1. Resolution (YoloV5P5, YoloX) = (640x640), (YoloV5P6) = (1280x1280)
2. max batch size = 16
3. preprocessing + inference + postprocessing
4. cuda10.2, cudnn8.2.2.26, TensorRT-8.0.1.6
5. RTX2080Ti
6. num of testing: take the average on the results of 100 times but excluding the first time for warmup
7. Testing log: [workspace/perf.result.std.log (workspace/perf.result.std.log)
8. code for testing: src/application/app_yolo.cpp
9. images for testing: 6 images in workspace/inference
   • with resolution 810x1080，500x806，1024x684，550x676，1280x720，800x533 respetively
10. Testing method: load 6 images. Then do the inference on the 6 images, which will be repeated for 100 times. Note that each image should be preprocessed and postprocessed.

• Highlight: 0.5 ms faster without any loss in precision compared with the above. Specifically, we remove the Focus and some transpose nodes etc, and implement them in CUDA kenerl function. But the rest remains the same.
• Test log:workspace/perf.result.std.log
• Code for testing:src/application/app_yolo_fast.cpp
• Tips: you can do the modification while refering to the downloaded onnx. Any questions are welcomed through any kinds of contact.
• Conclusion: the main idea of this work is to optimize the pre-and-post processing. If you go for yolox, yolov5 small version, the optimization might help you.

• complied dependency (Out of the Box feature)
  • link for download [lean-tensorRT8.0.1.6-protobuf3.11.4-cudnn8.2.2.tar.gz] (http://zifuture.com:1556/fs/25.shared/lean-tensorRT8.0.1.6-protobuf3.11.4-cudnn8.2.2.tar.gz)

1. VSCode (highly recommended!)
2. Configure your path for cudnn, cuda, tensorRT8.0 and protobuf.
3. Configure the compute capability matched with your nvidia graphics card in Makefile/CMakeLists.txt
   • e.g. -gencode=arch=compute_75,code=sm_75. If you are using 3080Ti, that should be gencode=arch=compute_86,code=sm_86
   • reference for the table for GPU Compute Capability: https://developer.nvidia.com/cuda-gpus#compute
4. Configure your library path in .vscode/c_cpp_properties.json
5. CUDA version: CUDA10.2
6. CUDNN version: cudnn8.2.2.26. Note that dev(.h file) and runtime(.so file) should be downloaded.
7. tensorRT version：tensorRT-8.0.1.6-cuda10.2
8. protobuf version（for onnx parser）：protobufv3.11.4
   • if other version, refer to the ........
   • link for download: https://github.com/protocolbuffers/protobuf/tree/v3.11.4
   • download, compile and replace the path in Makefile/CMakeLists.txt with new path to protobuf3.11.4

• CMake:
  • mkdir build && cd build
  • cmake ..
  • make yolo -j8
• Makefile:
  • make yolo -j8

• compile and install
  • Makefile：
    • set use_python := true in Makefile
  • CMakeLists.txt:
    • set(HAS_PYTHON ON) in CMakeLists.txt
  • Type in make pyinstall -j8
  • Complied files are in python/trtpy/libtrtpyc.so

1. Please check the lean/README.md for the detailed dependency

2. In TensorRT.vcxproj, replace the <Import Project="$(VCTargetsPath)\BuildCustomizations\CUDA 10.0.props" /> with your own CUDA path

3. In TensorRT.vcxproj, replace the <Import Project="$(VCTargetsPath)\BuildCustomizations\CUDA 10.0.targets" /> with your own CUDA path

4. In TensorRT.vcxproj, replace the <CodeGeneration>compute_61,sm_61</CodeGeneration> with your compute capability.

   • refer to the table in https://developer.nvidia.com/cuda-gpus#compute

5. Configure your dependency or download it to the foler /lean. Configure VC++ dir (include dir and refence)

6. Configure your env, debug->environment

7. Compile and run the example, where 3 options are available.

1. Compile trtpyc.pyd. Choose python in visual studio to compile
2. Copy dll and execute 'python/copy_dll_to_trtpy.bat'
3. Execute the example in python dir by 'python test_yolov5.py'

• if installation is needed, switch to target env(e.g. your conda env) then 'python setup.py install', which has to be followed by step 1 and step 2.
• the compiled files are in python/trtpy/libtrtpyc.pyd

• in onnx/make_pb.sh, replace the path protoc=/data/sxai/lean/protobuf3.11.4/bin/protoc in protoc with the protoc of your own version

#cd the path in terminal to /onnxcd onnx #execuete the command to make pb files bash make_pb.sh

• CMake:
  • replace the set(PROTOBUF_DIR "/data/sxai/lean/protobuf3.11.4") in CMakeLists.txt with the same path of your protoc.

mkdir build &&cd build cmake .. make yolo -j64

• Makefile:
  • replace the path lean_protobuf := /data/sxai/lean/protobuf3.11.4 in Makefile with the same path of protoc

make yolo -j64

• The default is tensorRT8.x

1. Replace onnx_parser_for_7.x/onnx_parser to src/tensorRT/onnx_parser
   • bash onnx_parser/use_tensorrt_7.x.sh
2. Configure Makefile/CMakeLists.txt path to TensorRT7.x
3. Execute make yolo -j64

• The default is tensorRT8.x

1. Replace onnx_parser_for_8.x/onnx_parser to src/tensorRT/onnx_parser
   • bash onnx_parser/use_tensorrt_8.x.sh
2. Configure Makefile/CMakeLists.txt path to TensorRT8.x
3. Execute make yolo -j64

• if pytorch >= 1.7, and the model is 5.0+, the model is suppored by the framework
• if pytorch < 1.7 or yolov5(2.0, 3.0 or 4.0), minor modification should be done in opset.
• if you want to achieve the inference with lower pytorch, dynamic batchsize and other advanced setting, please check our blog (http://zifuture.com:8090)(now in Chinese) and scan the QRcode via Wechat to join us.

1. Download yolov5

git clone git@github.com:ultralytics/yolov5.git

2. Modify the code for dynamic batchsize

# line 55 forward function in yolov5/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=x[i].shape# x(bs,255,20,20) to x(bs,3,20,20,85)bs=-1ny=int(ny) nx=int(nx) x[i] =x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() # line 70 in yolov5/models/yolo.py# z.append(y.view(bs, -1, self.no))# modified into：z.append(y.view(bs, self.na*ny*nx, self.no)) # line 52 in yolov5/export.py# torch.onnx.export(dynamic_axes={'images':{0: 'batch', 2: 'height', 3: 'width'}, # shape(1,3,640,640)# 'output':{0: 'batch', 1: 'anchors'} # shape(1,25200,85) 修改为# modified into:torch.onnx.export(dynamic_axes={'images':{0: 'batch'}, # shape(1,3,640,640)'output':{0: 'batch'} # shape(1,25200,85) 

3. Export to onnx model

cd yolov5 python export.py --weights=yolov5s.pt --dynamic --include=onnx --opset=11

4. Copy the model and execute it

cp yolov5/yolov5s.onnx tensorRT_cpp/workspace/ cd tensorRT_cpp make yolo -j32

• download from: https://github.com/Megvii-BaseDetection/YOLOX
• If you don't want to export onnx by yourself, just make run in the repo of Megavii

1. Download YoloX

git clone git@github.com:Megvii-BaseDetection/YOLOX.git cd YOLOX

2. Modify the code The modification ensures a successful int8 compilation and inference, otherwise Missing scale and zero-point for tensor (Unnamed Layer* 686) will be raised.

# line 206 forward fuction in yolox/models/yolo_head.py. Replace the commented code with the uncommented code# self.hw = [x.shape[-2:] for x in outputs] self.hw= [list(map(int, x.shape[-2:])) forxinoutputs] # line 208 forward function in yolox/models/yolo_head.py. Replace the commented code with the uncommented code# [batch, n_anchors_all, 85]# outputs = torch.cat(# [x.flatten(start_dim=2) for x in outputs], dim=2# ).permute(0, 2, 1)proc_view=lambdax: x.view(-1, int(x.size(1)), int(x.size(2) *x.size(3))) outputs=torch.cat( [proc_view(x) forxinoutputs], dim=2 ).permute(0, 2, 1) # line 253 decode_output function in yolox/models/yolo_head.py Replace the commented code with the uncommented code#outputs[..., :2] = (outputs[..., :2] + grids) * strides#outputs[..., 2:4] = torch.exp(outputs[..., 2:4]) * strides#return outputsxy= (outputs[..., :2] +grids) *strideswh=torch.exp(outputs[..., 2:4]) *stridesreturntorch.cat((xy, wh, outputs[..., 4:]), dim=-1) # line 77 in tools/export_onnx.pymodel.head.decode_in_inference=True

3. Export to onnx

# download model wget https://github.com/Megvii-BaseDetection/YOLOX/releases/download/0.1.1rc0/yolox_m.pth # export python tools/export_onnx.py -c yolox_m.pth -f exps/default/yolox_m.py --output-name=yolox_m.onnx --dynamic --no-onnxsim

4. Execute the command

cp YOLOX/yolox_m.onnx tensorRT_cpp/workspace/ cd tensorRT_cpp make yolo -j32

• https://github.com/biubug6/Pytorch_Retinaface

1. Download Pytorch_Retinaface Repo

git clone git@github.com:biubug6/Pytorch_Retinaface.git cd Pytorch_Retinaface

2. Download model from the Training of README.md in https://github.com/biubug6/Pytorch_Retinaface#training .Then unzip it to the /weights . Here, we use mobilenet0.25_Final.pth

3. Modify the code

# line 24 in models/retinaface.py# return out.view(out.shape[0], -1, 2) is modified into returnout.view(-1, int(out.size(1) *out.size(2) *2), 2) # line 35 in models/retinaface.py# return out.view(out.shape[0], -1, 4) is modified intoreturnout.view(-1, int(out.size(1) *out.size(2) *2), 4) # line 46 in models/retinaface.py# return out.view(out.shape[0], -1, 10) is modified intoreturnout.view(-1, int(out.size(1) *out.size(2) *2), 10) # The following modification ensures the output of resize node is based on scale rather than shape such that dynamic batch can be achieved.# line 89 in models/net.py# up3 = F.interpolate(output3, size=[output2.size(2), output2.size(3)], mode="nearest") is modified intoup3=F.interpolate(output3, scale_factor=2, mode="nearest") # line 93 in models/net.py# up2 = F.interpolate(output2, size=[output1.size(2), output1.size(3)], mode="nearest") is modified intoup2=F.interpolate(output2, scale_factor=2, mode="nearest") # The following code removes softmax (bug sometimes happens). At the same time, concatenate the output to simplify the decoding.# line 123 in models/retinaface.py# if self.phase == 'train':# output = (bbox_regressions, classifications, ldm_regressions)# else:# output = (bbox_regressions, F.softmax(classifications, dim=-1), ldm_regressions)# return output# the above is modified into:output= (bbox_regressions, classifications, ldm_regressions) returntorch.cat(output, dim=-1) # set 'opset_version=11' to ensure a successful export# torch_out = torch.onnx._export(net, inputs, output_onnx, export_params=True, verbose=False,# input_names=input_names, output_names=output_names)# is modified into:torch_out=torch.onnx._export(net, inputs, output_onnx, export_params=True, verbose=False, opset_version=11, input_names=input_names, output_names=output_names) 

4. Export to onnx

python convert_to_onnx.py

5. Execute

cp FaceDetector.onnx ../tensorRT_cpp/workspace/mb_retinaface.onnx cd ../tensorRT_cpp make retinaface -j64

• https://github.com/dlunion/DBFace

make dbface -j64

• https://github.com/deepinsight/insightface/tree/master/detection/scrfd
• The know-how about exporting to onnx is comming. Before it is released, come and join us to disucss.

• https://github.com/deepinsight/insightface/tree/master/recognition/arcface_torch

auto arcface = Arcface::create_infer("arcface_iresnet50.fp32.trtmodel", 0); auto feature = arcface->commit(make_tuple(face, landmarks)).get(); cout << feature << endl; // 1x512

• In the example of Face Recognition, workspace/face/library is the set of faces registered.
• workspace/face/recognize is the set of face to be recognized.
• the result is saved in workspace/face/result和workspace/face/library_draw

check the great details in tutorial/2.0

• Just one line of code to export onnx and trtmodel. And save them for usage in the future.

importtrtpymodel=models.resnet18(True).eval() trtpy.from_torch( model, dummy_input, max_batch_size=16, onnx_save_file="test.onnx", engine_save_file="engine.trtmodel" )

• YoloX TensorRT Inference

importtrtpyyolo=tp.Yolo(engine_file, type=tp.YoloType.X) # engine_file is the trtmodel fileimage=cv2.imread("inference/car.jpg") bboxes=yolo.commit(image).get()

• Seamless Inference from Pytorch to TensorRT

importtrtpymodel=models.resnet18(True).eval().to(device) # pt modeltrt_model=tp.from_torch(model, input) trt_out=trt_model(input)

TRT::compile( TRT::Mode::FP32, // compile model in fp323, // max batch size"plugin.onnx", // onnx file"plugin.fp32.trtmodel", // save path{} // redefine the shape of input when needed );

• For fp32 compilation, all you need is offering onnx file whose input shape is allowed to be redefined.

• The in8 inference performs slightly worse than fp32 in precision(about -5% drop down), but stunningly faster. In the framework, we offer int8 inference

// define int8 calibration function to read data and handle it to tenor.auto int8process = [](int current, int count, vector<string>& images, shared_ptr<TRT::Tensor>& tensor){for(int i = 0; i < images.size(); ++i){// int8 compilation requires calibration. We read image data and set_norm_mat. Then the data will be transfered into the tensor.auto image = cv::imread(images[i]); cv::resize(image, image, cv::Size(640, 640)); float mean[] ={0, 0, 0}; float std[] ={1, 1, 1}; tensor->set_norm_mat(i, image, mean, std)} }; // Specify TRT::Mode as INT8auto model_file = "yolov5m.int8.trtmodel"; TRT::compile( TRT::Mode::INT8, // INT83, // max batch size"yolov5m.onnx", // onnx model_file, // saved filename{}, // redefine the input shape int8process, // the recall function for calibration".", // the dir where the image data is used for calibration""// the dir where the data generated from calibration is saved(a.k.a where to load the calibration data.) );

• We integrate into only one int8process function to save otherwise a lot of issues that might happen in tensorRT official implementation.

• We introduce class Tensor for easier inference and data transfer between host to device. So that as a user, the details wouldn't be annoying.

• class Engine is another facilitator.

// load model and get a shared_ptr. get nullptr if fail to load.auto engine = TRT::load_infer("yolov5m.fp32.trtmodel"); // print model info engine->print(); // load imageauto image = imread("demo.jpg"); // get the model input and output node, which can be accessed by name or indexauto input = engine->input(0); // or auto input = engine->input("images");auto output = engine->output(0); // or auto output = engine->output("output");// put the image into input tensor by calling set_norm_mat()float mean[] ={0, 0, 0}; float std[] ={1, 1, 1}; input->set_norm_mat(i, image, mean, std); // do the inference. Here sync(true) or async(false) is optional engine->forward(); // engine->forward(true or false)// get the outut_ptr, which can used to access the outputfloat* output_ptr = output->cpu<float>();

• You only need to define kernel function and inference process. The details of code(e.g the serialization, deserialization and injection of plugin etc) are under the hood.
• Easy to implement a new plugin in FP32 and FP16. Refer to HSwish.cu for details.

template<> __global__ voidHSwishKernel(float* input, float* output, int edge){KernelPositionBlock; float x = input[position]; float a = x + 3; a = a < 0 ? 0 : (a >= 6 ? 6 : a); output[position] = x * a / 6} intHSwish::enqueue(const std::vector<GTensor>& inputs, std::vector<GTensor>& outputs, const std::vector<GTensor>& weights, void* workspace, cudaStream_t stream){int count = inputs[0].count(); auto grid = CUDATools::grid_dims(count); auto block = CUDATools::block_dims(count); HSwishKernel <<<grid, block, 0, stream >>> (inputs[0].ptr<float>(), outputs[0].ptr<float>(), count); return0} RegisterPlugin(HSwish);