6.1.6.2. ModelZoo

6.1.6.2.1. Classification

HAT：

network	float	qat	quantization	dataset
MobileNetV1	74.10	73.69	73.67	ImageNet
MobileNetV2	72.64	72.13	72.09	ImageNet
ResNet 18	72.04	71.50	71.49	ImageNet
ResNet 50	77.37	76.65	76.67	ImageNet
VargNetV2	73.94	73.34	73.31	ImageNet
EfficientNet-B0	74.32	74.05	73.91	ImageNet
SwinTransformer	79.33	78.69	78.67	ImageNet

Torchvision(浮点模型来自社区)：

network	float	qat	quantization	dataset
ResNet 18	69.76	69.71	69.73	ImageNet
ResNet 50	76.13	76.07	76.06	ImageNet
MobileNetV2.py	71.88	71.27	71.27	ImageNet

6.1.6.2.2. Detection

RetinaNet

network	backbone	float	qat	quantization	dataset
Retinanet-vargnetv2	vargnetv2	31.53	31.52	31.58	MS COCO

YOLOv3

network	backbone	float	qat	quantization	dataset
YOLOv3-MobileNetv1	mobilenetv1	76.64	76.40	76.12	VOC
YOLOv3-VarGDarknet	VarGDarknet	33.9	33.6	33.6	COCO

FCOS

network	backbone	float	qat	quantization	dataset
FCOS-efficientnet	efficientnetb0	36.26	35.01	35.01	MS COCO
FCOS-efficientnet	efficientnetb2	45.35	44.97	44.97	MS COCO
FCOS-efficientnet	efficientnetb3	48.02	47.72	47.69	MS COCO

6.1.6.2.3. Segmentation

UNet

network	backbone	model	MeanIoU	MeanAcc	dataset
UNet	MobileNetV1	float	68.02	76.84	Cityscapes
UNet	MobileNetV1	qat	67.56	76.09	Cityscapes
UNet	MobileNetV1	quantization	67.52	76.00	Cityscapes

6.1.6.2.4. OpticalFlow

PwcNet

network	backbone	float	qat	quantization	dataset
PwcNet-lg	PwcNet	1.4114	1.4003	1.4138	FlyingChairs

6.1.6.2.5. Lidar

PointPillars

network	backbone	float	qat	quantization	dataset
PointPillars	SequentialBottleNeck	77.4475	76.8201	76.7314	KITTI

.. note::

PointPillars 的指标是 `Box3d Moderate` 这项。