BASTARD GAN PUNKS V2 is a very interesting NFT collection. Beyond being historically significant tasty glitch art, for the purposes of this exercise the collection is an excellent test subject with which to stress test the hypermap tech we are developing at Sparklewerk. As such, herein the collection will be thoroughly explored via hypermap techniques.
GAN art, by its nature, is much more complex than simple generative PFP collections such as CryptoPunks. As such it is more difficult for machine learning algorithms to recognize patterns within the former. Additionally, there is more going on in BASTARD GAN PUNKS V2 than simply letting a GAN do its thing. For a primer on the collection, see Mooncat888's Introduction to Bastard Gan Punk V2 NFT art collection.
All the hypermaps on this page were generated from open source code. Jump to the bottom of this page if you want to go directly to hacking on that code running live — for free — on Google's Colab.
Initial data visualization
The following is the sort of quick and easy plot one uses early in data exploration. If such a plot is "clumpy" or "stripy" then there may well be structure that the ML algorithm detected (or the instrument is making things up; I'm looking at you, t-SNE) so keep exploring with the more computationally expensive, finer tools. In this case the algorithm is UMAP and the data is low resolution.
Great. The above shows that even at low resolution images, UMAP can see some structure in the collection's data. Next let us take a look at those NFTs on the map, and crank up the resolution of the data fed to UMAP.
The following image is (4096, 4096) pixels, the maximum size that Twitter allows. The NFTs have been reduced to (96, 96). That still provides for an entertaining resolution if you zoom in on the image. Note: only the calm bgans are plotted.
3D Orbiting PCA
Orbiting 3D hypermaps can provide a good quick overview of an NFT collection. For example, in the following hypermap all of the calm @bganpunks are represented. The algorithm is PCA, and the NFTs are low resolution.
All of the above hypermap images were generated by the Jupyter notebook, hypermap_bganpunks.ipynb, which is part of the open source (MIT licensed) repository, hypermap. The code can be run for free on Google Colab with an optional free GPU.
The hypermaps on this page are simply static images. At the end of the above notebook there is an interactive hypermap explorer. (Note: that code is a bit ornery at this point. It is very cutting edge stuff, that is: a bit buggy.) Those who actual read what is written there will be well rewarded and will understand what the next generation of hypermap technology will look like soon.