Using Google Coral USB on Ubuntu 22.04 / Raspberry pi4

This project is a guide for using Raspberry Pi 4, Ubuntu 22.04 (aarch64) and Google Coral USB.

I followed Google's instructions, I was able to use Coral USB without any major problems but only on my PC (amd64/ubuntu 20.04). With Raspberry Pi 4 and ubuntu 22.04 aarch64, the build goes to failure or the result binary even after being built did not work normally from the ways of compiling including native / cross / docker / colab.

On Raspberry Pi 4, all operations were going well only on 32bit Debian Buster, and I wasted too much time for trying to use tpu on other 64bit OS.

I hope this page helps users who have been struggling as me.

1. Prerequisite

Platform: raspberry pi 4 / Ubuntu 22.04(64bit) installed on NVME/SSD without SD micro
- Refer to https://www.instructables.com/Raspberry-Pi-4-USB-Boot-No-SD-Card/
Kernel/Package
- Linux ubuntu 5.15.0-1029-raspi #31-Ubuntu SMP PREEMPT Sat Apr 22 12:26:40 UTC 2023 aarch64 aarch64 aarch64 GNU/Linux
- Ubuntu 22.04.2 LTS \n \l

Set bash as default & install minimals.

sudo dpkg-reconfigure dash
sudo apt install build-essential vim ssh git tig cmake samba curl

Set up bazelisk
- version 1.17.0 : download bazelisk binary at https://github.com/bazelbuild/bazelisk/releases
- copy it to /usr/local/bin/bazel and apply chmod +x

Set up pyenv

install pyenv

git clone htps://github.com/pyenv/pyenv.git ~/.pyenv
echo 'export PYENV_ROOT="$HOME/.pyenv"' >> ~/.bashrc
echo 'export PATH="$PYENV_ROOT/bin:$PATH"' >> ~/.bashrc
echo -e 'if command -v pyenv 1>/dev/null 2>&1; then\n  eval "$(pyenv init -)"\nfi' >> ~/.bashrc

sudo apt install zlib1g-dev bzip2 readline-common sqlite3 openssl libssl-dev xz-utils \
libbz2-dev libncurses5-dev libssl-dev libreadline-dev libffi-dev libsqlite3-dev liblzma-dev python3-tk tk-dev

. ~/.bashrc

set virtualenv v3.9.16

git clone https://github.com/pyenv/pyenv-virtualenv.git $(pyenv root)/plugins/pyenv-virtualenv
pyenv install 3.9.16
pyenv virtualenv 3.9.16 v3.9.16

required re-login

pyenv activate v3.9.16
pyenv global 3.9.16
pip install numpy==1.23 (for the time to run application)
pip install Pillow wheel

2. Build tensorflowlite version 2.5.0

sudo apt install libabsl-dev libusb-1.0-0-dev

 mkdir ~/github && cd github
 git clone git@github.com:tensorflow/tensorflow.git tf250
 cd tf250 && git check out v2.5.0 -b v2.5.0
 ./configure
	## configure : check the python location whether indicate pyenv environments.

Build tensorflowlite
```
bazel build -c opt //tensorflow/lite:tensorflowlite
```
You may have a build error ${bazel cache}/external/ruy/ruy/block_map.cc, therefore, you shall add next statement into the file.
```
#include <limits>
```
When you used upper version than 2.5.0 and the error occurred related with bazel version, you need to declare export USE_BAZEL_VERSION=5.4.0 . But, it could produce another issues regarding edgetpu.

Installing libflatbuffers which used for build tensorflowlite onto system library

cd ~
mkdir ~/tf250_lib ~/tf250_external && \
cp ~/github/tf250/bazel-tf250/bazel-out/aarch64-opt/bin/tensorflow/lite/libtensorflowlite.so ~/tf250_lib/
cp -rfl ~/github/tf250/bazel-tf250/external/* ~/tf250_external/

Go to the flatbuffers folder and build with cmake:

cd ~/tf250_external/flatbuffers && mkdir build_cmake && cd build_cmake
cmake -DCMAKE_CXX_FLAGS="-Wno-error=class-memaccess" ..
make -j4
sudo  make install
## You shall check a file /usr/local/lib/libflatbuffers.a, and  find a version:1.12.0 in /usr/local/include/flatbuffers/base.h

3. libedgetpu.so : for cpp & python

Cautions: Never try to build aarch64 libedgetpu.so for raspberry pi4. Native/CrossCompile/Docker/CoLab, any build ways from websites doesn't provide proper & exact solution you may need. Even though built the library successfully, there were so many fxxx terrible issues on Ubuntu 22.04 64bit.

echo "deb https://packages.cloud.google.com/apt coral-edgetpu-stable main" | sudo tee /etc/apt/sources.list.d/coral-edgetpu.list

curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | sudo apt-key add -

sudo apt update
sudo apt install libedgetpu1-std

4. python module : tflite_runtime : only for python

Build in tflite_runtime reference source code.

 cd ~/github/tf250/tensorflow/lite/tools/pip_package

Patch build_pip_package_with_bazel.sh with following statement

     @@ -48,6 +48,8 @@ cp -r "${TENSORFLOW_LITE_DIR}/tools/pip_package/debian" \
            "${TENSORFLOW_LITE_DIR}/python/interpreter_wrapper" \
            "${BUILD_DIR}"
      cp "${TENSORFLOW_LITE_DIR}/python/interpreter.py" \
     +   "${TENSORFLOW_LITE_DIR}/python/metrics_interface.py" \
     +   "${TENSORFLOW_LITE_DIR}/python/metrics_portable.py" \
         "${BUILD_DIR}/tflite_runtime"

Build pip packages

./build_pip_packages_with_bazel.sh
pip install --force-reinstall gen/tflite_pip/python3/dist/tflite_runtime-2.5.0-cp39-cp39-linux_aarch64.whl

The content below is not required. For reference only

Misc.

 cd tensorflow/lite/tools/make
 ./download_dependecies.sh && ./build_aarch64_lib.sh
 ##it could be errors in ruy/block_map.cc and absl/CMakeLists.txt

 ## 1. the errors would occurr in ruy/ruy/block_map.cc regarding for <limits>
 ##    it needs to add these headers into the source file
 ##     #include <stdexcept>
 ##     #include <limits>
 ## 2. CMakeList.txt
 ## 	 it needs to link system library libraries with renamed.

5. Running example / cpp

classification

	cd coral/tflite/cpp/examples/classification
	make
	./run_tpu.sh # or run.sh : cpu

lstpu

	cd coral/tflite/cpp/examples/lstpu
	./build.sh
	./lstpu

6. Running example / python

You should activate virtual environment before starting example.

	pyenv activate v3.9.16

classification

	cd coral/tflite/python/examples/classification
	python3 classify_image.py  \
  --model	models/mobilenet_v2_1.0_224_inat_bird_quant_edgetpu.tflite \
  --labels models/inat_bird_labels.txt \
  --input images/parrot.jpg

detection

	cd coral/tflite/python/examples/detection
	python3 detect_image.py \
  --model models/ssd_mobilenet_v2_coco_quant_postprocess_edgetpu.tflite \
  --labels models/coco_labels.txt   --input images/grace_hopper.bmp

6. Trouble shooting

No	Trouble Cases	Remark
1	RuntimeError: module compiled against API version 0x10 but this version of numpy is 0xd ImportError: numpy.core.multiarray failed to import	Python
	check the pynum version `pip lists numpy`. Probably, you need to re-install pynum:1.23 `pip install numpy==1.23`
2	ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts. tflite-runtime 2.5.0 requires numpy~=1.19.2, but you have numpy 1.23.0 which is incompatible.	Python
	You can ignore this error statement.

dig9 / coral_rpi4_ubuntu22