Model zoo performance

The Brainchip akida_models package offers a set of pre-built akida compatible models (e.g MobileNet, AkidaNet, ViT), pretrained weights for those models and training scripts. Please refer to the model zoo API reference for a complete list of the available models.

This page lists the performance of all models from the zoo reported for both Akida 1.0 and Akida 2.0. Please refer to:

• Akida 1.0 models for models targetting the Akida Neuromorphic Processor IP 1.0 and the AKD1000 reference SoC,

• Akida 2.0 models for models targetting the Akida Neuromorphic Processor IP 2.0,

• Upgrading to Akida 2.0 tutorial to understand the architectural differences between 1.0 and 2.0 models and their respective workflows.

Note

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

Akida 1.0 models

For 1.0 models, 4-bit accuracy is provided and is always obtained through a QAT phase.

Note

The “8/4/4” quantization scheme stands for 8-bit weights in the input layer, 4-bit weights in other layers and 4-bit activations.

Note

The NPs column provides the minimal number of neural processors required for the model excecution on the Akida IP. The numbers given are the result of the map operation using the Minimal MapMode targetting AKD1000 reference SoC.

Image domain

Classification

Architecture	Resolution	Dataset	#Params	Quantization	Top-1 accuracy	Example	Size (KB)	NPs	Download
AkidaNet 0.25	160	ImageNet	480K	8/4/4	42.58%		403.3	20
AkidaNet 0.5	160	ImageNet	1.4M	8/4/4	57.80%		1089.1	24
AkidaNet	160	ImageNet	4.4M	8/4/4	66.94%		4061.1	68
AkidaNet 0.25	224	ImageNet	480K	8/4/4	46.71%		409.1	22
AkidaNet 0.5	224	ImageNet	1.4M	8/4/4	61.30%		1202.2	32
AkidaNet	224	ImageNet	4.4M	8/4/4	69.65%		6294.0	116
AkidaNet 0.5 edge	160	ImageNet	4.0M	8/4/4	51.66%		2017.4	38
AkidaNet 0.5 edge	224	ImageNet	4.0M	8/4/4	54.03%		2130.5	46
AkidaNet 0.5	224	PlantVillage	1.1M	8/4/4	97.92%		1019.1	33
AkidaNet 0.25	96	Visual Wake Words	229K	8/4/4	84.77%		179.6	16
MobileNetV1 0.25	160	ImageNet	467K	8/4/4	36.05%		376.4	20
MobileNetV1 0.5	160	ImageNet	1.3M	8/4/4	54.59%		1007.0	24
MobileNetV1	160	ImageNet	4.2M	8/4/4	65.47%		3525.8	65
MobileNetV1 0.25	224	ImageNet	467K	8/4/4	39.73%		377.9	22
MobileNetV1 0.5	224	ImageNet	1.3M	8/4/4	58.50%		1065.3	32
MobileNetV1	224	ImageNet	4.2M	8/4/4	68.76%		5223.3	110
GXNOR	28	MNIST	1.6M	2/2/1	98.03%		412.8	3

Object detection

Architecture	Resolution	Dataset	#Params	Quantization	mAP	Example	Size (KB)	NPs	Download
YOLOv2	224	PASCAL-VOC 2007 - person and car classes	3.6M	8/4/4	41.51%		3061.4	71
YOLOv2	224	WIDER FACE	3.5M	8/4/4	77.63%		3053.1	71

Regression

Architecture	Resolution	Dataset	#Params	Quantization	MAE	Example	Size (KB)	NPs	Download
VGG-like	32	UTKFace (age estimation)	458K	8/2/2	6.1791		138.6	6

Face recognition

Architecture	Resolution	Dataset	#Params	Quantization	Accuracy	Size (KB)	NPs	Download
AkidaNet 0.5	112×96	CASIA Webface face identification	2.3M	8/4/4	70.18%	1930.1	21
AkidaNet 0.5 edge	112×96	CASIA Webface face identification	23.6M	8/4/4	71.13%	6980.2	34

Audio domain

Keyword spotting

Architecture	Dataset	#Params	Quantization	Top-1 accuracy	Example	Size (KB)	NPs	Download
DS-CNN	Google speech command	22.7K	8/4/4	91.72%		23.1	5

Point cloud

Classification

Architecture	Dataset	#Params	Quantization	Accuracy	Size (KB)	NPs	Download
PointNet++	ModelNet40 3D Point Cloud	602K	8/4/4	79.78%	490.9	12

Akida 2.0 models

For 2.0 models, both 8-bit PTQ and 4-bit QAT numbers are given. When not explicitely stated 8-bit PTQ accuracy is given as is (ie no further tuning/training, only quantization and calibration). The 4-bit QAT is the same as for 1.0.

Note

The digit for quantization scheme stands for both weights and activations bitwidth. Weights in the first layer are always quantized to 8-bit. When given, ‘edge’ means that the model backbone output (before classification layer) is quantized to 1-bit to allow Akida edge learning.

Image domain

Classification

CNNs

Architecture	Resolution	Dataset	#Params	Quantization	Accuracy	Download
AkidaNet 0.25	160	ImageNet	483K	8 4	48.50% 41.60%
AkidaNet 0.5	160	ImageNet	1.4M	8 4	61.86% 57.93%
AkidaNet	160	ImageNet	4.4M	8 4	69.94% 67.23%
AkidaNet 0.25	224	ImageNet	483K	8 4	52.39% 46.06%
AkidaNet 0.5	224	ImageNet	1.4M	8 4	64.88% 61.47%
AkidaNet	224	ImageNet	4.4M	8 4	72.16% 70.11%
AkidaNet 0.5	224	PlantVillage	1.2M	8 4	99.61% 98.25%
AkidaNet 0.25	96	Visual Wake Words	227K	8 4	87.03% 85.80%
AkidaNet18	160	ImageNet	2.4M	8	64.77%
AkidaNet18	224	ImageNet	2.4M	8	67.32%
MobileNetV1 0.25	160	ImageNet	469K	8 4	45.72% 37.51%
MobileNetV1 0.5	160	ImageNet	1.3M	8 4	60.27% 54.81%
MobileNetV1	160	ImageNet	4.2M	8 4	69.02% 65.28%
MobileNetV1 0.25	224	ImageNet	469K	8 4	49.63% 42.08%
MobileNetV1 0.5	224	ImageNet	1.3M	8 4	63.65% 59.20%
MobileNetV1	224	ImageNet	4.2M	8 4	71.18% 68.52%
GXNOR	28	MNIST	1.6M	4	98.81%

Transformers

Architecture	Resolution	Dataset	#Params	Quantization	Accuracy	Download
ViT	224	ImageNet	5.8M	8	73.79% [1]
DeiT-dist	224	ImageNet	6.0M	8	74.34% [1]

[1] (1,2)

PTQ accuracy boosted with 1 epoch of QAT.

Object detection

Architecture	Resolution	Dataset	#Params	Quantization	mAP 50	mAP 75	mAP	Download
YOLOv2 (AkidaNet 0.5 backbone)	224	PASCAL-VOC 2007	3.6M	8 4	50.96% 47.21%	27.40% 23.21%	27.79% 24.66%
CenterNet (AkidaNet18 backbone)	384	PASCAL-VOC 2007	2.4M	8	70.32%	47.30%	43.88%
YOLOv2 (AkidaNet 0.5 backbone)	224	WIDER FACE	3.6M	8 4	80.19% 78.60%

Regression

Architecture	Resolution	Dataset	#Params	Quantization	MAE	Download
VGG-like	32	UTKFace (age estimation)	458K	8 4	6.0304 5.8858

Face recognition

Architecture	Resolution	Dataset	#Params	Quantization	Accuracy	Download
AkidaNet 0.5	112×96	CASIA Webface face identification	2.3M	8 4	72.83% 69.79%

Segmentation

Architecture	Resolution	Dataset	#Params	Quantization	Binary IOU	Download
AkidaUNet 0.5	128	Portrait128	1.1M	8	0.9057 [2]

[2]

PTQ accuracy boosted with 1 epoch QAT.

Audio domain

Keyword spotting

Architecture	Dataset	#Params	Quantization	Top-1 accuracy	Download
DS-CNN	Google speech command	23.8K	8 4 4 + edge	92.87% 92.67% 90.61%

Classification

Architecture	Resolution	Dataset	#Params	Quantization	Accuracy	Download
ViT (1 block)	224	Urbansound8k	539.9K	8	97.73% [3]

[3]

PTQ accuracy boosted with 5 epochs QAT.

Point cloud

Classification

Architecture	Dataset	#Params	Quantization	Accuracy	Download
PointNet++	ModelNet40 3D Point Cloud	605K	8 4	80.88% [4] 81.56%

[4]

PTQ accuracy boosted with 5 epochs QAT.