test_umap

A UMAP projection fit with cuML and saved as plain PyTorch tensors (model.pt, written by utils.umap_utils.save_umap). New vectors are projected with the PyTorch port in utils.umap_utils.transform, on CPU or GPU, without cuML.

Fit on data/feed_50k.pt (50000 x 768).

setting	value
training points	50000
input dim	768
PCA components	none
output dim	2
metric	cosine
n_neighbors	50
min_dist	0.2
spread	1.0
n_epochs	500
negative_sample_rate	10
repulsion_strength	1.0
a, b	1.2621, 1.003

Usage

from utils.hf_hub import pull_umap
from utils.umap_utils import transform

proj = pull_umap("Ty-Perez/test_umap")            # UMAPProjection on CPU
xy = transform(proj, vectors)            # vectors: (N, 768) float tensor

params.json, when present, is the full fit record (parameters, timing, versions).