Deeper DCGAN with AE stabilization
Parallel training of generative adversarial network as an autoencoder with dedicated losses for each stage. Generator class has conditional .forward() method for enhanced ergonomics. Autoencoding pass seems to avoid mode collapse and recover faster if Generator is not doing well.
Has been used successfully with as few as ~200 images in the source folder.
Builds on the DCGAN PyTorch demo. This one generates images upto 1024x1024 so it can use a lot of VRAM.
Should work when the "dataroot" is configured ImageNet style: ".../a_dir_of_images/what_would_be_a_label" or ".../cat_pics/cute_cats/cat_001.jpg"
Have fun!
Examples of generated drawings here
A description of the project for which this was developed can be found here.