Ludwig: Feature request: GAN support

[bitcoinmeetups]

Hello,

Does Ludwig support GAN's?

How about a feature where the user inputs a list of images and receives as output a bunch of cool GAN generated images?

BM

[w4nderlust]

Well at prediction time you don't need to provide anything, you just sample from noise and you produce images from there :)
Anyway, yes it's on out list, there needs to be a little piece before it, the image output feature so that first Ludwig can output images, then we can add a loss that is a GAN loss which contains also a discriminator.
More thought needs to be put into it, but it's definitely something that we eventually want to support.

[sk1210]

Hi @w4nderlust , I would like to work on adding this feature, Can you guide me in that, like how to start or any reference.

[bitcoinmeetups]

Awesome @w4nderlust!

Following this!

[w4nderlust]

@sk1210 the first step would be to implement an image output feature, with a way to produce an image (deconvolution for instance), a way to calculate loss and a way to compute measures.
I would use the cetegory_feature.py module as a template to copy from.

[EdwardFKit]

Has there been progress on this - this would be of great use!

[w4nderlust]

@EdwardFKit This is currently not being worked on, after the release of v0.3 (a month or so) we will reassess all users requests to figure out priorities for v0.4, but we did not forget about this, it's just that with the resources at our disposal, it has not been prioritized yet.

[sk1210]

@w4nderlust, @EdwardFKit I have now started working on this, I will keep you updated.

[EdwardFKit]

Great – please do – FYI what we are working on is not image / gradient based – it is more related to NLP / Keyword / Phrase / special purpose text generation – basically to generate automated Keyword software test cases for software systems -

[w4nderlust]

@w4nderlust, @EdwardFKit I have now started working on this, I will keep you updated.

@sk1210 that sound great, thank you for being willing to help, that is hreatly appreciated.
Assing GANs to Ludwig is non trivial at the moment because of its architectures. Some of the changes we are making for v0.3 will definitely make it easier.
Because it's non trivial I suggest you to fisrt study in detail the codebase, the best way being training a model by going through it with a debugger, line by line.
After you have done that, you can come un with a design proposal and we can discuss together about it. For it to be something I woyld want to include in the codebase, it will need to be a generic solution. So I encourage you to get in touch before steaming ahead and implementing things to avoid the risk you'll come up with a solution I would not want to include.
Does this make sense to you?