mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 08:37:06 +08:00

Files

DefTruth 5b143219ce [Docs] Pick seg fastdeploy docs from PaddleSeg (#1482 )

* [Docs] Pick seg fastdeploy docs from PaddleSeg

* [Docs] update seg docs

* [Docs] Add c&csharp examples for seg

* [Docs] Add c&csharp examples for seg

* [Doc] Update paddleseg README.md

* Update README.md

2023-03-17 11:22:46 +08:00

5.3 KiB

Raw Blame History

English | 简体中文

PaddleSeg 语义分割模型在华为昇腾上部署方案-FastDeploy

1. 说明

PaddleSeg支持通过FastDeploy在华为昇腾上部署Segmentation相关模型

2. 使用预导出的模型列表

模型	参数文件大小	输入Shape	mIoU	mIoU (flip)	mIoU (ms+flip)
PP-LiteSeg-B(STDC2)-cityscapes-with-argmax \| PP-LiteSeg-B(STDC2)-cityscapes-without-argmax	31MB	1024x512	79.04%	79.52%	79.85%
PP-HumanSegV1-Lite-with-argmax(通用人像分割模型) \| PP-HumanSegV1-Lite-without-argmax(通用人像分割模型)	543KB	192x192	86.2%	-	-
PP-HumanSegV2-Lite-with-argmax(通用人像分割模型) \| PP-HumanSegV2-Lite-without-argmax(通用人像分割模型)	12MB	192x192	92.52%	-	-
PP-HumanSegV2-Mobile-with-argmax(通用人像分割模型) \| PP-HumanSegV2-Mobile-without-argmax(通用人像分割模型)	29MB	192x192	93.13%	-	-
PP-HumanSegV1-Server-with-argmax(通用人像分割模型) \| PP-HumanSegV1-Server-without-argmax(通用人像分割模型)	103MB	512x512	96.47%	-	-
Portait-PP-HumanSegV2-Lite-with-argmax(肖像分割模型) \| Portait-PP-HumanSegV2-Lite-without-argmax(肖像分割模型)	3.6M	256x144	96.63%	-	-
FCN-HRNet-W18-cityscapes-with-argmax \| FCN-HRNet-W18-cityscapes-without-argmax(暂时不支持ONNXRuntime的GPU推理)	37MB	1024x512	78.97%	79.49%	79.74%
Deeplabv3-ResNet101-OS8-cityscapes-with-argmax \| Deeplabv3-ResNet101-OS8-cityscapes-without-argmax	150MB	1024x512	79.90%	80.22%	80.47%
SegFormer_B0-cityscapes-with-argmax \| SegFormer_B0-cityscapes-without-argmax	15MB	1024x1024	76.73%	77.16%	-

补充说明：

文件名标记了without-argmax的模型，导出方式为：不指定--input_shape，指定--output_op none
文件名标记了with-argmax的模型导出方式为：不指定--input_shape，指定--output_op argmax

3. 自行导出PaddleSeg部署模型

3.1 模型版本

支持PaddleSeg高于2.6版本的Segmentation模型，如果部署的为PP-Matting、PP-HumanMatting以及ModNet请参考Matting模型部署。目前FastDeploy测试过成功部署的模型:

3.2 模型导出

PaddleSeg模型导出，请参考其文档说明模型导出，注意：PaddleSeg导出的模型包含model.pdmodel、model.pdiparams和deploy.yaml三个文件，FastDeploy会从yaml文件中获取模型在推理时需要的预处理信息

3.3 导出须知

请参考模型导出中output_op参数的说明，获取您部署所需的模型，比如是否带argmax或softmax算子

5.3 KiB Raw Blame History Unescape Escape