pull

.devcontainer/devcontainer.json

@@ -1,7 +1,11 @@
 // See this page for reference of options: https://containers.dev/implementors/json_reference
 {
 	"name": "Existing Dockerfile",
+<<<<<<< Updated upstream
 	"image": "ghcr.io/tritonuas/obcpp:x86",
+=======
+	"image": "ghcr.io/tritonuas/obcpp:jetson",
+>>>>>>> Stashed changes
 	// enable when need to connect over USB to pixhawk
 	// also: need to run obcpp with sudo or add tuas user to dialout group with
 	// 	   `sudo usermod -aG dialout tuas && newgrp && bash`

docker/Dockerfile.jetson

@@ -1,4 +1,12 @@
+<<<<<<< Updated upstream
 FROM dustynv/l4t-pytorch:r36.2.0
+=======
+# FROM tritonuas/jetson-base:r36.2.0
+FROM dustynv/l4t-pytorch:r36.2.0
+
+# this base image came from dusty-nv/jetson-containers
+# ./build.sh --name=tritonuas/jetson-base pytorch:2.1 torchvision opencv 
+>>>>>>> Stashed changes
 
 ARG USERNAME=tuas USER_UID=1000 USER_GID=1000 DEBIAN_FRONTEND=noninteractive
 
@@ -50,6 +58,14 @@ RUN --mount=target=/var/lib/apt/lists,type=cache,sharing=locked \
                            libopenblas-dev \
                            ninja-build
 
+# RUN sudo dpkg --remove libopencv-dev
+# RUN sudo apt install -f
+# RUN --mount=target=/var/lib/apt/lists,type=cache,sharing=locked \
+#     --mount=target=/var/cache/apt,type=cache,sharing=locked \
+#     rm -f /etc/apt/apt.conf.d/docker-clean \ 
+#     && apt-get update \
+#     && apt-get install -y libopencv-dev
+
 RUN pip3 install typing-extensions PyYAML cpplint
 
 RUN echo $USERNAME ALL=\(root\) NOPASSWD:ALL > /etc/sudoers.d/$USERNAME \
@@ -65,6 +81,7 @@ RUN git clone --depth 1 https://github.com/mavlink/MAVSDK.git --branch v2.9.1 --
 # pull and build torchvision
 # refer to this page for version compatibilty with pytorch (libtorch) https://github.com/pytorch/pytorch/wiki/PyTorch-Versions
 ARG TORCHVISION_VERSION=0.17.0
+<<<<<<< Updated upstream
 # Space separated list of CUDA architecture versions.
 # The version nubmers depend on the NVIDIA GPU model we're using and the installed CUDA version. 
 # For the Jetson Orin Nano with the "Ampere" architecture and CUDA 12.1 we can use version 8.6 (written as 86 in CUDA_ARCH_LIST).
@@ -94,16 +111,43 @@ RUN gdown 1VtBji-cWfetM5nXZwt55JuHPWPGahQOH -O ${ARENA_TAR_PATH}
 RUN tar -xvzf ${ARENA_TAR_PATH}
 WORKDIR ${ARENA_EXTRACTED_PATH}
 RUN sh Arena_SDK_ARM64.conf
+=======
+#ARG TORCHVISION_INSTALL_DIR=/torchvision-tmp
+WORKDIR ${TORCHVISION_INSTALL_DIR} 
+RUN wget "https://github.com/pytorch/vision/archive/refs/tags/v${TORCHVISION_VERSION}.zip" \
+    && unzip "v${TORCHVISION_VERSION}.zip" \
+    && cd vision-0.17.0 \
+    && mkdir build \
+    && cd build \
+    && cmake -DWITH_CUDA=1 -DCUDA_HAS_FP16=1 -DCUDA_NO_HALF_OPERATORS=1 -DCUDA_NO_HALF_CONVERSIONS=1 -DCUDA_NO_HALF2_OPERATORS=1 -DCMAKE_BUILD_TYPE=Release -DCMAKE_PREFIX_PATH="/usr/local/lib/python3.10/dist-packages/torch/share/cmake/Torch" .. \
+    && make -j4 \
+    && make install
+
+# # Install g++10 and replace the older version. For some reason some c++ 20 features aren't working with g++9 even though
+# # we have CMake configured to use c++ 20 https://stackoverflow.com/questions/6903.1073/why-am-i-missing-c20-headers-and-how-do-i-fix-this
+# RUN apt-get update && apt-get install -y g++-10 gcc-10
+# RUN update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 10
+# RUN update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 10
+# RUN update-alternatives --set gcc /usr/bin/gcc-10
+# RUN update-alternatives --set g++ /usr/bin/g++-10
+>>>>>>> Stashed changes
 
 WORKDIR /obcpp
 COPY . .
 
 RUN rm -rf /obcpp/build
 WORKDIR /obcpp/build
 ENV CMAKE_PREFIX_PATH="/usr/local/lib/python3.10/dist-packages/torch/share/cmake/Torch;/usr/local/share/cmake/TorchVision"
+<<<<<<< Updated upstream
 RUN GITHUB_ACTIONS=true cmake -DCMAKE_PREFIX_PATH="/usr/local/lib/python3.10/dist-packages/torch/share/cmake/Torch;/usr/local/share/cmake/TorchVision" -DCMAKE_MODULE_PATH="/usr/local/share/cmake/TorchVision" -DCMAKE_BUILD_TYPE="Release" ..
 
 RUN ninja obcpp 
+=======
+RUN GITHUB_ACTIONS=true cmake -DCMAKE_PREFIX_PATH="/usr/local/lib/python3.10/dist-packages/torch/share/cmake/Torch;/usr/local/share/cmake/TorchVision" -DCMAKE_MODULE_PATH="/usr/local/share/cmake/TorchVision" -DCMAKE_BUILD_TYPE="Release" -DCMAKE_JOB_POOLS="j=2" ..
+
+# RUN make obcpp cuda_check load_torchvision_model VERBOSE=1
+RUN ninja obcpp
+>>>>>>> Stashed changes
 
 # login as non-root user
 # USER $USERNAME

docker/Makefile

@@ -26,3 +26,5 @@ stop-jetson-pixhawk-compose:
 run-jetson-cuda-check:
 	docker run -it --rm --runtime nvidia -i tritonuas/obcpp:nvidia /obcpp/build/bin/cuda_check
 
+jetson-develop:
+	cd .. && docker run -it --net=host --runtime=nvidia --volume=./:/obcpp -i tritonuas/obcpp:jetson /bin/bash

docker/jetson-pixhawk-compose.yml

@@ -1,7 +1,7 @@
 version: "3"
 services:
   obcpp:
-    image: tritonuas/obcpp:jetson
+    image: ghcr.io/tritonuas/obcpp:debug-jetson-docker-jetpack36
     runtime: nvidia
     network_mode: "host"
     devices:

include/cv/saliency.hpp

@@ -36,6 +36,7 @@ class Saliency {
  private:
     std::string modelPath;              // path to prediction model
     torch::jit::script::Module module;  // the loaded model
+   //  c10::Device device;                 // 
 };
 
 #endif  // INCLUDE_CV_SALIENCY_HPP_

src/cv/CMakeLists.txt

@@ -14,6 +14,7 @@ set(FILES
 set(LIB_DEPS
     obcpp_protos
     obcpp_utilities
+    obcpp_camera
 )
 
 add_library(${LIB_NAME} STATIC

src/cv/saliency.cpp

@@ -34,15 +34,19 @@ std::vector<CroppedTarget> Saliency::salience(cv::Mat image) {
     tensor = tensor.toType(c10::kFloat).div(255);
     // swap axis
     tensor = Saliency::transpose(tensor, { (2), (0), (1) });
-    auto input_to_net = ToInput(tensor);
+
+    c10::Device device = torch::cuda::is_available() ? torch::kCUDA : torch::kCPU;  // eventually add device as member of Saliency
+    auto tensor_cuda = tensor.to(device);
+
+    auto input_to_net = ToInput(tensor_cuda);
 
     /* 
      * forward() runs an inference on the input image using the provided model
      * and returns predictions as List[Dict[Tensor]], where the fields of Dict
      * that we want are : a) boxes (FloatTensor[N, 4]): the predicted boxes, and
      * b) scores (Tensor[N]): the scores of each detection.
      */
-
+    
     // output is a tuple of (losses, detections)
     auto output = module.forward(input_to_net);
     c10::ivalue::Tuple& tuple = output.toTupleRef();

tests/integration/cv_saliency.cpp

@@ -9,12 +9,14 @@
 
 // expected arguments: <path-to-model> <path-to-image> 
 int main(int argc, const char* argv[]) {
+
   if (argc != 3) {
     std::cerr << "usage: example-app <path-to-model> <path-to-image>\n";
     return -1;
   }
 
   // convert image to tensor
+
   const char* modelPath = argv[1];
   Saliency sal(modelPath);
   const char* imgPath = argv[2];
@@ -36,6 +38,7 @@ int main(int argc, const char* argv[]) {
   // cv::namedWindow("cropped targets", cv::WINDOW_FULLSCREEN);
   // cv::imshow("cropped targets", img);
   // cv::waitKey(0);  
+
   cv::imwrite("croppedTargets.jpg", img);
   LOG_F(INFO, "saved croppedTargets.jpg to build/");
   // testing: save input image to file path (cv::imsave?) with bounding boxes overlayed