WRITEUPA couple of good ways to deal with limited datasets....

So you have only got a little bit of data... inside the 100s of images. Well, I had that problem too... here's how I've managed to work around it.

  1. I duplicated the images that I wanted my GANs to reproduce. This increases the likelihood that your GANs converge towards generating those images (but that's kind of the point anyway.)

  2. I used DiffAugment (https://github.com/mit-han-lab/data-efficient-gans), specifically the transform & cutout policies.

  3. Use the Pytorch torchvision image transforms liberally. I've found that randomly applying greyscale, and horizontal flips work well. I'm sure that colour jitter is also good, but it doesn't work in my use case(s).

Covers: implementation of Limited Datasets
Estimated time needed: 20 minutes
Questions this item adddesses:
  • What happens when I don't have enough data?

Quick tips to tune your GANs

Ashley Beattie,Total time needed: ~20 minutes
Learning Objectives
Help you get started in developing and tuning GANs which generate high-resolution artwork
Potential Use Cases
image generation of high-resolution artwork
Target Audience
INTERMEDIATEPython developers with some experience with Deep Neural Nets and want to try their hands at generating artwork
Go through the following annotated items in order:
WRITEUP 1. GAN - Ways to improve GAN performance
  • What are the common ways to tune GANs?
10 minutes
WRITEUP 2. A couple of good ways to deal with limited datasets....
  • What happens when I don't have enough data?
20 minutes
REPO 3. Use DeepSpeed
  • What are the ways that I can parallelize and speed up my GANs' training?
10 minutes
WRITEUP 4. Use Dilations on your Discriminator to increase the receptive field size of your convolutional filters.
  • What does increasing the dilation of your convolutional filters do?
10 minutes

Concepts Convered