Skip to content

Latest commit

 

History

History
19 lines (15 loc) · 749 Bytes

todo.md

File metadata and controls

19 lines (15 loc) · 749 Bytes

TO DO

  • Plug VAE sampling to WGAN create_real_samples func
  • Use distribution of sim genomes in PCA as stop metric for WGAN

TO TEST

  • Try WGAN-GP
  • Like, all the other models mentioned in the manuscript
  • Split genome with no max size, play around with min size / min dist between hotspots
  • Sleep the VAEs to reduce latent space (!! watch out for mode searching !!)

WHAT I WOULD TRY IF STARTING OVER

  • Take ~1K top hotspots, with 10-100kb sparsity around them
  • Split with no maxsize
  • VAE 100x with sleeping, on 10?% chunks to optimise hyperparams (A LOT to try), test KLD to reduce mode searching?
  • WGAN-GP (with VAE sampling) on ~10 chunks to see if works + approx training resources needed
  • Scale up
  • Try multi chr?