English | 简体中文
📌AiDB : A toolbox for deep learning model deployment using C++. Abstract mainstream deep learning inference frameworks into unified interfaces, including ONNXRUNTIME, MNN, NCNN, TNN, PaddleLite, and OpenVINO. Provide deployment demo for multiple scenarios and languages. This project is not only used to record learning, but also a summary of my work over the past few years. If this project inspires you or helps you, welcome ⭐Star⭐ to support me, which is the driving force for me to keep updating! 🍔
wasm_capture.mp4
wasm_webcam.mp4
- rich backend: integrating mainstream AI inference frameworks, including OnnxRuntime, MNN, NCNN, TNN, PaddleLite, and OpenVINO;
- user friendly: abstracting all inference frameworks into a unified interface and selecting them through configuration files for ease of use;
- multiple scenarios: support Linux、MacOS、Android(Win64 coming soon、IOS--poverty has limited my work)，provide PC Demo(Qt) 、 Android Demo(Kotlin) 、 Lua Demo and minimal server deployment demo( Go Zeros and Python FastAPI )
- multiple languages: provide calling instances for Python, Lua, and Go;
-
Try out the web demo:
-
Run the C++ demo on Google Colab:
-
Run the Python demo on Google Colab:
-
Run the Go demo on Google Colab:
-
Run the Lua demo on Google Colab:
- Features
- Demo
- Demo Repo
- FrameWork
- Quick Start
- Sample Usage
- Pitfalls
- Maintainers
- Contributing
- LICENSE
There are two modes available, S mode(C api) and H mode(C++ api)
Recommend using Docker
docker pull mister5ive/ai.deploy.boxgit lfs clone https://github.com/TalkUHulk/ai.deploy.box.git
cd ai.deploy.box.git
docker build -t aidb-dev .
docker run -it --name aidb-test aidb-devgit lfs clone https://github.com/TalkUHulk/ai.deploy.box.git
cd ai.deploy.box.git
mkdir build && cd build
cmake .. -DC_API={ON/OFF} -DBUILD_SAMPLE={ON/OFF} -DBUILD_PYTHON={ON/OFF} -DBUILD_LUA={ON/OFF} -DENGINE_NCNN_WASM={ON/OFF} -DOPENCV_HAS_FREETYPE={ON/OFF} -DENGINE_MNN={ON/OFF} -DENGINE_ORT={ON/OFF} -DENGINE_NCNN={ON/OFF} -DENGINE_TNN={ON/OFF} -DENGINE_OPV={ON/OFF} -DENGINE_PPLite={ON/OFF}
make -j8- C_API: compile C library;
- BUILD_PYTHON: compile Python api;
- BUILD_LUA: compile Lua api;
- OPENCV_HAS_FREETYPE: opencv-contrib complied with freetype or not. cv::putText can put chinese;
- BUILD_SAMPLE: compile sample;
- ENGINE_NCNN_WASM: compile sample;
- ENGINE_MNN: enable mnn;
- ENGINE_ORT: enable onnxruntim;
- ENGINE_NCNN: enable ncnn;
- ENGINE_TNN: enable tnn;
- ENGINE_OPV: enable openvino;
- ENGINE_PPLite: enable paddle-lite;
Model Lite: MEGA | Baidu: 92e8
Model List
Example: use scrfd detect face by mnn:
#include <opencv2/opencv.hpp>
#include "Interpreter.h"
#include "utility/Utility.h"
auto interpreter = AIDB::Interpreter::createInstance("scrfd_500m_kps", "mnn");
auto bgr = cv::imread("./doc/test/face.jpg");
cv::Mat blob = *det_ins << bgr;
std::vector<std::vector<float>> outputs;
std::vector<std::vector<int>> outputs_shape;
det_ins->forward((float*)blob.data, det_ins->width(), det_ins->height(), det_ins->channel(), outputs, outputs_shape);
std::vector<std::shared_ptr<AIDB::FaceMeta>> face_metas;
assert(face_detect_input->scale_h() == face_detect_input->scale_w());
AIDB::Utility::scrfd_post_process(outputs, face_metas, det_ins->width(), det_ins->height(), det_ins->scale_h());
In linux, run source set_env.sh before test.
Sample Usage
Face Detect
./build/samples/FaceDetect model_name backend type inputfile-
model_name
- scrfd_10g_kps
- scrfd_2.5g_kps
- scrfd_500m_kps
-
backend
- ONNX
- MNN
- NCNN
- OpenVINO
- TNN
- PaddleLite
-
type
- 0 - image
- 1 - video
-
inputfile: 0 is webcam
Face landmark
./build/samples/FaceDetectWithLandmark model_name backend pfpld backend type inputfile-
model_name
- scrfd_10g_kps
- scrfd_2.5g_kps
- scrfd_500m_kps
-
backend
- ONNX
- MNN
- NCNN
- OpenVINO
- TNN
- PaddleLite
-
type
- 0 - image
- 1 - video
-
inputfile: 0 is webcam
Face Alignment
./build/samples/FaceDetectWith3DDFA det_model_name backend tddfa_model_name backend type inputfile-
det_backend
- scrfd_10g_kps
- scrfd_2.5g_kps
- scrfd_500m_kps
-
tddfa_model_name
- 3ddfa_mb1_bfm_base
- 3ddfa_mb1_bfm_dense
- 3ddfa_mb05_bfm_base
- 3ddfa_mb05_bfm_dense
-
backend
- ONNX
- MNN
- NCNN
- OpenVINO
- TNN
- PaddleLite
-
type
- 0 - image
- 1 - video
-
inputfile: 0 is webcam
Face Parsing
./build/samples/FaceParsing bisenet backend type inputfile-
backend
- ONNX
- MNN
- NCNN
- OpenVINO
- TNN
- PaddleLite
-
type
- 0 - image
- 1 - video
-
inputfile: 0 is webcam
Ocr
./build/samples/PPOcr ppocr_det det_backend ppocr_cls cls_backend ppocr_ret rec_backend type inputfile-
det_backend/cls_backend/rec_backend
- ONNX
- MNN
- NCNN
- OpenVINO
- PaddleLite
-
type
- 0 - image
- 1 - video
-
inputfile: 0 is webcam
YoloX
./build/samples/YoloX model_name backend type inputfile-
model_name
- yolox_tiny
- yolox_nano
- yolox_s
- yolox_m
- yolox_l
- yolox_x
- yolox_darknet
-
backend
- ONNX
- MNN
- NCNN
- OpenVINO
- TNN
- PaddleLite
-
type
- 0 - image
- 1 - video
-
inputfile: 0 is webcam
YoloV7
./build/samples/YoloV7 model_name backend type inputfile-
model_name
- yolov7_tiny
- yolov7_tiny_grid
- yolov7
- yolov7_grid
- yolov7x
- yolov7x_grid
- yolov7_d6_grid
- yolov7_e6_grid
-
backend
- ONNX
- MNN
- NCNN
- OpenVINO
- TNN
- PaddleLite
-
type
- 0 - image
- 1 - video
-
inputfile: 0 is webcam
YoloV8
./build/samples/YoloV8 model_name backend type inputfile-
model_name
- yolov8n
- yolov8s
- yolov8m
- yolov8l
- yolov8x
-
backend
- ONNX
- MNN
- NCNN
- OpenVINO
- TNN
- PaddleLite
-
type
- 0 - image
- 1 - video
-
inputfile: 0 is webcam
MobileVit
./build/samples/MobileViT model_name backend inputfile-
model_name
- mobilevit_xxs
- mobilevit_s
-
backend
- ONNX
- MNN
- OpenVINO
MoveNet
./build/samples/Movenet movenet backend type inputfile-
backend
- ONNX
- MNN
- NCNN
- OpenVINO
- TNN
- PaddleLite
-
type
- 0 - image
- 1 - video
-
inputfile: 0 is webcam
MobileStyleGan
./build/samples/MobileStyleGan mobilestylegan_mappingnetwork map_backend mobilestylegan_synthesisnetwork syn_backend- map_backend/syn_backend
- ONNX
- MNN
- OpenVINO
AnimeGan
./build/samples/AnimeGan model_name backend 0 inputfile-
model_name
- animeganv2_celeba_distill
- animeganv2_celeba_distill_dynamic
- animeganv2_face_paint_v1
- animeganv2_face_paint_v1_dynamic
- animeganv2_face_paint_v2
- animeganv2_face_paint_v2_dynamic
- animeganv2_paprika
- animeganv2_paprika_dynamic
-
backend
- ONNX
- MNN
- OpenVINO
Pitfalls
- Android-rtti
Q:
XXXX/3rdparty/yaml-cpp/depthguard.h:54:9: error: cannot use 'throw' with exceptions disabled
In file included from XXXX/opencv/native/jni/include/opencv2/opencv.hpp:65:
In file included from XXXX/opencv/native/jni/include/opencv2/flann.hpp:48:
In file included from XXXX/opencv/native/jni/include/opencv2/flann/flann_base.hpp:41:
In file included from XXXX/opencv/native/jni/include/opencv2/flann/params.h:35:
XXXX/opencv/native/jni/include/opencv2/flann/any.h:60:63: error: use of typeid requires -frttiA: app->build.gradle add "-fexceptions"
externalNativeBuild {
ndkBuild {
cppFlags "-std=c++11", "-fexceptions"
arguments "APP_OPTIM=release", "NDK_DEBUG=0"
abiFilters "arm64-v8a"
}
}if complie with ncnn, disable rtti.
cmake -DCMAKE_TOOLCHAIN_FILE=../../android-ndk-r25c/build/cmake/android.toolchain.cmake -DANDROID_ABI="arm64-v8a" -DANDROID_PLATFORM=android-24 -DNCNN_SHARED_LIB=ON -DANDROID_ARM_NEON=ON -DNCNN_DISABLE_EXCEPTION=OFF -DNCNN_DISABLE_RTTI=OFF ..reference: issues/3231
- Android-mnn
Q:
I/MNNJNI: Can't Find type=3 backend, use 0 instead A:
java load so 时，显式调 System.load("libMNN_CL.so") 。在 CMakeLists.txt 里面让你的 so link libMNN_CL 貌似是不行的。
init {
System.loadLibrary("aidb")
System.loadLibrary("MNN");
try {
System.loadLibrary("MNN_CL")
System.loadLibrary("MNN_Express")
System.loadLibrary("MNN_Vulkan")
} catch (ce: Throwable) {
Log.w("MNNJNI", "load MNN GPU so exception=%s", ce)
}
System.loadLibrary("mnncore")
}- Android Studio
Q:
Out of memory:Java heap space
A:
gradle.properties->org.gradle.jvmargs=-Xmx4g -Dfile.encoding=UTF-8- Android paddle-lite
Q: because kernel for 'calib' is not supported by Paddle-Lite. A: 使用带fp16标签的库
Q: kernel for 'conv2d' is not supported by Paddle-lite. A: 转模型--valid_targets =arm, 打开fp16，opt\lib版本对应
- Android OpenVINO
Q: How to deploy models in android with openvino(reference).
A:
-
step1
- build openvino library. Here are my compilation instructions:
cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=android-ndk-r25c/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=30 -DANDROID_STL=c++_shared -DENABLE_SAMPLES=OFF -DENABLE_OPENCV=OFF -DENABLE_CLDNN=OFF -DENABLE_VPU=OFF -DENABLE_GNA=OFF -DENABLE_MYRIAD=OFF -DENABLE_TESTS=OFF -DENABLE_GAPI_TESTS=OFF -DENABLE_BEH_TESTS=OFF ..
- build openvino library. Here are my compilation instructions:
-
step2
- Put openvino library(*.so) to assets.(plugin you need)
-
step3
- If your device is not root, put libc++.so and libc++_shared.so to assets.
Q:
dlopen failed: library “libc++_shared.so“ not foundA:
cmakelist.txt
add_library(libc++_shared STATIC IMPORTED)
set_target_properties(libc++_shared PROPERTIES IMPORTED_LOCATION ${CMAKE_CURRENT_LIST_DIR}/../libs/android/opencv/native/libs/${ANDROID_ABI}/libc++_shared.so)Q:
java.lang.UnsatisfiedLinkError: dlopen failed: cannot locate symbol "__emutls_get_address" referenced by "/data/app/~~DMBfRqeqFvKzb9yUIkUZiQ==/com.hulk.aidb_demo-HrziaiyGs2adLTT-aQqemg==/lib/arm64/libopenvino_arm_cpu_plugin.so"...A: Build openvino arm library, and put *.so in app/libs/${ANDROID_ABI}/. (need add jniLibs.srcDirs = ['libs'] in build.gradle)
Q:
library "/system/lib64/libc++.so" ("/system/lib64/libc++.so") needed or dlopened by "/apex/com.android.art/lib64/libnativeloader.so" is not accessible for the namespace: [name="classloader-namespace", ld_library_paths="", default_library_paths="/data/app/~~IWBQrjWXHt7o71mstUGRHA==/com.hulk.aidb_demo-cfq3aSk8DN62UDtoKV4Vfg==/lib/arm64:/data/app/~~IWBQrjWXHt7o71mstUGRHA==/com.hulk.aidb_demo-cfq3aSk8DN62UDtoKV4Vfg==/base.apk!/lib/arm64-v8a", permitted_paths="/data:/mnt/expand:/data/data/com.hulk.aidb_demo"]A: Put libc++.so in android studio app/libs/${ANDROID_ABI}/. (need add jniLibs.srcDirs = ['libs'] in build.gradle)
Feel free to dive in! Open an issue or submit PRs.
AiDB follows the Contributor Covenant Code of Conduct.
This project exists thanks to all the people who contribute.
MIT © Hulk Wang