Creating a StyleGAN2-ADA Custom Dataset with Google Colab

After experimenting with simple style transfers using two images, and projecting from popular models like ffhq, it was finally time to train my first custom dataset. Having very little idea of what to do, I decided to scrape some images from instagram and pinterest.

What does "done" look like?

After 168 total ticks, I decided to stop training the model on this first dataset. If it had kept training, the results would probably start improving more and more.

In fact after just 9 ticks the generated fakes were already looking quite impressive.

The difference between Colab free tier and Colab Pro

After ~24 hours of training, my Colab free tier ran out, and I was faced with either switching to a different account, waiting for the cap to be reset, or signing up for Colab Pro. For just $10 it seemed worth trying out.

Colab Free (K80)

The Colab Free tier allocates you with a K80 (a ~$300 card) Using this to train on my dataset of 440 images I got ~17 ticks per 9-12 hour session.

Colab Pro (P100 / V100)

On Colab Pro you get either a P100 (~$2,000) or V100 (~$7,000). So far I've been lucky and got the V100 more than half the time, which translates to ~4x the performance of the K80. So instead of ~17 tickets overnight I get ~80.


The results above are interesting already, but the real fun comes with exploring various checkpoints with something like GANSpace