1. 量化#

1.1 图象分类#

数据集：ImageNet1000类

评价指标：Top-1/Top-5准确率

Model FP32 离线量化 量化训练
MobileNetV1 FP32 70.99%/89.68% xx%/xx% xx%/xx%
MobileNetV2 FP32 72.15%/90.65%
ResNet50 FP32 76.50%/93.00%

量化训练前后，模型大小的变化对比如下：

Model FP32 离线量化 量化训练
MobileNetV1 17M xx%/xx% xx%/xx%
MobileNetV2 xxM
ResNet50 99M

1.2 目标检测#

数据集：COCO 2017

Model 输入尺寸 Image/GPU FP32 BoxAP 离线量化 BoxAP 量化训练 BoxAP
MobileNet-V1-YOLOv3 608 8 29.3 xx xx
MobileNet-V1-YOLOv3 416
MobileNet-V1-YOLOv3 320
R50-dcn-YOLOv3 608 41.4
R50-dcn-YOLOv3 416
R50-dcn-YOLOv3 320

数据集：WIDER-FACE

评价指标：Easy/Medium/Hard mAP

Model 输入尺寸 Image/GPU FP32 离线量化 量化训练
BlazeFace 640 8 0.915/0.892/0.797 xx/xx/xx xx/xx/xx
BlazeFace-Lite 640 0.909/0.885/0.781
BlazeFace-NAS 640 0.837/0.807/0.658

1.3 图像分割#

数据集：Cityscapes

Model FP32 mIoU 离线量化 mIoU 量化训练 mIoU
DeepLabv3+/MobileNetv1 63.26 xx xx
DeepLabv3+/MobileNetv2 69.81

2. 剪枝#

2.1 图像分类#

数据集：ImageNet1000类

Model Top-1/Top-5
MobileNetV1 70.99%/89.68%
MobileNetV1 uniform -50%
MobileNetV1 sensitive -xx%
MobileNetV2 72.15%/90.65%
MobileNetV2 uniform -50%
MobileNetV2 sensitive -xx%
ResNet34 74.57%/92.14%
ResNet34 uniform -50%
ResNet34 auto -50%

2.2 目标检测#

数据集：Pasacl VOC & COCO 2017

Model 数据集 输入尺寸 Image/GPU baseline mAP 敏感度剪枝 mAP
MobileNet-V1-YOLOv3 Pasacl VOC 608 8 76.2 77.59 (-50%)
MobileNet-V1-YOLOv3 Pasacl VOC 416 76.7 xx
MobileNet-V1-YOLOv3 Pasacl VOC 320 75.2 xx
MobileNet-V1-YOLOv3 COCO 608 29.3 29.56 (-20%)
MobileNet-V1-YOLOv3 COCO 416 29.3 xx
MobileNet-V1-YOLOv3 COCO 320 27.1 xx
R50-dcn-YOLOv3 COCO 608 41.4 37.8 (-30%)
R50-dcn-YOLOv3 COCO 416
R50-dcn-YOLOv3 COCO 320

2.3 图像分割#

数据集：Cityscapes

Model Baseline mIoU xx剪枝 mIoU
DeepLabv3+/MobileNetv2 69.81 xx

3. 蒸馏#

3.1 图象分类#

数据集：ImageNet1000类

评价指标：Top-1/Top-5准确率

Model baseline 蒸馏后
MobileNetV1 70.99%/89.68% 72.79%/90.69% (teacher: ResNet50_vd1)
MobileNetV2 72.15%/90.65% 74.30%/91.52% (teacher: ResNet50_vd)
ResNet50 76.50%/93.00% 77.40%/93.48% (teacher: ResNet1012)

Note

  [1]：ResNet50_vd预训练模型Top-1/Top-5准确率分别为79.12%/94.44%

带_vd后缀代表开启了Mixup训练，Mixup相关介绍参考mixup: Beyond Empirical Risk Minimization

[2]：ResNet101预训练模型Top-1/Top-5准确率分别为77.56%/93.64%

3.2 目标检测#

数据集：Pasacl VOC & COCO 2017

Model 数据集 输入尺寸 Image/GPU baseline 蒸馏后 mAP
MobileNet-V1-YOLOv3 Pasacl VOC 640 16 76.2 79.0 (teacher: ResNet34-YOLOv3-VOC3)
MobileNet-V1-YOLOv3 Pasacl VOC 416 76.7 78.2
MobileNet-V1-YOLOv3 Pasacl VOC 320 75.2 75.5
MobileNet-V1-YOLOv3 COCO 640 29.3 31.0 (teacher: ResNet34-YOLOv3-COCO4)
MobileNet-V1-YOLOv3 COCO 416
MobileNet-V1-YOLOv3 COCO 320

Note

[3]：ResNet34-YOLOv3-VOC预训练模型的Box AP为82.6

[4]：ResNet34-YOLOv3-COCO预训练模型的Box AP为36.2