ENH reduce memory usage in make_blobs - #22412

I think that we can avoid adding a test to check the performance. Instead, @ogrisel and @jeremiedbb think that we could add such utility in the ASV benchmark suite.

@MaxwellLZH Could you still add an entry in the changelog to acknowledge the no-copy behaviour.

Thanks @MaxwellLZH

Hi @glemaitre @jeremiedbb , I noticed that we can also improve the memory usage for datasets.make_s_curve in a similar way:

# before
x = np.sin(t)
y = 2.0 * generator.uniform(size=(1, n_samples))
z = np.sign(t) * (np.cos(t) - 1)
X = np.concatenate((x, y, z))

# after
X = np.empty(shape=(n_samples, 3), dtype=np.float64)
X[:, 0] = np.sin(t)
X[:, 1] = 2.0 * generator.uniform(size=n_samples)
X[:, 2] = np.sign(t) * (np.cos(t) - 1)

I've tested on my machine this indeed uses less memory (from 694MiB to 541MiB when n_samples = 10000000), shall I open a separate PR for this ?

I've tested on my machine this indeed uses less memory (from 694MiB to 541MiB when n_samples = 10000000), shall I open a separate PR for this ?

Yes, it is better to do in a separate PR. Feel free to open one.