Page 1 of 1

cross-compiling ffmpeg with --enable-libtensorflow ?

Posted: Tue Jul 03, 2018 2:02 pm
by hydra3333
I noticed that ffmpeg has an interesting feature
--enable-libtensorflow enable TensorFlow as a DNN module backend for DNN based filters like sr [no]
( with an associated ffmpeg check "enabled libtensorflow && require libtensorflow tensorflow/c/c_api.h TF_Version -ltensorflow" )
which apparently enables "AI" type filters to be used eg per libavfilter/vf_sr.c
"srcnn", "Super-Resolution Convolutional Neural Network model
"espcn", "Efficient Sub-Pixel Convolutional Neural Network model"
The github source https://github.com/tensorflow doesn't really provide instructions on how to cross-compile libtensorflow as a Windows library to use in building a static Windows x64 ffmpeg eg along the lines of dependencies built in rdp's cross-compile script https://github.com/rdp/ffmpeg-windows-build-helpers or in deadsix27's cross-compile tool https://github.com/DeadSix27/python_cro ... ile_script or indeed in Zeranoe's builds.

Thoughts and hints on how to cross-compile a static x64 ffmpeg with --enable-libtensorflow (cpu based and maybe even optionally an opencl based tensorflow if at all possible) would be much appreciated.

Zeranoe, I wonder if it would it be possible for you to foresee a time when --enable-libtensorflow may be enabled in your builds ?

Re: cross-compiling ffmpeg with --enable-libtensorflow ?

Posted: Wed Oct 31, 2018 12:01 pm
by cepesh
I would also like to join hydra3333 in this request. Thanks

Re: cross-compiling ffmpeg with --enable-libtensorflow ?

Posted: Thu Feb 28, 2019 4:41 am
by cjplay
Agree. Super Resolution scale to 3x or 4x, then downscale to make an SD->HD conversion. It should be sharper, but have no platform to test it.