Files

T

MediaPipe Teamandmgyong 48bcbb115f Project import generated by Copybara.

GitOrigin-RevId: 50714fe28298d7b707eff7304547d89d6ec34a54

2019-11-21 13:20:47 -08:00

5.4 KiB

Raw Blame History

Coral Dev Board Setup (experimental)

Dislaimer: Running MediaPipe on Coral is experimental, and this process may not be exact and is subject to change. These instructions have only been tested on the coral dev board with OS version mendel day, and may vary for different devices and workstations.

This file describes how to prepare a Google Coral Dev Board and setup a linux Docker container for building MediaPipe applications that run on Edge TPU.

Before creating the Docker

(on host machine) run setup.sh from MediaPipe root directory
```
  sh mediapipe/examples/coral/setup.sh
```
Setup the coral device via here, and ensure the mdt command works

(on coral device) prepare MediaPipe

  cd ~
  sudo apt-get install git
  git clone https://github.com/google/mediapipe.git
  mkdir mediapipe/bazel-bin

(on coral device) install opencv 3.2

  sudo apt-get update && apt-get install -y libopencv-dev

(on coral device) find all opencv libs

  find /usr/lib/aarch64-linux-gnu/ -name 'libopencv*so'

(on host machine) copy core opencv libs from coral device to a local folder inside MediaPipe checkout:

  # in root level mediapipe folder #
  mdt pull /usr/lib/aarch64-linux-gnu/libopencv_core.so opencv32_arm64_libs
  mdt pull /usr/lib/aarch64-linux-gnu/libopencv_calib3d.so opencv32_arm64_libs
  mdt pull /usr/lib/aarch64-linux-gnu/libopencv_features2d.so opencv32_arm64_libs
  mdt pull /usr/lib/aarch64-linux-gnu/libopencv_highgui.so opencv32_arm64_libs
  mdt pull /usr/lib/aarch64-linux-gnu/libopencv_imgcodecs.so opencv32_arm64_libs
  mdt pull /usr/lib/aarch64-linux-gnu/libopencv_imgproc.so opencv32_arm64_libs
  mdt pull /usr/lib/aarch64-linux-gnu/libopencv_video.so opencv32_arm64_libs
  mdt pull /usr/lib/aarch64-linux-gnu/libopencv_videoio.so opencv32_arm64_libs

(on host machine) Create and start the docker environment

  # from mediapipe root level directory #
  docker build -t coral .
  docker run -it --name coral coral:latest

Inside the Docker environment

Update library paths in /mediapipe/third_party/opencv_linux.BUILD

(replace 'x86_64-linux-gnu' with 'aarch64-linux-gnu')

  "lib/aarch64-linux-gnu/libopencv_core.so",
  "lib/aarch64-linux-gnu/libopencv_calib3d.so",
  "lib/aarch64-linux-gnu/libopencv_features2d.so",
  "lib/aarch64-linux-gnu/libopencv_highgui.so",
  "lib/aarch64-linux-gnu/libopencv_imgcodecs.so",
  "lib/aarch64-linux-gnu/libopencv_imgproc.so",
  "lib/aarch64-linux-gnu/libopencv_video.so",
  "lib/aarch64-linux-gnu/libopencv_videoio.so",

Attempt to build hello world (to download external deps)

  bazel build -c opt --define MEDIAPIPE_DISABLE_GPU=1 mediapipe/examples/desktop/hello_world:hello_world

Edit /mediapipe/bazel-mediapipe/external/com_github_glog_glog/src/signalhandler.cc

on line 78, replace

  return (void*)context->PC_FROM_UCONTEXT;

with

  return NULL;

Edit /edgetpu/libedgetpu/BUILD

to add this build target

   cc_library(
     name = "lib",
     srcs = [
         "libedgetpu.so",
     ],
     visibility = ["//visibility:public"],
   )

Edit tflite_inference_calculator.cc BUILD rules:

  sed -i 's/\":tflite_inference_calculator_cc_proto\",/\":tflite_inference_calculator_cc_proto\",\n\t\"@edgetpu\/\/:header\",\n\t\"@libedgetpu\/\/:lib\",/g' mediapipe/calculators/tflite/BUILD

The above command should add

  "@edgetpu//:header",
  "@libedgetpu//:lib",

to the _deps_ of tflite_inference_calculator.cc

Now try cross-compiling for device

Object detection demo

  bazel build -c opt --crosstool_top=@crosstool//:toolchains --compiler=gcc --cpu=aarch64 --define MEDIAPIPE_DISABLE_GPU=1 --copt -DMEDIAPIPE_EDGE_TPU --copt=-flax-vector-conversions mediapipe/examples/coral:object_detection_cpu

Copy object_detection_cpu binary to the MediaPipe checkout on the coral device

    # outside docker env, open new terminal on host machine #
    docker ps
    docker cp <container-id>:/mediapipe/bazel-bin/mediapipe/examples/coral/object_detection_cpu /tmp/.
    mdt push /tmp/object_detection_cpu /home/mendel/mediapipe/bazel-bin/.

Face detection demo

  bazel build -c opt --crosstool_top=@crosstool//:toolchains --compiler=gcc --cpu=aarch64 --define MEDIAPIPE_DISABLE_GPU=1 --copt -DMEDIAPIPE_EDGE_TPU --copt=-flax-vector-conversions mediapipe/examples/coral:face_detection_cpu

Copy face_detection_cpu binary to the MediaPipe checkout on the coral device

    # outside docker env, open new terminal on host machine #
    docker ps
    docker cp <container-id>:/mediapipe/bazel-bin/mediapipe/examples/coral/face_detection_cpu /tmp/.
    mdt push /tmp/face_detection_cpu /home/mendel/mediapipe/bazel-bin/.

On the coral device (with display)

 # Object detection
 cd ~/mediapipe
 chmod +x bazel-bin/object_detection_cpu
 export GLOG_logtostderr=1
 bazel-bin/object_detection_cpu --calculator_graph_config_file=mediapipe/examples/coral/graphs/object_detection_desktop_live.pbtxt

 # Face detection
 cd ~/mediapipe
 chmod +x bazel-bin/face_detection_cpu
 export GLOG_logtostderr=1
 bazel-bin/face_detection_cpu --calculator_graph_config_file=mediapipe/examples/coral/graphs/face_detection_desktop_live.pbtxt

5.4 KiB Raw Blame History