• demo.py

In demo.py, the simplest usage of a PS neuron is demonstrated. You can easily replace the activation function of an artificial neural network with a PS neuron using the replace_activation_with_Psactivation function, effectively converting it into a spiking neural network. The fitting performance of the activation function is visually demonstrated through an image.

• config.json

  The config.json file contains some default parameters, where you can configure the default storage path, fitting interval, fitting accuracy, and more.

• prepare_hdT.py

  This script implements the parameter iteration process based on the algorithm described in the paper. The results of the parameter iterations are saved in the default storage path. You can specify any activation function to fit using the target_func_name parameter. The dy, l, and r parameters can also be specified simultaneously, though they are optional.

• PSActivation.py

  This file contains the main code for the PSNeuron class, including how to perform the ANN-to-SNN conversion using the iterated parameters.The count of PS neurons and the number of spikes generated are also tracked here.

• utils.py

  This file includes various tools and commonly used activation functions.

As shown in demo.py, you can directly use the replace_activation_with_Psactivation function from PSActivation.py to replace the activation functions in an ANN, enabling the ANN-to-SNN conversion. If the corresponding pre-iterated parameters are not found in the default storage location, the find_hdT function from prepare_hdT.py will be automatically invoked to generate a set of parameters based on the algorithm and save them. During this process, you can manually control the activation function to be generated, as well as its precision and range, by adjusting activation_name, dy, l, and r.

Alternatively, you can directly use prepare_hdT.py to generate a different set of parameters and observe the parameter iteration process.

The code runs under Python 3.12.4, and the required dependencies are as follows:

matplotlib==3.9.2 numpy==2.1.1 scipy==1.14.1 torch==2.3.1+cu121 transformers==4.42.0

The experiments were conducted on a Windows 11 system using the PyTorch framework with a publicly available pre-trained model. The setup included CUDA Version 12.3 and an NVIDIA GeForce RTX 3050 GPU.