This was a project training a StyleGAN network to generate a potentially infinite series of parallel universe microfossils, based on the amazing collection of samples at Endless Forams
Data available here (open access, with bulk download of entire training set planned but not currently high priority) http://endlessforams.org/
Download resource pattern goes like this:
http://endlessforams.org/randomizer/download/[classification]/[total]?download=capsule.zip
Very convenient list of classifications and number of available samples here: http://endlessforams.org/summary
Most of this was based on the notebooks and training routines in the stylegan-art project.
It's now shifted to StyleGAN2 re-using the NVLabsTensorflow implementation.
Running the code requires a checkout of stylegan2 in one's PYTHONPATH.
The main addition is a bit of image processing, trying to extract just the foram shape from the original sources, which also include a lot of text metadata. This was a rough and ready approach involving Yen thresholding from scikit-image, falling back to Otsu thresholding if Yen didn't extract a region with the right area, then looking for the most square of the largest regions in the image.
This was the output from a test run on a Google Colab GPU up to the end of the free 12 hour runtime. Hopefully a longer run would produce much more compellingly detailed images of unreal forams.
- Do a much longer training run to improve image quality
- Generate a lot of parallel universe forams, hierarchically cluster them into an imaginary taxonomy, and generate names for the imaginary taxonmy. Then train a microfossil classifier on that data to identify imaginary forams