add readme for neural rendering

include/rt/rt_trainer.h

@@ -43,7 +43,7 @@ namespace rt_sample
 	// generate random datas randomly
 	RayParam SampleRandom(size_t num);
 	// generate random datas randomly within a sphere
-	RayParam SampleSphere(size_t num);
+	RayParam SampleSphereFixhit(size_t num);
 	// generate uniform datas within on a sphere
 	RayParam UniformSphere(size_t num);
 	// generate datas with random points and fix vec
@@ -52,8 +52,8 @@ namespace rt_sample
 	RayParam SampleSphereFixVec(size_t num, std::vector<fastf_t> vec);
 	// generate datas with random points and fix vec on sphere surface, rays are forced to hit sphere
 	RayParam SampleFixVecHit(size_t num, std::vector < fastf_t> vec);
-	// generate datas with random points fix range and fix vec on sphere surface, rays are forced to hit sphere
-	RayParam RangeFixVecHit(size_t num, fastf_t max, fastf_t min, std::vector < fastf_t> vec);
+	// generate datas with random points fix range, rays are forced to hit sphere
+	RayParam RangeHit(size_t num, fastf_t max, fastf_t min);
 }
 
 namespace rt_neu

include/rt/torch_runner.h

@@ -1,6 +1,7 @@
 #ifndef RT_TORCH_RUNNER_H
 #define RT_TORCH_RUNNER_H
 void set_model_path(const char* model_path);
+void set_model_type(int num);
 typedef enum {
     normal = 1,
     neu_coordinate = 2,

src/rt/CMakeLists.txt

@@ -52,7 +52,7 @@ if(NOT TORCHVISION_FOUND)
     execute_process(COMMAND ${Python_EXECUTABLE} -m pip install torchvision)
 endif()
 
-# 设置 Torch 的路径
+# set libTorch path
 set(Torch_DIR "C:/lib/VS/libtorch/share/cmake/Torch")
 set(Torch_INCLUDE_DIRS "C:/lib/VS/libtorch/include")
 set(Torch_INCLUDE_LIBS "C:/lib/VS/libtorch/lib")

src/rt/README.md

@@ -0,0 +1,16 @@
+## Dependence
+[libtorch](https://pytorch.org/cppdocs/installing.html)
+used for loading model
+do remember to change the absolute path in:src/rt/CMakeLists.txt Line 56-58
+[nlohmann](https://github.com/nlohmann/json)
+used for write json file
+## Project structure
+Add two projects:
+### rt_trainneural
+used for sampling and rendering
+new files:train_neural.cpp(main) rt/neu_util.h rt/rt_trainer.h
+neural.c rt_trainer.cpp
+
+see train_neural.cpp for usage
+### torch_runner
+used for run torch model

src/rt/rt_trainer.cpp

@@ -28,7 +28,6 @@ vect_t ambient_color = { 1, 1, 1 };	/* Ambient white light */
 int ibackground[3] = { 0 };		/* integer 0..255 version */
 int inonbackground[3] = { 0 };		/* integer non-background */
 fastf_t gamma_corr = 0.0;		/* gamma correction if !0 */
-const std::string model_path = "C:\\works\\soc\\rainy\\test\\model.pt";
 
 
 namespace convert	
@@ -46,29 +45,30 @@ namespace convert
 	{
 		RayParamSph res;
 		res.reserve(datas.size());
-		fastf_t elevation(0.0);
-		fastf_t azimuth(0.0);
-		fastf_t pelevation(0.0);
-		fastf_t pazimuth(0.0);
+		fastf_t theta;
+		fastf_t phi;
+		fastf_t ptheta;
+		fastf_t pphi;
 		for (auto& data : datas)
 		{
 			fastf_t x = data.first[0] - origin[0];
 			fastf_t y = data.first[1] - origin[1];
 			fastf_t z = data.first[2] - origin[2];
 			fastf_t z_abs = abs(z);
 			fastf_t r_abs = abs(r);
-			elevation = acos(z / r); //0-pi
-			azimuth = SIGN(y)*acos(x / (r * sin(elevation))); //-pi-pi
+
+			theta = acos(z / r); //0-pi
+			phi = SIGN(y)*acos(x / (r * sin(theta))); //-pi-pi
 			if (pow(data.second[0], 2) + pow(data.second[1], 2) + pow(data.second[2], 2) != 1)
 			{
 				fastf_t square_sum = pow(pow(data.second[0], 2) + pow(data.second[1], 2) + pow(data.second[2], 2), 0.5);
 				data.second[0] /= square_sum;
 				data.second[1] /= square_sum;
 				data.second[2] /= square_sum;
 			}
-			pelevation = acos(data.second[2]);
-			pazimuth = SIGN(data.second[1]) * acos(data.second[0] / (1 * sin(pelevation)));
-			res.push_back(std::make_pair(std::make_pair(elevation, azimuth), std::make_pair(pelevation, pazimuth)));
+			ptheta = acos(data.second[2]);
+			pphi = SIGN(data.second[1]) * acos(data.second[0] / (1 * sin(ptheta)));
+			res.push_back(std::make_pair(std::make_pair(theta,phi), std::make_pair(ptheta, pphi)));
 		}
 		return res;
 	}
@@ -197,7 +197,7 @@ namespace rt_sample
 		}
 		return res;
 	}
-	RayParam SampleSphere(size_t num)
+	RayParam SampleSphereFixhit(size_t num)
 	{
 		RayParam res;
 		point_t center{ 0 };
@@ -206,21 +206,28 @@ namespace rt_sample
 		fastf_t square_sum(0);
 		std::vector<fastf_t> p;
 		std::vector<fastf_t> d;
+		vect_t center_v;
 		for (int i = 0; i < num; ++i) {
 			p.clear();
 			d.clear();
-			fastf_t theta = RandomNum(0, 2 * M_PI); 
-			fastf_t phi = acos(2 * RandomNum(0, 1) - 1); 
-			p.push_back(center[0] + radius * sin(phi) * cos(theta));
-			p.push_back(center[1] + radius * sin(phi) * sin(theta));
-			p.push_back(center[2] + radius * cos(phi));
-			d.push_back(center[0] - p[0]);
-			d.push_back(center[1] - p[1]);
-			d.push_back(center[2] - p[2]);
+			fastf_t theta = RandomNum(0, M_PI);
+			fastf_t phi = RandomNum(-M_PI, M_PI);
+			p.push_back(center[0] + radius * sin(theta) * cos(phi));
+			p.push_back(center[1] + radius * sin(theta) * sin(phi));
+			p.push_back(center[2] + radius * cos(theta));
+			d.push_back(RandomNum(0, 1));
+			d.push_back(RandomNum(0, 1));
+			d.push_back(RandomNum(0, 1));
 			square_sum = pow(pow(d[0], 2) + pow(d[1], 2) + pow(d[2], 2), 0.5);
 			d[0] /= square_sum;
 			d[1] /= square_sum;
 			d[2] /= square_sum;
+			VSUB2(center_v, center, p);
+			if (VDOT(center_v, d) <= 0)
+			{
+				i--;
+				continue;
+			}
 			res.push_back(std::make_pair(p, d));
 		}
 		return res;
@@ -315,22 +322,30 @@ namespace rt_sample
 		}
 		return res;
 	}
-	RayParam RangeFixVecHit(size_t num, fastf_t max, fastf_t min, std::vector < fastf_t> vec)
+	RayParam RangeHit(size_t num, fastf_t max, fastf_t min)
 	{
 		RayParam res;
 		point_t center{ 0 }, random_point{ 0 };
 		fastf_t radius = APP.a_rt_i->rti_radius;
 		VADD2SCALE(center, APP.a_rt_i->rti_pmin, APP.a_rt_i->rti_pmax, 0.5);
 		std::vector<fastf_t> p;
-		std::vector<fastf_t> d = vec;
+		std::vector<fastf_t> d;
 		vect_t center_v;
+		fastf_t square_sum;
+		fastf_t theta;
+		fastf_t phi;
 		for (int i = 0; i < num; ++i) {
-			p.clear();
-			fastf_t theta = RandomNum(0, 2 * M_PI);
-			fastf_t phi = acos(2 * RandomNum(0, 1) - 1);
-			VSET(p, center[0] + radius * sin(phi) * cos(theta), center[1] + radius * sin(phi) * sin(theta), center[2] + radius * cos(phi));
+			p = std::vector<fastf_t>(3, 0);
+			d = std::vector<fastf_t>(3, 0);
+			theta = RandomNum(0,M_PI);
+			theta = theta == 0 ? 1e-5 : theta;
+			phi = RandomNum(-M_PI,M_PI);
+			fastf_t ptheta = RandomNum(0.5*M_PI, 0.85*M_PI);
+			fastf_t pphi = RandomNum(-0.85*M_PI, -0.5*M_PI);
+			VSET(p, center[0] + radius * sin(theta) * cos(phi), center[1] + radius * sin(theta) * sin(phi), center[2] + radius * cos(theta));
 			VSUB2(center_v, center, p);
-			if (VDOT(center_v, vec) <= 0)
+			VSET(d, 1 * sin(ptheta) * cos(pphi), 1 * sin(ptheta) * sin(pphi), 1 * cos(ptheta));
+			if (VDOT(center_v, d) <= 0)
 			{
 				i--;
 				continue;

src/rt/torch_runner.cpp

@@ -3,6 +3,7 @@
 #include "torch_runner.h"
 #include <iostream>
 std::string global_model_path;
+int model_type;
 typedef enum {
     normal = 1,
     neu_coordinate = 2,
@@ -12,6 +13,10 @@ void set_model_path(const char* model_path)
 {
     global_model_path = std::string(model_path);
 }
+void set_model_type(int num)
+{
+    model_type = num;
+}
 static torch::jit::script::Module& get_model() {
     static torch::jit::script::Module module;
     static bool is_loaded = false;
@@ -49,16 +54,22 @@ void run_torch(double* para, int* res)
     else if (int(para[0]) == neu_sphere)
     {
         para++;
-        for (int i = 0; i < 5; i++)
+        for (int i = 0; i < 6; i++)
         {
             input_vec.push_back(*para);
             para++;
         }
-        inputs.push_back(torch::tensor({ {input_vec[0], input_vec[1]} }, torch::kFloat).to(at::kCUDA));
+        if(model_type==2)
+            inputs.push_back(torch::tensor({ {input_vec[0], input_vec[1]} }, torch::kFloat).to(at::kCUDA));
+        else if (model_type == 4)
+        {
+            /*inputs.push_back(torch::tensor({ {input_vec[0], input_vec[1]} }, torch::kFloat).to(at::kCUDA));
+            inputs.push_back(torch::tensor({ {input_vec[2], input_vec[3]} }, torch::kFloat).to(at::kCUDA));*/
+            inputs.push_back(torch::tensor({ {input_vec[0], input_vec[1],input_vec[2], input_vec[3]} }, torch::kFloat).to(at::kCUDA));
+        }
     }
     // inputs.push_back(input_tensor);
     at::Tensor output = module.forward(inputs).toTensor().to(torch::kCPU);
-    auto tmp = output.data_ptr<float>();
     std::vector<double> output_vector(output.data_ptr<float>(), output.data_ptr<float>() + output.numel());
     res[0] = int(output_vector[0]);
     if (res[0] > 255)

src/rt/train_neural.cpp

@@ -29,35 +29,51 @@
 render_type rt_render_type;
 int main(int argc, char* argv[])
 {
-	fastf_t a = 74.49;
-	fastf_t b = -a;
+	// database name
 	const char* db = "C:\\works\\soc\\rainy\\brlcad\\build\\share\\db\\moss.g";
+	// object name
 	const char* ob = "all.g";
 	struct rt_i* rtip = NULL;
+	// set size
 	set_size(64);
 	rt_tool::init_rt(db, ob, rtip);
-	// do_ae(10, 10);
-	//rt_perspective = 90;
+	// do_ae(30, 25);
+	// rt_perspective = 120;
 
+//rendering 
 #if 1
+	// choose rendering type:normal , neu_coordinate(Cartesian coordinate) neu_sphere
 	set_type(neu_sphere);
-	set_model_path("C:\\works\\soc\\rainy\\Rendernn\\models\\model_sph5.pt");
+	// choose model's input number: 2 for pos, 4 for both pos and dir
+	set_model_type(2);
+	// choose your model type
+	set_model_path("C:\\works\\soc\\rainy\\Rendernn\\models\\model_sph_grid6.pt");
+	// set_model_path("C:\\works\\soc\\rainy\\Rendernn\\models\\dir_model_sph_grid2.pt");
+	// set_model_path("C:\\works\\soc\\rainy\\Rendernn\\models\\test.pt");
+	// begin rendering
 	rt_neu::render();
 #endif
 
+//sampling
 # if 0
+	// there are many different sample methods:
 	// auto ray_list = rt_sample::RangeFixVec(100000, 400, -100, { -0.742403865,-0.519836783,-0.422618270 });
 	// auto ray_list = rt_sample::SampleRandom(1000);
 	// auto ray_list = rt_sample::SampleFixVecHit(100000, { -0.74240387650610373, -0.51983679072568467, -0.42261826174069961 });
 	// auto ray_list = rt_sample::RangeFixVecHit(1000, 100,-100,{ -0.74240387650610373, -0.51983679072568467, -0.42261826174069961 });
-	auto ray_list = rt_sample::SampleFixVecHit(1000000, { -0.74240387650610373, -0.51983679072568467, -0.42261826174069961 });
+	// auto ray_list = rt_sample::SampleFixVecHit(1000000, { -0.74240387650610373, -0.51983679072568467, -0.42261826174069961 });
+	auto ray_list = rt_sample::RangeHit(1000,M_PI,0.75*M_PI);
 	point_t center{ 0 };
 	get_center(center);
+	// shoot these points
 	auto ray_res = rt_tool::ShootSamples(ray_list);
+	// convert Cartesian coordinate to Spherical coordinate
 	auto ray_list_sph = convert::cert_to_sph(ray_list, center, get_r());
 	// test for write json
 	// util::write_json(ray_list, ray_res, "C:\\works\\soc\\rainy\\Rendernn\\datas\\e.json");
-	util::write_sph_json(ray_list_sph, ray_res, "C:\\works\\soc\\rainy\\Rendernn\\datas\\sph_6.json");
+	ray_list.clear();
+	// write to a local file
+	util::write_sph_json(ray_list_sph, ray_res, "C:\\works\\soc\\rainy\\Rendernn\\datas\\all_dir_sph_range_3.json");
 #endif
 	return 0;
 }