1
$\begingroup$

I'm trying to generate images at minimum of size 128 x 128 with a Generative Adversarial Network. I already tried a SAGAN pytorch implementation, but I'm not very happy with results. The images look cool but and I see some correct shape but without explanation you wouldn't know what the images are about. I have a dataset of 4000 images. Lightness, colors and shapes vary a lot, but they are similar in style and on what they portray.

  • With a Google Cloud V100 GPU the GAN would run a week to two with default parameters. Does this sound realistic time for this kind of dataset? It's definitely not feasible for me.
  • Is 4000 images enough to train a GAN from scratch?
  • Is there any implementation with pytorch/keras that would be good to get nice results with?
$\endgroup$
1
$\begingroup$

With a Google Cloud V100 GPU the GAN would run a week to two with default parameters. Does this sound realistic time for this kind of dataset? It's definitely not feasible for me.

Yes, V100s are quite beefy. You shouldn't even need a week. Obviously this is based on my experience with various problems, rather than a concrete calculation.

Is 4000 images enough to train a GAN from scratch?

For the size you want to generate, it is still on the edge of what would constitute a decent training set. You will get some results(depending on architecture) but will probably want to grab more data if at all possible.

Is there any implementation with pytorch/keras that would be good to get nice results with?

I would check out this link: https://github.com/eriklindernoren/Keras-GAN. It has some nice implementations in Keras. As far as the particular GAN to use, I would start out with a vanilla GAN that fits your purposes and focus on toying with hyperparameters, and if that doesn't work, look into one of the other variations that correlate well with your problem.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.