HydroGAN

TODO

• Data

  • Change the subcube sampling procedure to let non-128 multiple cubes to be selected as well
    • Current way = (2048/128 - 2)^3 = 2744 different subcubes only
    • New way = (2048 - 128*2)^3 = 5.7 billion different subcube combinations
  • Change the subcube sampling procedure
    • Currently, the subsampling subsamples all the training samples and stores them.
      • in the self.samples under HydrogenDataset2
    • Change this to quickly sample when the get_samples() function is called.
      • incorporate get_samples into "__getitem__"
      • get_samples should open f itself without reading the whole 2048 cube wholly, sample using the coordinates and f.close()
  • Calculate the total number of different subcubes we can sample from

• Cube transformations/Scaling

• VAE

  • Investigate why VAE is not producing any hydrogen masses? (Probably an issue with the decoder part of the VAE). In the decode() of the VAE class, create multiple checkpoints: sum all the values in the out variable and plot the evolution of the sum. Compare with the sum of the input subcubes.

    • No Transformation Output
      • First Convolution Output
      • First ReLU Output
      • First MaxUnpool Output
      • Second Convolution Output
      • Second ReLU Output
      • Second MaxUnpool Output
      • Third Convolution Output
      • Third ReLU Output

• GAN

  • Plotting the evolution of the Generator(noise) subcubes during training
  • Check whether the inputs are in the range of (-1,1) -> https://github.com/soumith/ganhacks
  • We don't have any downsampling/upsampling components in both the discriminator and the generator, maybe try those (Downsampling : Average Pooling, Upsampling: PixelShuffle) -> https://github.com/soumith/ganhacks
  • Try Label smoothing? -> https://github.com/soumith/ganhacks
  • Add script to check norms of gradients - if they are over 100 things are screwing up - https://github.com/soumith/ganhacks
  • while lossD > A: train D, while lossG > B: train G - https://github.com/soumith/ganhacks
  • MMD-GAN
    • Does a high number of 0's hinder the training procedure?
    • try square root instead of log for transformation
    • log(a + 1) doesnt transform the distribution to normal!
      • how to overcome this?
      • Can MMD replicate this non-normal distribution?
    • Add f_dec_X_D & f_dec_Y_D to see whether the AE is working correctly
      • Or maybe try to train with L1 loss vs. L2 (default)
    • Add 3D plots!

• Hybrid Models

  • https://github.com/soumith/ganhacks -> if you cant use DCGANs and no model is stable, use a hybrid model : KL + GAN or VAE + GAN
  • How to use VAE + GAN?

• Validation

  • 1 PDF
  • Log Histograms (compare the real distribution vs. the generated ones)
    • VAE: the output of the decoder
    • GAN: The output of Generator(noise)
  • Power spectrum
  • 3D Plot comparisons

Data

sample data is in data folder where a .h5 file is put. sample_32.h5 is 32 of randomly sampled cubes with dimensions ?x?x?