mmdeploy/docs/zh_cn/03-benchmark/quantization.md
hanrui1sensetime 3b6e1ba34d
[Docs] Sync mmcls mmedit docstring 2.0 (#1347)
* update docs

* fix title

* update docstring of mmcls and mmedit

* fix mdformat

* fix lint

* fix docstring

* fix lint
2022-11-14 16:52:34 +08:00

37 lines
2.9 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# 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/1.x/configs/resnet/resnet18_8xb16_cifar10.py) | Cifar10 | 94.82 | 94.83 |
| [ResNeXt-32x4d-50](https://github.com/open-mmlab/mmclassification/blob/1.x/configs/resnext/resnext50-32x4d_8xb32_in1k.py) | ImageNet-1k | 77.90 | 78.20\* |
| [MobileNet V2](https://github.com/open-mmlab/mmclassification/blob/1.x/configs/mobilenet_v2/mobilenet-v2_8xb32_in1k.py) | ImageNet-1k | 71.86 | 71.43\* |
| [HRNet-W18\*](https://github.com/open-mmlab/mmclassification/blob/1.x/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/1.x/configs/textdet/panet/panet_resnet18_fpem-ffm_600e_icdar2015.py) | ICDAR2015 | 0.795 | 0.792 @thr=0.9 |
| [TextSnake](https://github.com/open-mmlab/mmocr/blob/1.x/configs/textdet/textsnake/textsnake_resnet50_fpn-unet_1200e_ctw1500.py) | CTW1500 | 0.817 | 0.818 |
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
### 姿态检测任务
| model | dataset | fp32 AP | int8 AP |
| :---------------------------------------------------------------------------------------------------------------------------------------------------: | :------: | :-----: | :-----: |
| [Hourglass](https://github.com/open-mmlab/mmpose/blob/1.x/configs/body_2d_keypoint/topdown_heatmap/coco/td-hm_hourglass52_8xb32-210e_coco-256x256.py) | COCO2017 | 0.717 | 0.713 |
备注:测试转换后的模型精度时,对于 mmpose 模型,在模型配置文件中 `flip_test` 需设置为 `False`