ENH add sparse_matmul_to_dense #31952
Conversation
A little benchmarkSummary: Roughly up to a factor of 2 as compared to sparse-sparse A @ B, never worse. I guess that the csc @ csc case could be improved by constructing Detailsfrom time import time
import numpy as np
import scipy.sparse as sp
from sklearn.utils.sparsefuncs import sparse_matmul_to_dense
def bench(n1=100, n2=100, n3=100, sparseness=0.5, rng=123):
rng = np.random.default_rng(rng)
a_dense = rng.standard_normal((n1, n2))
b_dense = rng.standard_normal((n2, n3))
p = [1 - sparseness, sparseness]
a_dense.flat[rng.choice([False, True], size=n1 * n2, p=p)] = 0
b_dense.flat[rng.choice([False, True], size=n2 * n3, p=p)] = 0
t0 = time()
a_dense @ b_dense
t = time() - t0
print(f"dense @ dense = {t}s")
for af in ("csr", "csc"):
a = sp.csr_array(a_dense) if af == "csr" else sp.csc_array(a_dense)
for bf in ("csr", "csc"):
b = sp.csr_array(b_dense) if bf == "csr" else sp.csc_array(b_dense)
t0 = time()
a @ b
t = time() - t0
print(f"{af} @ {bf} time = {t}s")
t0 = time()
sparse_matmul_to_dense(a, b)
t = time() - t0
print(f"sparse_matmul_to_dense time = {t}s\n") |
* csr_matmul_csc_to_dense does not improve performance vs. converting to csr
78d79e6 to
7bc3bd1
Compare
OmarManzoor
left a comment
There was a problem hiding this comment.
Thank you for the PR @lorentzenchr. I left a few comments
doc/whats_new/upcoming_changes/sklearn.utils/31952.efficiency.rst
Outdated
Show resolved
Hide resolved
OmarManzoor
left a comment
There was a problem hiding this comment.
LGTM. Thank you @lorentzenchr
| ) | ||
|
|
||
|
|
||
| def sparse_matmul_to_dense(A, B, out=None): |
There was a problem hiding this comment.
Do you see using out somewhere out side of testing code?
There was a problem hiding this comment.
So far, the only call in non-tests is via safe_sparse_dot - on purpose. safe_sparse_dot does not have an out parameter. But there are a few places where it could be used, e.g. in sandwich_dot inside linear_models._linear_loss.py.
| n1, n3 = n3, n1 | ||
| else: | ||
| # It seems best to just convert to csr. | ||
| A = A.tocsr() |
There was a problem hiding this comment.
Doe this use more memory than the implementation on main?
(Same question for the B conversion below: B = B.tocsr())
There was a problem hiding this comment.
Scipy (at least for CSR and CSC) converts the "other" matrix/array to the same format, here https://github.com/scipy/scipy/blob/f762ab3dddf1541da7475580f16d5a4b8da31fea/scipy/sparse/_compressed.py#L432C32-L432C37
Therefore the code of this PR uses the same amount of memory as pure scipy A @ B, only difference is the returned dense array.
Reference Issues/PRs
Fixes #516.
What does this implement/fix? Explain your changes.
This adds a dedicated Cython routine to compute
dense_C = sparse_A @ sparse_B.Any other comments?