Improve GAN labels

StyleGAN2 is now live!

This release makes some significant improvements to the new wine bottles including:

• Using new Stylegan2 images (improved from Stylegan1)
• Warping label perspective to better fit on the bottle template
• Using both light and dark bottles depending on the specific category of wine

Descriptions trained from GPT2-XL 1.5b

GPT2-XL is now live!

This release sees the result of migrating from a standard LSTM network to something much more coherent. While the name is still generated from a simple LSTM network (because I want the words to be completely original and made up), the category, origin, and descriptions are all trained on the 130,000 sample from www.wine.com I scraped and then primed with the LSTM name. Training was accomplished via DeepSpeed on an azure VM with a P100 for about 12 hours of fine-tuning.

fix: generator crashing

The release fixes the issue with the generator sometimes not producing an output and crashing

Generate GPT2_XL Samples

Script generate.py is now fully functional for generating samples from the pretrained deepspeed GPT2_XL model. Still have slight issues with PyTorch on certain samples but a simple try/except block catches them and just moves on for now.

initial release

A checkpoint of all the notebooks and code up to this point