DelRec

This repository contains the code to reproduce the experiments presented in the article DelRec: learning delays in recurrent spiking neural networks, by Alexandre Queant, Ulysse Rançon, Benoit R Cotterau and Timothée Masquelier.

Datasets

Create a Datasets folder at the root of the repository.

For the SSC dataset

The Spiking Speech Commands (SSC) dataset contains 35 classes from a larger number of speakers. The number of examples in the train, validation and test splits are 75466, 9981 and 20382 respectively.

The datasets folder should have a subfolder SSC/containing ssc_train.h5, ssc_valid.h5 and ssc_test.h5, which are all downloadable at: https://zenkelab.org/datasets. The dataloader used comes from https://github.com/dgxdn/ASRC-SNN.

For the SHD dataset

The Spiking Heidelberg Digits (SHD) dataset containing spoken digits from 0 to 9 in both English and German (20 classes). The train and test sets contain 8332 and 2088 examples respectively (there is no validation set provided).

Just create a subfolder SHD/in the Datasets/ folder, then the Spiking Jelly dataloader we use for this dataset will download and process the data automatically, as seen in https://github.com/Thvnvtos/SNN-delays.

Finally, the Datasets/ folder should have one subfolder SHD/ and one subfolder SSC/.

Then, make sure the configurations in configs/ have the right dataset_pathfor the corresponding dataset. For example, for the SSC dataset, you need to have:

datasets_path = 'Datasets/SSC'

in configs/perf_SSC.py.

The PS-MNIST dataset us directly derived from the MNIST dataset, which is already in Pytorch.

Requirements

This code is based on Pytorch. To install the required libraries, do:

pip install -r requirements.txt

Repository Structure

configs/           # Parameter & performance configs
notebooks/         # Jupyter notebook for Fig.3
src/               # Core source code

On top of that, we provide the scripts for the main experiments:

perf_dataset.py     # Train our model on the given dataset
equiparam_SHD.py    # Training for Fig.3 B and C (top)
penalize_spikes.py  # Training for Fig.3 C (bottom)

Details

Main Scripts

Script	Description
perf_SHD.py	Train and evaluate the model on the SHD dataset. To reproduce the accuracies displayed in Table 2, run this script with the corresponding SNN model on seeds = [i for i in range(10)].
perf_PSMNIST.py	Train and evaluate the model on the PSMNIST dataset. To reproduce the accuracies displayed in Table 1, run this script with the corresponding SNN model on seeds = [0, 1, 2].
perf_SSC.py	Train and evaluate the model on the SSC dataset. To reproduce the accuracies displayed in Table 1, run this script with the corresponding SNN model on seeds = [0, 1, 2].
equiparam_SHD.py	Training the different models in order to obtain the evolution of accuracy on the SHD as a function of the number of parameters. To reproduce Fig.3 B and C (top), run this script on seeds = [0, 1, 2].
penalize_spikes.py	Training the different models while gradually penalizing spikes, in order to have model accuracy as a function of mean firing rate. To reproduce Fig.3 C (bottom), run this script on seeds = [0, 1, 2]. The models in this experiment use the same configuration as for equiparam_SHD.

To run one of these scripts, just use :

python perf_SSC.py

for example, in the case of the SSC dataset.

We use Weights and Biases in order to log our metrics and the evolution of parameters. Please add:

os.environ["WANDB_MODE"] = "disabled"

at the beggining of the script if you don't want to use wandb. Otherwise, enter your wandb key inn the place-holder:

WANDB_KEY = None # Your key here

Source Code (src/)

File/Folder	Description
datasets.py	Dataset loading scripts for SHD, SSC, PSMNIST.
recurrent_neurons.py	Implementation of our class of neurons with delays in recurrent connections.
utils.py	Helper functions (metrics, seeding, logging ...)
SSC/ PSMNIST/ SHD/	Dataset-specific processing scripts for each dataset. In each subfolder, you have snn provides the implementation of the different networks that can be used on the dataset, and trainer implements the function used for the training of the networks. SSC/and PSMNIST/use the same snn script.

Training time

On a NVIDIA A100 GPU, the training time on one seed for the number of epochs detailed in the corresponding config is of the order of:

• SSC ~ 1 day for 100 epochs.
• PSMNIST ~ 2 days for 200 epochs.
• SHD ~ 1 hour for 150 epochs.

License

This project is licensed under the MIT License. See the LICENSE file for details.