Quantized Spiking Neural Networks

This repository contains the models and training scripts used in the papers: "Quantizing Spiking Neural Networks with Integers" (ICONS 2020) and "Memory Organization for Energy-Efficient Learning and Inference in Digital Neuromorphic Accelerators" (ISCAS 2020).

Requiremnts

• Python
• PyTorch
• torchvision
• NumPy
• pickle
• argparse

Quantized SNNs for Spatio-Temporal Patterns

All relevant code for the experiments from the ISCAS paper is contained in qsnn_precise.py, quantization.py and qsnn_util.py. To run the experiments execute:

python qsnn_precise.py

You can specify desired setting either as command-line arguments or within qsnn_precise.py.

Optional arguments:

Argument	Description
--input INPUT	Input pickle file (default: ./data/input_700_250_25.pkl)
--target TARGET	Target pattern pickle (default: ./data/smile95.pkl)
--global_wb GLOBAL_WB	Weight bitwidth (default: 2)
--global_ab GLOBAL_AB	Membrane potential, synapse state bitwidth (default: 8)
--global_gb GLOBAL_GB	Gradient bitwidth (default: 8)
--global_eb GLOBAL_EB	Error bitwidth (default: 8)
--global_rb GLOBAL_RB	Gradient RNG bitwidth (default: 16)
--time_step TIME_STEP	Simulation time step size (default: 0.001)
--nb_steps NB_STEPS	Simulation steps (default: 250)
--nb_epochs NB_EPOCHS	Simulation steps (default: 10000)
--tau_mem TAU_MEM	Time constant for membrane potential (default: 0.01)
--tau_syn TAU_SYN	Time constant for synapse (default: 0.005)
--tau_vr TAU_VR	Time constant for Van Rossum distance (default: 0.005)
--alpha ALPHA	Time constant for synapse (default: 0.75)
--beta BETA	Time constant for Van Rossum distance (default: 0.875)
--nb_inputs NB_INPUTS	Spatial input dimensions (default: 700)
--nb_hidden NB_HIDDEN	Spatial hidden dimensions (default: 400)
--nb_outputs NB_OUTPUTS	Spatial output dimensions (default: 250)

Quantized SNNs for Gesture Detection with Local Learning

Download and extract the DVS Slow Poker and DVS Gesture data set.

To prepare the data run the following commands in the respective directories (e.g. in the directory of the DVS Poker data or the DVS Gesture data).

python prepPoker.py

python prepGesture.py

All relevant code for the experiments from the ICONS paper is contained in qsnn_decolle.py, quantization.py and localQ.py. To run the experiments execute:

python qsnn_decolle.py

You can specify desired setting either as command-line arguments or within qsnn_decolle.py.

Optional arguments:

Argument	Description
--data-set DATA_SET	Input date set: Poker/Gesture (default: Gesture)
--global_wb GLOBAL_WB	Weight bitwidth (default: 8)
--global_qb GLOBAL_QB	Synapse bitwidth (default: 10)
--global_pb GLOBAL_PB	Membrane trace bitwidth (default: 12)
--global_rfb GLOBAL_RFB	Refractory bitwidth (default: 2)
--global_sb GLOBAL_SB	Learning signal bitwidth (default: 6)
--global_gb GLOBAL_GB	Gradient bitwidth (default: 10)
--global_eb GLOBAL_EB	Error bitwidth (default: 6)
--global_ub GLOBAL_UB	Membrane Potential bitwidth (default: 6)
--global_ab GLOBAL_AB	Activation bitwidth (default: 6)
--global_sig GLOBAL_SIG	Sigmoid bitwidth (default: 6)
--global_rb GLOBAL_RB	Gradient RNG bitwidth (default: 16)
--global_lr GLOBAL_LR	Learning rate for quantized gradients (default: 1)
--global_lr_sgd GLOBAL_LR_SGD	Learning rate for SGD (default: 1e-09)
--global_beta GLOBAL_BETA	Beta for weight init (default: 1.5)
--delta_t DELTA_T	Time step in ms (default: 0.001)
--input_mode INPUT_MODE	Spike processing method (default: 0)
--ds DS	Downsampling (default: 4)
--epochs EPOCHS	Epochs for training (default: 320)
--lr_div LR_DIV	Learning rate divide interval (default: 80)
--batch_size BATCH_SIZE	Batch size (default: 72)
--PQ_cap PQ_CAP	Value cap for membrane and synpase trace (default: 1)
--weight_mult WEIGHT_MULT	Weight multiplier (default: 4e-05)
--dropout_p DROPOUT_P	Dropout probability (default: 0.5)
--lc_ampl LC_AMPL	Magnitude amplifier for weight init (default: 0.5)
--l1 L1	Regularizer 1 (default: 0.001)
--l2 L2	Regularizer 2 (default: 0.001)
--tau_mem_lower TAU_MEM_LOWER	Tau mem lower bound (default: 5)
--tau_mem_upper TAU_MEM_UPPER	Tau mem upper bound (default: 35)
--tau_syn_lower TAU_SYN_LOWER	Tau syn lower bound (default: 5)
--tau_syn_upper TAU_SYN_UPPER	Tau syn upper bound (default: 10)
--tau_ref TAU_REF	Tau ref (default: 2.857142857142857)