Machine learning packages like TensorFlow perform much better on a 64-bit OS than on a 32-bit OS.
If you are going to use TensorFlow on Raspberry Pi, make sure to run it on a 64-bit OS.
In a previous article, I looked at how to cross-build the Raspberry Pi OS 64-bit version of TensorFlow Lite on your x86 PC and then install it on the Raspberry Pi. Since TensorFlow Lite can be installed not only on Raspberry Pi but also on Android devices using arm CPU, cross-build is often used. Isn't it too difficult to download and build the TensorFlow Lite source code on your Android smartphone?
In this article, we will look at how to build a TensorFlow Python on Raspberry Pi OS 64-bit . To be honest, I tried to cross-build the TensorFlow package, but it failed due to several issues.
For Raspberry Pi, 8GB model must be prepared. The reason is that the Basel build requires a lot of memory. If a build error occurs even when using 8GB of memory,
Increase memory using ZRAM
If an error occurs during the Basel build process during the following tasks despite using the 8GB model, increase the memory using zram. It is also faster than using a swap file. And since it reduces the IO of the SD card that occurs when using the swap file, it also helps the life of the SD card.
install zram tool
Since there is already a well-made installation script file, download it from git.
$ git clone https://github.com/StuartIanNaylor/zram-swap-config \ && cd zram-swap-config $ sudo ./install.sh
zram settings
Edit the /etc/zram-swap-config.conf file. The description of the item values is as follows.
- MEM_FACTOR = Percentage of available ram to allocate to all zram swap devices which is divided equally by swap_devices number
- DRIVE_FACTOR = Virtual uncompressed zram drive size estimate approx alg compression ratio
- COMP_ALG lz4 is faster than lzo but some distro's show compile and operational difference and in use lzo depending on binary may be faster. Compression rates list below are minimums and generally far bigger in operation but dependent on content.
- SWAP_DEVICES = number of indivial drives sharing memeory provided by MEM_FACTOR each device support multiple streams 1 large drive is generally better
- SWAP_PRI = swap_priority for each drive 75 is a high order preference and should be well above other swap drives
- PAGE_CLUSTER default page cluster is 3 which caches fetches in batches of 8 and helps with HDD paging, with zram mem 0 forces single page fetches This can help reduce latency and increase performance
- SWAPPINESS default swappiness is 60 but with increased performance of zram swap 80 garners overall performance gain without excessive load Because zram uses compression load is created and even if minimal at intense load periods such as boot any extra load is unwanted Unfortunately there is no dynamic load balancing of swappiness as with zram in general operation SWAPINESS=100 will garner performance benefit If the overall load is reasonable at high load it will cause load to accumulate. If you check my repo there will also be a simple dynamic load based SWAPPINESS governor that will get of the hurdle of a static compromise on swappiness
<from https://github.com/StuartIanNaylor/zram-swap-config>
The following is an example of setting in Raspberry Pi 4B with 8GB memory. Zram is applied to 40% of the memory.
root@raspberrypi:~# cat /etc/zram-swap-config.conf MEM_FACTOR=40 DRIVE_FACTOR=300 COMP_ALG=lz4 SWAP_DEVICES=1 SWAP_PRI=75 PAGE_CLUSTER=0 SWAPPINESS=90 pi@raspberrypi:~ $ sudo cat /proc/swaps Filename Type Size Used Priority /dev/zram0 partition 9600608 0 75 /var/swap file 102396 0 -2 pi@raspberrypi:~ $ free -m total used free shared buff/cache available Mem: 7813 358 7012 36 441 7306 Swap: 9475 0 9475
Change the /var/swap swap file to use only zram. This operation is performed as the root user (you can omit this operation).
pi@raspberrypi:~ $ sudo su - SSH is enabled and the default password for the 'pi' user has not been changed. This is a security risk - please login as the 'pi' user and type 'passwd' to set a new password. root@raspberrypi:~# root@raspberrypi:~# root@raspberrypi:~# dphys-swapfile swapoff \ && dphys-swapfile uninstall \ && update-rc.d dphys-swapfile disable root@raspberrypi:~# root@raspberrypi:~# root@raspberrypi:~# root@raspberrypi:~# free -m total used free shared buff/cache available Mem: 7813 363 7004 36 445 7302 Swap: 9375 0 9375
Reboot for zram changes to take effect.
Install required packages to build tensorflow
For reference, the build work was carried out on the Raspberry Pi OS 64-bit Bullseye Desktop version.
First, install the packages required for the build task.
apt-get update apt-get install -y \ build-essential \ curl \ git \ libcurl3-dev \ libfreetype6-dev \ libhdf5-serial-dev \ libzmq3-dev \ pkg-config \ rsync \ software-properties-common \ sudo \ unzip \ zip \ zlib1g-dev
apt-get install -y openjdk-11-jdk \ openjdk-11-jre-headlessapt-get install -y python3-pippython3 -m pip --no-cache-dir install --upgrade \ "pip<20.3" \ setuptools apt-get install -y \ python3-dev \ virtualenv \ swig \ gfortran \ libblas-dev \ liblapack-dev python3 -m pip --no-cache-dir install \ Pillow \ keras_preprocessing \ h5py \ matplotlib \ mock \ scipy \ sklearn \ pandas \ portpicker \ enum34
And install Basel.
mkdir bazel
cd bazel
curl -fSsL -O https://github.com/bazelbuild/bazel/releases/download/4.2.1/bazel-4.2.1-dist.zip
unzip bazel-4.2.1-dist.zip
bash ./compile.sh
cp output/bazel /usr/local/bin/
If Basel is installed properly, you can check as follows.
pi@raspberrypi:~/src/bazel $ bazel --version
bazel 4.2.1- (@non-git)Copy the TensorFlow source code.
pi@raspberrypi64:~ $ cd src pi@raspberrypi64:~/src $git clone -b v2.8.0 https://github.com/tensorflow/tensorflow.git cd tensorflow
It's time to build the python wheel
The following command takes more than 10 hours. Therefore, please avoid working on remote ssh if possible. When the ssh connection is disconnected, the build process is terminated.
pi@raspberrypi:~/src/tensorflow $ sudo bazel build \ --config=monolithic \ --config=noaws \ --config=nohdfs \ --config=nonccl \ --config=v2 \ --define=tflite_pip_with_flex=true \ --define=tflite_with_xnnpack=true \ --ui_actions_shown=64 \ //tensorflow/tools/pip_package:build_pip_package
The build ended after about 10 hours.
............Target //tensorflow/tools/pip_package:build_pip_package up-to-date: bazel-bin/tensorflow/tools/pip_package/build_pip_package INFO: Elapsed time: 53061.319s, Critical Path: 3604.28s INFO: 11853 processes: 1175 internal, 10678 local. INFO: Build completed successfully, 11853 total actions
Finally, build the TensorFlow Python wheel.
pi@raspberrypi:~/src/tensorflow $ ./tensorflow/tools/pip_package/build_pip_package.sh /tmp/tensorflow_pkg /home/pi/src/tensorflow /tmp/tmp.wiZjwBgJfY/tensorflow/include /home/pi/src/tensorflow /home/pi/src/tensorflow Sat 26 Feb 14:27:09 KST 2022 : === Building wheel warning: no files found matching 'README' warning: no files found matching '*.pyd' under directory '*' warning: no files found matching '*.pyi' under directory '*' warning: no files found matching '*.pd' under directory '*' warning: no files found matching '*.so.[0-9]' under directory '*' warning: no files found matching '*.dylib' under directory '*' warning: no files found matching '*.dll' under directory '*' warning: no files found matching '*.lib' under directory '*' warning: no files found matching '*.csv' under directory '*' warning: no files found matching '*.h' under directory 'tensorflow/include/tensorflow' warning: no files found matching '*.proto' under directory 'tensorflow/include/tensorflow' warning: no files found matching '*' under directory 'tensorflow/include/third_party' /usr/local/lib/python3.9/dist-packages/setuptools/command/install.py:34: SetuptoolsDeprecationWarning: setup.py install is deprecated. Use build and pip and other standards-based tools. warnings.warn( Sat 26 Feb 14:31:58 KST 2022 : === Output wheel file is in: /tmp/tensorflow_pkg
You can see the message that the tensorflow python wheel has been created in the tmp/tensorflow_pkg directory.
pi@raspberrypi:~/src/tensorflow $ ls -al /tmp/tensorflow_pkg/ total 279968 drwxr-xr-x 2 root root 4096 Feb 26 14:31 . drwxrwxrwt 735 root root 36864 Feb 26 14:32 .. -rw-r--r-- 1 root root 286646178 Feb 26 14:31 tensorflow-2.8.0-cp39-cp39-linux_aarch64.whl
The reason the Python version of wheel is 3.9 is because the Python version installed on the Raspberry Pi OS 64-bit Bullseye version is 3.9. If you want to make a different version of Python wheel such as 3.8, install Python 3.8 and connect the python3 symbolic link in the /usr/bin directory to python3.8.
pi@raspberrypi:~/src/tensorflow $ ls -al /usr/bin/python3 lrwxrwxrwx 1 root root 9 Apr 5 2021 /usr/bin/python3 -> python3.9
Install the Tensorflow wheel
To install the wheel file you just built, you must first install the packages required to use TensorFlow.
$ sudo apt-get install -y libhdf5-dev libc-ares-dev libeigen3-dev gcc gfortran libgfortran5 \ libatlas3-base libatlas-base-dev libopenblas-dev libopenblas-base libblas-dev \ liblapack-dev cython3 libatlas-base-dev openmpi-bin libopenmpi-dev $ sudo pip3 install pip --upgrade $ sudo pip3 install keras_applications==1.0.8 --no-deps $ sudo pip3 install keras_preprocessing==1.1.0 --no-deps $ sudo pip3 install h5py==3.1.0 $ sudo pip3 install pybind11 $ pip3 install -U --user six wheel mock
It's time to install the python wheel
Finally, it's time to install the Python TensorFlow package.
pi@raspberrypi:~/src/tensorflow $ pip3 install /tmp/tensorflow_pkg/tensorflow-2.8.0-cp39-cp39-linux_aarch64.whl ...... Successfully installed absl-py-1.0.0 astunparse-1.6.3 cachetools-5.0.0 flatbuffers-20181003210633 gast-0.5.3 google-auth-2.6.0 google-auth-oauthlib-0.4.6 google-pasta-0.2.0 grpcio-1.44.0 importlib-metadata-4.11.1 keras-2.8.0 keras-preprocessing-1.1.2 libclang-13.0.0 markdown-3.3.6 opt-einsum-3.3.0 protobuf-3.19.4 pyasn1-0.4.8 pyasn1-modules-0.2.8 rsa-4.8 tensorboard-2.8.0 tensorboard-data-server-0.6.1 tensorboard-plugin-wit-1.8.1 tensorflow-2.8.0 tensorflow-io-gcs-filesystem-0.24.0 termcolor-1.1.0 tf-estimator-nightly-2.8.0.dev2021122109
Let's check if TensorFlow is installed properly. You should not test in the tensorflow source directory. Because the tensorflow directory exists in this directory, an error occurs in import tensorflow.
pi@raspberrypi:~/src/tensorflow $ cd .. pi@raspberrypi:~/src $ python3 Python 3.9.2 (default, Feb 28 2021, 17:03:44) [GCC 10.2.1 20210110] on linux Type "help", "copyright", "credits" or "license" for more information. >>> import tensorflow as tf >>> tf.__version__ '2.8.0'
Wrapping up
On Github, many developers have already created and provided a Python wheel for Raspberry Pi OS 64-bit. You can download the wheel they made and use it, but it's also a good job to build a TensorFlow package yourself that matches the Python version and OS version you want.
댓글 없음:
댓글 쓰기