Performance

Time

Times on an RTX 4070 Ti with the default settings.

Points	Input	Fit
5,000	50 to 2,048 columns	about 2 minutes
100,000	50 principal components	about 4 minutes
250,000	10-dimensional latent, padded to 30	about 6 minutes

Each step trains on the whole neighbour graph up to 262,144 edges, and on a random batch of that many above, so the cost of a step stops growing with the number of points. Between 50,000 and 500,000 points only the heaviest quarter of the graph edges is kept.

max_iter trades time for quality. Fewer steps train faster and give a rougher layout.

Memory

A fit of 250,000 points peaks at about 2.7 GB of GPU memory. In version 0.1.0 the last step of fit passes all points through the flow at once, which on a 12 GB GPU ran out of memory at 1.77 million points.

CPU

Training runs on the CPU when no GPU is available or with device="cpu", several times slower than on a GPU. FloDR limits PyTorch to the performance cores of the CPU.

Reproducibility

On the CPU a fixed random_state reproduces a fit exactly. On CUDA the training step is compiled and runs its hidden layers in half precision, which is about twice as fast but not bitwise reproducible between runs. The eager path in full precision runs with

FloDR(advanced={"compile_edges": False, "half_edges": ""})

Checking a fit

In rare cases a fit returns coordinates that are not finite. This was seen once, at w=3, and a refit with the same seed was finite. Check the layout and refit with another seed if it happens.

import numpy as np

assert np.isfinite(model.embedding_).all()