GPU acceleration

[Drakula44]

Hi. I have working on my project when I wanted to use compute_svd_entropy but it was taking too much time.
I stumbled upon a library cupy that is replacement for numpy for running on cuda.
I modified locally compute_svd_entropy like this:

def compute_svd_entropy(data, tau=2, emb=10, n_jobs=1):
...
    import numpy as np
    if n_jobs == "cuda":
        import cupy as cp
        np = cp
        data = cp.asarray(data)
    _, sv, _ = np.linalg.svd(_embed(data, d=emb, tau=tau, n_jobs=n_jobs))
    m = np.sum(sv, axis=-1)
    sv_norm = np.divide(sv, m[:, None])
    out = -np.sum(np.multiply(sv_norm, np.log2(sv_norm)), axis=-1)
    if n_jobs == "cuda":
        out = cp.asnumpy(out)
    return out

And for my case it speeded up from 37s per chunk that I was passing to 3-4s per chunk.
I can create PR for that for every function that is applicable will you be willing to merge something like that. I know that it introduces one more dependency (it can be optional) so I wanted just to check first before submitting PR. Thanks in advance.

PS. n_jobs was motivated by the mne that is using n_jobs in that way.

[agramfort]

you can look how sklearn does it with array API. See https://github.com/scikit-learn/scikit-learn/blob/main/sklearn/discriminant_analysis.py#L108 with the get_namespace function

…

On Fri, Aug 9, 2024 at 6:18 PM Drakula44 ***@***.***> wrote: Hi. I have working on my project when I wanted to use compute_svd_entropy but it was taking too much time. I stumbled upon a library cupy that is replacement for numpy for running on cuda. I modified locally compute_svd_entropy like this: def compute_svd_entropy(data, tau=2, emb=10, n_jobs=1): ... import numpy as np if n_jobs == "cuda": import cupy as cp np = cp data = cp.asarray(data) _, sv, _ = np.linalg.svd(_embed(data, d=emb, tau=tau, n_jobs=n_jobs)) m = np.sum(sv, axis=-1) sv_norm = np.divide(sv, m[:, None]) out = -np.sum(np.multiply(sv_norm, np.log2(sv_norm)), axis=-1) if n_jobs == "cuda": out = cp.asnumpy(out) return out And for my case it speeded up from 37s per chunk that I was passing to 3-4s per chunk. I can create PR for that for every function that is applicable will you be willing to merge something like that. I know that it introduces one more dependency (it can be optional) so I wanted just to check first before submitting PR. Thanks in advance. PS. n_jobs was motivated by the mne that is using n_jobs in that way. — Reply to this email directly, view it on GitHub <#96>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AABHKHF2XOPFUKPBW5HYOG3ZQTTVFAVCNFSM6AAAAABMIYFNPSVHI2DSMVQWIX3LMV43ASLTON2WKOZSGQ2TQMRVHE4DENA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[Drakula44]

Thank you for suggestion.

I looked into it a bit and dont think that there is much currently to do.
I looked into other functions and saw that they are not that straight forward as this one (I got quite lucky).
So in my opinion sklearn array API should mature a bit before it could be implemented here as I saw mne_features relays on sklern.

I still can make PR for this function or if some similar but dont think there is currently need for that?

[agramfort]

the clean way for me would be to rely on sklearn backend system. a constraint: - mne-features should still work if you don't have cupy or pytorch. Message ID: ***@***.***>

…