Here's the dataset, model training, and output phases for a generative adversarial network I wrote that basically learned about...Me, and subsequently created a custom social media avatar.

I wrote the damn thing and it still couldn't figure it out. I'm too complex. My therapist was right.

  • 1
    So you want to generate an avatar?
  • 0
    @mr-user more or less. Basically a proof on concept idea I had over a weekend to see if a neural network could learn from a small dataset of pics of me, and produce a semi-coherent image, suitable for an avatar.


    But in the grand scheme of things (and hilariously): FAIL. :)
  • 1

    To be honest I haven't work with GAN before but you could try to fix your data.

    I figure you do data argumentation to produce more sample and some of the sample look totally unusable (in my option). In one photo your nose look totally out of places.

    If the GAN is like any other network you can check for bias and variance. What error percentage are you getting on training set and what data are you using in development set?
  • 0
    @mr-user You're right...data augmentation would help provide more samples, but I was testing a recent concept in which GANs can learn from a handful of samples. It's tricky for multimedia, but it provides decent, if not rock solid, results.

    My dataset was just a bunch of pictures of myself I curated mainly from Flickr, Facebook and Twitter over the years.
Add Comment