Fixed-Point-Training

Low-bit & Hardware-aware Quantization Training

Experiment Results

CIFAR10

Methods	Models	Acc	Precision	Dataset
FP	ResNet-20	92.32	fp32	CIFAR10
FP8	ResNet-20	92.21	fp8	CIFAR10
Unified	ResNet-20	91.95	int8	CIFAR10
Distributed	ResNet-20	92.76	int8	CIFAR10

Methods	Models	Acc	Precision	Dataset
FP	MobileNetV2	94.39	fp32	CIFAR10
DoReFa	MobileNetV2	91.03	int8	CIFAR10
WAGEUBN	MobileNetV2	92.32	int8	CIFAR10
SBM	MobileNetV2	93.57	int8	CIFAR10
CPT	MobileNetV2	93.76	int8	CIFAR10
Unified	MobileNetV2	93.38	int8	CIFAR10
Distributed	MobileNetV2	94.37	int8	CIFAR10

Methods	Models	Acc	Precision	Dataset
FP	InceptionV3	94.89	fp32	CIFAR10
Unified	InceptionV3	95.00	int8	CIFAR10
Distributed	InceptionV3	95.21	int8	CIFAR10

CIFAR100

Methods	Models	Acc	Precision	Dataset
DoReFa	MobileNetV2	70.17	int8	CIFAR100
WAGEUBN	MobileNetV2	71.45	int8	CIFAR100
SBM	MobileNetV2	75.28	int8	CIFAR100
CPT	MobileNetV2	75.65	int8	CIFAR100

Methods	Models	Acc	Precision	Dataset
DoReFa	ResNet-74	69.31	int8	CIFAR100
WAGEUBN	ResNet-74	69.61	int8	CIFAR100
SBM	ResNet-74	71.44	int8	CIFAR100
CPT	ResNet-74	72.35	int8	CIFAR100

ImageNet

Todo-List

• full-precision baseline for CIFAR10/CIFAR100
• A Simple int8 quantization training Framework
• CPT baseline for CIFAR10/CIFAR100