XMAS Project — Part 1: Status Quo Getting into AI, ML, DNNs, …
I’m working now since 15yrs. Mainly in Tech (Engineering, DevOps, SRE), Product Management, (Technical) Sales Consultant and Advisory Board roles for Startups & Corporates.
The last years I’ve watched every thing around AI, ML and … from the sideline. You could see how large organisations have made it commodity. You have seen how many said they do it (and at the end just had a SQL).
Through my Advisory Board role for Product & Tech at yoona.ai, a small startup which is changing the world how fashion is designed, build from Designers for Designers — I took the chance to dive deeper into the topic and I have to say I’ve been blown away since then.
Besides the thousands buzzwords, the world of AI has improved so massively, that it is hard to find a spot where to start. For a tech-oriented person the space is like a big shop full with toys. You want to have it all! Now! All of it! :)
Tip 1: Think about the problem you want to solve. Yes I know sounds cheesy, but do you want to learn about Text Generation, Image Generation, Image Classification, Anomaly Detection, Text to Image, … ? Are you going to predict the price of a house, based on a photo? (Link) Do you want to detect which letter has been written? (Link) … towards data science is quite a nice side, where you can find a lot of articles around these topics.
If you don’t pick one, you will find so many different things and you can easily get lost. And never ever do one mistake, pick a technology and apply it to your problem. Know your problem and pick the right technology to solve it! “If I have a hammer, everything looks like a nail!”
yoona.ai is in the Computer Vision space. They are generating new designs for clothes, accessories, textiles, … with the help of Generative Adversarial Networks. Simples description of GANs? I put a lot of pictures into a machine, the machine learns how to recreate them and at the end is able to generate new images, which don’t exists. This found a lot of interests in the Anime-space. Have a look at This Anime Doesn’t Exist or This Person Doesn’t Exist
If you are interested how to do it, you will stumble about Progressive GANs, CycleGANs, … and all the versions of StyleGAN, StyleGAN2, StyleGAN3. Especially the StyleGANx sponsored and implemented by NVIDIA. (github)
Just watch out. NVIDIA has put all StyleGAN versions 2 and future under their non-commercial license. Means you can’t use it for commercial products. You can request a license from them, but they are going to tell you, that they are not offering it yet. I guess they just want to know, who is using it (and in worst case, sue or request a license fee later)
The network used for StyleGAN itself, can’t be licensed, so there are a few implementations out there, which are not using any of the original code (let’s hope). StyleGAN2 — a pytorch implementation (Link), LightweightGAN — a minimalised StyleGAN (Link) or my personal favorite GANsformer (Link)
Most important thing what I’ve learned? It is all about the data. The cleaner it is, the easier will it for you. (otherwise you will fail often, because your data is messed up)
If you don’t have an own dataset, look at platforms like Kaggle, where you can some to get started with.
Secondly you don’t have to train a model from scratch. Especially in the Image Classification and Object Detection space, there are several Models implemented and pre-trained.
I’ll show you in one of my next articles, how get started with it.
