Anupam comments

Results 33 comments of


                                            Anupam

Enable dice metric for UMAP and Pairwise Distances

Yup @divyegala, closing now

Spectral Embedding

Thank you for the review @jcrist! I have addressed your comments.

Spectral Embedding

Thank you for another review @jcrist! I have addressed your comments.

cupyx.scipy.sparse.linalg.eigsh not returning correct eigenvalues

We first came across this issue when trying to implement the Lanczos Solver from cupy in rapids/raft. Specifically, using this code when searching for < 258 eigenvalues it works fine....

cupyx.scipy.sparse.linalg.eigsh not returning correct eigenvalues

More observations: The orthogonality of lanczos vectors seems to deteriorate quickly. I measured this by doing `np.dot(u, V[i - 1])` Potential mitigation tactics: 1. Reduce `ncv` parameter 2. Add additional...

cupyx.scipy.sparse.linalg.eigsh not returning correct eigenvalues

Hi, I have found the root cause problem and solution here. https://github.com/cupy/cupy/blob/d717917e1d72c4318fcdcb73b4c9b29fbda176ff/cupyx/scipy/sparse/linalg/_eigen.py#L293-L296 Lanczos vector `u` may contain values that are very small

Spectral Embedding API

Moving Lanczos solver api section here #6034

Enable dice metric for UMAP and pairwise_distances

Blocked on constraints around cuvs binary size.

Enable dice metric for UMAP and pairwise_distances

Currently this is a P2+ priority but basically, we need to refactor the cuvs side PR to avoid using cutlass which bloats the binary size.

Enable dice metric for UMAP and pairwise_distances

Yes https://github.com/rapidsai/cuvs/pull/971, also in this PR description.