Hi author here. Training on ADE20K was performed on a V100x8 instance on AWS, which costs $24.48/hour (or did last I checked). Training time was 13h 12m 13s, so total cost was $320-ish.
GANs are expensive to train from scratch. Fine-tuning on the Bob Ross image corpus was comparatively cheap.
Yeah I've been noticing that. If I were to do a V2 of this project I would definitely revisit the segmentation maps to see if I can sharpen the edges of the trees. If that didn't work I'd probably try adding a transition region, like you suggest here.
4 comments
[ 2.7 ms ] story [ 17.1 ms ] threadit mentioned 8x GPUs in the blog to run a GAUGAN for this. What would that cost and how long did it take to run?
GANs are expensive to train from scratch. Fine-tuning on the Bob Ross image corpus was comparatively cheap.