Som estimator seed

[TarikExner]

Hey,

thank you very much for this great implementation!

When running two consecutive runs of the FlowSOMEstimator.fit_predict() method, the SOM codes and therefore cluster annotations were different.

Reproducible example (fresh conda environment using python=3.10):

from flowsom.models import FlowSOMEstimator
import numpy as np

arr = np.random.randint(0,1000,2000).reshape(200,10)
fse = FlowSOMEstimator(cluster_kwargs = {"xdim": 10, "ydim": 10, "seed": 1}, metacluster_kwargs = {"n_clusters": 15})
fse2 = FlowSOMEstimator(cluster_kwargs = {"xdim": 10, "ydim": 10, "seed": 1}, metacluster_kwargs = {"n_clusters": 15})
c1 = fse.fit_predict(arr)
c2 = fse2.fit_predict(arr)
all(c1 == c2)
>>> False

The bug was caused by a missing numpy seed for the random selection of data points in the code initialization step.

This PR fixes this issue by introducing a seed-set if SOMEstimator.seed is not None.

I intentionally did not remove the call to numpy.random.seed in the SOM function due to differences in setting seeds in numba environments as described here. In fact, removal of the seed setting in the SOM function also results in above mentioned error, even with the proposed commits.

This commit also includes tests for the SOMEstimator class and the FlowSOMEstimator class for reproducibility.

Please let me know if there is anything missing from this commit.

Best,
Tarik

[berombau]

Hi Tarik,

Thank you for investigating this issue. I added your fix and test functions and merged them with some changes to passing the input parameters. This is an important fix, so I'll merge it to main. We're working on some parallellization improvements the coming months, which will likely also change the Numba code and the random number generation (e.g. like in https://blog.scientific-python.org/numpy/numpy-rng/).

Best,
Benjamin

[TarikExner]

Thank you very much for the quick merge!

Best,
Tarik