Automatic generation of ambigrams

• I wonder if it’s possible to create an Anagram using the approach of Deep Learning.

• I want to try it with University of Tokyo 1S Information Alpha.

  • If I have time, I also want to try Ambigram of sound.

• For example, when generating human faces using GAN, if we feed both the original version and the upside-down version to the Discriminator every time, we can create an anagram (?) of human face photos.

• If we do the same thing with Conditional GAN, is it possible to generate characters?

• @nbaji9: An anagram that combines letters and pictures, like “dear / read”.

• #anagramart https://t.co/wmuFtsGaTV

  • It would be nice to have a model that can be used to generate this kind of artwork.

• Log 20220606

  • I think it will be difficult without changing the data, so let’s start by successfully training DCGAN with MNIST.
  • After that, I’ll try using notMNIST, which seems promising.
  • DCGAN
    • It works well when only rotated.
    • When anagrammed, it gets stuck in a bad state.
      • I think it’s because it hasn’t generalized well, so I’ll try increasing the Dropout.
      • What happens if it’s a simple one like 0, 1, or 2? (blu3mo)
        • 0 is successful.
        • 1 is successful once out of two attempts.
          • 
            • Unlike MNIST, it’s not always tilted to the right, which is evidence that it’s working well (blu3mo).
    • notMNIST + DCGAN
      • 
        • 1000 iterations
        • It’s okay.
      • 
        • 5000 iterations
        • Various styles are coming out, so it was worth using nonMNIST.
      • What if we do an anagram with this?
        • 
        • It’s impossible.
      • What could be the problem?
        • https://gangango.com/2018/11/16/post-322/
          • The Discriminator is too smart, and the gradients that the Generator can learn have disappeared.
            • That seems to be it (blu3mo).
            • Maybe we should make the Discriminator dumb.

• Log 20220523

  • For now, let’s train a GAN to generate “9” using MNIST.
    • 400
    • It worked ✅.
  • Then, let’s try feeding the output of the generator to the Discriminator with a flipped version.
    • We can easily do this using Data Augmentation with RandomFlip.
    • Add RandomFlip to the end of the Generator.
      • Huh, that’s strange.
        • 1500
        • Only one-directional “9” is coming out normally…?
        • I gave up on making it work with both directions and settled for only half of the “9” being generated with a 50% success rate.
          • No, that’s not true.
          • Even if that were the case, there should still be a flipped “9” in the output.
        • https://www.tensorflow.org/api_docs/python/tf/keras/layers/RandomFlip

        • During inference time, the output will be identical to input. Call the layer with training=True to flip the input.

          • That seems to be it (blu3mo).
      • I found another way.
        • 3500
        • Hmm…w
        • I appreciate the effort (blu3mo).
        • Let’s try with other numbers.
          • 5
            • 
      • Let’s try with DCGAN.
        • Normal 5:
          • Adjustments seem to be necessary.
        • Anagram 5:
          • It’s completely unsuccessful, haha.
    • Add RandomFlip to the beginning of the Discriminator.
      • This means that the flipped version of “9” will also be recognized as “9”.
      • 
        • Well, that’s what happens.- It doesn’t matter which one, just choose one.

• This won’t work (blu3mo)

• For now, let’s leave DCGAN and use a regular GAN to increase the dataset.

• LRDS https://paperswithcode.com/dataset/letter

  • Doesn’t seem suitable.
  • This dataset seems to be extracted features of character images, not actual character images.

• #might implement