add PP-OCRv4 support

hhx518 · Jul 24, 2024 · 65e7480 · 65e7480
1 parent 712782d
commit 65e7480
Show file tree

Hide file tree

Showing 44 changed files with 13,857 additions and 3 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -49,6 +49,19 @@ link_directories(
 
 set(CMAKE_CXX_FLAGS  "${CMAKE_CXX_FLAGS} -std=c++11 -Wall -O0 -Wfatal-errors -pthread -w -g")
 set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -std=c++11 -O0 -Xcompiler -fPIC -g -w ${CUDA_GEN_CODE}")
+
+option(BUILD_PANGO_BACKEND "Build with Pango backend" OFF)
+option(BUILD_TRUETYPE_BACKEND "Build with TrueType backend" ON)
+
+if(BUILD_PANGO_BACKEND)
+    set(PANGO_LIBS ${PANGO_LIBS} pango-1.0 cairo pangocairo-1.0 glib-2.0 gobject-2.0)
+    add_definitions(-DENABLE_TEXT_BACKEND_PANGO)
+endif()
+
+if(BUILD_TRUETYPE_BACKEND)
+    add_definitions(-DENABLE_TEXT_BACKEND_STB)
+endif()
+
 file(GLOB_RECURSE cpp_srcs ${PROJECT_SOURCE_DIR}/src/*.cpp)
 file(GLOB_RECURSE cuda_srcs ${PROJECT_SOURCE_DIR}/src/*.cu)
 cuda_add_library(plugin_list SHARED ${cuda_srcs})
@@ -77,6 +90,7 @@ target_link_libraries(pro protobuf pthread plugin_list)
 target_link_libraries(pro ${OpenCV_LIBS})
 target_link_libraries(pro opencv_core opencv_imgproc opencv_videoio opencv_highgui opencv_imgcodecs)
 # target_link_libraries(pro dl)
+# target_link_libraries(pro ${PANGO_LIBS})
 
 add_custom_target(
     yolo
@@ -134,6 +148,13 @@ add_custom_target(
     COMMAND ./pro rtmo
 )
 
+add_custom_target(
+    ppocr
+    DEPENDS pro
+    WORKING_DIRECTORY ${PROJECT_SOURCE_DIR}/workspace
+    COMMAND ./pro ppocr
+)
+
 add_custom_target(
     test_yolo_map
     DEPENDS pro

diff --git a/Makefile b/Makefile
@@ -1,6 +1,9 @@
 cc        := g++
 nvcc      = ${lean_cuda}/bin/nvcc
 
+build_pango_backend    := no
+build_truetype_backend := yes
+
 lean_protobuf  := /home/jarvis/protobuf
 # lean_tensor_rt := /opt/TensorRT-8.4.1.5
 # RT-DETR 必须指定高版本的 tensorRT
@@ -44,6 +47,15 @@ link_librarys := opencv_core opencv_imgproc opencv_videoio opencv_highgui opencv
 			cuda cublas cudart cudnn \
 			stdc++ protobuf dl
 
+ifeq ($(build_pango_backend), yes)
+link_librarys += pango-1.0 cairo pangocairo-1.0 glib-2.0 gobject-2.0
+cppdefine += -DENABLE_TEXT_BACKEND_PANGO
+endif
+
+ifeq ($(build_truetype_backend), yes)
+cppdefine += -DENABLE_TEXT_BACKEND_STB
+endif
+
 empty         :=
 export_path   := $(subst $(empty) $(empty),:,$(library_paths))
 
@@ -52,7 +64,7 @@ include_paths := $(foreach item,$(include_paths),-I$(item))
 library_paths := $(foreach item,$(library_paths),-L$(item))
 link_librarys := $(foreach item,$(link_librarys),-l$(item))
 
-cpp_compile_flags := -std=c++11 -g -w -O0 -fPIC -pthread -fopenmp
+cpp_compile_flags := -std=c++11 -g -w -O0 -fPIC -pthread -fopenmp $(cppdefine)
 cu_compile_flags  := -std=c++11 -g -w -O0 -Xcompiler "$(cpp_compile_flags)" $(cuda_arch)
 link_flags        := -pthread -fopenmp -Wl,-rpath='$$ORIGIN'
 
@@ -119,6 +131,9 @@ rtdetr : workspace/pro
 rtmo : workspace/pro
 	@cd workspace && ./pro rtmo
 
+ppocr : workspace/pro
+	@cd workspace && ./pro ppocr
+
 test_yolo_map : workspace/pro
 	@cd workspace && ./pro test_yolo_map
 
@@ -131,5 +146,5 @@ clean :
 .PHONY : clean yolo  debug
 
 # 导出符号，使得运行时能够链接上
-# export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/zhouwenguang/lean/TensorRT-8.5.1.7/lib
+# export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/jarvis/lean/TensorRT-8.6.1.6/lib
 export LD_LIBRARY_PATH:=$(export_path):$(LD_LIBRARY_PATH)
diff --git a/README.md b/README.md
@@ -3,7 +3,7 @@
 
 该仓库基于 [shouxieai/tensorRT_Pro](https://github.com/shouxieai/tensorRT_Pro)，并进行了调整以支持 YOLOv8 的各项任务。
 
-* 目前已支持 YOLOv8、YOLOv8-Cls、YOLOv8-Seg、YOLOv8-OBB、YOLOv8-Pose、RT-DETR、ByteTrack、YOLOv9、YOLOv10、RTMO 高性能推理！！！🚀🚀🚀
+* 目前已支持 YOLOv8、YOLOv8-Cls、YOLOv8-Seg、YOLOv8-OBB、YOLOv8-Pose、RT-DETR、ByteTrack、YOLOv9、YOLOv10、RTMO、PP-OCRv4 高性能推理！！！🚀🚀🚀
 * 基于 tensorRT8.x，C++ 高级接口，C++ 部署，服务器/嵌入式使用
 
 <div align=center><img src="./assets/output.jpg" width="50%" height="50%"></div>
@@ -20,8 +20,14 @@
 - 🔥 [MMPose-RTMO推理详解及部署实现（上）](https://blog.csdn.net/qq_40672115/article/details/139364023)
 - 🔥 [MMPose-RTMO推理详解及部署实现（下）](https://blog.csdn.net/qq_40672115/article/details/139375752)
 - 🔥 [LayerNorm Plugin的使用与说明](https://blog.csdn.net/qq_40672115/article/details/140246052)
+- 🔥 [PaddleOCR-PP-OCRv4推理详解及部署实现（上）](https://blog.csdn.net/qq_40672115/article/details/140571346)
+- 🔥 [PaddleOCR-PP-OCRv4推理详解及部署实现（中）](https://blog.csdn.net/qq_40672115/article/details/140585830)
+- 🔥 [PaddleOCR-PP-OCRv4推理详解及部署实现（下）]
 
 ## Top News
+- **2024/7/24**
+  - PP-OCRv4 支持
+  - cuOSD 支持，代码 copy 自 [Lidar_AI_Solution/libraries/cuOSD](https://github.com/NVIDIA-AI-IOT/Lidar_AI_Solution/tree/master/libraries/cuOSD)
 - **2024/7/7**
   - LayerNorm Plugin 支持，代码 copy 自 [CUDA-BEVFusion/src/plugins/custom_layernorm.cu](https://github.com/NVIDIA-AI-IOT/Lidar_AI_Solution/blob/master/CUDA-BEVFusion/src/plugins/custom_layernorm.cu)
   - 提供 ONNX 模型下载（[Baidu Drive](https://pan.baidu.com/s/1MbPYzUEkONsjCPOudiTt1A?pwd=onnx)），方便大家测试使用
@@ -1214,6 +1220,48 @@ ${TRTEXEC} \
   > trtexec_output.log 2>&1
 ```
 
+</details>
+
+<details>
+<summary>PP-OCRv4支持</summary>
+
+1. 导出环境搭建
+
+```shell
+conda create --name paddleocr python=3.9
+conda activate paddleocr
+pip install shapely scikit-image imgaug pyclipper lmdb tqdm numpy==1.26.4 rapidfuzz onnxruntime
+pip install "opencv-python<=4.6.0.66" "opencv-contrib-python<=4.6.0.66" cython "Pillow>=10.0.0" pyyaml requests
+pip install paddlepaddle paddleocr paddle2onnx
+```
+
+2. 项目克隆
+
+```shell
+git clone https://github.com/PaddlePaddle/PaddleOCR.git
+```
+
+3. 预训练权重下载
+
+- 参考：[🛠️ PP-OCR 系列模型列表（更新中）](https://github.com/PaddlePaddle/PaddleOCR?tab=readme-ov-file#%EF%B8%8F-pp-ocr-%E7%B3%BB%E5%88%97%E6%A8%A1%E5%9E%8B%E5%88%97%E8%A1%A8%E6%9B%B4%E6%96%B0%E4%B8%AD)
+
+4. 导出 onnx 模型，具体流程请参考：[PaddleOCR-PP-OCRv4推理详解及部署实现（上）](https://blog.csdn.net/qq_40672115/article/details/140571346)
+
+5. engine 生成
+
+- **方案一**：利用 **TRT::compile** 接口，HardSwish 算子解析问题可以通过插件或者替换 onnxparser 解析器解决
+- **方案二**：利用 **trtexec** 工具生成 engine (**recommend**)
+
+```shell
+cd tensorRT_Pro-YOLOv8/workspace
+bash ocr_build.sh
+```
+
+6. 执行
+
+```shell
+make ppocr -j64
+```
 
 </details>
 

diff --git a/python/change_hardswish.py b/python/change_hardswish.py
@@ -0,0 +1,21 @@
+import onnx
+import onnx_graphsurgeon as gs
+
+# 加载 ONNX 模型
+input_model_path = "workspace/ppocr_det.sim.onnx"
+output_model_path = "workspace/ppocr_det.sim.plugin.onnx"
+graph = gs.import_onnx(onnx.load(input_model_path))
+
+# 遍历图中的所有节点
+for node in graph.nodes:
+    if node.op == "HardSwish":
+        node.op = "Plugin"
+        # 添加自定义属性
+        node.attrs["name"] = "HSwish"
+        node.attrs["info"] = "This is custom HardSwish node"
+
+# 删除无用的节点和张量
+graph.cleanup()
+
+# 导出修改后的模型
+onnx.save(gs.export_onnx(graph), output_model_path)