Tensorflow-cpu_docker-builder

Compilation of Tensorflow-cpu from source via docker, customized for old processors or those that require specific flags.

Prerequisites

• Docker
• Git
• Python >= 3.10
• At least 16 GB of RAM
• At least 8 cpu-cores
• Linux environment

Get started

From Linux terminal - load repository:

git clone https://github.com/fab2112/Tensorflow-cpu_Docker-builder.git

cd Tensorflow-cpu_Docker-builder

Build docker image:

docker build --tag tf-builder-from-source .

Run docker container:

• Define the Python version (--env TF_PYTHON_VERSION=3.xx) for compilation.
• Defines the number of cpus (--cpus) according to machine resources.
• This process can take several hours depending on the number of cpus.

docker run --rm -it \
    --name tf-builder-from-source \
    --cpus=4 \
    --volume $(pwd)/tf_wheel:/tf_wheel:rw \
    --env TF_PYTHON_VERSION=3.11 \
    tf-builder-from-source

• After compilation, the .whl file will be available in the /tf_wheel directory.
• Install the compiled .whl file in the python environment via pip.

# Example
pip install ./tf_wheel/tensorflow_cpu-2.18.0-cp311-cp311-linux_x86_64.whl

Native configurations

• Configuration parameters during docker runtime
• Check your CPU "Optimization Flags"

Parameters	Settings
Python Location	/usr/bin/python3
Python Library	/usr/lib/python3/dist-packages
Tensorflow with ROCm	N
Tensorflow with CUDA	N
CLang as Compiler	Y
Optimization Flags	-march=native
Android Builds	N

Check Optimization Flags via terminal

grep flags -m1 /proc/cpuinfo | cut -d ":" -f 2 | tr '[:upper:]' '[:lower:]' | { read FLAGS; OPT="-march=native"; for flag in $FLAGS; do case "$flag" in "sse4_1" | "sse4_2" | "ssse3" | "fma" | "cx16" | "popcnt" | "avx" | "avx2") OPT+=" -m$flag";; esac; done; MODOPT=${OPT//_/\.}; echo "$MODOPT"; }

Ecosystem

Componente	Versão
Docker Engine	20.10.0+

• Docker image

App	Docker Tag
Ubuntu	ubuntu:22.04

License

MIT