[BUG] simplicial_set_embedding() broken

[willleeney]

Describe the bug
A clear and concise description of what the bug is.

simplicial_set_embedding() requests some keyword arguments are positional
simplicial_set_embedding() requests some arguments not in the source or in the docs
simplicial_set_embedding() doesn't work with random_state=int

Steps/Code to reproduce bug

import numpy as np
from cuml.manifold.umap import fuzzy_simplicial_set, simplicial_set_embedding
from umap.spectral import spectral_layout
from umap.umap_  import find_ab_params

X = np.random.rand(1000, 50).astype(np.float32)
graph = fuzzy_simplicial_set(X, n_neighbors=15, metric='euclidean')
layout = spectral_layout(X, graph, dim=2, random_state=42)
a, b = find_ab_params(1.0, 0.1)
embedding = simplicial_set_embedding(
            X, 
            graph,
            init=layout,
            a=a, b=b,
            n_epochs=200,
            n_components=2,
            random_state=4,
            verbose=False,
)

if you include these parameters:

      initial_alpha=1.0,
      gamma=1.0,
      negative_sample_rate=5,
      metric="euclidean",
      metric_kwds=None,
      output_metric="euclidean",
      output_metric_kwds=None

it will request these to also be included:

      densmap=False,
      densmap_kwds=None,
      output_dens=None,

as these as detailed in the docs are not allowed

      repulsion_strength=None, 
      convert_dtype=True,

and to actually get it to work i need to pass random_state as np.random.RandomState(42)

When it does work, it returns a tuple not an array as expected

Environment details:

• Environment location: Docker
• Linux Distro/Architecture: Ubuntu amd64
• no GPU
• conda install -c rapidsai -c nvidia -c conda-forge cuml-cpu=24.12 numpy-base=1.26.4

[viclafargue]

Thank you for opening this issue. You can find the reference implementations for fuzzy_simplicial_set and simplicial_set_embedding in the umap-learn library. We do not offer perfect equivalency for now, notably only the graph is returned in the regular cuML version, whereas the CPU version might return the sigmas and rhos values, as well as the KNN distances. These values are not exposed in the Python layer of the version of cuML running on GPU.

The repulsion_strength and gamma arguments can be used interchangeably in the regular cuML library, whereas only gamma would be available in the CPU version.

Again, thank you for opening this issue. We should work on solving this by either following umap-learn behavior exactly or by creating a wrapper for cuml-cpu.

cc @dantegd

[willleeney]

Thanks for the prompt reply and helpful response. It would be great if this could be solved for easy testing