0

We all know that GAN can be used to augment and expand our dataset Can a GAN be used for data augmentation?. But my question is, is it effective and fast enough?

For example I have done experiment using this repo to train PGAN (Progressive Growing of GAN - state-of-the-art result).

  1. The training using large dataset such as CelebA and Cifar-10 took days using multi-GPU (Gtx1080 and 940) to produce desirable result (can't be distinguished as fake), and using my own dataset consists of 75.000 images (128x128) split into 3 classes also took days.
  2. After hours or days of training the model only generates few (less than 100) images.

Meanwhile I can do manual annotation and get more images during those days of training. So, is GAN effective and fast enough to replace image augmentation or manual annotation method? Or is there any a more effective GAN model for expanding dataset than the one that I've tried?

gameon67
  • 161
  • 1
  • 5
  • "the model only generates few (less than 100) images" -- because you only chose to sample 100 images? i can't imagine that PGAN sampling would be that slow. – shimao Dec 03 '19 at 06:19
  • @shimao yes, my point is, generating a few images took me days, I can't imagine how much computational resource that I need to generate thousand of images – gameon67 Dec 03 '19 at 06:22
  • something sounds horribly wrong with your code then. sampling shouldn't take more than a fraction of a second per image. – shimao Dec 03 '19 at 06:26
  • @shimao sorry I think I have to edit my post. I mean to generate a good image that can't be distinguished as fake by human eyes took days, not a sample image. If you read the paper, I think they trained for several days and they used a super high-end GPU like Tesla P100 – gameon67 Dec 03 '19 at 06:29
  • yes, but what i mean is, after you've spent a few days training, you should be able to easily sample tens of thousands of images, not just 100. – shimao Dec 03 '19 at 06:32
  • @shimao I see, I think I miss-understood the training and eval concept of GAN. I thought I could only generate images through training. Thanks for the clarification, I'll be back after trying to eval my checkpoint model – gameon67 Dec 03 '19 at 06:37

0 Answers0