Skip to content

Latest commit

 

History

History
26 lines (20 loc) · 958 Bytes

File metadata and controls

26 lines (20 loc) · 958 Bytes
  • Changed lr and wd to match paper (1e-4) but this gave bad results
  • Why is hidden dim 250 instead of 256?
  • Is batch size 32 always correct?
  • Alpha gradients are going to zero
  • Loss just stalls and doesn't go anywhere

Try to set threshold/sigma so you have about ~25% of edges at init (or use elbow method) sigma=10

Look at HEIST for Cell Clustering

Look at perturbseq data

For unsupervised cell-level embedding

  • Use contrastive learning based on spatial + cell type loss
  • Use autoencoder with reconstructive loss

Should we try looking at PC.T matrix for point-cloud level embedding?

TODO

  • change diffusion to timepoint
  • sections: take 3 different tissues from the brain
    • create different point clouds for each section
    • we get around ~70 point clouds across whole SEA dataset
    • want to train node embedding autoencoder over all these point clouds.
    • each file in Hiren upload is a separate section, need to preprocess each