- 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?
- 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