ZeroDivisionError in `inverse_transform()`

[jvpeetz]

Dear Leland McInnes,
trying to use the inverse_transform() method of umap-learn, following the tutorial on Inverse transforms, when calling inv_transformed_points = mapper.inverse_transform(test_pts), this exception is hit:

---------------------------------------------------------------------------
ZeroDivisionError                         Traceback (most recent call last)
Cell In[5], line 1
----> 1 inv_transformed_points = mapper.inverse_transform(test_pts)

File /usr/lib/python3/dist-packages/umap/umap_.py:3332, in UMAP.inverse_transform(self, X)
   3329 tail = graph.col
   3330 weight = graph.data
-> 3332 inv_transformed_points = optimize_layout_inverse(
   3333     inv_transformed_points,
   3334     self._raw_data,
   3335     head,
   3336     tail,
   3337     weight,
   3338     self._sigmas,
   3339     self._rhos,
   3340     n_epochs,
   3341     graph.shape[1],
   3342     epochs_per_sample,
   3343     self._a,
   3344     self._b,
   3345     rng_state,
   3346     self.repulsion_strength,
   3347     self._initial_alpha / 4.0,
   3348     self.negative_sample_rate,
   3349     self._inverse_distance_func,
   3350     tuple(self._metric_kwds.values()),
   3351     verbose=self.verbose,
   3352     tqdm_kwds=self.tqdm_kwds,
   3353 )
   3355 return inv_transformed_points

File /usr/lib/python3/dist-packages/umap/layouts.py:850, in optimize_layout_inverse(head_embedding, tail_embedding, head, tail, weight, sigmas, rhos, n_epochs, n_vertices, epochs_per_sample, a, b, rng_state, gamma, initial_alpha, negative_sample_rate, output_metric, output_metric_kwds, verbose, tqdm_kwds, move_other)
    847     tqdm_kwds["disable"] = not verbose
    849 for n in tqdm(range(n_epochs), **tqdm_kwds):
--> 850     optimize_fn(
    851         epochs_per_sample,
    852         epoch_of_next_sample,
    853         head,
    854         tail,
    855         head_embedding,
    856         tail_embedding,
    857         output_metric,
    858         output_metric_kwds,
    859         weight,
    860         sigmas,
    861         dim,
    862         alpha,
    863         move_other,
    864         n,
    865         epoch_of_next_negative_sample,
    866         epochs_per_negative_sample,
    867         rng_state,
    868         n_vertices,
    869         rhos,
    870         gamma,
    871     )
    872     alpha = initial_alpha * (1.0 - (float(n) / float(n_epochs)))
    874 return head_embedding

ZeroDivisionError: division by zero

A minimal Pythone code to reproduce the crash is:

import numpy as np
import sklearn.datasets
import umap

data, labels = sklearn.datasets.fetch_openml('mnist_784', version=1, return_X_y=True)

mapper = umap.UMAP(random_state=42).fit(data)

corners = np.array([
    [-5, -10],  # 1
    [-7, 6],  # 7
    [2, -8],  # 2
    [12, 4],  # 0
])

test_pts = np.array([
    (corners[0]*(1-x) + corners[1]*x)*(1-y) +
    (corners[2]*(1-x) + corners[3]*x)*y
    for y in np.linspace(0, 1, 10)
    for x in np.linspace(0, 1, 10)
])

inv_transformed_points = mapper.inverse_transform(test_pts)

Some informations regarding the test system, a Debian testing installation:

Python implementation: CPython
Python version:        3.13.9
IPython version:       8.35.0

Jupyter:               8.6.3

numba:        0.61.2
numpy:        2.3.4
pynndescent:  0.5.13
scipy:        1.16.3
sklearn:      1.7.2
tqdm:         4.67.1
umap:         0.5.9.post2

Any idea why it crashes?

Regards,
Jörg.

[nachoDRT]

I also experienced the same. @jvpeetz, did you find a solution? Thanks!