Model zoo performances

This page lets you discover all of Akida model zoo machine learning models with their respective performances.

Note

The download links provided point towards standard Tensorflow Keras models that must be converted to Akida model using cnn2snn.convert.

Image domain

Classification

Architecture	Resolution	Dataset	Quantization	Top-1 accuracy	Example	#Params	Size (KB)	NPs	Download
AkidaNet 0.25	160	ImageNet	8/4/4	42.58%		480K	392.3	23
AkidaNet 0.5	160	ImageNet	8/4/4	57.80%		1.4M	1099.4	30
AkidaNet	160	ImageNet	8/4/4	66.94%		4.4M	4090.2	81
AkidaNet 0.25	224	ImageNet	8/4/4	46.71%		480K	398.1	25
AkidaNet 0.5	224	ImageNet	8/4/4	61.30%		1.4M	1214.4	38
AkidaNet	224	ImageNet	8/4/4	69.65%		4.4M	6322.6	129
AkidaNet 0.5 edge	160	ImageNet	8/4/4	51.66%		4.0M	2017.1	38
AkidaNet 0.5 edge	224	ImageNet	8/4/4	54.03%		4.0M	2130.1	46
AkidaNet 0.5	160	Cats vs dogs	8/4/4	96.60%		868K	698.4	24
AkidaNet 0.25	224	Imagenette	8/4/4	91.54%		227K	203.9	22
AkidaNet 0.5	224	Imagenette	8/4/4	95.67%		873K	815.5	32
AkidaNet	224	Imagenette	8/4/4	97.58%		3.4M	5544.2	116
AkidaNet 0.5	224	SIIM-ISIC Melanoma Classification	8/4/4	98.31% - AUROC 0.7969		868K	811.4	32
AkidaNet 0.5	224	ODIR-5K Ocular disease recognition	8/4/4	92.42% - AUROC 0.9847		870K	811.4	32
AkidaNet 0.5	224	Retinal OCT ocular disease recognition	8/4/4	81.10% - AUROC 0.9662		870K	811.4	32
AkidaNet 0.5	224	PlantVillage	8/4/4	97.92%		1.1M	1018.8	33
AkidaNet 0.5	224	CIFAR10	8/4/4	92.89%		896K	829.3	33
AkidaNet 0.25	96	Visual Wake Words	8/4/4	84.77%		229K	179.2	16
MobileNetV1 0.25	160	ImageNet	8/4/4	40.86%		467K	365.4	23
MobileNetV1 0.5	160	ImageNet	8/4/4	55.94%		1.3M	1017.1	30
MobileNetV1	160	ImageNet	8/4/4	66.40%		4.2M	3554.5	78
MobileNetV1 0.25	224	ImageNet	8/4/4	45.12%		467K	366.9	25
MobileNetV1 0.5	224	ImageNet	8/4/4	59.76%		1.3M	1075.4	38
MobileNetV1	224	ImageNet	8/4/4	69.53%		4.2M	5251.8	123
MobileNetV1 0.5 edge	160	ImageNet	8/4/4	49.69%		3.9M	1935.1	38
MobileNetV1 0.5 edge	224	ImageNet	8/4/4	51.83%		3.9M	1993.4	46
VGG11	224	ImageNet	8/4/4	52.22%		47.1M	34825.2	21
GXNOR	28	MNIST	2/2/1	99.20%		1.6M	412.6	3

Object detection

Architecture	Resolution	Dataset	Quantization	mAP	Example	#Params	Size (KB)	NPs	Download
YOLOv2	224	PASCAL-VOC 2007 - person and car classes	8/4/4	38.85%		3.6M	3061.0	71
YOLOv2	224	WIDER FACE	8/4/4	73.81%		3.5M	3052.7	71

Regression

Architecture	Resolution	Dataset	Quantization	MAE	Example	#Params	Size (KB)	NPs	Download
VGG-like	32	UTKFace (age estimation)	8/2/2	6.1791		458K	139.8	6

Face recognition

Architecture	Resolution	Dataset	Quantization	Accuracy	#Params	Size (KB)	NPs	Download
AkidaNet 0.5	112x96	CASIA Webface face identification	8/4/4	70.18%	2.3M	1929.8	21
AkidaNet 0.5 edge	112x96	CASIA Webface face identification	8/4/4	71.13%	23.6M	6979.6	35
AkidaNet 0.5	112x96	LFW face verification	8/4/4	97.25%	933K	691.2	20

Audio domain

Keyword spotting

Architecture	Dataset	Quantization	Top-1 accuracy	Example	#Params	Size (KB)	NPs	Download
DS-CNN	Google speech command	8/4/4	91.72%		22.7K	22.8	5

Time domain

Fault detection

Architecture	Dataset	Quantization	Accuracy	#Params	Size (KB)	NPs	Download
Convtiny	CWRU Electric Motor Ball Bearing Fault Diagnosis	8/2/4	99.3%	59K	25.3	3

Classification

Architecture	Resolution	Dataset	Quantization	Accuracy	#Params	Size (KB)	NPs	Download
AkidaNet 0.5	224	Physionet2017 ECG classification	8/4/4	73.50% - AUROC 0.7940	1.1M	1008.4	36

Point cloud

Classification

Architecture	Dataset	Quantization	Accuracy	Input scaling	#Params	Size (KB)	NPs	Download
PointNet++	ModelNet40 3D Point Cloud	8/4/4	84.76%	(127, 127)	602K	528.5	17