Update ppseg with eigen functions (#238)

* Update ppseg backend support type * Update ppseg preprocess if condition * Update README.md * Update README.md * Update README.md * Update ppseg with eigen functions * Delete old argmax function * Update README.md * Delete apply_softmax in ppseg example demo * Update ppseg code with createFromTensor function * Delete FDTensor2CVMat function * Update README.md * Update README.md * Update README.md * Update README.md * Update ppseg model.cc with transpose&&softmax in place convert * Update segmentation_result.md * Update model.cc * Update README.md * Update README.md Co-authored-by: Jason <jiangjiajun@baidu.com>
2025-10-05 16:48:03 +08:00 · 2022-09-22 21:21:47 +08:00
parent 5ad7f64a3a
commit 625845c7d6
13 changed files with 237 additions and 204 deletions
--- a/examples/vision/segmentation/paddleseg/README.md
+++ b/examples/vision/segmentation/paddleseg/README.md
@@ -21,16 +21,19 @@ PaddleSeg模型导出，请参考其文档说明[模型导出](https://github.co

 ## 下载预训练模型

-为了方便开发者的测试，下面提供了PaddleSeg导出的部分模型（导出方式为：**不指定**`input_shape`和`with_softmax`，**指定**`without_argmax`），开发者可直接下载使用。
+为了方便开发者的测试，下面提供了PaddleSeg导出的部分模型（导出方式为：**不指定**`--input_shape`，**指定**`--output_op none`），开发者可直接下载使用。

 | 模型                                                               | 参数文件大小    |输入Shape |  mIoU | mIoU (flip) | mIoU (ms+flip) |
 |:---------------------------------------------------------------- |:----- |:----- | :----- | :----- | :----- |
 | [Unet-cityscapes](https://bj.bcebos.com/paddlehub/fastdeploy/Unet_cityscapes_without_argmax_infer.tgz) | 52MB | 1024x512 | 65.00% | 66.02% | 66.89% |
 | [PP-LiteSeg-T(STDC1)-cityscapes](https://bj.bcebos.com/paddlehub/fastdeploy/PP_LiteSeg_T_STDC1_cityscapes_without_argmax_infer.tgz)  | 31MB  | 1024x512 |73.10% | 73.89% | - |
-| [PP-HumanSegV1-Lite](https://bj.bcebos.com/paddlehub/fastdeploy/PP_HumanSegV1_Lite_infer.tgz) |  543KB | 192x192 | 86.2% | - | - |
-| [PP-HumanSegV1-Server](https://bj.bcebos.com/paddlehub/fastdeploy/PP_HumanSegV1_Server_infer.tgz) |  103MB | 512x512 | 96.47% | - | - |
+| [PP-HumanSegV1-Lite(通用人像分割模型)](https://bj.bcebos.com/paddlehub/fastdeploy/PP_HumanSegV1_Lite_infer.tgz) |  543KB | 192x192 | 86.2% | - | - |
+| [PP-HumanSegV2-Lite(通用人像分割模型)](https://bj.bcebos.com/paddle2onnx/libs/PP_HumanSegV2_Lite_192x192_infer.tgz) |  12MB | 192x192 | 92.52% | - | - |
+| [PP-HumanSegV2-Mobile(通用人像分割模型)](https://bj.bcebos.com/paddlehub/fastdeploy/PP_HumanSegV2_Mobile_192x192_infer.tgz) |  29MB | 192x192 | 93.13% | - | - |
+| [PP-HumanSegV1-Server(通用人像分割模型)](https://bj.bcebos.com/paddlehub/fastdeploy/PP_HumanSegV1_Server_infer.tgz) |  103MB | 512x512 | 96.47% | - | - |
+| [Portait-PP-HumanSegV2_Lite(肖像分割模型)](https://bj.bcebos.com/paddlehub/fastdeploy/Portrait_PP_HumanSegV2_Lite_256x144_infer.tgz) |  3.6M | 256x144 | 96.63% | - | - |
 | [FCN-HRNet-W18-cityscapes](https://bj.bcebos.com/paddlehub/fastdeploy/FCN_HRNet_W18_cityscapes_without_argmax_infer.tgz) |  37MB | 1024x512 | 78.97% | 79.49% | 79.74% |
-| [Deeplabv3-ResNet50-OS8-cityscapes](https://bj.bcebos.com/paddlehub/fastdeploy/Deeplabv3_ResNet50_OS8_cityscapes_without_argmax_infer.tgz) |  150MB | 1024x512 | 79.90% | 80.22% | 80.47% |
+| [Deeplabv3-ResNet101-OS8-cityscapes](https://bj.bcebos.com/paddlehub/fastdeploy/Deeplabv3_ResNet101_OS8_cityscapes_without_argmax_infer.tgz) |  150MB | 1024x512 | 79.90% | 80.22% | 80.47% |

 ## 详细部署文档

--- a/examples/vision/segmentation/paddleseg/cpp/README.md
+++ b/examples/vision/segmentation/paddleseg/cpp/README.md
@@ -7,7 +7,7 @@
 - 1. 软硬件环境满足要求，参考[FastDeploy环境要求](../../../../../docs/environment.md)  
 - 2. 根据开发环境，下载预编译部署库和samples代码，参考[FastDeploy预编译库](../../../../../docs/quick_start)

-以Linux上CPU推理为例，在本目录执行如下命令即可完成编译测试
+以Linux上推理为例，在本目录执行如下命令即可完成编译测试

 ```bash
 wget https://bj.bcebos.com/fastdeploy/release/cpp/fastdeploy-linux-x64-gpu-0.2.1.tgz
@@ -25,16 +25,16 @@ wget https://paddleseg.bj.bcebos.com/dygraph/demo/cityscapes_demo.png


 # CPU推理
-./infer_demo Unet_cityscapes_without_argmax_infer Unet_cityscapes_without_argmax_infer cityscapes_demo.png 0
+./infer_demo Unet_cityscapes_without_argmax_infer cityscapes_demo.png 0
 # GPU推理
-./infer_demo Unet_cityscapes_without_argmax_infer Unet_cityscapes_without_argmax_infer cityscapes_demo.png 1
+./infer_demo Unet_cityscapes_without_argmax_infer cityscapes_demo.png 1
 # GPU上TensorRT推理
-./infer_demo Unet_cityscapes_without_argmax_infer Unet_cityscapes_without_argmax_infer cityscapes_demo.png 2
+./infer_demo Unet_cityscapes_without_argmax_infer cityscapes_demo.png 2
 ```

 运行完成可视化结果如下图所示
 <div  align="center">  
-<img src="https://user-images.githubusercontent.com/16222477/184588768-45ee673b-ef1f-40f4-9fbd-6b1a9ce17c59.png", width=512px, height=256px />
+<img src="https://user-images.githubusercontent.com/16222477/191712880-91ae128d-247a-43e0-b1e3-cafae78431e0.jpg", width=512px, height=256px />
 </div>

 以上命令只适用于Linux或MacOS, Windows下SDK的使用方式请参考:  
@@ -80,10 +80,10 @@ PaddleSegModel模型加载和初始化，其中model_file为导出的Paddle模
 #### 预处理参数
 用户可按照自己的实际需求，修改下列预处理参数，从而影响最终的推理和部署效果

-> > * **is_vertical_screen**(bool): PP-HumanSeg系列模型通过设置此参数为`True`表明输入图片是竖屏，即height大于width的图片
+> > * **is_vertical_screen**(bool): PP-HumanSeg系列模型通过设置此参数为`true`表明输入图片是竖屏，即height大于width的图片

 #### 后处理参数
-> > * **with_softmax**(bool): 当模型导出时，并未指定`with_softmax`参数，可通过此设置此参数为`True`，将预测的输出分割标签（label_map）对应的概率结果(score_map)做softmax归一化处理
+> > * **appy_softmax**(bool): 当模型导出时，并未指定`apply_softmax`参数，可通过此设置此参数为`true`，将预测的输出分割标签（label_map）对应的概率结果(score_map)做softmax归一化处理

 - [模型介绍](../../)
 - [Python部署](../python)
--- a/examples/vision/segmentation/paddleseg/cpp/infer.cc
+++ b/examples/vision/segmentation/paddleseg/cpp/infer.cc
@@ -26,6 +26,7 @@ void CpuInfer(const std::string& model_dir, const std::string& image_file) {
  auto config_file = model_dir + sep + "deploy.yaml";
  auto model = fastdeploy::vision::segmentation::PaddleSegModel(
      model_file, params_file, config_file);
+
  if (!model.Initialized()) {
    std::cerr << "Failed to initialize." << std::endl;
    return;
@@ -40,6 +41,7 @@ void CpuInfer(const std::string& model_dir, const std::string& image_file) {
    return;
  }

+  std::cout << res.Str() << std::endl;
  auto vis_im = fastdeploy::vision::Visualize::VisSegmentation(im_bak, res);
  cv::imwrite("vis_result.jpg", vis_im);
  std::cout << "Visualized result saved in ./vis_result.jpg" << std::endl;
@@ -54,6 +56,7 @@ void GpuInfer(const std::string& model_dir, const std::string& image_file) {
  option.UseGpu();
  auto model = fastdeploy::vision::segmentation::PaddleSegModel(
      model_file, params_file, config_file, option);
+
  if (!model.Initialized()) {
    std::cerr << "Failed to initialize." << std::endl;
    return;
@@ -68,6 +71,7 @@ void GpuInfer(const std::string& model_dir, const std::string& image_file) {
    return;
  }

+  std::cout << res.Str() << std::endl;
  auto vis_im = fastdeploy::vision::Visualize::VisSegmentation(im_bak, res);
  cv::imwrite("vis_result.jpg", vis_im);
  std::cout << "Visualized result saved in ./vis_result.jpg" << std::endl;
@@ -85,6 +89,7 @@ void TrtInfer(const std::string& model_dir, const std::string& image_file) {
                          {1, 3, 2048, 2048});
  auto model = fastdeploy::vision::segmentation::PaddleSegModel(
      model_file, params_file, config_file, option);
+
  if (!model.Initialized()) {
    std::cerr << "Failed to initialize." << std::endl;
    return;
@@ -99,6 +104,7 @@ void TrtInfer(const std::string& model_dir, const std::string& image_file) {
    return;
  }

+  std::cout << res.Str() << std::endl;
  auto vis_im = fastdeploy::vision::Visualize::VisSegmentation(im_bak, res);
  cv::imwrite("vis_result.jpg", vis_im);
  std::cout << "Visualized result saved in ./vis_result.jpg" << std::endl;
--- a/examples/vision/segmentation/paddleseg/python/README.md
+++ b/examples/vision/segmentation/paddleseg/python/README.md
@@ -27,7 +27,7 @@ python infer.py --model Unet_cityscapes_without_argmax_infer --image cityscapes_

 运行完成可视化结果如下图所示
 <div  align="center">  
-<img src="https://user-images.githubusercontent.com/16222477/184588768-45ee673b-ef1f-40f4-9fbd-6b1a9ce17c59.png", width=512px, height=256px />
+<img src="https://user-images.githubusercontent.com/16222477/191712880-91ae128d-247a-43e0-b1e3-cafae78431e0.jpg", width=512px, height=256px />
 </div>

 ## PaddleSegModel Python接口
@@ -69,7 +69,7 @@ PaddleSeg模型加载和初始化，其中model_file, params_file以及config_fi
 > > * **is_vertical_screen**(bool): PP-HumanSeg系列模型通过设置此参数为`true`表明输入图片是竖屏，即height大于width的图片

 #### 后处理参数
-> > * **with_softmax**(bool): 当模型导出时，并未指定`with_softmax`参数，可通过此设置此参数为`true`，将预测的输出分割标签（label_map）对应的概率结果(score_map)做softmax归一化处理
+> > * **apply_softmax**(bool): 当模型导出时，并未指定`apply_softmax`参数，可通过此设置此参数为`true`，将预测的输出分割标签（label_map）对应的概率结果(score_map)做softmax归一化处理

 ## 其它文档