Add ncnn deployment examples (#145)

* Add ncnn deployment examples

* fix:add space to depth function (#146)

* Fixing lint

* Fixing C++ return bug

* Fixing lint and add more tests for space_to_depth

* Fixing TypeError of torch.Size

* Adding onnx export tools

* Refactor YOLODeployFriendly

* Move export_onnx.py to ncnn/tools

* Adapt to yolov5

* Remove tools

* Add yolort ncnn param examples and minor fixes

* Rename to yolort-opt.param

* Fixing lint

Co-authored-by: xiguadong <55774832+xiguadong@users.noreply.github.com>

deployment/ncnn/CMakeLists.txt

@@ -0,0 +1,25 @@
+cmake_minimum_required(VERSION 3.14)
+
+project(yolort_ncnn)
+
+find_package(OpenCV REQUIRED)
+
+# If the package has been found, several variables will
+# be set, you can find the full list with descriptions
+# in the OpenCVConfig.cmake file.
+# Print some message showing some of them
+message(STATUS "OpenCV library status:")
+message(STATUS "    config: ${OpenCV_DIR}")
+message(STATUS "    version: ${OpenCV_VERSION}")
+message(STATUS "    libraries: ${OpenCV_LIBS}")
+message(STATUS "    include path: ${OpenCV_INCLUDE_DIRS}")
+
+find_package(ncnn REQUIRED)
+
+FILE(GLOB YOLO_SOURCE_FILES *.cpp)
+
+add_executable(yolort_ncnn ${YOLO_SOURCE_FILES})
+
+target_compile_features(yolort_ncnn PUBLIC cxx_range_for)
+
+target_link_libraries(yolort_ncnn ncnn ${OpenCV_LIBS})

deployment/ncnn/README.md

@@ -0,0 +1,48 @@
+# Ncnn Inference
+
+The ncnn inference for `yolort`, both GPU and CPU are supported.
+
+## Dependencies
+
+- Ubuntu 18.04
+- ncnn
+- OpenCV 3.4+
+
+## Usage
+
+1. First, Setup the environment variables.
+
+    ```bash
+    export TORCH_PATH=$(dirname $(python -c "import torch; print(torch.__file__)"))
+    export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$TORCH_PATH/lib/
+    ```
+
+1. First, compile `ncnn` using the following scripts.
+
+    ```bash
+    git clone --recursive git@github.com:Tencent/ncnn.git
+    cd ncnn
+    mkdir build && cd build
+    cmake -DCMAKE_BUILD_TYPE=Release -DNCNN_SYSTEM_GLSLANG=ON -DNCNN_BUILD_EXAMPLES=ON .. # Set -DNCNN_VULKAN=ON if you're using VULKAN
+    make -j4
+    make install
+    ```
+
+    Or follow the [official instructions](https://github.com/Tencent/ncnn/wiki/how-to-build) to install ncnn.
+
+1. Then compile the source code.
+
+    ```bash
+    cd deployment/ncnn
+    mkdir build && cd build
+    cmake .. -Dncnn_DIR=<ncnn_install_dir>/lib/cmake/ncnn/
+    make
+    ```
+
+_Note: you have to change <ncnn_install_dir> to your machine's directory, it is the directory that contains ncnnConfig.cmake, if you are following the above operations, you should set it to <./ncnn/build/install>_
+
+1. Now, you can infer your own images with ncnn.
+
+    ```bash
+    ./yolort_ncnn ../../../test/assets/zidane.jpg
+    ```

deployment/ncnn/export_onnx.py

@@ -0,0 +1,62 @@
+# Copyright (c) 2021, Zhiqiang Wang. All Rights Reserved.
+import argparse
+import torch
+from tools.yolort_deploy_friendly import yolov5s_r40_deploy_ncnn
+
+
+def get_parser():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--weights', type=str, default='./yolov5s.pt',
+                        help='weights path')
+    parser.add_argument('--output_path', type=str, default='./yolov5s.onnx',
+                        help='path of exported onnx')
+    parser.add_argument('--img_size', nargs='+', type=int, default=[640, 640],
+                        help='image (height, width)')
+    parser.add_argument('--num_classes', type=int, default=80,
+                        help='number of classes')
+    parser.add_argument('--batch_size', type=int, default=1,
+                        help='batch size')
+    parser.add_argument('--device', default='cpu',
+                        help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--half', action='store_true',
+                        help='FP16 half-precision export')
+    parser.add_argument('--dynamic', action='store_true',
+                        help='ONNX: dynamic axes')
+    parser.add_argument('--simplify', action='store_true',
+                        help='ONNX: simplify model')
+    parser.add_argument('--opset', type=int, default=11,
+                        help='ONNX: opset version')
+    return parser
+
+
+def cli_main():
+    parser = get_parser()
+    args = parser.parse_args()
+    print(args)
+    export_onnx(args)
+
+
+def export_onnx(args):
+
+    model = yolov5s_r40_deploy_ncnn(
+        pretrained=True,
+        num_classes=args.num_classes,
+    )
+    img = torch.rand(args.batch_size, 3, 640, 640)
+    outputs = model(img)
+    assert len(outputs) == 3
+
+    torch.onnx.export(
+        model,
+        img,
+        args.output_path,
+        verbose=False,
+        opset_version=args.opset,
+        do_constant_folding=True,
+        input_names=['images'],
+        output_names=['h1', 'h2', 'h3'],
+    )
+
+
+if __name__ == "__main__":
+    cli_main()