本文基于官方文档Installing TensorFlow For Jetson Platform :: NVIDIA Deep Learning Frameworks Documentation
https://docs.nvidia.com/deeplearning/frameworks/install-tf-jetson-platform/index.html编译安装 TensorFlow 1 （1.15.2）+ GPU 方法另请参考AArch64编译安装特定版本TensorFlow及Bazel_yihuajack的博客-CSDN博客ALBERT 的 requirements.txt 要求 tensorflow==1.15.2，而这显然是不能通过 Ubuntu apt 安装的，也根本没有为 aarch64 架构编译好的 binary，所以采用编译安装。首先在 tensorflow 的 GitHub Release 中找到 1.15.2 版本，下载 Assets 中的 Source code (tar.gz)，然后解压。这时如果直接执行 sudo ./configure 会报错找不到 bazel。参考最新版本的 bazelbuild https://blog.csdn.net/yihuajack/article/details/121045347?spm=1001.2014.3001.5501
编译安装 TensorFlow 2 （2.7）+ GPU 方法另请参考
sudo apt update
sudo apt install libhdf5-serial-dev hdf5-tools libhdf5-dev zlib1g-dev zip libjpeg8-dev liblapack-dev libblas-dev gfortran
sudo apt install python3-pip
sudo pip3 install -U pip testresources setuptools==49.6.0
sudo pip3 install -U --no-deps numpy==1.19.4 future==0.18.2 mock==3.0.5 keras_preprocessing==1.1.2 keras_applications==1.0.8 gast==0.4.0 protobuf pybind11 cython pkgconfig
sudo env H5PY_SETUP_REQUIRES=0 pip3 install -U h5py==3.1.0
sudo pip3 install --pre --extra-index-url https://developer.download.nvidia.com/compute/redist/jp/v46 tensorflow
sudo apt install virtualenv
python3 -m virtualenv -p python3 <chosen_venv_name>
source <chosen_venv_name>/bin/activate
pip3 install -U numpy grpcio absl-py py-cpuinfo psutil portpicker six mock requests gast h5py astor termcolor protobuf keras-applications keras-preprocessing wrapt google-pasta setuptools testresources对于 tensorflow 2.6.0+nv21.9，要求的 dependency 版本为：
absl-py==0.12.0
gast==0.4.0
six~=1.15.0
wrapt~=1.12.1
如果?import tensorflow 报错
[1]? ? 8707 illegal hardware instruction (core dumped)? python3
sudo pip3 install numpy==1.19.4或
sudo pip3 install --no-binary :all: numpy==1.19.5安装 Anaconda 3
wget https://repo.anaconda.com/archive/Anaconda3-2021.05-Linux-aarch64.sh
sudo bash ~/Anaconda3-2021.05-Linux-aarch64.sh但是执行 conda 命令（conda info，conda list）会报错
[1]? ? 11034 illegal hardware instruction (core dumped)? conda info
在虚拟环境中安装 h5py 时报错
Using cached h5py-3.1.0.tar.gz (371?kB)
Installing build dependencies ... done
Getting requirements to build wheel ... done
Installing backend dependencies ... error
ERROR: Command errored with exit status 1:
comand: /home/yihua/tf1/bin/python /home/yihua/tf1/lib/python3.6/site-packages/pip install --ignore-installed --no-user -prefix /tmp/pip-build-env-dpuj0w44/normal --no-warn-script-location --no-binary :none: --only-binary :none: -i https://pypi.org/simple -- pkgconfig 'numpy==1.14.5; python_version == "3.7"'?'numpy==1.17.5; python_version == "3.8"'?'numpy==1.19.3; python_version >= "3.9"' 'Cython>=0.29; python_version < "3.8"'?'numpy==1.12; python_version == "3.6"'?'Cython>=0.29.14; python_version >=?"3.8"'
cwd: None
Complete output (2509 lines):
Ignoring numpy: markers 'python_version == "3.7"' don't match your environment
Ignoring numpy: markers 'python_version == "3.8"' don't match your environment
Ignoring numpy: markers 'python_version >= "3.9"' don't match your environment
Ignoring Cython: markers 'python_version == "3.8"' don't match your environment
Collecting pkgconfig
Using cached pkgconfig-1.5.5-py3-none-any.whl (6.7 kB)
Collecting Cython>=0.29
Using cached Cython-0.29.24-cp36-cp36m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl (1.9 MB)
Collecting numpy==1.12
Using cached numpy-1.12.0.zip (4.8 MB)
Preparing metadata (setup.py): started
Preparing metadata (setup.py): finished with status ‘done’
Building wheels for collected packages: numpy
Building wheel for numpy (setup.py): started
Building wheel for numpy (setup.py): still running...
Building wheel for numpy (setup.py): still running...
Building wheel for numpy (setup.py): finished with status 'error'
ERROR: Command errored out with exit status 1:...
ERROR: Command errored out with exit status 1
...
WARNING: Discarding ...
Using cached h5py-3.0.0.tar.gz (370 kB)
Installing build dependencies ... done
Getting requirements to build wheel ... done
Installing backend?dependencies ... error
ERROR: Command errored out with exit status 1:
comand: /home/yihua/tf1/bin/python /home/yihua/tf1/lib/python3.6/site-packages/pip install --ignore-installed --no-user -prefix /tmp/pip-build-env-dpuj0w44/normal --no-warn-script-location --no-binary :none: --only-binary :none: -i https://pypi.org/simple -- pkgconfig 'numpy==1.14.5; python_version == "3.7"'?'numpy==1.17.5; python_version == "3.8"'?'numpy==1.19.3; python_version >= "3.9"' 'Cython>=0.29; python_version < "3.8"'?'numpy==1.12; python_version == "3.6"'?'Cython>=0.29.14; python_version >=?"3.8"'
cwd: None
Complete output (2509 lines):
Ignoring numpy: markers 'python_version == "3.7"' don't match your environment
Ignoring numpy: markers 'python_version == "3.8"' don't match your environment
Ignoring numpy: markers 'python_version >= "3.9"' don't match your environment
Ignoring Cython: markers 'python_version == "3.8"' don't match your environment
Collecting pkgconfig
Using cached pkgconfig-1.5.5-py3-none-any.whl (6.7 kB)
Collecting Cython>=0.29
Using cached Cython-0.29.24-cp36-cp36m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl (1.9 MB)
Collecting numpy==1.12
Using cached numpy-1.12.0.zip (4.8 MB)
Preparing metadata (setup.py): started
Preparing metadata (setup.py): finished with status ‘done’
Building wheels for collected packages: numpy
Building wheel for numpy (setup.py): started
Building wheel for numpy (setup.py): still running...
Building wheel for numpy (setup.py): still running...
Building wheel for numpy (setup.py): finished with status 'error'
ERROR: Command errored out with exit status 1:...
虚拟环境中不能使用 sudo，env 命令也会失效，在 build wheel 的时候会提示 ModuleNotFoundError: No Module Named 'Cython'，即使你已经安装了 Cython。如果在主环境中不加 env 就会产生同意的错误。参考了?Cannot install h5py on Jetson Xavier NX - Jetson & Embedded Systems / Jetson Xavier NX - NVIDIA Developer Forums、Can't install h5py on Jetpack 4.3 - Jetson & Embedded Systems / Jetson TX2 - NVIDIA Developer Forums、cython: compilation error while building h5py (solved by workaround) · Issue #1533 · h5py/h5py (github.com)
都不会解决该问题。最后想到一个办法，因为在主环境中是可以正常用 sudo env 编译安装 h5py的，所以参考?How to use Python's pip to download and keep the zipped files for a package? - Stack Overflow 在主环境中执行
sudo env H5PY_SETUP_REQUIRES=0 pip3 download -d /ssddata/h5py h5py==3.1.0会在 /ssddata/h5py 目录下下载生成 cached_property-1.5.2-py2.py3-none-any.whl、h5py-3.1.0.tar.gz、numpy-1.19.5-cp36-cp36m-manylinux2014_aarch64.whl 三个文件。h5py-3.1.0是最后一个支持 Python 3.6 的版本。然后执行
sudo env H5PY_SETUP_REQUIRES=0 pip3 wheel -w /ssddata/h5py h5py=3.1.0会在 /ssddata/h5py 目录下生成 h5py-3.1.0-cp36-cp36m-linux_aarch64.whl。然后切换到虚拟环境中执行
pip3 install /ssddata/h5py/h5py-3.1.0-cp36-cp36m-linux_aarch64.whl
pip3 install -U numpy grpcio absl-py py-cpuinfo psutil portpicker six mock requests gast astor termcolor protobuf keras-applications keras-preprocessing wrapt google-pasta setuptools testresources
pip3 install --pre --extra-index-url https://developer.download.nvidia.com/compute/redist/jp/v46 'tensorflow<2'可以自行到 extra-index-url 的 URL 地址查看包含哪些版本的 wheel。在安装 tensorflow<2 时它还会自动下载编译安装 numpy-1.18.5 并试图编译 h5py 并报错
running build_ext
ERROR: Failed building wheel for h5py
Running setup.py clean for h5py
然后生成 numpy 的 wheel，提示 numpy 构建成功、h5py 构建失败，然后安装一系列包，其中要卸载 numpy-1.19.5、h5py-3.1.0（Running setup.py install for h5py ... done）、gast-0.5.2。最后安装的包有：
astunparse-1.6.3 dataclasses-0.8 gast-0.3.3 h5py-2.10.0
importlib-metadata-4.8.2 markdown-3.3.4 numpy-1.18.5 opt-einsum-3.3.0
tensorboard-1.15.0 tensorflow-1.15.5+nv21.9 tensorflow-estimator-1.15.1
typing-extensions-3.10.0.2 werkzeug-2.0.2 zipp-3.6.0