Pytorch Implementation of Spiking Neural Networks Calibration, ICML 2021

Last update: Dec 27, 2022

Related tags

Overview

SNN_Calibration

Pytorch Implementation of Spiking Neural Networks Calibration, ICML 2021

Feature Comparison of SNN calibration:

Features	SNN Direct Training	ANN-SNN Conversion	SNN Calibration
Accuract (T<100)	High	Low	High
Scalability to ImageNet	Tiny	Large	Large
Training Speed	Slow	Fast	Fast
# Required Data	Full-set (1.2M For ImageNet)	~1000	~1000
Inference Speed	Fast	Slow	Fast

Requirements

Pytorch 1.8

For ImageNet experiments, please be sure that you can initialize distributed environments

For CIFAR experiments, one GPU would suffice.

Pre-training ANN on CIFAR10&100

Train an ANN model with main_train.py

python CIFAR/main_train.py --dataset CIFAR10 --arch VGG16 --dpath PATH/TO/DATA --usebn

Pre-trained results:

Dataset	Model	Random Seed	Accuracy
CIFAR10	VGG16	1000	95.76
CIFAR10	ResNet-20	1000	95.68
CIFAR100	VGG16	1000	77.98
CIFAR100	ResNet-20	1000	76.52

SNN Calibration on CIFAR10&100

Calibrate an SNN with main_calibration.py.

python CIFAR/main_calibration.py --dataset CIFAR10 --arch VGG16 --T 16 --usebn --calib advanced --dpath PATH/TO/DATA

--T is the time step, --calib is the calibration method, please use none, light, advanced for experiments.

The calibration will run for 5 times, and return the mean accuracy as well as the standard deviation.

Example results:

Architecture	Datset	T	Random Seed	Calibration	Mean Acc	Std.
VGG16	CIFAR10	16	1000	None	64.52	4.12
VGG16	CIFAR10	16	1000	Light	93.30	0.08
VGG16	CIFAR10	16	1000	Advanced	93.65	0.25
ResNet-20	CIFAR10	16	1000	None	67.88	3.63
ResNet-20	CIFAR10	16	1000	Light	93.89	0.20
ResNet-20	CIFAR10	16	1000	Advanced	94.33	0.12
VGG16	CIFAR100	16	1000	None	2.69	0.76
VGG16	CIFAR100	16	1000	Light	65.26	0.99
VGG16	CIFAR100	16	1000	Advanced	70.91	0.65
ResNet-20	CIFAR100	16	1000	None	39.27	2.85
ResNet-20	CIFAR100	16	1000	Light	73.89	0.15
ResNet-20	CIFAR100	16	1000	Advanced	74.48	0.16

Pre-training ANN on ImageNet

To be updaed

Pytorch Implementation of Spiking Neural Networks Calibration, ICML 2021

Related tags

Overview

SNN_Calibration

Requirements

Pre-training ANN on CIFAR10&100

SNN Calibration on CIFAR10&100

Pre-training ANN on ImageNet

Owner

Yuhang Li

Sequence-to-Sequence learning using PyTorch

Source code of NeurIPS 2021 Paper ''Be Confident! Towards Trustworthy Graph Neural Networks via Confidence Calibration''

Points2Surf: Learning Implicit Surfaces from Point Clouds (ECCV 2020 Spotlight)

CAPRI: Context-Aware Interpretable Point-of-Interest Recommendation Framework

The fastai deep learning library

Towards Representation Learning for Atmospheric Dynamics (AtmoDist)

LWCC: A LightWeight Crowd Counting library for Python that includes several pretrained state-of-the-art models.

The Hailo Model Zoo includes pre-trained models and a full building and evaluation environment

YOLOv5 🚀 is a family of object detection architectures and models pretrained on the COCO dataset

Deep Learning applied to Integral data analysis

SciFive: a text-text transformer model for biomedical literature

Image reconstruction done with untrained neural networks.

本步态识别系统主要基于GaitSet模型进行实现

A TensorFlow 2.x implementation of Masked Autoencoders Are Scalable Vision Learners

This is the code used in the paper "Entity Embeddings of Categorical Variables".

A texturizer that I just made. Nothing special here.

Streamlit Tutorial (ex: stock price dashboard, cartoon-stylegan, vqgan-clip, stylemixing, styleclip, sefa)

TResNet: High Performance GPU-Dedicated Architecture

LF-YOLO (Lighter and Faster YOLO) is used to detect defect of X-ray weld image.

FluxTraining.jl gives you an endlessly extensible training loop for deep learning