Training a GAN on the EMNIST alphanumeric handwriting dataset, performing inversion, and exporting to ONNX. Network and training code copied almost verbatim from Diego Gomez's great PyTorch implementation.
I've written about this project in more detail here.
EMNIST from PyTorch is mirrored and rotated 90 degrees, so I fix the data with some transforms.
Nothing new to see here.
Gradient-descent search through the generator's inputs in search of a good example of each character.
Worked just as advertised.