Flux.jl: Flux and Images

[MikeInnes]

It's unhelpful that ML kernels require the WHCN or CWHN, neither of which is compatible with Julia's more standard CHWN. (Aside from the fact that thinking about this stuff is hard work to begin with.)

We should find a way to have (batches of) images that store metadata about their axes, either via functionality from Images.jl, AxisArrays, or something else. If we hook this up to conv we can aim to always do the right thing with regard to necessary permutations and choice of algorithm.

[datnamer]

@evizero has a bunch of julia ml and Julia image experience so may be interested in this.

[Evizero]

Thanks for the ping. I think this issue was actually the result of a slack discussion Mike and I had the other day.

concerning Image axis: AFAIK as of 0.6 Images.jl implicitly assumes CHW everywhere

[ChuckNielsen]

https://faceswap.dev// looks like the A B image pulls with C is whats needed

It's intended to take FACE A (folder of 1 person face) and swap with B
it can also find other faces in a crawl that allows A to go to other places it visually looks correct.

Also, this new Colab release for BERT called Tapas uses 4 ML Libs to judge the NLP. The best one is based on non-checked user input from what I can tell (correct me if I am wrong) but the wiki lib? BS Sells in conversational length, not the quality of info, potentially just extensions of nothing using user input as a worth parameter to pick the next statement.

I would like to see this get put into AR in the long run with fun tools.

This looks like a 3 point access using an image tool already on ML.

Thoughts?