[Backend And DOC] 改进ppseg文档 + 为RKNPU2后端新增对多输入模型的支持 (#491)

* 11-02/14:35
* 新增输入数据format错误判断
* 优化推理过程,减少内存分配次数
* 支持多输入rknn模型
* rknn模型输出shape为三维时,输出将被强制对齐为4纬。现在将直接抹除rknn补充的shape,方便部分对输出shape进行判断的模型进行正确的后处理。

* 11-03/17:25
* 支持导出多输入RKNN模型
* 更新各种文档
* ppseg改用Fastdeploy中的模型进行转换

* 11-03/17:25
* 新增开源头

* 11-03/21:48
* 删除无用debug代码,补充注释
This commit is contained in:
Zheng_Bicheng
2022-11-04 09:39:23 +08:00
committed by GitHub
parent a36d49a973
commit ce828ecb38
12 changed files with 303 additions and 175 deletions

View File

@@ -2,7 +2,7 @@
在部署前,需确认以下两个步骤
- 1. 软硬件环境满足要求,参考[FastDeploy环境要求](../../../../../../docs/cn/build_and_install/rknpu2.md)
- 1. 软硬件环境满足要求,参考[FastDeploy环境要求](../../../../../../docs/cn/build_and_install/rknpu2.md)
【注意】如你部署的为**PP-Matting**、**PP-HumanMatting**以及**ModNet**请参考[Matting模型部署](../../../matting)
@@ -13,17 +13,13 @@
git clone https://github.com/PaddlePaddle/FastDeploy.git
cd FastDeploy/examples/vision/segmentation/paddleseg/python
# 下载模型
wget https://bj.bcebos.com/fastdeploy/models/rknn2/human_pp_humansegv2_lite_192x192_pretrained_3588.tgz
tar xvf human_pp_humansegv2_lite_192x192_pretrained_3588.tgz
# 下载图片
wget https://paddleseg.bj.bcebos.com/dygraph/pp_humanseg_v2/images.zip
unzip images.zip
# 推理
python3 infer.py --model_file ./human_pp_humansegv2_lite_192x192_pretrained_3588/human_pp_humansegv2_lite_192x192_pretrained_3588.rknn \
--config_file ./human_pp_humansegv2_lite_192x192_pretrained_3588/deploy.yaml \
python3 infer.py --model_file ./Portrait_PP_HumanSegV2_Lite_256x144_infer/Portrait_PP_HumanSegV2_Lite_256x144_infer_rk3588.rknn \
--config_file ./Portrait_PP_HumanSegV2_Lite_256x144_infer/deploy.yaml \
--image images/portrait_heng.jpg
```