I am using UMAP (https://umap-learn.readthedocs.io/en/latest/#) to reduce the dimensions in my data. My dataset contains 4700 samples with 1,2 million features each (which I would like to reduce). However, this takes quite a while despite using 32 CPUs and 120GB RAM. Especially the construction of the embedding is slow and the verbose output did not change in the last 3.5 hours:
UMAP(dens_frac=0.0, dens_lambda=0.0, low_memory=False, n_neighbors=10,
verbose=True)
Construct fuzzy simplicial set
Mon Jul 5 09:43:28 2021 Finding Nearest Neighbors
Mon Jul 5 09:43:28 2021 Building RP forest with 59 trees
Mon Jul 5 10:06:10 2021 metric NN descent for 20 iterations
1 / 20
2 / 20
3 / 20
4 / 20
5 / 20
Stopping threshold met -- exiting after 5 iterations
Mon Jul 5 10:12:14 2021 Finished Nearest Neighbor Search
Mon Jul 5 10:12:25 2021 Construct embedding
Are there any ways to make this process faster. I am already using a sparse matrix (scipy.sparse.lil_matrix) as described here: https://umap-learn.readthedocs.io/en/latest/sparse.html. Additionally I have installed pynndescent (as described here: https://github.com/lmcinnes/umap/issues/416). My Code is as follows:
from scipy.sparse import lil_matrix
import numpy as np
import umap.umap_ as umap
term_dok_matrix = np.load('term_dok_matrix.npy')
term_dok_mat_lil = lil_matrix(term_dok_matrix, dtype=np.float32)
test = umap.UMAP(a=None, angular_rp_forest=False, b=None,
force_approximation_algorithm=False, init='spectral', learning_rate=1.0,
local_connectivity=1.0, low_memory=False, metric='euclidean',
metric_kwds=None, n_neighbors=10, min_dist=0.1, n_components=2, n_epochs=None,
negative_sample_rate=5, output_metric='euclidean',
output_metric_kwds=None, random_state=None, repulsion_strength=1.0,
set_op_mix_ratio=1.0, spread=1.0, target_metric='categorical',
target_metric_kwds=None, target_n_neighbors=-1, target_weight=0.5,
transform_queue_size=4.0, unique=False, verbose=True).fit_transform(term_dok_mat_lil)
Are there any tricks or ideas how to make the computation faster? Can I change some parameters? Does it help that my matrix consists only of zeros and ones (meaning all non-zeros entries in my matrix are a one).
With 1.2 million features and only 4700 samples you are going to be better off just precomputing the full distance matrix and passing that in with metric="precomputed". Currently it is expending a lot of work computing nearest neighbors of those 1.2million long vectors. Just brute force will be a lot better.