28 lines
2.0 KiB
Markdown
28 lines
2.0 KiB
Markdown
# 量化测试结果
|
||
|
||
目前 mmdeploy 支持 ncnn 量化
|
||
|
||
## ncnn 量化
|
||
|
||
### 分类任务
|
||
|
||
| model | dataset | fp32 top-1 (%) | int8 top-1 (%) |
|
||
| :--------------------------------------------------------------------------------------------------------------------------: | :---------: | :------------: | :------------: |
|
||
| [ResNet-18](https://github.com/open-mmlab/mmclassification/blob/master/configs/resnet/resnet18_8xb16_cifar10.py) | Cifar10 | 94.82 | 94.83 |
|
||
| [ResNeXt-32x4d-50](https://github.com/open-mmlab/mmclassification/blob/master/configs/resnext/resnext50-32x4d_8xb32_in1k.py) | ImageNet-1k | 77.90 | 78.20\* |
|
||
| [MobileNet V2](https://github.com/open-mmlab/mmclassification/blob/master/configs/mobilenet_v2/mobilenet-v2_8xb32_in1k.py) | ImageNet-1k | 71.86 | 71.43\* |
|
||
| [HRNet-W18\*](https://github.com/open-mmlab/mmclassification/blob/master/configs/hrnet/hrnet-w18_4xb32_in1k.py) | ImageNet-1k | 76.75 | 76.25\* |
|
||
|
||
备注：
|
||
|
||
- 因为 imagenet-1k 数据量很大、ncnn 未正式发布 Vulkan int8 版本，考虑到 CPU 运行时间，仅用部分测试集（4000/50000）
|
||
- 量化后精度会有差异，分类模型涨点 1% 以内是正常情况
|
||
|
||
### OCR 检测任务
|
||
|
||
| model | dataset | fp32 hmean | int8 hmean |
|
||
| :---------------------------------------------------------------------------------------------------------------: | :-------: | :--------: | :------------: |
|
||
| [PANet](https://github.com/open-mmlab/mmocr/blob/main/configs/textdet/panet/panet_r18_fpem_ffm_600e_icdar2015.py) | ICDAR2015 | 0.795 | 0.792 @thr=0.9 |
|
||
|
||
备注：[mmocr](https://github.com/open-mmlab/mmocr) 使用 `shapely` 计算 IoU，实现方法会导致轻微的精度差异
|