[KohakuKirisame]

Add VanillaSort as an external sorter through its public vanillasort.sort() API (requires VanillaSort >= 0.2.1). The wrapper provides lightweight discovery/version checks, 17 serializable parameters, optional CUDA, progress logging, and reloadable multisegment results. Python gate_fn callbacks remain available through the direct VanillaSort API.

Detailed parameters and examples and wrapper-specific regression tests are maintained in VanillaSort. SpikeInterface retains the shared SorterCommonTestSuite subclass, its standard installation entry, and external documentation links.

Validation:

• Shared sorter suite: 2 passed using the unchanged 60-second, 25 kHz fixture and published checkpoints on CPU.
• VanillaSort package/integration suite: 21 passed, including direct/wrapper equivalence, saved results/logs, parameter forwarding, fixed-K policies, and 20/25/30 kHz round trips.
• Focused Sphinx build, parameter/default consistency, Black/Ruff, and distribution checks passed.

The 25 kHz common-suite run used VanillaSort main at 46970dc. Its removal of the fixed 30 kHz check is unreleased; PyPI 0.2.1 retains that earlier check.

Repository · Models · Paper

[zm711]

Few comments to improve performance of our package overall as we add the features. Let me know if you need more guidance.

[alejoe91]

Hi @KohakuKirisame

I tested this on my laptop with a small GPU, but it seems to get stuck. I only see this output ofter more than one hour.

import spikeinterface.full as si

rec, sort = si.generate_ground_truth_recording(num_channels=64, durations=[60], sampling_frequency=30_000)
rec = rec.save(folder="rec_cache", n_jobs=-1, overwrite=True)
sorting_vanilla = si.run_sorter("vanillasort", rec, verbose=True, remove_existing_folder=True)

>>> Warning! The recording is already filtered, but vanillasort filter is enabled

Any idea what could be going wrong? I see my GPU being used...

[KohakuKirisame]

Hi @alejoe91 , the wrapper was redirecting all progress to a log even with verbose=True. Your example also processes 62 overlapping four-channel neighborhoods, each scanning the full recording, which makes inference expensive on smaller GPUs. I’ve updated the wrapper to display progress while retaining the log; new version of VanillaSort (0.2.0) provides the attention optimization and improved CUDA batch retries, try pip install --upgrade vanillasort. On Windows with an RTX 5080 (16 GB), Python 3.12.14, PyTorch 2.13.0+cu132, and SpikeInterface 0.105.1, your 64-channel/60-second example (fixed seed 0) completed in 279.90 s with VanillaSort 0.1.0 and 217.50 s with 0.2.0 (22.3% less time). All 62 neighborhoods and final outputs matched exactly, and CUDA retries under a restricted memory budget succeeded. I reproduced the hidden-progress issue, but not the hour-long runtime locally.

[alejoe91]

Hi @alejoe91 , the wrapper was redirecting all progress to a log even with verbose=True. Your example also processes 62 overlapping four-channel neighborhoods, each scanning the full recording, which makes inference expensive on smaller GPUs. I’ve updated the wrapper to display progress while retaining the log; new version of VanillaSort (0.2.0) provides the attention optimization and improved CUDA batch retries, try pip install --upgrade vanillasort. On Windows with an RTX 5080 (16 GB), Python 3.12.14, PyTorch 2.13.0+cu132, and SpikeInterface 0.105.1, your 64-channel/60-second example (fixed seed 0) completed in 279.90 s with VanillaSort 0.1.0 and 217.50 s with 0.2.0 (22.3% less time). All 62 neighborhoods and final outputs matched exactly, and CUDA retries under a restricted memory budget succeeded. I reproduced the hidden-progress issue, but not the hour-long runtime locally.

I tested on my laptop, with a 4GB NVIDIA GeForce RTX 3050...So the new version runs relatively quickly, but I tested now on an 8-channel recording with 10 units. The issue now is that vanilla sort outputs 131 units versus the 10 ground truth ones.

The ground-truth units are found, with some oversplits. But is the extra number of units expected? Is it controlled by the components method?

Regarding the verbose output, I still got no std out:

sorting_vanilla = si.run_sorter("vanillasort", rec, verbose=True, remove_existing_folder=True)

>>> Warning! The recording is already filtered, but vanillasort filter is enabled
vanillasort run time 150.47s
[/home/alessio/Documents/Codes/spike_sorting/spikeinterface/spikeinterface/src/spikeinterface/core/core_tools.py:947](http://localhost:8888/home/alessio/Documents/Codes/spike_sorting/spikeinterface/spikeinterface/src/spikeinterface/core/core_tools.py#line=946): UserWarning: The extractor is not serializable to file. The provenance will not be saved.
  warnings.warn("The extractor is not serializable to file. The provenance will not be saved.")

[KohakuKirisame]

Hi @alejoe91 , the wrapper was redirecting all progress to a log even with verbose=True. Your example also processes 62 overlapping four-channel neighborhoods, each scanning the full recording, which makes inference expensive on smaller GPUs. I’ve updated the wrapper to display progress while retaining the log; new version of VanillaSort (0.2.0) provides the attention optimization and improved CUDA batch retries, try pip install --upgrade vanillasort. On Windows with an RTX 5080 (16 GB), Python 3.12.14, PyTorch 2.13.0+cu132, and SpikeInterface 0.105.1, your 64-channel/60-second example (fixed seed 0) completed in 279.90 s with VanillaSort 0.1.0 and 217.50 s with 0.2.0 (22.3% less time). All 62 neighborhoods and final outputs matched exactly, and CUDA retries under a restricted memory budget succeeded. I reproduced the hidden-progress issue, but not the hour-long runtime locally.

I tested on my laptop, with a 4GB NVIDIA GeForce RTX 3050...So the new version runs relatively quickly, but I tested now on an 8-channel recording with 10 units. The issue now is that vanilla sort outputs 131 units versus the 10 ground truth ones.

The ground-truth units are found, with some oversplits. But is the extra number of units expected? Is it controlled by the components method?

Regarding the verbose output, I still got no std out:

sorting_vanilla = si.run_sorter("vanillasort", rec, verbose=True, remove_existing_folder=True)

>>> Warning! The recording is already filtered, but vanillasort filter is enabled
vanillasort run time 150.47s
[/home/alessio/Documents/Codes/spike_sorting/spikeinterface/spikeinterface/src/spikeinterface/core/core_tools.py:947](http://localhost:8888/home/alessio/Documents/Codes/spike_sorting/spikeinterface/spikeinterface/src/spikeinterface/core/core_tools.py#line=946): UserWarning: The extractor is not serializable to file. The provenance will not be saved.
  warnings.warn("The extractor is not serializable to file. The provenance will not be saved.")

Hi @alejoe91, components sets K independently for each four-channel neighborhood, so the combined output can contain excessive units and oversplits as the channel count increases. VanillaSort 0.2.1 is now available (pip install --upgrade vanillasort) and adds optional fixed-K merging. With the updated wrapper, you can pass target_units=10 through run_sorter(); postprocessing={"on_shortfall": "return_available"} allows fewer units when reliable anchors are insufficient. Adjusting components may also help, though these settings need tuning for each recording. I’ve updated the parameter documentation to explain all available options. Regarding stdout, please update your SpikeInterface checkout to the latest PR revision: 72a80d3 already forwards progress to the console with verbose=True. Updating VanillaSort alone does not update the wrapper.

[alejoe91]

Thanks @KohakuKirisame

Can you remove the vanillasort-specific tests and docs? Those should live in the vanillasort repo

[KohakuKirisame]

Thanks @alejoe91! I've updated the PR as suggested. The detailed documentation and sorter-specific tests now live in the VanillaSort repository, with links from SpikeInterface.

The test module now only inherits SorterCommonTestSuite, without custom fixtures or setUp overrides. Both common-suite tests pass against the current VanillaSort main using the default recording fixture and published checkpoints. All 21 VanillaSort package/integration tests also pass.