Differences in results between the old and new versions

[KnSun99]

New version:

Old version:

The code and the data are the same:
umap_model = umap.UMAP(n_components=2, n_neighbors=150, min_dist=0.051, metric='euclidean', n_jobs=-1, verbose=True)
reduction = umap_model.fit_transform(matrix)

Could you please guide me on how to adjust to obtain results similar to those of the old version?
Thank you

[KnSun99]

conda list:

channels:

• conda-forge
• defaults
  dependencies:
• aiodns=3.5.0=py312haa95532_0
• aiohappyeyeballs=2.6.1=py312haa95532_0
• aiohttp=3.12.15=py312h02ab6af_0
• aiosignal=1.4.0=py312haa95532_0
• alabaster=0.7.16=py312haa95532_0
• arrow=1.4.0=py312haa95532_0
• astroid=3.3.11=py312haa95532_0
• asttokens=3.0.0=py312haa95532_0
• asyncssh=2.17.0=py312haa95532_0
• atomicwrites=1.4.0=py_0
• attrs=25.4.0=py312haa95532_0
• autopep8=2.0.4=pyhd3eb1b0_0
• babel=2.16.0=py312haa95532_0
• bcrypt=4.3.0=py312h636fa0f_0
• beautifulsoup4=4.13.5=py312haa95532_0
• binaryornot=0.4.4=pyhd3eb1b0_1
• black=25.9.0=py312haa95532_0
• bleach=6.2.0=py312haa95532_0
• brotlicffi=1.0.9.2=py312h5da7b33_1
• bzip2=1.0.8=h0ad9c76_8
• c-ares=1.34.5=h731ff69_0
• ca-certificates=2025.10.5=h4c7d964_0
• cattrs=24.1.2=py312haa95532_0
• certifi=2025.10.5=pyhd8ed1ab_0
• cffi=2.0.0=py312h02ab6af_0
• chardet=5.2.0=py312haa95532_0
• charset-normalizer=3.3.2=pyhd3eb1b0_0
• click=8.2.1=py312haa95532_0
• cloudpickle=3.1.1=py312haa95532_0
• colorama=0.4.6=py312haa95532_0
• comm=0.2.1=py312haa95532_0
• cookiecutter=2.6.0=py312haa95532_1
• cryptography=46.0.2=py312h114bc41_0
• debugpy=1.8.16=py312h885b0b7_0
• decorator=5.2.1=py312haa95532_0
• defusedxml=0.7.1=pyhd3eb1b0_0
• deprecated=1.2.18=py312haa95532_0
• diff-match-patch=20200713=pyhd3eb1b0_0
• dill=0.4.0=py312haa95532_0
• docstring-to-markdown=0.17=py312haa95532_0
• docutils=0.21.2=py312haa95532_0
• executing=2.2.1=py312haa95532_0
• ezc3d=1.6.0=np2py312h6d06127_0
• flake8=7.1.1=py312haa95532_0
• frozenlist=1.8.0=py312h885b0b7_0
• fzf=0.55.0=h7d69c46_0
• icu=73.1=h6c2663c_0
• idna=3.7=py312haa95532_0
• imagesize=1.4.1=py312haa95532_0
• importlib-metadata=8.5.0=py312haa95532_0
• inflection=0.5.1=py312haa95532_1
• intervaltree=3.1.0=pyhd3eb1b0_0
• ipykernel=6.30.1=py312haa95532_0
• ipython=9.5.0=py312haa95532_0
• ipython_pygments_lexers=1.1.1=py312haa95532_0
• isort=6.1.0=py312haa95532_0
• jaraco.classes=3.4.0=py312haa95532_0
• jaraco.context=6.0.0=py312haa95532_0
• jaraco.functools=4.1.0=py312haa95532_0
• jedi=0.19.2=py312haa95532_0
• jellyfish=1.1.3=py312h8ecf97c_0
• jinja2=3.1.6=py312haa95532_0
• jpeg=9f=ha349fce_0
• jsonschema=4.25.0=py312haa95532_0
• jsonschema-specifications=2023.7.1=py312haa95532_0
• jupyter_client=8.6.3=py312haa95532_0
• jupyter_core=5.8.1=py312haa95532_0
• jupyterlab_pygments=0.3.0=py312haa95532_0
• keyring=25.6.0=py312haa95532_0
• krb5=1.20.1=h5b6d351_0
• libblas=3.9.0=35_h5709861_mkl
• libcblas=3.9.0=35_h2a3cdd5_mkl
• libclang13=14.0.6=default_h8e68704_2
• libexpat=2.7.1=hac47afa_0
• libffi=3.5.2=h52bdfb6_0
• libhwloc=2.12.1=default_h88281d1_1000
• libiconv=1.18=hc1393d2_2
• libkrb5=1.21.3=h885b0b7_4
• liblapack=3.9.0=35_hf9ab0e9_mkl
• liblzma=5.8.1=h2466b09_2
• liblzma-devel=5.8.1=h2466b09_2
• libpng=1.6.50=h46444df_0
• libpq=17.6=h652a1e2_0
• libsodium=1.0.20=h83e8143_0
• libspatialindex=1.9.3=h6c2663c_0
• libsqlite=3.50.4=hf5d6505_0
• libuv=1.48.0=h827c3e9_0
• libwinpthread=12.0.0.r4.gg4f2fc60ca=h57928b3_10
• libxml2=2.13.8=h741aa76_1
• libzlib=1.3.1=h2466b09_2
• llvm-openmp=21.1.4=hfa2b4ca_0
• lsprotocol=2025.0.0=py312haa95532_0
• lz4-c=1.9.4=h2bbff1b_1
• markdown-it-py=4.0.0=py312haa95532_0
• markupsafe=3.0.2=py312h827c3e9_0
• matplotlib-inline=0.1.7=py312haa95532_0
• mccabe=0.7.0=pyhd3eb1b0_0
• mdurl=0.1.2=py312haa95532_0
• mistune=3.1.2=py312haa95532_0
• mkl=2024.2.2=h57928b3_16
• more-itertools=10.8.0=py312haa95532_0
• multidict=6.7.0=py312h02ab6af_0
• mypy_extensions=1.0.0=py312haa95532_0
• nbclient=0.10.2=py312haa95532_0
• nbconvert=7.16.6=py312haa95532_0
• nbconvert-core=7.16.6=py312haa95532_0
• nbconvert-pandoc=7.16.6=py312haa95532_0
• nbformat=5.10.4=py312haa95532_0
• nest-asyncio=1.6.0=py312haa95532_0
• numpydoc=1.9.0=py312haa95532_0
• openssl=3.5.4=h725018a_0
• packaging=25.0=py312haa95532_1
• pandoc=3.8=haa95532_0
• pandocfilters=1.5.1=py312haa95532_0
• parso=0.8.4=py312haa95532_0
• pathspec=0.12.1=py312haa95532_0
• pexpect=4.9.0=py312haa95532_0
• pickleshare=0.7.5=pyhd3eb1b0_1003
• pip=25.2=pyh8b19718_0
• platformdirs=4.3.7=py312haa95532_0
• pluggy=1.5.0=py312haa95532_0
• ply=3.11=py312haa95532_1
• prompt-toolkit=3.0.52=py312haa95532_0
• prompt_toolkit=3.0.52=hd3eb1b0_0
• propcache=0.3.1=py312h827c3e9_0
• psutil=7.0.0=py312h02ab6af_0
• ptyprocess=0.7.0=pyhd3eb1b0_2
• pure_eval=0.2.3=py312haa95532_0
• pycares=4.10.0=py312he29e43d_0
• pycodestyle=2.12.1=py312haa95532_0
• pycparser=2.23=py312haa95532_0
• pydocstyle=6.3.0=py312haa95532_0
• pyflakes=3.2.0=py312haa95532_0
• pygithub=2.4.0=py312haa95532_0
• pygments=2.19.1=py312haa95532_0
• pyjwt=2.10.1=py312haa95532_0
• pylint=3.3.8=py312haa95532_0
• pylint-venv=3.0.3=py312haa95532_0
• pyls-spyder=0.4.0=pyhd3eb1b0_0
• pynacl=1.6.0=py312h581ae37_0
• pyqt=5.15.10=py312h5da7b33_1
• pyqt5-sip=12.13.0=py312h827c3e9_1
• pyqtwebengine=5.15.10=py312h5da7b33_1
• pysocks=1.7.1=py312haa95532_0
• python=3.12.12=h0159041_1_cpython
• python-dateutil=2.9.0post0=py312haa95532_2
• python-fastjsonschema=2.20.0=py312haa95532_0
• python-lsp-black=2.0.0=py312haa95532_1
• python-lsp-jsonrpc=1.1.2=pyhd3eb1b0_0
• python-lsp-ruff=2.3.0=py312haa95532_0
• python-lsp-server=1.13.1=py312hbc747e5_0
• python-slugify=8.0.4=py312haa95532_0
• python-tzdata=2025.2=pyhd3eb1b0_0
• python_abi=3.12=1_cp312
• pytokens=0.2.0=py312haa95532_0
• pytoolconfig=1.2.6=py312haa95532_0
• pyuca=1.2=py312haa95532_1
• pywin32=311=py312h885b0b7_0
• pywin32-ctypes=0.2.2=py312haa95532_0
• pyyaml=6.0.2=py312h827c3e9_0
• pyzmq=27.1.0=py312h7149c55_0
• qdarkstyle=3.2.3=pyhd3eb1b0_0
• qstylizer=0.2.2=py312haa95532_0
• qt-main=5.15.2=h19c9488_12
• qt-webengine=5.15.9=h5bd16bc_7
• qtawesome=1.4.0=py312haa95532_0
• qtconsole=5.7.0=py312haa95532_0
• qtconsole-base=5.7.0=pyha770c72_0
• qtpy=2.4.3=py312haa95532_0
• referencing=0.30.2=py312haa95532_0
• requests=2.32.5=py312haa95532_0
• rich=14.2.0=py312haa95532_0
• roman-numerals-py=3.1.0=py312haa95532_0
• rope=1.13.0=py312haa95532_0
• rpds-py=0.22.3=py312h636fa0f_0
• rtree=1.4.1=py312h1e5ca3c_0
• ruff=0.12.0=py312h1361809_0
• setuptools=80.9.0=pyhff2d567_0
• sip=6.7.12=py312h5da7b33_1
• six=1.17.0=py312haa95532_0
• snowballstemmer=3.0.1=py312haa95532_0
• sortedcontainers=2.4.0=pyhd3eb1b0_0
• soupsieve=2.5=py312haa95532_0
• sphinx=8.2.3=py312h827c3e9_0
• sphinxcontrib-applehelp=2.0.0=pyhd3eb1b0_1
• sphinxcontrib-devhelp=2.0.0=pyhd3eb1b0_0
• sphinxcontrib-htmlhelp=2.1.0=pyhd3eb1b0_0
• sphinxcontrib-jsmath=1.0.1=pyhd3eb1b0_0
• sphinxcontrib-qthelp=2.0.0=pyhd3eb1b0_1
• sphinxcontrib-serializinghtml=2.0.0=pyhd3eb1b0_0
• spyder=6.1.0=hd8ed1ab_0
• spyder-base=6.1.0=py312h2e8e312_0
• spyder-kernels=3.1.0=py312hbc747e5_0
• sqlite=3.50.2=hda9a48d_1
• stack_data=0.6.3=py312haa95532_0
• superqt=0.7.6=py312hdfb3b6f_0
• text-unidecode=1.3=pyhd3eb1b0_0
• textdistance=4.6.3=py312hbc747e5_1
• three-merge=0.1.1=pyhd3eb1b0_0
• tinycss2=1.4.0=py312haa95532_0
• tk=8.6.13=h2c6b04d_2
• tomlkit=0.13.2=py312haa95532_0
• tornado=6.5.1=py312h827c3e9_0
• traitlets=5.14.3=py312haa95532_0
• typing-extensions=4.15.0=py312haa95532_0
• typing_extensions=4.15.0=py312haa95532_0
• tzdata=2025b=h78e105d_0
• ucrt=10.0.26100.0=h57928b3_0
• ujson=5.10.0=py312h5da7b33_1
• urllib3=2.5.0=py312haa95532_0
• vc=14.3=h2b53caa_32
• vc14_runtime=14.44.35208=h818238b_32
• vcomp14=14.44.35208=h818238b_32
• vs2015_runtime=14.44.35208=h38c0c73_32
• watchdog=6.0.0=py312haa95532_0
• wcwidth=0.2.13=py312haa95532_0
• webencodings=0.5.1=py312haa95532_2
• whatthepatch=1.0.7=py312haa95532_0
• wheel=0.45.1=pyhd8ed1ab_1
• win_inet_pton=1.1.0=py312haa95532_0
• winloop=0.2.2=py312h913537a_0
• wrapt=1.17.0=py312h827c3e9_0
• xz=5.8.1=h208afaa_2
• xz-tools=5.8.1=h2466b09_2
• yaml=0.2.5=he774522_0
• yapf=0.43.0=py312haa95532_0
• yarl=1.18.0=py312h827c3e9_0
• zeromq=4.3.5=h6c54ac7_1
• zipp=3.21.0=py312haa95532_0
• zlib=1.3.1=h2466b09_2
• zstd=1.5.7=h56299aa_0
• pip:
  • absl-py==2.3.1
  • addict==2.4.0
  • astunparse==1.6.3
  • blosc2==3.10.2
  • c3d==0.6.0
  • contourpy==1.3.3
  • cycler==0.12.1
  • cython==3.1.6
  • daal==2025.9.0
  • filelock==3.19.1
  • flatbuffers==25.9.23
  • fonttools==4.60.1
  • fsspec==2025.9.0
  • gast==0.6.0
  • google-pasta==0.2.0
  • grpcio==1.76.0
  • h5py==3.15.1
  • imageio==2.37.2
  • joblib==1.5.2
  • keras==3.12.0
  • kiwisolver==1.4.9
  • lazy-loader==0.4
  • libclang==18.1.1
  • llvmlite==0.45.1
  • markdown==3.9
  • matplotlib==3.10.7
  • ml-dtypes==0.5.3
  • mpmath==1.3.0
  • msgpack==1.1.2
  • multiprocess==0.70.18
  • namex==0.1.0
  • narwhals==2.10.0
  • natsort==8.4.0
  • ndindex==1.10.0
  • networkx==3.5
  • numba==0.62.1
  • numexpr==2.14.1
  • numpy==2.3.5
  • opencv-python==4.12.0.88
  • opt-einsum==3.4.0
  • optree==0.17.0
  • pandas==2.3.3
  • pathos==0.3.4
  • pillow==11.3.0
  • plotly==6.3.1
  • pox==0.3.6
  • ppft==1.7.7
  • protobuf==6.33.0
  • py-cpuinfo==9.0.0
  • pynndescent==0.5.13
  • pyparsing==3.2.5
  • pytz==2025.2
  • scikit-image==0.25.2
  • scikit-learn==1.7.2
  • scikit-learn-intelex==2025.9.0
  • scipy==1.16.2
  • seaborn==0.13.2
  • sympy==1.14.0
  • tables==3.10.2
  • tbb==2022.3.0
  • tcmlib==1.4.1
  • tensorboard==2.20.0
  • tensorboard-data-server==0.7.2
  • tensorflow==2.20.0
  • termcolor==3.2.0
  • threadpoolctl==3.6.0
  • tifffile==2025.10.16
  • torch==2.8.0+cu128
  • torchaudio==2.8.0+cu128
  • torchvision==0.23.0+cu128
  • tqdm==4.67.1
  • umap-learn==0.5.9.post2
  • werkzeug==3.1.3

[lmcinnes]

I don't know which old version you are referring to. Not much has changed in layout optimization for a couple of semi-major versions, so it seems surprising that there would be this big a difference unless you were previously using version 0.3 or earlier.

[KnSun99]

The old version was 0.5.6.

However, upon my re-test, I found that it was not an issue with the umap version, but rather with the numba version.

I downgraded numba from version 0.62.1 to version 0.61.2, and then recalculated and everything returned to normal.
0.62.1 -> 0.61.2

Would it be necessary to update the umap version based on the new version of numba?

[viclafargue]

It looks like smooth_knn_dist may be setting some sigmas to infinity when using the latest version of Numba. Probably because of the use of fastmath here that may have an effect on this hi == NPY_INFINITY check.

repro.py

import numpy as np
from sklearn.datasets import make_blobs
from sklearn.neighbors import NearestNeighbors
from umap.umap_ import smooth_knn_dist

n_samples = 200
n_features = 10
n_neighbors = 15

X, _ = make_blobs(n_samples=n_samples, n_features=n_features, random_state=42)
X = X.astype(np.float32)

knn = NearestNeighbors(n_neighbors=n_neighbors, metric="euclidean")
knn.fit(X)
knn_dists, _ = knn.kneighbors(X)

sigmas, rhos = smooth_knn_dist(knn_dists.astype(np.float32), float(n_neighbors))

inf_mask = np.isinf(sigmas)
n_inf = inf_mask.sum()

print(f"Number of inf sigmas: {n_inf} / {sigmas.shape[0]}")

With numba version 0.61.2 :

$ python repro.py
> Number of inf sigmas: 0 / 200

With numba version 0.62.0 :

$ python repro.py
> Number of inf sigmas: 5 / 200