skip to content W3C SVG

How I trained my Rangoli Art(GAN)

Publish date: Dec 7, 2020
Tags:

Materials I used :

  • Data-set comprising of rangoli images ( ~13k)
  • StyleGan2 - Tensorflow
  • A little patience
  • Google Colab (Tesla P100 ~24gb ram)

Firstly I thouht of using pre-made data-sets, but decided to build one. Making a data-set is painful process but rewarding. I used instagram and search engine scrappers to download rangoli images. Found many on web, really thankful to the Indian tradition and people.

real images

Image Preprocessing

Downloaded images, now what?

Ans: Trash Unwanted

Why?

Ans: Images downloaded may have lot of unrelated items in them.

My cleaning involved these following things:

  • removed duplicates

  • hand curated images

  • trashed video files, gif files came along

Image Size Conversion

It’s a rule that image size must be multiple of 2. (i.e, 64, 256, 512, 1024). So I Converted images to size 256*256 (keeping colab in mind).

Tool Used for conversion : Image magik (recommended)

Model training

Out there many forks are available for Stylegan2, good one is Peter Baylies fork.

I used this fork by woctezuma, which allows you to resume training from where you left. Converted data into tfrecord format (size matters in colab). You can see generated fake images for every tick (1tick = 5000 imgs). Each takes 15 min time for P100 gpu.

fake images

​

Calculate the fid metrics (takes about 15min) so that you know how model is performing and when to stop training.