add yolov7 support

Makefile

@@ -6,7 +6,7 @@ lean_tensor_rt := /datav/lean/TensorRT-8.2.3.0-cuda11.4-cudnn8.2
 lean_cudnn     := /datav/lean/cudnn8.2.4.15-cuda11.4
 lean_opencv    := /datav/lean/opencv-4.2.0
 lean_cuda      := /datav/lean/cuda-11.2
-use_python     := false
+use_python     := true
 python_root    := /datav/software/anaconda3
 
 # python_root指向的lib目录下有个libpython3.9.so，因此这里写python3.9
@@ -201,6 +201,9 @@ pycenternet : pytrtc
 pyyolov5 : pytrtc
 	@cd example-python && python test_yolov5.py
 
+pyyolov7 : pytrtc
+	@cd example-python && python test_yolov7.py
+
 pyyolox : pytrtc
 	@cd example-python && python test_yolox.py
 

README.md

@@ -390,6 +390,73 @@ make yolo -j32
 </details>
 
 
+<details>
+<summary>YoloV7 Support</summary>
+1. Download yolov7 and pth
+
+```bash
+# from cdn
+# or wget https://github.com/WongKinYiu/yolov7/releases/download/v0.1/yolov7.pt
+
+wget https://cdn.githubjs.cf/WongKinYiu/yolov7/releases/download/v0.1/yolov7.pt
+git clone git@github.com:WongKinYiu/yolov7.git
+```
+
+2. Modify the code for dynamic batchsize
+```python
+# line 45 forward function in yolov7/models/yolo.py 
+# bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
+# x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
+# modified into:
+
+bs, _, ny, nx = map(int, x[i].shape)  # x(bs,255,20,20) to x(bs,3,20,20,85)
+bs = -1
+x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
+
+# line 52 in yolov7/models/yolo.py
+# y = x[i].sigmoid()
+# y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
+# y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
+# z.append(y.view(bs, -1, self.no))
+# modified into：
+y = x[i].sigmoid()
+xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
+wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i].view(1, -1, 1, 1, 2)  # wh
+classif = y[..., 4:]
+y = torch.cat([xy, wh, classif], dim=-1)
+z.append(y.view(bs, self.na * ny * nx, self.no))
+
+# line 57 in yolov7/models/yolo.py
+# return x if self.training else (torch.cat(z, 1), x)
+# modified into:
+return x if self.training else torch.cat(z, 1)
+
+
+# line 52 in yolov7/models/export.py
+# output_names=['classes', 'boxes'] if y is None else ['output'],
+# dynamic_axes={'images': {0: 'batch', 2: 'height', 3: 'width'},  # size(1,3,640,640)
+#               'output': {0: 'batch', 2: 'y', 3: 'x'}} if opt.dynamic else None)
+# modified into:
+output_names=['classes', 'boxes'] if y is None else ['output'],
+dynamic_axes={'images': {0: 'batch'},  # size(1,3,640,640)
+              'output': {0: 'batch'}} if opt.dynamic else None)
+
+```
+3. Export to onnx model
+```bash
+cd yolov7
+python models/export.py --dynamic --grid --weight=yolov7.pt
+```
+4. Copy the model and execute it
+```bash
+cp yolov7/yolov7.onnx tensorRT_cpp/workspace/
+cd tensorRT_cpp
+make yolo -j32
+```
+
+</details>
+
+
 <details>
 <summary>YoloX Support</summary>
 

example-python/test_yolov7.py

@@ -0,0 +1,62 @@
+import os
+import cv2
+import numpy as np
+import pytrt as tp
+
+# change current workspace
+os.chdir("../workspace/")
+
+# 如果执行出错，请删掉 ~/.pytrt 的缓存模型
+# rm -rf ~/.pytrt，重新下载
+device_id = 7
+engine_file = "yolov7.fp32.trtmodel"
+if not os.path.exists(engine_file):
+    tp.compile_onnx_to_file(5, tp.onnx_hub("yolov7"), engine_file, device_id=device_id)
+
+yolo   = tp.Yolo(engine_file, type=tp.YoloType.V5, device_id=device_id)
+image  = cv2.imread("inference/car.jpg")
+bboxes = yolo.commit(image).get()
+print(f"{len(bboxes)} objects")
+print(bboxes)
+
+for box in bboxes:
+    left, top, right, bottom = map(int, [box.left, box.top, box.right, box.bottom])
+    cv2.rectangle(image, (left, top), (right, bottom), tp.random_color(box.class_label), 5)
+
+os.makedirs("single_inference", exist_ok=True)
+saveto = "single_inference/yolov5.car.jpg"
+print(f"Save to {saveto}")
+
+cv2.imwrite(saveto, image)
+
+
+try:
+    import torch
+
+    image  = cv2.imread("inference/car.jpg")
+    device = 0
+
+    gpu_image = torch.from_numpy(image).to(device)
+    bboxes = yolo.commit_gpu(
+        pimage = gpu_image.data_ptr(),
+        width = image.shape[1],
+        height = image.shape[0],
+        device_id = device,
+        imtype = tp.ImageType.GPUBGR,
+        stream = torch.cuda.current_stream().cuda_stream
+    ).get()
+    print(f"{len(bboxes)} objects")
+    print(bboxes)
+
+    for box in bboxes:
+        left, top, right, bottom = map(int, [box.left, box.top, box.right, box.bottom])
+        cv2.rectangle(image, (left, top), (right, bottom), tp.random_color(box.class_label), 5)
+
+    os.makedirs("single_inference", exist_ok=True)
+    saveto = "single_inference/yolov7-gpuptr.car.jpg"
+    print(f"Save to {saveto}")
+
+    cv2.imwrite(saveto, image)
+
+except Exception as e:
+    print("GPUPtr test failed.", e)

example-restful_server/src/infer/simple_yolo.cu

@@ -1813,7 +1813,7 @@ namespace SimpleYolo{
 
         virtual bool startup(const string& file, Type type, int gpuid, float confidence_threshold, float nms_threshold){
 
-            if(type == Type::V5){
+            if(type == Type::V5 || type == Type::V3 || type == Type::V7){
                 normalize_ = Norm::alpha_beta(1 / 255.0f, 0.0f, ChannelType::SwapRB);
             }else if(type == Type::X){
                 //float mean[] = {0.485, 0.456, 0.406};
@@ -1993,7 +1993,7 @@ namespace SimpleYolo{
     void image_to_tensor(const cv::Mat& image, shared_ptr<Tensor>& tensor, Type type, int ibatch){
 
         Norm normalize;
-        if(type == Type::V5){
+        if(type == Type::V5 || type == Type::V3 || type == Type::V7){
             normalize = Norm::alpha_beta(1 / 255.0f, 0.0f, ChannelType::SwapRB);
         }else if(type == Type::X){
             //float mean[] = {0.485, 0.456, 0.406};

example-simple_yolo/Makefile

@@ -17,10 +17,14 @@ cu_mk   := $(cu_objs:.cuo=.cumk)
 #    tensorRT下载GA版本（通用版、稳定版），EA（尝鲜版本）不要
 # 3. cuda10.2，也可以是11.x看搭配（请自行下载安装）
 
-lean_tensor_rt := /data/sxai/lean/TensorRT-8.0.1.6-cuda10.2-cudnn8.2
-lean_cudnn     := /data/sxai/lean/cudnn8.2.2.26
-lean_opencv    := /data/sxai/lean/opencv4.2.0
-lean_cuda      := /data/sxai/lean/cuda-10.2
+# lean_tensor_rt := /data/sxai/lean/TensorRT-8.0.1.6-cuda10.2-cudnn8.2
+# lean_cudnn     := /data/sxai/lean/cudnn8.2.2.26
+# lean_opencv    := /data/sxai/lean/opencv4.2.0
+# lean_cuda      := /data/sxai/lean/cuda-10.2
+lean_tensor_rt := /datav/lean/TensorRT-8.2.3.0-cuda11.4-cudnn8.2
+lean_cudnn     := /datav/lean/cudnn8.2.4.15-cuda11.4
+lean_opencv    := /datav/lean/opencv-4.2.0
+lean_cuda      := /datav/lean/cuda-11.2
 
 # lean_tensor_rt := /data/sxai/lean/TensorRT-7.0.0.11
 # lean_cudnn     := /data/sxai/lean/cudnn7.6.5.32-cuda10.2

example-simple_yolo/src/main.cpp

@@ -96,7 +96,7 @@ static double timestamp_now_float() {
 
 bool requires_model(const string& name) {
 
-    auto onnx_file = cv::format("%s_dynamic.onnx", name.c_str());
+    auto onnx_file = cv::format("%s.onnx", name.c_str());
     if (!exists(onnx_file)) {
         printf("Auto download %s\n", onnx_file.c_str());
         system(cv::format("wget http://zifuture.com:1556/fs/25.shared/%s", onnx_file.c_str()).c_str());
@@ -189,8 +189,8 @@ static void test(SimpleYolo::Type type, SimpleYolo::Mode mode, const string& mod
     if(!requires_model(name))
         return;
 
-    string onnx_file = cv::format("%s_dynamic.onnx", name);
-    string model_file = cv::format("%s_dynamic.%s.trtmodel", name, mode_name);
+    string onnx_file = cv::format("%s.onnx", name);
+    string model_file = cv::format("%s.%s.trtmodel", name, mode_name);
     int test_batch_size = 16;
 
     if(!exists(model_file)){
@@ -211,11 +211,11 @@ void direct_test(){
     printf("TRTVersion: %s\n", SimpleYolo::trt_version());
 
     int device_id = 0;
-    string model = "yolox_s";
+    string model = "yolox_s_dynamic";
     auto type = SimpleYolo::Type::X;
     auto mode = SimpleYolo::Mode::FP32;
-    string onnx_file = cv::format("%s_dynamic.onnx", model.c_str());
-    string model_file = cv::format("%s_dynamic.%s.trtmodel", model.c_str(), SimpleYolo::mode_string(mode));
+    string onnx_file = cv::format("%s.onnx", model.c_str());
+    string model_file = cv::format("%s.%s.trtmodel", model.c_str(), SimpleYolo::mode_string(mode));
     SimpleYolo::set_device(device_id);
 
     if(!requires_model(model)){
@@ -257,27 +257,28 @@ void direct_test(){
 int main(){
 
     //direct_test();
-    test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP32, "yolov5s");
-    //test(SimpleYolo::Type::X, SimpleYolo::Mode::INT8, "yolox_s");
+    test(SimpleYolo::Type::V7, SimpleYolo::Mode::FP32, "yolov7");
+    //test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP32, "yolov5s_dynamic");
+    //test(SimpleYolo::Type::X, SimpleYolo::Mode::INT8, "yolox_s_dynamic");
 
-    //test(SimpleYolo::Type::X, SimpleYolo::Mode::FP32, "yolox_s");
-    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP32, "yolox_x");
-    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP32, "yolox_l");
-    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP32, "yolox_m");
-    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP32, "yolox_s");
-    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP16, "yolox_x");
-    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP16, "yolox_l");
-    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP16, "yolox_m");
-    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP16, "yolox_s");
-
-    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP32, "yolov5x");
-    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP32, "yolov5l");
-    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP32, "yolov5m");
-    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP32, "yolov5s");
-
-    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP16, "yolov5x");
-    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP16, "yolov5l");
-    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP16, "yolov5m");
-    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP16, "yolov5s");
+    //test(SimpleYolo::Type::X, SimpleYolo::Mode::FP32, "yolox_s_dynamic");
+    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP32, "yolox_x_dynamic");
+    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP32, "yolox_l_dynamic");
+    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP32, "yolox_m_dynamic");
+    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP32, "yolox_s_dynamic");
+    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP16, "yolox_x_dynamic");
+    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP16, "yolox_l_dynamic");
+    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP16, "yolox_m_dynamic");
+    // test(SimpleYolo::Type::X, SimpleYolo::Mode::FP16, "yolox_s_dynamic");
+
+    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP32, "yolov5x_dynamic");
+    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP32, "yolov5l_dynamic");
+    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP32, "yolov5m_dynamic");
+    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP32, "yolov5s_dynamic");
+
+    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP16, "yolov5x_dynamic");
+    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP16, "yolov5l_dynamic");
+    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP16, "yolov5m_dynamic");
+    // test(SimpleYolo::Type::V5, SimpleYolo::Mode::FP16, "yolov5s_dynamic");
     return 0;
 }

example-simple_yolo/src/simple_yolo.cu

@@ -1771,6 +1771,8 @@ namespace SimpleYolo{
     const char* type_name(Type type){
         switch(type){
         case Type::V5: return "YoloV5";
+        case Type::V3: return "YoloV3";
+        case Type::V7: return "YoloV7";
         case Type::X: return "YoloX";
         default: return "Unknow";
         }
@@ -1814,7 +1816,7 @@ namespace SimpleYolo{
 
         virtual bool startup(const string& file, Type type, int gpuid, float confidence_threshold, float nms_threshold){
 
-            if(type == Type::V5){
+            if(type == Type::V5 || type == Type::V3 || type == Type::V7){
                 normalize_ = Norm::alpha_beta(1 / 255.0f, 0.0f, ChannelType::SwapRB);
             }else if(type == Type::X){
                 //float mean[] = {0.485, 0.456, 0.406};
@@ -1994,7 +1996,7 @@ namespace SimpleYolo{
     void image_to_tensor(const cv::Mat& image, shared_ptr<Tensor>& tensor, Type type, int ibatch){
 
         Norm normalize;
-        if(type == Type::V5){
+        if(type == Type::V5 || type == Type::V3 || type == Type::V7){
             normalize = Norm::alpha_beta(1 / 255.0f, 0.0f, ChannelType::SwapRB);
         }else if(type == Type::X){
             //float mean[] = {0.485, 0.456, 0.406};

example-simple_yolo/src/simple_yolo.hpp

@@ -17,7 +17,9 @@ namespace SimpleYolo{
 
     enum class Type : int{
         V5 = 0,
-        X  = 1
+        X  = 1,
+        V3 = 2,
+        V7 = 3
     };
 
     enum class Mode : int {

src/application/app_high_performance/yolo_high_perf.cpp

@@ -104,7 +104,7 @@ namespace YoloHighPerf{
     public:
         virtual bool startup(const string& file, Type type, int gpuid, float confidence_threshold, float nms_threshold){
 
-            if(type == Type::V5){
+            if(type == Type::V5 || type == Type::V3 || type == Type::V7){
                 normalize_ = CUDAKernel::Norm::alpha_beta(1 / 255.0f, 0.0f, CUDAKernel::ChannelType::Invert);
             }else if(type == Type::X){
                 //float mean[] = {0.485, 0.456, 0.406};

src/application/app_high_performance/yolo_high_perf.hpp

@@ -19,7 +19,9 @@ namespace YoloHighPerf{
 
     enum class Type : int{
         V5 = 0,
-        X  = 1
+        X  = 1,
+        V3 = 2,
+        V7 = 3
     };    
 
     struct Box{

src/application/app_python/interface.cpp

@@ -568,6 +568,7 @@ PYBIND11_MODULE(libpytrtc, m) {
 	py::enum_<YoloGPUPtr::Type>(m, "YoloType")
 		.value("V5", YoloGPUPtr::Type::V5)
 		.value("V3", YoloGPUPtr::Type::V3)
+		.value("V7", YoloGPUPtr::Type::V7)
 		.value("X", YoloGPUPtr::Type::X);
 
 	py::enum_<YoloGPUPtr::NMSMethod>(m, "NMSMethod")

src/application/app_yolo.cpp

@@ -174,7 +174,8 @@ void multi_gpu_test(){
 
 int app_yolo(){
 
-    test(Yolo::Type::V5, TRT::Mode::FP32, "yolov5s");
+    test(Yolo::Type::V7, TRT::Mode::FP32, "yolov7");
+    //test(Yolo::Type::V5, TRT::Mode::FP32, "yolov5s");
     //test(Yolo::Type::V3, TRT::Mode::FP32, "yolov3");
 
     // multi_gpu_test();

src/application/app_yolo/yolo.cpp

@@ -17,6 +17,8 @@ namespace Yolo{
     const char* type_name(Type type){
         switch(type){
         case Type::V5: return "YoloV5";
+        case Type::V3: return "YoloV3";
+        case Type::V7: return "YoloV7";
         case Type::X: return "YoloX";
         default: return "Unknow";
         }
@@ -165,7 +167,7 @@ namespace Yolo{
             NMSMethod nms_method, int max_objects,
             bool use_multi_preprocess_stream
         ){
-            if(type == Type::V5){
+            if(type == Type::V5 || type == Type::V3 || type == Type::V7){
                 normalize_ = CUDAKernel::Norm::alpha_beta(1 / 255.0f, 0.0f, CUDAKernel::ChannelType::Invert);
             }else if(type == Type::X){
                 //float mean[] = {0.485, 0.456, 0.406};
@@ -400,7 +402,7 @@ namespace Yolo{
     void image_to_tensor(const cv::Mat& image, shared_ptr<TRT::Tensor>& tensor, Type type, int ibatch){
 
         CUDAKernel::Norm normalize;
-        if(type == Type::V5){
+        if(type == Type::V5 || type == Type::V3 || type == Type::V7){
             normalize = CUDAKernel::Norm::alpha_beta(1 / 255.0f, 0.0f, CUDAKernel::ChannelType::Invert);
         }else if(type == Type::X){
             //float mean[] = {0.485, 0.456, 0.406};

src/application/app_yolo/yolo.hpp

@@ -21,7 +21,8 @@ namespace Yolo{
     enum class Type : int{
         V5 = 0,
         X  = 1,
-        V3 = V5
+        V3 = 2,
+        V7 = 3
     };
 
     enum class NMSMethod : int{