ONNX for SNN(Spiking Neural Networks)

What is ONNX-SNN

• ONNX is an open format to represent deep learning models.
  • ONNX protocol - https://github.com/onnx/onnx/blob/master/onnx/onnx.proto
• ONNX-SNN is an extension for Spiking Neural Networks.
• Its main purpose is to translate DNN model to SNN model.

Necessary Libraries

• Tensorflow 1.10.0
• Keras 2.2.2
• Nengo 2.8.0
• NengoDL 2.2.0
• ONNX 1.6.0
• Numpy 1.14.5
• Protobuf 3.6.0
• Cudatoolkit 9.0
• Cudnn 7.6.4

Features

• ONNX-SNN
  • DNN's neuron -> spike neuron
  • Neuron activation (Sigmoid, Relu, tanh, etc.) -> Spike (LIF, softLIF, etc.)
• NengoDL with ONNX-SNN
  • Build Deep Spiking neural networks with ONNX-SNN
  • Training weights with NengoDL (Rate neuron)
  • Writing the trained model to ONNX-SNN

source

• convert_snnOnnx.py
  • ONNX -> ONNX-SNN
• onnxToNengoCode.py
  • ONNX-SNN -> Nengo Code
• onnxToNengoModel.py
  • ONNX-SNN -> Nengo Model

Progress

• Converting ONNX of DNN to ONNX-SNN
• Reading ONNX-SNN and building Nengodl code

To do list

• Training, prediction target data --> Mnist

---

• convert_conv2d - padding(=border_mode) problem --> O
• convert_pool2d - kind deal(Max, Average) --> O
• convert_flatten function --> O
• convert_matmul function --> O
• convert_batchnormalization function --> O
• softmax --> O
• training, simulation --> O
• Run neuron type(LIF, LIFRate, AdaptiveLIF, AdaptiveLIFRate, Izhikevich) --> O
• -->11/13/2019

---

• nengo_dl support only Sequential network(ex vggnet, alexnet)

---

• Apply to different models(vgg16, vgg19, alexnet) --> O
• nengo_dl model -> onnx-snn -> onnx -> keras model --> X
• onnx model, weight -> onnx-snn -> nengo_dl model --> X

---