[Bug Fix] Fixed Ppclas Bugs (PaddlePaddle#1009)

* 更新ppclas * 更新ppclas * 更新ppclas * 更新ppclas
war3gu · Dec 29, 2022 · 0401580 · 0401580
1 parent ab09296
commit 0401580
Show file tree

Hide file tree

Showing 5 changed files with 63 additions and 32 deletions.
diff --git a/docs/cn/faq/rknpu2/rknpu2.md b/docs/cn/faq/rknpu2/rknpu2.md
@@ -23,4 +23,4 @@ ONNX模型不能直接调用RK芯片中的NPU进行运算，需要把ONNX模型
 | Segmentation   | [PP-HumanSegV2Lite](../../../../examples/vision/segmentation/paddleseg/rknpu2/README.md) | portrait(int8)           | 133/43             |
 | Segmentation   | [PP-HumanSegV2Lite](../../../../examples/vision/segmentation/paddleseg/rknpu2/README.md) | human(int8)              | 133/43             |
 | Face Detection | [SCRFD](../../../../examples/vision/facedet/scrfd/rknpu2/README.md)                      | SCRFD-2.5G-kps-640(int8) | 108/42             |
-| Classification | [ResNet](../../../../examples/vision/classification/paddleclas/rknpu2/README.md)         | ResNet50_vd              | -/92               |
+| Classification | [ResNet](../../../../examples/vision/classification/paddleclas/rknpu2/README.md)         | ResNet50_vd              | -/33               |
diff --git a/examples/vision/classification/paddleclas/rknpu2/README.md b/examples/vision/classification/paddleclas/rknpu2/README.md
@@ -3,6 +3,7 @@
 ## 转换模型
 下面以 ResNet50_vd为例子，教大家如何转换分类模型到RKNN模型。
 
+### 导出ONNX模型
 ```bash
 # 安装 paddle2onnx
 pip install paddle2onnx
@@ -17,34 +18,59 @@ paddle2onnx --model_dir ResNet50_vd_infer  \
             --params_filename inference.pdiparams  \
             --save_file ResNet50_vd_infer/ResNet50_vd_infer.onnx  \
             --enable_dev_version True  \
-            --opset_version 12  \
+            --opset_version 10  \
             --enable_onnx_checker True
 
 # 固定shape，注意这里的inputs得对应netron.app展示的 inputs 的 name，有可能是image 或者 x
 python -m paddle2onnx.optimize --input_model ResNet50_vd_infer/ResNet50_vd_infer.onnx \
                                --output_model ResNet50_vd_infer/ResNet50_vd_infer.onnx \
                                --input_shape_dict "{'inputs':[1,3,224,224]}"
-```                               
+```  
 
- ### 编写模型导出配置文件
-以转化RK3588的RKNN模型为例子，我们需要编辑tools/rknpu2/config/ResNet50_vd_infer_rknn.yaml，来转换ONNX模型到RKNN模型。                              
+### 编写模型导出配置文件
+以转化RK3588的RKNN模型为例子，我们需要编辑tools/rknpu2/config/ResNet50_vd_infer_rknn.yaml，来转换ONNX模型到RKNN模型。
 
-默认的 mean=0, std=1是在内存做normalize，如果你需要在NPU上执行normalize操作，请根据你的模型配置normalize参数，例如:
+如果你需要在NPU上执行normalize操作，请根据你的模型配置normalize参数，例如:
 ```yaml
-model_path: ./ResNet50_vd_infer.onnx
-output_folder: ./
-target_platform: RK3588
-normalize:
-  mean: [[0.485,0.456,0.406]]
-  std: [[0.229,0.224,0.225]]
-outputs: []
-outputs_nodes: []
+model_path: ./ResNet50_vd_infer/ResNet50_vd_infer.onnx
+output_folder: ./ResNet50_vd_infer
+mean:
+  -
+    - 123.675
+    - 116.28
+    - 103.53
+std:
+  -
+    - 58.395
+    - 57.12
+    - 57.375
+outputs_nodes:
 do_quantization: False
-dataset: 
+dataset: "./ResNet50_vd_infer/dataset.txt"
 ```
 
+**在CPU上做normalize**可以参考以下yaml：
+```yaml
+model_path: ./ResNet50_vd_infer/ResNet50_vd_infer.onnx
+output_folder: ./ResNet50_vd_infer
+mean:
+  -
+    - 0
+    - 0
+    - 0
+std:
+  -
+    - 1
+    - 1
+    - 1
+outputs_nodes:
+do_quantization: False
+dataset: "./ResNet50_vd_infer/dataset.txt"
+```
+这里我们选择在NPU上执行normalize操作.
+
 
-# ONNX模型转RKNN模型
+### ONNX模型转RKNN模型
 ```shell
 python tools/rknpu2/export.py \
         --config_path tools/rknpu2/config/ResNet50_vd_infer_rknn.yaml \
@@ -54,4 +80,4 @@ python tools/rknpu2/export.py \
 ## 其他链接
 - [Cpp部署](./cpp)
 - [Python部署](./python)
-- [视觉模型预测结果](../../../../../docs/api/vision_results/)
+- [视觉模型预测结果](../../../../../docs/api/vision_results/)
diff --git a/fastdeploy/vision/classification/ppcls/preprocessor.cc b/fastdeploy/vision/classification/ppcls/preprocessor.cc
@@ -57,7 +57,7 @@ bool PaddleClasPreprocessor::BuildPreprocessPipelineFromConfig() {
       int height = op.begin()->second["size"].as<int>();
       processors_.push_back(std::make_shared<CenterCrop>(width, height));
     } else if (op_name == "NormalizeImage") {
-      if (!disable_normalize) {
+      if (!disable_normalize_) {
         auto mean = op.begin()->second["mean"].as<std::vector<float>>();
         auto std = op.begin()->second["std"].as<std::vector<float>>();
         auto scale = op.begin()->second["scale"].as<float>();
@@ -67,7 +67,7 @@ bool PaddleClasPreprocessor::BuildPreprocessPipelineFromConfig() {
         processors_.push_back(std::make_shared<Normalize>(mean, std));
       }
     } else if (op_name == "ToCHWImage") {
-      if (!disable_permute) {
+      if (!disable_permute_) {
         processors_.push_back(std::make_shared<HWC2CHW>());
       }
     } else {
@@ -83,14 +83,14 @@ bool PaddleClasPreprocessor::BuildPreprocessPipelineFromConfig() {
 }
 
 void PaddleClasPreprocessor::DisableNormalize() {
-  this->disable_normalize = true;
+  this->disable_normalize_ = true;
   // the DisableNormalize function will be invalid if the configuration file is loaded during preprocessing
   if (!BuildPreprocessPipelineFromConfig()) {
     FDERROR << "Failed to build preprocess pipeline from configuration file." << std::endl;
   }
 }
 void PaddleClasPreprocessor::DisablePermute() {
-  this->disable_permute = true;
+  this->disable_permute_ = true;
   // the DisablePermute function will be invalid if the configuration file is loaded during preprocessing
   if (!BuildPreprocessPipelineFromConfig()) {
     FDERROR << "Failed to build preprocess pipeline from configuration file." << std::endl;

diff --git a/fastdeploy/vision/classification/ppcls/preprocessor.h b/fastdeploy/vision/classification/ppcls/preprocessor.h
@@ -59,9 +59,9 @@ class FASTDEPLOY_DECL PaddleClasPreprocessor {
   // GPU device id
   int device_id_ = -1;
   // for recording the switch of hwc2chw
-  bool disable_permute = false;
+  bool disable_permute_ = false;
   // for recording the switch of normalize
-  bool disable_normalize = false;
+  bool disable_normalize_ = false;
   // read config file
   std::string config_file_;
 };

diff --git a/tools/rknpu2/config/ResNet50_vd_infer_rknn.yaml b/tools/rknpu2/config/ResNet50_vd_infer_rknn.yaml
@@ -1,10 +1,15 @@
-model_path: ./ResNet50_vd_infer.onnx
-output_folder: ./
-target_platform: RK3588
-normalize:
-  mean: [[0, 0, 0]]
-  std: [[1, 1, 1]]
-outputs: []
-outputs_nodes: []
+model_path: ./ResNet50_vd_infer/ResNet50_vd_infer.onnx
+output_folder: ./ResNet50_vd_infer
+mean:
+  -
+    - 123.675
+    - 116.28
+    - 103.53
+std:
+  -
+    - 58.395
+    - 57.12
+    - 57.375
+outputs_nodes:
 do_quantization: False
-dataset: 
+dataset: "./ResNet50_vd_infer/dataset.txt"