mirrors/mmdeploy
1
0
Fork 0
mirror of https://github.com/open-mmlab/mmdeploy.git synced 2025-01-14 08:09:43 +08:00
Code Issues Projects Releases Wiki Activity
mmdeploy/docs/zh_cn/03-benchmark/quantization.md
tpoisonooo ea0a9e5d43
feat(tools/deploy.py): support ncnn quantization (#476) 
* feat(tools): add onnx2ncnn_quant_table

* feat(tools): add quantization image dataset

* feat(tools): add image dataset

* feat(tools/deploy.py): support quant

* fix(CI): lint

* fix(CI): format

* docs(zh_cn): add quantization usage

* docs(zh_cn): add benchmark

* feat(tools): add onnx2ncnn_quant_table

* docs(zh_cn): add more test result

* CI(github): add quant script

* CI(.github/scripts): add test quant

* fix(CI): remove pushd and popd

* feat(CI): debug

* fix(CI): path error

* fix(CI): fix path

* fix(CI): install wget

* fix(CI): review advices

* improvement(mmdeploy): review advice

* fix(tools): rename to onnx2ncnn_quant_table.py

* improvement(tools): rename file

* improvement(test): remove useless

* fix(tools/quant_image_dataset): remove loadFile in test.pipeline

* docs(quantization.md): update description

* fix(CI): protobuf version

* fix(CI): pip install

* docs(quantization): review advice

* fix(CI): revert mmcv version

* fix(CI): udpate pb version

* fix(CI): update
2022-05-26 19:53:56 +08:00

1.3 KiB
Raw Blame History

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 Cifar10 94.82 94.83
ResNeXt-32x4d-50 ImageNet-1k 77.90 78.20*
MobileNet V2 ImageNet-1k 71.86 71.43*
HRNet-W18* 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 ICDAR2015 0.795 0.792 @thr=0.9

备注：mmocr 使用 shapely 计算 IoU，实现方法会导致轻微的精度差异