pros-sketch 0.1.0__tar.gz

pros_sketch-0.1.0/.gitignore

@@ -0,0 +1,16 @@
+__pycache__/
+*.py[cod]
+.pytest_cache/
+.ruff_cache/
+.coverage
+htmlcov/
+build/
+dist/
+*.egg-info/
+.venv/
+
+# Generated parameter-experiment artifacts are reproducible locally.
+test_para/data/*.npz
+test_para/results/*.csv
+test_para/results/*.md
+test_para/results/*.html

pros_sketch-0.1.0/CITATION.cff

@@ -0,0 +1,17 @@
+cff-version: 1.2.0
+title: "PROS: Partitioned and Refined Oversampling Sketches"
+message: "If you use PROS in your work, please cite this software."
+type: software
+abstract: "A Python package for partitioned, oversampled, and globally refined geometric data sketches."
+authors:
+  - family-names: Li
+    given-names: Lei
+repository-code: "https://github.com/LeiLi-Uchicago/PROS"
+license: MIT
+keywords:
+  - geometric sketching
+  - representative sampling
+  - k-center
+  - coreset
+version: 0.1.0
+date-released: 2026-08-15

pros_sketch-0.1.0/LICENSE

@@ -0,0 +1,22 @@
+MIT License
+
+Copyright (c) 2026 Lei Li
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+

pros_sketch-0.1.0/PKG-INFO

@@ -0,0 +1,222 @@
+Metadata-Version: 2.5
+Name: pros-sketch
+Version: 0.1.0
+Summary: Partitioned and Refined Oversampling Sketches for certified geometric sketching.
+Project-URL: Homepage, https://github.com/LeiLi-Uchicago/PROS
+Project-URL: Issues, https://github.com/LeiLi-Uchicago/PROS/issues
+Author: Lei Li
+License: MIT License
+        
+        Copyright (c) 2026 Lei Li
+        
+        Permission is hereby granted, free of charge, to any person obtaining a copy
+        of this software and associated documentation files (the "Software"), to deal
+        in the Software without restriction, including without limitation the rights
+        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+        copies of the Software, and to permit persons to whom the Software is
+        furnished to do so, subject to the following conditions:
+        
+        The above copyright notice and this permission notice shall be included in all
+        copies or substantial portions of the Software.
+        
+        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+        SOFTWARE.
+        
+License-File: LICENSE
+Keywords: coreset,k-center,single-cell,sketching,subsampling
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Bio-Informatics
+Classifier: Topic :: Scientific/Engineering :: Information Analysis
+Classifier: Typing :: Typed
+Requires-Python: >=3.10
+Requires-Dist: numpy>=1.23
+Requires-Dist: scikit-learn>=1.2
+Requires-Dist: scipy>=1.9
+Provides-Extra: adata
+Requires-Dist: anndata>=0.9; extra == 'adata'
+Provides-Extra: all
+Requires-Dist: anndata>=0.9; extra == 'all'
+Requires-Dist: mkdocs-material>=9.5; extra == 'all'
+Requires-Dist: mkdocs>=1.6; extra == 'all'
+Requires-Dist: pytest>=8; extra == 'all'
+Requires-Dist: ruff>=0.6; extra == 'all'
+Requires-Dist: scanpy>=1.9; extra == 'all'
+Requires-Dist: scsampler; extra == 'all'
+Provides-Extra: dev
+Requires-Dist: pytest>=8; extra == 'dev'
+Requires-Dist: ruff>=0.6; extra == 'dev'
+Provides-Extra: docs
+Requires-Dist: mkdocs-material>=9.5; extra == 'docs'
+Requires-Dist: mkdocs>=1.6; extra == 'docs'
+Provides-Extra: scanpy
+Requires-Dist: anndata>=0.9; extra == 'scanpy'
+Requires-Dist: scanpy>=1.9; extra == 'scanpy'
+Provides-Extra: scsampler
+Requires-Dist: scsampler; extra == 'scsampler'
+Description-Content-Type: text/markdown
+
+# PROS
+
+PROS (Partitioned and Refined Oversampling Sketches) selects a small,
+diversity-preserving subset of rows from a large real-valued coordinate matrix.
+It builds an oversampled candidate pool within partitions and performs one
+global refinement pass over that pool. The returned subset is represented by
+indices into the original matrix.
+
+The package is domain-agnostic. It can be used with reduced single-cell
+embeddings, as well as other feature matrices where Euclidean distance is a
+meaningful geometry.
+
+## Installation
+
+PROS requires Python 3.10 or newer. The distribution name is `pros-sketch`;
+the Python import remains `pros`. The commands below install this checkout.
+
+```bash
+python -m pip install -e .
+```
+
+Optional AnnData support:
+
+```bash
+python -m pip install -e ".[adata]"
+```
+
+For testing and linting:
+
+```bash
+python -m pip install -e ".[dev]"
+```
+
+## Quick start
+
+```python
+import numpy as np
+from pros import sketch
+
+X = np.random.default_rng(0).random((5_000, 50))
+result = sketch(X, n=500, seed=0)
+
+indices = result.indices
+X_sketch = X[indices]
+print(result.timings)
+```
+
+`result.indices` is a sorted, unique `int64` array of length `n`. By default,
+PROS uses k-means partitioning, water-filling allocation, an oversampling ratio
+of `r=10.0`, and global farthest-first refinement. Pass an explicit `seed` to
+make a stochastic configuration reproducible.
+
+## Certification
+
+Set `certify=True` to compute the final covering radius, the candidate-pool
+radius, and bounds derived from a full-data farthest-first traversal:
+
+```python
+result = sketch(X, n=500, seed=0, certify=True)
+
+print(result.radius)       # R(X, S): final covering radius
+print(result.rho)          # R(X, P): candidate-pool covering radius
+print(result.opt_lower)    # lower bound on the optimal n-centre radius
+print(result.ratio_upper)  # upper bound on R(X, S) / OPT_n(X)
+```
+
+Certification performs additional full-data work and is not included in
+`result.timings["total"]`. Degenerate data may have `opt_lower == 0`; in that
+case a finite approximation-ratio bound is not defined.
+
+For an existing set of indices, use `pros.certificate`. `pros.opt_bounds`,
+`pros.estimate_opt_scale`, and `pros.choose_r` are available for certificate
+and oversampling-ratio workflows.
+
+## AnnData
+
+```python
+from pros import sketch_adata
+
+result = sketch_adata(adata, n=5_000, use_rep="X_pca", seed=0)
+subset = adata[result.indices].copy()
+```
+
+`sketch_adata` also accepts an `.h5ad` path and reads it in backed mode before
+extracting the requested representation.
+
+## Configuration
+
+```python
+result = sketch(
+    X,
+    n=1_000,
+    partitioner="kmeans",
+    n_blocks="auto",
+    allocator="water_filling",
+    r=10.0,
+    selector="fft",
+    refiner="fft",
+    seed=0,
+)
+```
+
+The `partitioner`, within-partition `selector`, `allocator`, and global
+`refiner` are modular. See [`docs/configuration.md`](docs/configuration.md) for
+the supported values and [`docs/theory.md`](docs/theory.md) for the covering
+objective and certificate definitions.
+
+## Repository layout
+
+- `pros/`: installable library code.
+- `tests/`: automated package tests.
+- `examples/`: small runnable examples using synthetic data in `data/`.
+- `test_para/`: standalone parameter experiments and their generated outputs;
+  this reproduction material is intentionally not packaged with the library.
+- `docs/`: MkDocs source pages.
+
+Run the examples from the repository root after installation:
+
+```bash
+python examples/01_basic_numpy.py
+python examples/03_certification.py
+```
+
+## Citation
+
+See [`CITATION.cff`](CITATION.cff) for software citation metadata.
+
+## License
+
+PROS is distributed under the [MIT License](LICENSE).
+
+## Validated behavior and limits
+
+- Water-filling is fused with FFT and requires `selector="fft"`. Choose another
+  allocator to use random or scSampler selection. Missing or broken scSampler
+  installations raise an error; no alternate algorithm is substituted.
+- `mix>0` currently requires `refiner="fft"` so reserved points are retained.
+  `refiner="none"` uniformly downsamples the candidate pool to `n` rows.
+- `certificate` computes the a posteriori ratio for any valid sketch. Its
+  `theory_bound` and `bound_slack` are NaN unless FFT refinement is explicitly
+  asserted; `sketch` sets this assertion only for unmixed FFT refinement.
+  Caches are tied to the exact data and sketch size. Floating-point results
+  are numerical bounds, not interval-arithmetic certificates.
+- Internal stage timings exclude input conversion, validation, AnnData I/O,
+  and certification. Measure an external wall clock for runtime comparisons.
+- `sketch_adata(..., return_adata=True, copy=False)` returns an in-memory view
+  without attaching metadata. Backed subset returns require `copy=True`.
+  Sparse `use_rep="X"` is rejected; supply a dense reduced embedding instead.
+- Distances use SciPy directly and nearest-center evaluation tiles both axes.
+  Full input and block copies still require memory proportional to `N*d`;
+  optional swap refiners may allocate much larger matrices.
+
+After installing development dependencies, run `python -m pytest`. Install
+`.[dev,adata]` to include the optional AnnData integration regression test.

pros_sketch-0.1.0/README.md

@@ -0,0 +1,154 @@
+# PROS
+
+PROS (Partitioned and Refined Oversampling Sketches) selects a small,
+diversity-preserving subset of rows from a large real-valued coordinate matrix.
+It builds an oversampled candidate pool within partitions and performs one
+global refinement pass over that pool. The returned subset is represented by
+indices into the original matrix.
+
+The package is domain-agnostic. It can be used with reduced single-cell
+embeddings, as well as other feature matrices where Euclidean distance is a
+meaningful geometry.
+
+## Installation
+
+PROS requires Python 3.10 or newer. The distribution name is `pros-sketch`;
+the Python import remains `pros`. The commands below install this checkout.
+
+```bash
+python -m pip install -e .
+```
+
+Optional AnnData support:
+
+```bash
+python -m pip install -e ".[adata]"
+```
+
+For testing and linting:
+
+```bash
+python -m pip install -e ".[dev]"
+```
+
+## Quick start
+
+```python
+import numpy as np
+from pros import sketch
+
+X = np.random.default_rng(0).random((5_000, 50))
+result = sketch(X, n=500, seed=0)
+
+indices = result.indices
+X_sketch = X[indices]
+print(result.timings)
+```
+
+`result.indices` is a sorted, unique `int64` array of length `n`. By default,
+PROS uses k-means partitioning, water-filling allocation, an oversampling ratio
+of `r=10.0`, and global farthest-first refinement. Pass an explicit `seed` to
+make a stochastic configuration reproducible.
+
+## Certification
+
+Set `certify=True` to compute the final covering radius, the candidate-pool
+radius, and bounds derived from a full-data farthest-first traversal:
+
+```python
+result = sketch(X, n=500, seed=0, certify=True)
+
+print(result.radius)       # R(X, S): final covering radius
+print(result.rho)          # R(X, P): candidate-pool covering radius
+print(result.opt_lower)    # lower bound on the optimal n-centre radius
+print(result.ratio_upper)  # upper bound on R(X, S) / OPT_n(X)
+```
+
+Certification performs additional full-data work and is not included in
+`result.timings["total"]`. Degenerate data may have `opt_lower == 0`; in that
+case a finite approximation-ratio bound is not defined.
+
+For an existing set of indices, use `pros.certificate`. `pros.opt_bounds`,
+`pros.estimate_opt_scale`, and `pros.choose_r` are available for certificate
+and oversampling-ratio workflows.
+
+## AnnData
+
+```python
+from pros import sketch_adata
+
+result = sketch_adata(adata, n=5_000, use_rep="X_pca", seed=0)
+subset = adata[result.indices].copy()
+```
+
+`sketch_adata` also accepts an `.h5ad` path and reads it in backed mode before
+extracting the requested representation.
+
+## Configuration
+
+```python
+result = sketch(
+    X,
+    n=1_000,
+    partitioner="kmeans",
+    n_blocks="auto",
+    allocator="water_filling",
+    r=10.0,
+    selector="fft",
+    refiner="fft",
+    seed=0,
+)
+```
+
+The `partitioner`, within-partition `selector`, `allocator`, and global
+`refiner` are modular. See [`docs/configuration.md`](docs/configuration.md) for
+the supported values and [`docs/theory.md`](docs/theory.md) for the covering
+objective and certificate definitions.
+
+## Repository layout
+
+- `pros/`: installable library code.
+- `tests/`: automated package tests.
+- `examples/`: small runnable examples using synthetic data in `data/`.
+- `test_para/`: standalone parameter experiments and their generated outputs;
+  this reproduction material is intentionally not packaged with the library.
+- `docs/`: MkDocs source pages.
+
+Run the examples from the repository root after installation:
+
+```bash
+python examples/01_basic_numpy.py
+python examples/03_certification.py
+```
+
+## Citation
+
+See [`CITATION.cff`](CITATION.cff) for software citation metadata.
+
+## License
+
+PROS is distributed under the [MIT License](LICENSE).
+
+## Validated behavior and limits
+
+- Water-filling is fused with FFT and requires `selector="fft"`. Choose another
+  allocator to use random or scSampler selection. Missing or broken scSampler
+  installations raise an error; no alternate algorithm is substituted.
+- `mix>0` currently requires `refiner="fft"` so reserved points are retained.
+  `refiner="none"` uniformly downsamples the candidate pool to `n` rows.
+- `certificate` computes the a posteriori ratio for any valid sketch. Its
+  `theory_bound` and `bound_slack` are NaN unless FFT refinement is explicitly
+  asserted; `sketch` sets this assertion only for unmixed FFT refinement.
+  Caches are tied to the exact data and sketch size. Floating-point results
+  are numerical bounds, not interval-arithmetic certificates.
+- Internal stage timings exclude input conversion, validation, AnnData I/O,
+  and certification. Measure an external wall clock for runtime comparisons.
+- `sketch_adata(..., return_adata=True, copy=False)` returns an in-memory view
+  without attaching metadata. Backed subset returns require `copy=True`.
+  Sparse `use_rep="X"` is rejected; supply a dense reduced embedding instead.
+- Distances use SciPy directly and nearest-center evaluation tiles both axes.
+  Full input and block copies still require memory proportional to `N*d`;
+  optional swap refiners may allocate much larger matrices.
+
+After installing development dependencies, run `python -m pytest`. Install
+`.[dev,adata]` to include the optional AnnData integration regression test.

pros_sketch-0.1.0/data/README.md

@@ -0,0 +1,13 @@
+# Example Data
+
+`synthetic_clusters.npz` contains a small synthetic coordinate matrix for
+examples and smoke tests.
+
+Fields:
+
+- `X`: `(900, 8)` float64 coordinates.
+- `labels`: `(900,)` integer cluster labels.
+
+The dataset is generated from Gaussian clusters and is not intended as a
+scientific benchmark.
+

pros_sketch-0.1.0/docs/api.md

@@ -0,0 +1,16 @@
+# API
+
+Public API:
+
+- `pros.sketch(X, n, **kwargs)`: sketch a NumPy-compatible coordinate matrix.
+- `pros.sketch_adata(adata, n, use_rep="X_pca", **kwargs)`: sketch an AnnData
+  embedding or `.h5ad` path.
+- `pros.certificate(X, sketch_indices, pool_indices=None)`: compute a
+  certificate for an existing sketch.
+- `pros.choose_r(X, n, tol=0.5)`: choose an oversampling ratio from a grid.
+- `pros.estimate_opt_scale(X, n)`: estimate a target covering scale on a
+  subsample.
+
+Lower-level implementation modules are intentionally not re-exported from
+`pros`. They may be useful for local experiments, but are not part of the
+stable public API.

pros_sketch-0.1.0/docs/configuration.md

@@ -0,0 +1,48 @@
+# Configuration
+
+The default configuration is intended to be a strong starting point:
+
+```python
+res = sketch(
+    X,
+    n=1000,
+    partitioner="kmeans",
+    n_blocks="auto",
+    allocator="water_filling",
+    r=10.0,
+    selector="fft",
+    refiner="fft",
+    seed=0,
+)
+```
+
+Useful options:
+
+- `partitioner`: `"kmeans"`, `"pc_tree"`, `"random"`, or `"none"`.
+- `allocator`: `"water_filling"`, `"proportional"`, `"power"`, `"volume"`, or
+  `"uniform"`.
+- `selector`: `"fft"`, `"scsampler_maximin"`, or `"random"`.
+- `refiner`: `"fft"`, `"maximin"`, `"local_swap"`, or `"none"`.
+- `r`: oversampling ratio. Larger values make a denser pool and cost more.
+- `certify`: compute certificate fields for evaluation.
+
+For a single global FFT selection, use `partitioner="none", r=1,
+refiner="none"`: stage 1 selects n points and stage 2 retains the whole pool.
+With r>1, the single-block configuration still runs two selection stages.
+
+Water-filling requires `selector="fft"`; incompatible combinations raise an
+error. Other allocators support the other selectors. scSampler failures are
+propagated, never replaced silently with FFT. Fixed seeds are reproducible
+within a fixed numerical/software environment; record dependency versions and
+thread settings for experiments.
+
+`mix>0` requires FFT refinement to preserve the uniform reserve. With
+`mix_source="data"`, reserve points may enlarge the pool beyond the stage-1
+`budget`; `m_per_block` still describes stage 1. `block_radii` and
+`rho_blockwise_max` are available only with water-filling. An unfunded nonempty
+block has infinite local covering radius.
+
+The pool budget is `min(ceil(r*n), N)`. Larger r need not monotonically improve
+the greedy final radius. `choose_r` reports `tolerance_met` and certifies only
+its returned sketch. Its pool-radius evaluations can still be expensive.
+

pros_sketch-0.1.0/docs/examples.md

@@ -0,0 +1,18 @@
+# Examples
+
+The `examples/` directory contains small runnable scripts:
+
+- `01_basic_numpy.py`: basic matrix sketching.
+- `02_ann_data.py`: AnnData workflow.
+- `03_certification.py`: certificate fields.
+- `04_compare_configs.py`: simple configuration comparison.
+
+Run an example from the repository root:
+
+```bash
+python examples/01_basic_numpy.py
+```
+
+The checkout examples use `partitioner="pc_tree"` to keep their behaviour
+simple. The installed package includes `scikit-learn`, which is required by
+the default `partitioner="kmeans"` workflow.

pros_sketch-0.1.0/docs/index.md

@@ -0,0 +1,10 @@
+# PROS
+
+PROS builds diversity-preserving sketches of large coordinate matrices. It
+first partitions the data, oversamples a candidate pool inside those blocks,
+and then refines globally over the pool.
+
+The returned sketch is an array of row indices into the original matrix. For
+evaluation runs, PROS can also attach a certificate with the sketch covering
+radius, pool covering radius, and a proven approximation-ratio upper bound.
+

pros_sketch-0.1.0/docs/installation.md

@@ -0,0 +1,19 @@
+# Installation
+
+Install from a local checkout:
+
+```bash
+python -m pip install -e .
+```
+
+Optional extras:
+
+```bash
+python -m pip install -e ".[adata]"
+python -m pip install -e ".[dev]"
+python -m pip install -e ".[docs]"
+```
+
+The core package depends on `numpy` and `scikit-learn`. AnnData and Scanpy are
+optional so matrix-only users do not need the single-cell stack.
+

pros_sketch-0.1.0/docs/publishing.md

@@ -0,0 +1,61 @@
+# Publishing pros-sketch
+
+The distribution is `pros-sketch`, while the Python import is `pros`.
+The workflow `.github/workflows/publish.yml` publishes to PyPI only when a
+GitHub Release is published. Manual workflow runs perform validation and build
+artifacts without uploading. No PyPI API token secret is required.
+
+## One-time PyPI setup
+
+For the first release, sign in to https://pypi.org/manage/account/publishing/
+and add a pending publisher under GitHub:
+
+| Field | Value |
+| --- | --- |
+| PyPI Project Name | `pros-sketch` |
+| Owner | `LeiLi-Uchicago` |
+| Repository name | `PROS` |
+| Workflow name | `publish.yml` |
+| Environment name | `pypi` |
+
+Use the workflow filename, not its display title or full path. A pending
+publisher does not reserve the project name; successful first upload creates
+it. If this project already belongs to your account, add the publisher from
+the project's Publishing settings instead.
+
+In GitHub repository Settings → Environments, create the `pypi` environment.
+Its name must match both the workflow and PyPI configuration. Optional review
+and deployment restrictions can be configured there.
+
+## Release procedure
+
+1. Commit and push the reviewed code and publishing workflow to GitHub.
+2. Run **Publish to PyPI** manually from Actions. This tests Python 3.10–3.12,
+   builds the sdist and wheel, checks metadata, and tests the installed wheel
+   outside the checkout. Confirm the run succeeds.
+3. Ensure `pyproject.toml` and `pros/__init__.py` specify the same release
+   version. For the first release this is `0.1.0`.
+4. Create a GitHub Release with tag `v0.1.0` pointing to the validated commit.
+   Publishing it triggers the workflow, repeats validation, and uploads the
+   built distributions through Trusted Publishing. A draft release does not.
+5. Verify the project page and install `pros-sketch==0.1.0` in a new environment.
+   For AnnData support, install `pros-sketch[adata]==0.1.0`.
+
+For later releases, update both version values and use the corresponding tag.
+PyPI does not allow overwriting previously uploaded distribution filenames.
+If an upload fails, inspect the logs and which artifacts reached PyPI before
+retrying; this workflow intentionally does not hide existing-file errors.
+
+TestPyPI is not configured by this workflow. It requires a separate account
+and trusted publisher if a rehearsal upload is wanted.
+
+## Scope of validation
+
+The workflow tests core and AnnData functionality. The scSampler adapter is
+covered by contract tests, not an end-to-end installation of scSampler. Large
+benchmark datasets and manuscript runtime reproduction are not run in CI.
+
+References:
+- https://docs.pypi.org/trusted-publishers/creating-a-project-through-oidc/
+- https://docs.pypi.org/trusted-publishers/adding-a-publisher/
+- https://github.com/pypa/gh-action-pypi-publish

pros_sketch-0.1.0/docs/quickstart.md

@@ -0,0 +1,30 @@
+# Quickstart
+
+```python
+import numpy as np
+from pros import sketch
+
+rng = np.random.default_rng(0)
+X = rng.random((5000, 50))
+
+res = sketch(X, n=500, seed=0)
+X_sketch = X[res.indices]
+```
+
+The result is dict-like and also supports attribute access:
+
+```python
+print(res["indices"])
+print(res.indices)
+print(res.timings)
+```
+
+For AnnData:
+
+```python
+from pros import sketch_adata
+
+res = sketch_adata(adata, n=5000, use_rep="X_pca")
+subset = adata[res.indices]
+```
+