Efficient-spiking-networks - GitLite

KO | EN

gitlite — search

#python #llm #golang #android #machine-learning #rust #linux #docker #macos #cli #kubernetes #awesome-list #security #mcp

Efficient-spiking-networks

(★ 89)

Accurate and efficient Spiking recurrent networks on ECG,SHD,SSC,SOLI,(p)SMNIST,TIMIT

File Explorer

Download Latest Version (.zip)

- QTDB_test.mat
- QTDB_train.mat
- 0.9791-432--relu-MG.pth
- srnn_class_cnn_enc.ipynb
- srnn_class_scnn_enc.ipynb
- adapt_srnn_0.844_gpu-v1.py
- - spike_cnn.py
  - spike_dense.py
  - spike_neuron.py
  - spike_rnn.py
- 0.91398-gscv1-_class.pth
- data.py
- optim.py
- srnn_fin.py
- utils.py
- - .DS_Store
  - model_0.8589141409979619_36bptt-fr_MG.pth
  - visualization.ipynb
- - model_87.41166077738517-readout-2layer-v1-12Feb[128,128].pth
  - vis.ipynb
- - model_94.73_Task-psmnist||Time-07-12-2022 14:26:15||EC_f--rbf||DC_f--adp-spike||multiinput-multi_input.pth
  - model_98.21_Task-smnist||Time-30-06-2020 16:33:05||EC_f--rbf||DC_f--adp-spike||multiinput-multi_input.pth
  - ps_data.zip
  - psmnist-gpu-tbptt-v1.py
  - psmnist_vis.ipynb
  - smnist_vis.ipynb
- - srnn_vis.ipynb
- - model_74.18800902757334-v3-[400, 400]-2layer_MG.pth
  - vis.ipynb
- - 0.66292-bi-srnn-v3_MN-v1.pth
  - vis.ipynb
- .DS_Store
- generate_dataset.py
- model_90.7-readout-2layer-v2-4ms.pth
- SHD_2layer_ALIF.py
- .DS_Store
- data.py
- generate_dataset-intro.py
- generate_dataset_exm.py
- grad_search.sh
- new_dataset.py
- npy_dataset.py
- srnn_frame-grad.py
- srnn_frame-v3_finalized.py
- ssc_dataset_f.py
- ssc_generate_dataset.py
- 0.66292-bi-srnn-v3_MN-v1.pth
- bi_srnn-Dense-v3-multiG.py
- data_intro.py
- generate_dataset.py
- generate_dataset_f40_t100.py
- make_data_set4.py
- README.md
.DS_Store
LICENSE
README.md
req.txt

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Adaptive (SRNN) Spiking Recurrent  Neural network 
**Update:** 'upload Truncated-BPTT code to speed up the training'

-----
This code implements the adaptive spiking recurrent network with learnable parameters on [Pytorch]([PyTorch](https://pytorch.org/)) for various tasks. 
This is scientific software, and as such subject to many modifications; we aim to further improve the software to become more user-friendly and extendible in the future.

For those wanting to try it out: The best place to start is the Jupyter notebook. It’s a quick demonstration of spiking neural networks.
This is what you need:
1) A working version of python and Pytorch This should be easy: either use the Google Colab facilities, or do a simple installation on your laptop could probabily using pip. ([Start Locally | PyTorch](https://pytorch.org/get-started/locally/))
2) The datasets: 
	1. ECG: The original QTDB dataset was downloaded from [PhysioNet](https://physionet.org/content/qtdb/1.0.0/). The encoded dataset is attached in /data/ folder 
  	2. (P)S-MNIST: This dataset can easily be found in torchvision.datasets.MNIST
 	3. SHD and SSC: the Spiking Heidelberg dataset can be downloaded from the [official sit]( https://compneuro.net/datasets/)
	4. SoLi dataset: The official   [dataset](https://polybox.ethz.ch/index.php/s/wG93iTUdvRU8EaT) can be easily downloaded and other necessary information is well documented at the official git[GitHub - simonwsw/deep-soli: Gesture Recognition Using Neural Networks with Google’s Project Soli Sensor](https://github.com/simonwsw/deep-soli)
	5. Google [Speech Commands Dataset](http://download.tensorflow.org/data/speech_commands_v0.01.tar.gz) 
	6. TIMIT Acoustic-Phonetic Continuous Speech Corpus can be found here: LDC (https://catalog.ldc.upenn.edu/LDC93S1)

3) Pre-requisites

- Python 3: HDF5, OpenCV 2 interfaces for python.
- Pytorch>=1.5.1
- Preprocessing packages: [librosa]([librosa — librosa 0.8.0 documentation](https://librosa.org/doc/latest/index.html)),[tdqm]([GitHub - tqdm/tqdm: A Fast, Extensible Progress Bar for Python and CLI](https://github.com/tqdm/tqdm)),[tables]([tables · PyPI](https://pypi.org/project/tables/)),[wfdb]([wfdb — wfdb 3.3.0 documentation](https://wfdb.readthedocs.io/en/latest/)),[klepto]([klepto · PyPI](https://pypi.org/project/klepto/)) and [Scipy]([SciPy.org — SciPy.org](https://www.scipy.org/index.html)),[Jupyter Notebook]([Project Jupyter | Home](https://jupyter.org/))
- Optional: SSC and SHD dataset require more than 32GB RAM to hold the dataset. Training on GPU may accelerate the evaluation.

4) Code running
* Data preprocessing. The datasets(ECG,SHD,SSC,SoLi,TIMIT) are required to arrange the dataset for further training. 
‘’’
mkdir  data
python generate_dataset.py
‘’’
Note that you can also change the sampling frequency to generate a higher precision dataset. 
* Model training. 
‘’’
mkdir  model
python train.py
‘’’
* Visualization of the results(Demo). The results can be visualized by [Jupyter Notebook]([Project Jupyter | Home](https://jupyter.org/))
* Pre-trained models are attached in the folder of each dataset. Because the training takes a quite long time and

Finally, we’d love to hear from you if you have any comments or suggestions.

### References

[1]. Bojian Yin, Federico Corradi, Sander M. Bohté. **Accurate and efficient time-domain classification with adaptive spiking recurrent neural networks**

[2]. Bojian Yin, Federico Corradi, Sander M. Bohté. **Effective and efficient computation with multiple-timescale spiking recurrent neural networks**