💚

Author: springkim

3rdparty/src/pthread.c

@@ -13,9 +13,13 @@ typedef void (*windows_thread)(void *);
 int pthread_create(pthread_t *thread, const pthread_attr_t *attr, void
 				   *(*start_routine)(void *), void *arg)
 {
-	uintptr_t handle = _beginthread((windows_thread)start_routine,0,arg);
+	DWORD dwThreadId = 1;
+
+	HANDLE handle = CreateThread(NULL, 0, start_routine, arg, 0, &dwThreadId);
+
+	//uintptr_t handle = _beginthread((windows_thread)start_routine,0,arg);
 	thread->handle = (HANDLE)handle;
-	if(thread->handle == (HANDLE)-1){
+	if(thread->handle == (HANDLE)NULL){
 		return 1;
 	}else{
 		return 0;
@@ -36,6 +40,7 @@ void pthread_exit(void *value_ptr)
 int pthread_join(pthread_t thread, void **value_ptr)
 {
 	DWORD retvalue = WaitForSingleObject(thread.handle,INFINITE);
+	CloseHandle(thread.handle);
 	if(retvalue == WAIT_OBJECT_0){
 		return 0;
 	}else{

CMakeLists.txt

@@ -74,6 +74,8 @@ set(SOURCE_FILES
         src/im2col.h
         src/image.c
         src/image.h
+		src/iseg_layer.c
+		src/iseg_layer.h
         src/l2norm_layer.c
         src/l2norm_layer.h
         src/layer.c

README.md

@@ -1,6 +1,12 @@
 YOLOv3_SpringEdition <img src="https://i.imgur.com/oYejfWp.png" title="Windows8" width="48">
 --------------------------------------------------------------------------------------------
 
+
+작성중....
+VS2017
+LINUX
+기타 등등...
+
 <img src="https://i.imgur.com/ElCyyzT.png" title="Windows8" width="48"><img src="https://i.imgur.com/O5bye0l.png" width="48"><img src="https://i.imgur.com/kmfOMZz.png" width="48"><img src="https://i.imgur.com/6OT8yM9.png" width="48">
 
 #### YOLOv3 C++ Windows and Linux interface library. (Train,Detect both)
@@ -22,21 +28,54 @@ std::vector<BoxSE> boxes = detector.Detect(img, 0.5F);
 
 ### 1. Setup for train.
 
-You need only 2 files for train that are **YOLOv3SE_Train.exe** and **cudnn64_5.dll** on Windows. If you are on Linux, then you need only **YOLOv3SE_Train**. This files are in `YOLOv3_SpringEdition/bin`.
+#### 1.1. Train detector
+You need only 2 files for train that are **YOLOv3SE_Train.exe** and **cudnn64_7.dll** on Windows. If you are on Linux, then you need only **YOLOv3SE_Train**. This files are in `YOLOv3_SpringEdition/bin`.
 
 The requirement interface not changed. Same as **[pjreddie/darknet](https://github.com/pjreddie/darknet)**.
 
 There is a example training directory `Yolov3_SpringEdition_Train/`. You can start training using above files.
 
 Actually, all the interfaces are same with YOLOv2. So you can easily train your own data.
 
-The **YOLOv3SE_Train.exe**'s arguments are [base directory],[data file path] and [cfg file path].
+The **YOLOv3SE_Train.exe**'s arguments are `[option]`,`[base directory]`,`[data file path]` and `[cfg file path]`.
 
 And YOLOv3SE_Train.exe is automatically choosing multi-gpu training. and select latest backup weights file.
 
+#### 1.2. Train classifier
+Example : [Yolov3_SpringEdition_Train/ClassifierExample/](Yolov3_SpringEdition_Train/ClassifierExample/)
+
+##### Sample directory structure with STL10 dataset.
+```
+┌ stl10
+│  ├ airplane
+│  ├ bird
+│  ├ car
+│  ...
+│  └ truck
+├ backup
+├ resnext50.cfg
+├ stl10.data
+├ stl10.names
+├ train.txt
+├ cudnn64_7.dll
+├ YOLOv3SE_ Train.exe
+```
+##### Sample run argument with STL10 dataset.
+```
+"YOLOv3SE_Train.exe" classifier . stl10.data resnext50.cfg
+```
+
 ### 2. Setup for detect
 
-Just include **YOLOv3SE.h** and use it. See `YOLOv3_SpringEdition_Test/`. You need only **YOLOv3SE.h**, **libYOLOv3SE.dll** and **cudnn64_5.dll** for detect.
+**Do not change "batch" and "subdivisions" in cfg file. It automatically read those values as 1 when the network is on testing mode.**
+
+Just include **YOLOv3SE.h** and use it. See `YOLOv3_SpringEdition_Test/`. You need only **YOLOv3SE.h**, **libYOLOv3SE.dll** and **cudnn64_7.dll** for detect.
+
+###### 1. Go to [Yolov3_SpringEdition_Test](Yolov3_SpringEdition_Test)
+###### 2. Run "download_cudnn64_7.dll.bat" if you're in Windows.
+###### 3. Download "voc2007valid", "yolov3_darknet53.weights" for detection.
+###### Download "stl10valid", "resnext50_256_10000.weights" for classification.
+###### 4. Run VS solution or build.sh.
 
 ##### Reference
 
@@ -46,22 +85,40 @@ The class `YOLOv3` that in `YOLOv3SE.h` has 3 methods.
 void Create(std::string weights,std::string cfg,std::string names);
 ```
 
-This method load trained model(**weights**), network configuration(**cfg**) and class naming file(**names**\)* **Parameter** * **weights** : trained model path(e.g. "obj.weights") * **cfg** : network configuration file(e.g. "obj.cfg") * **names** : class naming file(e.g. "obj.names")
+This method load trained model(**weights**), network configuration(**cfg**) and class naming file(**names**\)*
+
+* **Parameter**
+	* **weights** : trained model path(e.g. "obj.weights") 
+	* **cfg** : network configuration file(e.g. "obj.cfg") 
+	* **names** : class naming file(e.g. "obj.names")
 
 ```cpp
 std::vector<BoxSE> Detect(cv::Mat img, float threshold);
 std::vector<BoxSE> Detect(std::string file, float threshold);
 std::vector<BoxSE> Detect(IplImage* img, float threshold);
+int Classify(IplImage* img);
+int Classify(cv::Mat img);
+int Classify(std::string file);
 ```
 
-This method is detecting objects of `file`,`cv::Mat` or `IplImage`.* **Parameter** * **file** : image file path * **img** : 3-channel image. * **threshold** : It removes predictive boxes if there score is less than threshold.
+This method is detecting objects or classify of `file`,`cv::Mat` or `IplImage`.
+* **Parameter** 
+	* **file** : image file path 
+	* **img** : 3-channel image. 
+	* **threshold** : It removes predictive boxes if there score is less than threshold.
+
 
 ```cpp
 void Release();
 ```
-
 Release loaded network.
 
+### 3. Setup for train(classification).
+
+
+
+
+
 Technical issue
 ---------------
 
@@ -73,11 +130,12 @@ change log
 **build_windows.bat** and **build_linux.sh** will download automatically correct version of cudnn. and build as cmake.
 
 ```
-Windows + 1080ti + CUDA8.0 + cudnn7.1 = 36FPS
-Windows + 1080ti + CUDA9.0 + cudnn7.1 = 36FPS
-Windows + 1080   + CUDA9.0 + cudnn7.1 = 27FPS
-Ubuntu  + 1080   + CUDA8.0 + cudnn7.1 = 30FPS
-Ubuntu  + 1080   + CUDA9.0 + cudnn7.1 = 30FPS
+Windows + 1080ti + CUDA8.0 + cudnn7.1 + yolov3      = 36FPS
+Windows + 1080ti + CUDA9.0 + cudnn7.1 + yolov3      = 36FPS
+Windows + 1080   + CUDA9.0 + cudnn7.1 + yolov3      = 27FPS
+Windows + 1080   + CUDA9.0 + cudnn7.1 + yolov3(spp) = 15FPS
+Ubuntu  + 1080   + CUDA8.0 + cudnn7.1 + yolov3      = 30FPS
+Ubuntu  + 1080   + CUDA9.0 + cudnn7.1 + yolov3      = 30FPS
 ```
 
 Software requirement
@@ -86,7 +144,7 @@ Software requirement
 -	CMake
 -	CUDA 8.0 or 9.0(9.1 is not working)
 -	OpenCV(for testing)
--	Visual Studio
+-	Visual Studio 2015
 
 Hardware requirement
 --------------------

Yolov3_SpringEdition_Test/.vs/Yolov3_SpringEdition_Test/v14/.suo

@@ -3,8 +3,8 @@
 *  YOLOv3_SE.h
 *  YOLOv3_SpringEdition
 *
-*  Created by kimbom on 2018. 03. 31...
-*  Copyright 2018 Sogang Univ. All rights reserved.
+*  Created by kimbomm on 2018. 03. 31...
+*  Copyright 2018 kimbomm. All rights reserved.
 *
 */
 #if !defined(YOLO_7E0_05_17_YOLOV3_H_INCLUDED)
@@ -69,11 +69,15 @@ class YOLOv3 {
 	using YoloDetectFromFileType = int(*)(char* img_path, int* _net, float threshold, float* result, int result_sz);
 	using YoloDetectFromImageType = int(*)(float* data, int w, int h, int c, int* _net, float threshold, float* result, int result_sz);
 	using YoloReleaseType = void(*)(int* net);
+	using YoloClassifyFromFileType = int(*)(char* img_path, int* _net, float* result);
+	using YoloClassifyFromImageType = int(*)(float* data, int w, int h, int c, int* _net, float* result);
 private:
 	YoloLoadType YoloLoad = nullptr;
 	YoloTrainType YoloTrain = nullptr;
 	YoloDetectFromFileType YoloDetectFromFile = nullptr;
 	YoloDetectFromImageType YoloDetectFromImage = nullptr;
+	YoloClassifyFromFileType YoloClassifyFromFile = nullptr;
+	YoloClassifyFromImageType YoloClassifyFromImage = nullptr;
 	YoloReleaseType YoloRelease=nullptr;
 protected:
 	int* m_network = nullptr;
@@ -151,25 +155,6 @@ class YOLOv3 {
 		return boxes;
 	}
 	std::vector<BoxSE> Detect(IplImage* img, float threshold) {
-		auto  ipl_to_image=[](IplImage* src)->image {
-			image out;
-			out.data = 0;
-			out.h = src->height;
-			out.w = src->width;
-			out.c = src->nChannels;
-			out.data = (float*)calloc(out.h*out.w*out.c, sizeof(float));
-			unsigned char *data = (unsigned char *)src->imageData;
-			int step = src->widthStep;
-			int i, j, k;
-			for (i = 0; i < out.h; ++i) {
-				for (k = 0; k < out.c; ++k) {
-					for (j = 0; j < out.w; ++j) {
-						out.data[k*out.w*out.h + i*out.w + j] = data[i*step + j*out.c + k] / 255.;
-					}
-				}
-			}
-			return out;
-		};
 		float result[6000] = { 0 };
 		image im = ipl_to_image(img);
 		int n = YoloDetectFromImage(im.data,im.w,im.h,im.c,this->m_network, threshold, result, 6000);
@@ -191,6 +176,22 @@ class YOLOv3 {
 		std::sort(boxes.begin(), boxes.end(), [](BoxSE a, BoxSE b)->bool { return a.m_score > b.m_score; });
 		return boxes;
 	}
+
+	int Classify(IplImage* img) {
+		image im = ipl_to_image(img);
+		int r=this->YoloClassifyFromImage(im.data, im.w, im.h, im.c, this->m_network, nullptr);
+		free(im.data);
+		return r;
+	}
+	int Classify(cv::Mat img) {
+		IplImage* iplimg = new IplImage(img);
+		int r = Classify(iplimg);
+		delete iplimg;
+		return r;
+	}
+	int Classify(std::string file) {
+		return YoloClassifyFromFile(const_cast<char*>(file.c_str()), this->m_network, nullptr);
+	}
 	std::vector<BoxSE> GroundTruth(std::string image_file) {
 		std::vector<BoxSE> ret;
 		std::string txt_file = image_file.substr(0, image_file.find_last_of(".")) + ".txt";
@@ -227,6 +228,8 @@ class YOLOv3 {
 		YoloTrain = (YoloTrainType)GetProcAddress(m_hmod, "YoloTrain");
 		YoloDetectFromFile = (YoloDetectFromFileType)GetProcAddress(m_hmod, "YoloDetectFromFile");
 		YoloDetectFromImage = (YoloDetectFromImageType)GetProcAddress(m_hmod, "YoloDetectFromImage");
+		YoloClassifyFromFile = (YoloClassifyFromFileType)GetProcAddress(m_hmod, "YoloClassifyFromFile");
+		YoloClassifyFromImage = (YoloClassifyFromImageType)GetProcAddress(m_hmod, "YoloClassifyFromImage");
 		YoloRelease = (YoloReleaseType)GetProcAddress(m_hmod, "YoloRelease");
 #else
 		m_hmod = dlopen("libYOLOv3SE.so",RTLD_LAZY);
@@ -245,6 +248,26 @@ class YOLOv3 {
 	~YOLOv3() {
 		this->Release();
 	}
+private:
+	image ipl_to_image(IplImage* src){
+		image out;
+		out.data = 0;
+		out.h = src->height;
+		out.w = src->width;
+		out.c = src->nChannels;
+		out.data = (float*)calloc(out.h*out.w*out.c, sizeof(float));
+		unsigned char *data = (unsigned char *)src->imageData;
+		int step = src->widthStep;
+		int i, j, k;
+		for (i = 0; i < out.h; ++i) {
+			for (k = 0; k < out.c; ++k) {
+				for (j = 0; j < out.w; ++j) {
+					out.data[k*out.w*out.h + i*out.w + j] = data[i*step + j*out.c + k] / 255.F;
+				}
+			}
+		}
+		return out;
+	}
 };
 
 #endif

Yolov3_SpringEdition_Test/YOLOv3SE.h

@@ -1,5 +1,7 @@
 @echo off
 title download_cudnn64_7.dll
+set cudir=%CD%
+cd %TEMP%
 nvcc -V | findstr /C:"release 8.0" > cuda_ver.txt
 set /p "ver80="<"cuda_ver.txt"
 if not "%ver80%" equ "" (
@@ -10,9 +12,11 @@ set /p "ver90="<"cuda_ver.txt"
 if not "%ver90%" equ "" (
 	powershell "[Net.ServicePointManager]::SecurityProtocol = [Net.SecurityProtocolType]::Tls12; (New-Object System.Net.WebClient).DownloadFile('https://github.com/springkim/YOLOv3_SpringEdition/releases/download/cudnn7.1cuda9.0/cudnn64_7.dll','cudnn64_7.dll')"
 )
+DEL cuda_ver.txt
+move cudnn64_7.dll "%cudir%"
 ::nvcc -V | findstr /C:"release 9.1" > cuda_ver.txt
 ::set /p "ver91="<"cuda_ver.txt"
 ::if not "%ver91%" equ "" (
 ::	powershell "[Net.ServicePointManager]::SecurityProtocol = [Net.SecurityProtocolType]::Tls12; (New-Object System.Net.WebClient).DownloadFile('https://www.dropbox.com/s/l0zsm9g78ciexzh/cudnn64_7.dll?dl=1','cudnn64_7.dll')"
 ::)
-DEL cuda_ver.txt
+pause

Yolov3_SpringEdition_Test/Yolov3_SpringEdition_CSTest/obj/Release/DesignTimeResolveAssemblyReferencesInput.cache

@@ -0,0 +1,4 @@
+@echo off
+title download_resnext50_256_10000.weights
+powershell "[Net.ServicePointManager]::SecurityProtocol = [Net.SecurityProtocolType]::Tls12; (New-Object System.Net.WebClient).DownloadFile('https://github.com/springkim/YOLOv3_SpringEdition/releases/download/image/resnext50_256_10000.weights','resnext50_256_10000.weights')"
+pause

Yolov3_SpringEdition_Test/download_cudnn64_7.dll.bat

@@ -0,0 +1 @@
+curl -L https://github.com/springkim/YOLOv3_SpringEdition/releases/download/image/resnext50_256_10000.weights -o resnext50_256_10000.weights