Generating 28x28 Twitch emotes with a Wasserstein GAN!
We used Colab to train the model, using their free K80 gpu. #ThanksGoogle
Next steps might be to initialize with Inception weights or even augment the data with ImageNet samples to help pull it toward more meaningful-looking objects.
It's already learned what edges are and how to deal with transparency.
It seems to be resolving things into objects, ish. An apple core?
Not much has improved; this seems converged.
A firefox?
A pokemon?
The training stopped in the middle of Epoch 42, due to Colab timing out the instance.
- downloader.go downloads all Twitch emotes in parallel. It's quite fast.
- png2hdf5.ipynb transforms the ones we want (28x28 RGBA, which is the most common dimensions) into an HDF5 file.
- dataset-first-50k is the first 50,000 28x28 RGBA Twitch emotes, which is what we're training on.
- gan.ipynb contains the actual training code.