main.py
├── manager.py ## Manage arguments under different experimental settings
├── models.py ## Backbones of INR
├── opt.py  ## Arugment parser
├── README.md
├── run.py ## Run this file
├── trainer
│   ├── base_trainer.py
│   ├── img_trainer.py  
└── util
    ├── io.py
    ├── logger.py
    ├── misc.py
    ├── recorder.py
    └── tensorboard.py

Code Execution

How to run experiments in this repo?

Please update code in run.py (~line 62) to run different experiments: The defualt is running demo (exp_000):

if __name__ == "__main__":

    exp = "000"
    param_sets = PAMRAM_SET[exp]
    gpu_list = [i for i in range(len(param_sets))]
    
    run_tasks(exp, param_sets, gpu_list)

For example, if you want to run experment 001, you can update it with exp = "001". Moreover, feel free to allocate tasks for different gpu:

if __name__ == "__main__":

    exp = "001"
    param_sets = PAMRAM_SET[exp]
    gpu_list = [3, 0] # assert len(param_sets) == len(gpu_list)
    
    run_tasks(exp, param_sets, gpu_list)

How to set different tasks & paramters?

For example, if you want to run sym_power and 01_norm in exp_001, please update config.py with:

PAMRAM_SET["001"] = (
            "01_norm",
            # "z_score",

            #"gamma_0.5",
            #"gamma_2.0",

            # "scale_0.5",
            # "scale_1.0",
            # "scale_2.0",

            # "inverse",
            # "rpp"
            # "box_cox",
            
            "sym_power",          
        )

How to register new task?

Feel free to register new experiment by adding new key in config.py:

PAMRAM_SET["999"] = (
            "xxx1",
            "xxx2",        
        )

and define corresponding function in manager.py:

def _set_exp_999(self, param_set):
    if param_set == "xxx1":
        self.p.xxx = xxx1
    elif param_set == "xxx2":
        self.p.xxx = xxx2

More Flexible Way

If you prefer a more flexible way to run this code, please refer to debug.py:

'''flexiblely set all arugments in opt.py'''
def debug(use_cuda=0):
    args = [
        "--model_type",
        "siren",
        "--input_path",
        "./data/div2k/test_data/00.png",
        "--eval_lpips",
        "--transform",
        "sym_power",
        "--tag",
        "debug_demo",
        "--lr",
        "0.0001",
        "--up_folder_name",
        "000_demo", 
    ]
    os.environ["CUDA_VISIBLE_DEVICES"] = str(use_cuda) 
    script = "python main.py " + " ".join(args)
    print(f"Running: {script}")
    os.system(script)

Then running debug.py:

python debug.py

Reproducing Results from the Paper

Comparison of Different Transformations (Table 1)

The setting of this experiments is correspoinding to _set_exp_001() in manager.py. Please run exp_001 following How to run experiments in this repo.

Chekckpoints can be found in here (log/001_trans).

Additional Related Research

Welcome to explore our related research. The source code for all works has been available.

• (CVPR'2025) EVOS: Efficient Implicit Neural Training via EVOlutionary Selector| [paper] | [project] | [code] |
• (ICME'2025) Expansive Supervision for Neural Radiance Fields| [paper] | [code] |
• Recent Progress of Implicit Neural Representations | [code]

Citation

Please consider leaving a ⭐ and citing our paper if you find this project helpful:

@inproceedings{sym-power-trans-inr,
  author       = {Weixiang Zhang and
                  Shuzhao Xie and
                  Chengwei Ren and
                  Shijia Ge and
                  Mingzi Wang and
                  Zhi Wang},
  title        = {Enhancing Implicit Neural Representations via Symmetric Power Transformation},
  booktitle    = {{AAAI}},
  pages        = {10157--10165},
  publisher    = {{AAAI} Press},
  year         = {2025}
}