[FFmpeg-devel] [FFmpeg-cvslog] Adds ESPCN super resolution filter merged with SRCNN filter.

Jean-Baptiste Kempf jb at videolan.org
Mon Jul 2 23:35:43 EEST 2018


On Mon, 2 Jul 2018, at 22:28, Pedro Arthur wrote:
> > Sorry, you cannot dump 10kB of array and not explain where they come from.
> We have the raw weights, the floats array, and we have a serialized
> TensorFlow model.
> The raw weights are use by our own "native" implementation, the
> serialized TF model is used in case the user enables the compilation
> with TF library.
> The TF model was constructed and serialized by our student Sergey.
> What may be done, if it is your concern, is to build the TF model on
> the fly using the raw weights.

I understand that. But you should share this model under a compatible license.

> Other that that I don't have any other solution for the large size of
> the filters, except disabling the sr filter in your build.

I don't care about "my" build. I do care about licenses issues.
Large stack of array of numbers is a common way to obfuscate code and is usually removed by all Linux distributions.

Either they are easy to reproduce (ln() table, for example), or are based on a spec that requires those numbers to be bit-exact.

But other numbers that are the output of an undocumented process is usually disliked.
The (L)GPL and its interpretations speak about the source, which is the core of the creation.
(also why GPL is mis-fitted for images, for example)

> > 9479955 was bad already with 3 convolution kernels dumped with almost no explanation.
> > And then basically reverted most of those kernels in bdf1bbd.
> > And then adding other unexplained numbers in d8c0bbb0.
> As stated previously, the data are the weights, when we achieve better
> image quality through training we update the weights.
> On the math side, the optimization problem can have multiple  minimum
> points, thus two completely different kernels may achieve the same or
> close quality (minimum point).
> As the training may be stochastic it is almost sure each training
> result will generate different kernels.

Sorry, but you must provide the way to reproduce those kernels.

Other people may have different datasets, you might disappear, someone might come and take maintenance of this code, a new type of animation might start be popular, etc...

(and you need to fix the license to LGPLv3)

-- 
Jean-Baptiste Kempf -  President
+33 672 704 734


More information about the ffmpeg-devel mailing list