snn-iir

PyTorch implementation of IJCAI 2020 paper Exploiting Neuron and Synapse Filter Dynamics in Spatial Temporal Learning of Deep Spiking Neural Network [arXiv] [IJCAI 2020]

Prerequisites

Install all the required Python packages:

pip install -r requirements.txt

Train Model

Run Python script to train the corresponding model:

python *.py --train

Prepare Trained Weights for Testing

To Use Your Own Trained Weights

Move them to the appropriate path.

To Use ZIP Weights

• Download the ZIP weights by initializing and updating the snn-iir-checkpoints git submodule;

  git submodule init
  git submodule update
  

• Unzip ZIP files to get trained weights;
• Move to appropriate path;

Test Model

• Modify test_checkpoint_path in .yaml config file;
• Run Python script to test the corresponding model with assigned weights: python *.py --test

Models

Details of the models for the following 3 tasks.

Associative Memory

experiment	network	states	filter	dataset	encoding	length
associative_memory	MLP	zero	dual exp iir	Pattern Dataset	original	300

Vision Tasks

experiment	network	states	filter	dataset	encoding	length
snn_mlp_1	MLP	zero	dual exp iir	MNIST	copy along time dimension	25
snn_mlp_1_non_zero	MLP	preserved	dual exp iir	MNIST	copy along time dimension	25
snn_mlp_1_poisson_input	MLP	zero	dual exp iir	MNIST	rate-based poisson	25
snn_mlp_2	MLP	zero	first order low pass	MNIST	copy along time dimension	25
snn_mlp_2_poisson_input	MLP	zero	first order low pass	MNIST	rate-based poisson	25
snn_conv_1_mnist	CNN	zero	dual exp iir	MNIST	copy along time dimension	25
snn_conv_1_mnist_poisson_input	CNN	zero	dual exp iir	MNIST	rate-based poisson	25
snn_conv_1_nmnist	CNN	zero	dual exp iir	N-MNIST	accumulate within time window(OR)	30
snn_conv_1_gesture	CNN	zero	dual exp iir	DVS128 Gesture Dataset	accumulate within time window(OR)	50
snn_conv_1_gesture_30	CNN	zero	dual exp iir	DVS128 Gesture Dataset	accumulate within time window(OR)	30
snn_conv_1_gesture_max	CNN	zero	dual exp iir	DVS128 Gesture Dataset	accumulate within time window(SUM)/frame(MAX)	30

Times Series Classification

Not implemented.

Results

The results of the following 3 tasks.

Associative Memory

experiment	train	dev	test	best epoch	paper
associative_memory	0.0031(93)	0.00369(92)	0.0042(92)	92	-

Vision Tasks

experiment	train	dev	test	best epoch	paper
snn_mlp_1	99.252(72)	98.58(72)	98.94(72)	72	-
snn_mlp_1_non_zero	99.116(93)	98.488(93)	98.858(93)	93	-
snn_mlp_1_poisson_input	99.208(98)	98.628(98)	98.928(98)	98	-
snn_mlp_2	99.3(72)	98.66(72)	98.96(72)	72	-
snn_mlp_2_poisson_input	99.284(96)	98.748(96)	98.978(96)	96	-
snn_conv_1_mnist	99.84(99)	99.47(99)	99.59(99)	99	-
snn_conv_1_mnist_poisson_input	99.822(93)	99.479(93)	99.519(93)	93	99.46
snn_conv_1_nmnist	99.998(51)	98.708(89)	98.558(89)	89	99.39
snn_conv_1_gesture	95.474(46)	85.156(46)	66.319(46)	46	96.09
snn_conv_1_gesture_30	96.094(59)	85.938(59)	68.75(59)	59	96.09
snn_conv_1_gesture_max	97.845(68)	75.781(68)	70.486(68)	68	96.09

Times Series Classification

Not implemented.

Author

Zhongyu Chen