Different result everytime I run snap.tl.spectral

[MathildaStigenberg]

Hi,
The result looks different everytime I run snap.tl.spectral on the same dataset. I run using this command: snap.tl.spectral(data, random_state=123). How can I get different result even though I have the same random_state? I want to have reproducible results if I have to re-run the tutorial.

Thanks in advance!

Best,
Mathilda

[kaizhang]

The embedding results should be consistent across different runs. We have a test case for this:

def test_reproducibility(mat):
    adata = ad.AnnData(X=csr_matrix(mat))
    embeddings = []
    for _ in range(3):
        embeddings.append(snap.tl.spectral(adata, features=None, random_state=0, inplace=False)[1])
    for x in embeddings:
        np.testing.assert_array_equal(x, embeddings[0])

But the UMAP embedding is not reproducible.

[MathildaStigenberg]

Hi,

Thanks for the help, much appreciated!
If I choose all features, then it is consistent. However, if I want to only choose the accessible features (from snap.pp.select_features) with the parameter features = "selected" in the snap.tl.spectral function, it is not consistent anymore. Why is that?

Furthermore, is it possible to make the UMAP embedding reproducible?

[MathildaStigenberg]

I try different values for the parameter n_features in snap.pp.select_features and then use the function you wrote. For some values, eg. 6000 or 15000, the spectral output is consistent, however for other values, eg. 4000 or 10000, the spectral output is not consistent.

[kaizhang]

I cannot reproduce the error based on what have described. If you can share the test data and code, I can see if I can reproduce the error and fix it.