dbsampler

Decision Boundary Sampler

Sample the decision boundary of classification problems
Blazingly fast geometric sampling
Built with Rust

Pepy Total Downlods

Installation
- Compiling from source
Usage
- Sparse
How does it work
Performance
Citing

DBSampler is a package to sample the decision boundary of binary or multiclass datasets. It is designed to remain efficient in high dimensions:

Lets the user request how many boundary points to sample. More points give broader coverage, while fewer points run faster.
Iteratively moves each point onto a boundary between Voronoi regions belonging to different classes.
Can keep one representative for each distinct boundary region it finds.

Installation

Pre-built packages for MacOS, Windows and Linux systems are available on PyPI and can be installed with:

pip install dbsampler

On uncommon architectures, you may need to first install Cargo before running pip install dbsampler.

Compilation from source

In order to compile from source you will need Rust/Cargo and maturin for the Python bindings. Maturin is best used within a Python virtual environment:

# activate your desired virtual environment first, then:
pip install maturin
git clone https://github.com/antonio-leitao/dbsampler.git
cd dbsampler
# build and install the package:
maturin develop --release

Usage

import dbsampler
import numpy as np

cover = dbsampler.dbs(
    data=X,
    y=y,
    n_points=1000,
    max_iter=100,
    tol=1e-6,
    sparse=True,
    parallel=True,
    seed=42,
    batch_size=256,
    max_batches=20,
)

Parameters:

data: NumPy array of shape (n, d) with dtype float32 or float64. The core computation uses normalized, contiguous float32 data.
y: 1-dimensional array or list of integer class labels, length n.
n_points: number of points to sample from the decision boundary. More points give a denser sample but increase runtime. Default 1000.
max_iter: maximum number of projection iterations. Default 100.
tol: per-point convergence threshold on squared displacement. Default 1e-6.
sparse: if True (default), aims to return n_points distinct boundary points by sampling additional batches when duplicates are found.
parallel: if True (default), uses rayon to parallelize the per-point nearest-neighbor search and projection steps across CPU cores. The BLAS matrix multiplications are multithreaded independently of this flag.
seed: optional integer seed for reproducible results. When None (default), initialization is random. Set to a fixed value (e.g. seed=42) for deterministic runs.
batch_size: maximum number of points processed together when sparse=True. Default 256.
max_batches: hard limit on the number of batches when sparse=True. Default 20.

Returns:

cover: list of lists, each of length d — the sampled boundary points (as float32 values). With sparse=True, the list can be shorter than n_points if the requested number of distinct boundary regions cannot be found before max_batches is reached.

Sparse

Passing sparse=True returns at most one representative for each distinct boundary region. The sampler processes points in batches and, when duplicates become common, uses the previous results to favor parts of the input space that have produced new regions efficiently. Exact nearest-neighbor search and convergence checks are still used for every selected point. If it cannot find n_points distinct regions before reaching max_batches, it returns those it found and emits a warning.

With sparse=False, the sampler generates n_points once without removing duplicates. Below are examples of dense sampling (left) and sparse sampling (right).

How does it work?

For an in-depth explanation check our paper. The algorithm samples the shared boundaries of Voronoi regions belonging to points of different classes. These boundary pieces form the decision boundary of a 1-nearest-neighbor classifier.

It starts by sampling points uniformly inside the bounds of the data. It then iteratively projects each point onto the bisecting hyperplane between its two nearest neighbors of different classes. With sparse=True, this work is split into batches so duplicates can be replaced by new samples.

Why the iteration moves toward the boundary. At each iteration, the sampler finds the nearest data point and the nearest point from a different class, then moves the query point onto the plane halfway between them. It checks the neighbours again after the move. If the same pair remains, the query point is on their shared boundary; otherwise the process continues with the new pair. A small movement can also stop a point approximately, and max_iter remains a hard limit.

Performance

DBSampler is written in Rust with BLAS-accelerated linear algebra (via Accelerate on macOS, OpenBLAS on Linux/Windows). The core dot-product and matrix-multiply operations use cblas_sgemm and cblas_sdot, and the algorithm automatically switches to a tiled iteration strategy when the score matrix would exceed 32 MB, avoiding allocation of the complete matrix for large datasets.

With parallel=True, the per-point nearest-neighbor search and bisector projection are distributed across CPU cores using rayon. The BLAS matrix multiplications are multithreaded independently. Each point stops independently once it meets the displacement tolerance or keeps the same neighboring pair after projection.

Pre-built binaries are available for Windows, macOS and most Linux distributions.

Citing

If you use DBSampler in your work or parts of the algorithm please consider citing:

@inproceedings{petri2020on,
               title={On The Topological Expressive Power of Neural Networks},
               author={Giovanni Petri and Ant{\'o}nio Leit{\~a}o},
               booktitle={NeurIPS 2020 Workshop on Topological Data Analysis and Beyond},
               year={2020},
               url={https://openreview.net/forum?id=I44kJPuvqPD}
}

In the paper above you can find the pseudocode of the algorithm and its convergence discussion.

License

DBSampler is distributed under the 3-clause BSD license.

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

dbsampler-0.4.0.tar.gz (1.6 MB view details)

Uploaded Aug 10, 2026 Source

Built Distributions

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

dbsampler-0.4.0-pp311-pypy311_pp73-manylinux_2_28_x86_64.whl (463.2 kB view details)

Uploaded Aug 10, 2026 PyPymanylinux: glibc 2.28+ x86-64

dbsampler-0.4.0-pp311-pypy311_pp73-manylinux_2_28_aarch64.whl (465.1 kB view details)

Uploaded Aug 10, 2026 PyPymanylinux: glibc 2.28+ ARM64

dbsampler-0.4.0-cp315-cp315t-manylinux_2_28_x86_64.whl (455.7 kB view details)

Uploaded Aug 10, 2026 CPython 3.15tmanylinux: glibc 2.28+ x86-64

dbsampler-0.4.0-cp315-cp315t-manylinux_2_28_aarch64.whl (459.9 kB view details)

Uploaded Aug 10, 2026 CPython 3.15tmanylinux: glibc 2.28+ ARM64

dbsampler-0.4.0-cp315-cp315-manylinux_2_28_x86_64.whl (454.2 kB view details)

Uploaded Aug 10, 2026 CPython 3.15manylinux: glibc 2.28+ x86-64

dbsampler-0.4.0-cp315-cp315-manylinux_2_28_aarch64.whl (457.1 kB view details)

Uploaded Aug 10, 2026 CPython 3.15manylinux: glibc 2.28+ ARM64

dbsampler-0.4.0-cp314-cp314t-manylinux_2_28_x86_64.whl (455.7 kB view details)

Uploaded Aug 10, 2026 CPython 3.14tmanylinux: glibc 2.28+ x86-64

dbsampler-0.4.0-cp314-cp314t-manylinux_2_28_aarch64.whl (459.8 kB view details)

Uploaded Aug 10, 2026 CPython 3.14tmanylinux: glibc 2.28+ ARM64

dbsampler-0.4.0-cp314-cp314-win_amd64.whl (246.3 kB view details)

Uploaded Aug 10, 2026 CPython 3.14Windows x86-64

dbsampler-0.4.0-cp314-cp314-manylinux_2_28_x86_64.whl (454.1 kB view details)

Uploaded Aug 10, 2026 CPython 3.14manylinux: glibc 2.28+ x86-64

dbsampler-0.4.0-cp314-cp314-manylinux_2_28_aarch64.whl (457.0 kB view details)

Uploaded Aug 10, 2026 CPython 3.14manylinux: glibc 2.28+ ARM64

dbsampler-0.4.0-cp314-cp314-macosx_11_0_arm64.whl (378.1 kB view details)

Uploaded Aug 10, 2026 CPython 3.14macOS 11.0+ ARM64

dbsampler-0.4.0-cp314-cp314-macosx_10_12_x86_64.whl (371.5 kB view details)

Uploaded Aug 10, 2026 CPython 3.14macOS 10.12+ x86-64

dbsampler-0.4.0-cp313-cp313-win_amd64.whl (246.4 kB view details)

Uploaded Aug 10, 2026 CPython 3.13Windows x86-64

dbsampler-0.4.0-cp313-cp313-manylinux_2_28_x86_64.whl (453.9 kB view details)

Uploaded Aug 10, 2026 CPython 3.13manylinux: glibc 2.28+ x86-64

dbsampler-0.4.0-cp313-cp313-manylinux_2_28_aarch64.whl (461.1 kB view details)

Uploaded Aug 10, 2026 CPython 3.13manylinux: glibc 2.28+ ARM64

dbsampler-0.4.0-cp313-cp313-macosx_11_0_arm64.whl (379.3 kB view details)

Uploaded Aug 10, 2026 CPython 3.13macOS 11.0+ ARM64

dbsampler-0.4.0-cp313-cp313-macosx_10_12_x86_64.whl (371.9 kB view details)

Uploaded Aug 10, 2026 CPython 3.13macOS 10.12+ x86-64

dbsampler-0.4.0-cp312-cp312-win_amd64.whl (246.8 kB view details)

Uploaded Aug 10, 2026 CPython 3.12Windows x86-64

dbsampler-0.4.0-cp312-cp312-manylinux_2_28_x86_64.whl (454.8 kB view details)

Uploaded Aug 10, 2026 CPython 3.12manylinux: glibc 2.28+ x86-64

dbsampler-0.4.0-cp312-cp312-manylinux_2_28_aarch64.whl (461.4 kB view details)

Uploaded Aug 10, 2026 CPython 3.12manylinux: glibc 2.28+ ARM64

dbsampler-0.4.0-cp312-cp312-macosx_11_0_arm64.whl (379.8 kB view details)

Uploaded Aug 10, 2026 CPython 3.12macOS 11.0+ ARM64

dbsampler-0.4.0-cp312-cp312-macosx_10_12_x86_64.whl (372.5 kB view details)

Uploaded Aug 10, 2026 CPython 3.12macOS 10.12+ x86-64

dbsampler-0.4.0-cp311-cp311-win_amd64.whl (250.3 kB view details)

Uploaded Aug 10, 2026 CPython 3.11Windows x86-64

dbsampler-0.4.0-cp311-cp311-manylinux_2_28_x86_64.whl (459.6 kB view details)

Uploaded Aug 10, 2026 CPython 3.11manylinux: glibc 2.28+ x86-64

dbsampler-0.4.0-cp311-cp311-manylinux_2_28_aarch64.whl (464.1 kB view details)

Uploaded Aug 10, 2026 CPython 3.11manylinux: glibc 2.28+ ARM64

dbsampler-0.4.0-cp311-cp311-macosx_11_0_arm64.whl (381.4 kB view details)

Uploaded Aug 10, 2026 CPython 3.11macOS 11.0+ ARM64

dbsampler-0.4.0-cp311-cp311-macosx_10_12_x86_64.whl (375.7 kB view details)

Uploaded Aug 10, 2026 CPython 3.11macOS 10.12+ x86-64

dbsampler-0.4.0-cp310-cp310-win_amd64.whl (250.0 kB view details)

Uploaded Aug 10, 2026 CPython 3.10Windows x86-64

dbsampler-0.4.0-cp310-cp310-manylinux_2_28_x86_64.whl (459.5 kB view details)

Uploaded Aug 10, 2026 CPython 3.10manylinux: glibc 2.28+ x86-64

dbsampler-0.4.0-cp310-cp310-manylinux_2_28_aarch64.whl (463.9 kB view details)

Uploaded Aug 10, 2026 CPython 3.10manylinux: glibc 2.28+ ARM64

dbsampler-0.4.0-cp310-cp310-macosx_11_0_arm64.whl (381.6 kB view details)

Uploaded Aug 10, 2026 CPython 3.10macOS 11.0+ ARM64

dbsampler-0.4.0-cp310-cp310-macosx_10_12_x86_64.whl (375.7 kB view details)

Uploaded Aug 10, 2026 CPython 3.10macOS 10.12+ x86-64

File details

Details for the file dbsampler-0.4.0.tar.gz.

File metadata

Download URL: dbsampler-0.4.0.tar.gz
Upload date: Aug 10, 2026
Size: 1.6 MB
Tags: Source
Uploaded using Trusted Publishing? No
Uploaded via: maturin/1.14.1

File hashes

Hashes for dbsampler-0.4.0.tar.gz
Algorithm	Hash digest
SHA256	`d61a2467332ec5cf13884411fd9f6324d4fe3c13dd5256be431a447548b4b79c`
MD5	`9ddc5622fb5a1d0c1a83a67ecf54f0ce`
BLAKE2b-256	`9cffdb156e80d4b3aa879c6188acd0883f8151da754dcdc4972e1f77ab9d1de4`

Algorithm	Hash digest
SHA256	`c851a2f4251906a999d492e71bc271a8380900254b2e8544631b48a00e024f07`
MD5	`899e0a983ce3a3c036c823e298c2f1a3`
BLAKE2b-256	`d53c7f7807c03795c4013e379bbfdb299698b8665da12901957f0aa3b3b3525a`

Algorithm	Hash digest
SHA256	`68027d9717034e5171e81c2998cff4f0c9107bff46423cff166e76e4109040ee`
MD5	`adaa62de493f1f725348b333edc65cff`
BLAKE2b-256	`171286706cf60f67eca04dc35228f87efa6e0afab5e1aac281d53eccf0d250db`

Algorithm	Hash digest
SHA256	`a667d6d7306b00a75f654da41523db6b177df014e16d2665006fe6871c931bd9`
MD5	`7965575038cca5febc311e5ff02ddaf6`
BLAKE2b-256	`8217c7e16e4504d6e4610ec2417132de586aa1006d0ccdbf370c456354478687`

Algorithm	Hash digest
SHA256	`e0801d3caa4dee0ba0381950f6e05a359607ab7f0f078c03fad1156d3e123937`
MD5	`ae48bd441077d816ca678bdaffe35d70`
BLAKE2b-256	`a1b9cadcbe843505bca597fa750122bf4afdcf6145c5c5a194a514a450d4b977`

Algorithm	Hash digest
SHA256	`8199f42d29a5bd8194c9026eb8110ec7183f55c229dc43ded0942fa6b18efe86`
MD5	`c6f62baa3c2f8295835eca4da1edc5a1`
BLAKE2b-256	`c2db8584dc7d3a4aed426b367167d24bd458c2e6ca98035703b19c904a806fbc`

Algorithm	Hash digest
SHA256	`6f6fb76b42fc6aca2423f1680d824d3219d5a0ef7fb7fd4bdf7fe01b5e36e790`
MD5	`b5e54ff1c6f5e6d5b9fdfda96f06af25`
BLAKE2b-256	`99a65f2327e75983ed8d690891e0c11d2a1d14e4f32b6fd8979dff2da9936472`

Algorithm	Hash digest
SHA256	`a3f978452b4b396d09f38e480242421becab34085c30dfb5a94860e1d5d4dfa3`
MD5	`61a1c9265a70811ccf80d22dada5ba29`
BLAKE2b-256	`573d33732012e805e6f1b717596fc607ea7e957f48aa604f4785b63adece1f65`

Algorithm	Hash digest
SHA256	`2ecf47b72c572babee85cbfe7a99b112df8ed4ad632c5c40a2f088b27a4eed44`
MD5	`2d04ab52f936e5e185499a2b0d1e4049`
BLAKE2b-256	`ccdcd636a2d16335001c41fb9446720e4c5bc2c9b004fc9fee4df7bc3c6d0b8b`

Algorithm	Hash digest
SHA256	`498d0e318e7701be0a49a7946194108603fab152d6b18866045040a1357f2877`
MD5	`2d28c7065f009063d75e0b0bbb4bde11`
BLAKE2b-256	`fd058aa2c997e79263f2b7dde1e1ffbe8de2d12cc9d653bd5bde76040dea8779`

Algorithm	Hash digest
SHA256	`f29c13f2d5ddeaa6aa46731272b36141aa9027745ae7a419c616c2d08dd79256`
MD5	`77b4c6998e8544fe0064bc9c06e1a111`
BLAKE2b-256	`16760ef76d53fa1d5710398a3753287d9152ddae97d939678f819b5ef33043ee`

Algorithm	Hash digest
SHA256	`bfa1f7ba37d392432b0267c9b4a311bb36cb65247235bcaafc0df3a01a7db699`
MD5	`82a2cce42fa9aa5b22bb3b617dad2237`
BLAKE2b-256	`53ca605fa0b3997c311076bc6fd3efe5d8fb1a52ef6d452d2550b6e30c3eadf4`

Algorithm	Hash digest
SHA256	`1bc1a371c6de3bc6d2935ef4ca86ac0c66650d05c463ad8c392c8138d6770a8f`
MD5	`69b99f721b1850880dcee6aa528b2bf2`
BLAKE2b-256	`b3c1c6a4a33b9d4c70626c84a64ca0e5ed0e9f8acebecb1eab077f8d86d18d79`

Algorithm	Hash digest
SHA256	`cb7de8697dc969019fbbc1e402459a1b28f6e7fb2629bc1e5a0139e63ef6139e`
MD5	`bc7dbec4b9f918bfbfafc6e9faf098b8`
BLAKE2b-256	`e1c0bbb3c74656a39e07671fd33a2d0c10d7ea42906a4bfeab3a92e29900bb50`

Algorithm	Hash digest
SHA256	`5119ec9729359f166c81c47c0d0ae9cde62a570213b4339e658f9dbdd4bdcd26`
MD5	`afca9efb235ebd2d5fd649a4585066fe`
BLAKE2b-256	`6f068f7e4d4188b248df8cabce6abf72fabaab801c24e24a444ea4dd5fe55a68`

Algorithm	Hash digest
SHA256	`13ca7fdc88f8a0240bf6c92feec27a1cc6c9ac60c296aa953531b0338b3812ab`
MD5	`bc2490bd0819a4986a9c952fe33a21d8`
BLAKE2b-256	`0f523b0e86c541dc8f91245453a6f877a47292d8ca0c925359d7201582bc602f`

Algorithm	Hash digest
SHA256	`591c8fee7f9d894be09ee6d74614f154ebbf0dca481f21ced050734d76c01efd`
MD5	`969e5f123ae730b4574e15f563d90d5f`
BLAKE2b-256	`25b3c60c00f7e8c477ced2b8c11ca57747b9156977fee17aaafc576b3dc92ee2`

Algorithm	Hash digest
SHA256	`cbb96b5a8b3d947a4697d9c9c503dccfe41adc076d106b4c273429febb3d3003`
MD5	`c5256c583b340bdb1f8c2caf3513b6f6`
BLAKE2b-256	`bec4bb199e048ba5db0dc365efe67145f458b931536c0a2ad96286eeaa14ba57`

Algorithm	Hash digest
SHA256	`78b8be0fd598b33a91a33b76d77d97cc4d96435c16690ada132bf6129c36ece6`
MD5	`fcf55276d83cf091f4efed776514b26a`
BLAKE2b-256	`db418a2efef2dbbd41d16b1a770e4745d5fc444704cc43a01ee29b9471550b75`

Algorithm	Hash digest
SHA256	`f3af4b4001d33166e520c77d2b6a4db0c705c6e87c1123b5e3228fc28fbf4428`
MD5	`ae2a351999b0e7bf108458aa2c4e09ed`
BLAKE2b-256	`a52ccd9ab2d9d4678e377796ecefeabc8491d7378c3abf7ac21163548b576e64`

Algorithm	Hash digest
SHA256	`fd535c2744a60e328af5483bcb06045a7f1c1012df629841761008149cd61f1b`
MD5	`b2f470d214572d21500fe415cbb70cf6`
BLAKE2b-256	`86a1adcde181df8d067fb52bb1c891d5f1261d44437f23744b1f4cc95f8c7d0a`

dbsampler 0.4.0

Decision Boundary Sampler

Contents

Installation

Compilation from source

Usage

Sparse

How does it work?

Performance

Citing

License

Project links

Key dates

1 maintainer

Credits

Requires

Classifiers

Download files

Source Distribution

Built Distributions

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

Algorithm	Hash digest
SHA256	`d3cc04d073b6560f442b4b833a567df98b7e399d2bf47b8d4240dfeeb0729677`
MD5	`dd06f493a94ff326634a9db7e120189b`
BLAKE2b-256	`6ff0dd3a1ac97bec588402d1e76af8e2ee8039907222dff68d5504b318ad9029`

Algorithm	Hash digest
SHA256	`14ee84d58cd4f6e7603f211463834c97e58cd79d03ed49b81bfc56b045e2fe68`
MD5	`2f8b17499a54d056ef431fa0a1786c63`
BLAKE2b-256	`4e11e1501c2813a4d48b4bca6f5b670efa979a1ef07a47a1b4b69b003c95ba41`

Algorithm	Hash digest
SHA256	`f83403f5bcb3829551becf3227cfcbb099a24fc0db73929809a51b1177a4b1b0`
MD5	`948e6c84572549c5e085f0008e1e4e9c`
BLAKE2b-256	`8d51ae8bbb9ee14af3fed43fbf6697d440797bf05752f8a8fc5a51c8f1f6f935`

Algorithm	Hash digest
SHA256	`c6c5321c44103421d6278411d3ef98ee9a5612845028c9acadc0b587a4291b16`
MD5	`6a43cfbfcfea3b3d5d6032c6b893b0b7`
BLAKE2b-256	`e1b985fc38957460913d0b6cdf5076d7840a635846d74ce745c45f4009f4f87b`

Algorithm	Hash digest
SHA256	`f1c1ec70a67b89aacc8668132b4a2486394cef941a1b7ad5ac51e322f95e8536`
MD5	`5b8674045d2d3131f79ce86c35d8b660`
BLAKE2b-256	`9a502a3604f83b467468a4d8c0dd0ea7bece4ac1146016c8526636ddd8584657`

Algorithm	Hash digest
SHA256	`8738c0f54773ec56d287d9121a621e501143425752fa355b47cd02f60d0383dd`
MD5	`494fc0f0929d12023a7e767631f705e1`
BLAKE2b-256	`fdcb0494cb636543b32a6ce176649969f1a4d77b80ff25cdab266cc10784f54d`

Algorithm	Hash digest
SHA256	`92369e4e9ecf711c9001399a808fd738aad36ed82beb72869ad32815ee4e22d5`
MD5	`8722599bb34bd2c3b210746e9934cc00`
BLAKE2b-256	`18cbf28186ba3255b752a272223d64f3c451be7b72543e6fa7a9201db1d99a20`

Algorithm	Hash digest
SHA256	`64bc45b9a0a5d575bc0ddffa84768524f09b3fe1d5154d4fd90bb382a718439c`
MD5	`a34de86b3fa2b98e593bb8894c9c2a43`
BLAKE2b-256	`b89b08f0963eee94823e6d09b21ac2419e4c1d3c83a1e615140ef6effa269752`

Algorithm	Hash digest
SHA256	`0050e1bd729afebe0dc91ccdd5db1f385eeb86a8bc04fcc5d5bb82e590ad451d`
MD5	`44525c26dab3b7b5596ca9916f9ac1f2`
BLAKE2b-256	`23efa1b1ea63e30892417f7c07c6b11c0851479501785e90b60ead253f68adcf`

Algorithm	Hash digest
SHA256	`324df40683cd37d2f69bfe3d752019dfd9244363586d406971fc3cb45c6328b3`
MD5	`2a65ac594954a2c43ae247d616d07104`
BLAKE2b-256	`ef170d12ff22e61e504d787f3b17c8f78c0e9caeac34bcfd64cc6938a114f207`