Skip to content

Faster OTD computation (up to 10x).

Compare
Choose a tag to compare
@saibalmars saibalmars released this 09 Jun 17:22
· 45 commits to master since this release
db43d65
  • Switch OTD computation from cvxpy to ot.emd2 since it's way faster (up to 10x) for OTD cases within neighbor counts 2000 x 2000.
  • Add a new computation method OTDSinkhornMix that uses ot.emd2 to compute OTD for smaller neighbors (less than _OTDSinkhorn_threshold=2000 in default setting). This is because Sinkhorn works faster for larger cases, for smaller cases, the iterating process makes it slower than directly compute the ot.emd2 written in C.
  • Update the default nbr_topk to be 3000 because of the faster computations.
  • Suppressed edge weight to zero message from warning to trace since the Ricci flow "converges" faster with exact Wasserstein distance.