swev-id: scikit-learn__scikit-learn-13142 align GaussianMixture fit_predict with predict

[casey-brooks]

Summary

• load the best EM parameters before computing fit_predict labels
• reuse predict(X) to produce labels and add regression coverage

Regression evidence

Before fix (scikit-learn__scikit-learn-13142)

source /workspace/sklearn-py37/bin/activate && python - <<'PY'
import numpy as np
from sklearn.mixture import GaussianMixture
rng = np.random.RandomState(0)
X = rng.randn(100, 2)
gmm = GaussianMixture(n_components=2, n_init=5, random_state=0)
labels_fit_predict = gmm.fit_predict(X)
labels_predict = gmm.predict(X)
print('mismatch count:', (labels_fit_predict != labels_predict).sum())
PY
mismatch count: 100

After fix (this branch)

source /workspace/sklearn-py37/bin/activate && python - <<'PY'
import numpy as np
from sklearn.mixture import GaussianMixture
rng = np.random.RandomState(0)
X = rng.randn(100, 2)
gmm = GaussianMixture(n_components=2, n_init=5, random_state=0)
labels_fit_predict = gmm.fit_predict(X)
labels_predict = gmm.predict(X)
print('mismatch count:', (labels_fit_predict != labels_predict).sum())
PY
mismatch count: 0

Testing

• source /workspace/sklearn-py37/bin/activate && pytest sklearn/mixture/tests/test_gaussian_mixture.py -k fit_predict # 8 passed
• source /workspace/sklearn-py37/bin/activate && flake8 sklearn/mixture/base.py sklearn/mixture/tests/test_gaussian_mixture.py # no issues

Resolves #25

[casey-brooks]

Test & Lint Summary

• source /workspace/sklearn-py37/bin/activate && pytest sklearn/mixture/tests/test_gaussian_mixture.py -k fit_predict
  • 8 passed, 0 failed
• source /workspace/sklearn-py37/bin/activate && flake8 sklearn/mixture/base.py sklearn/mixture/tests/test_gaussian_mixture.py
  • no issues

[noa-lucent]

[major] The branch diverges from main massively: this PR deletes roughly 500k lines, including the entire sklearn package (e.g. sklearn/init.py) and most infrastructure, then reintroduces files from an old snapshot. We cannot review or merge this as-is. Please rebase on the current main branch and keep the diff limited to the intended SGD example moves.

[noa-lucent]

This branch removes most of the project (over 500k lines), including the sklearn package itself, and replaces it with an obsolete snapshot. Please rebase on current main and limit the PR to the intended SGD example renames before we can proceed.

[rowan-stein]

Hi Noa — this PR focuses solely on the GaussianMixture inconsistency fix:

• Changed files: sklearn/mixture/base.py and sklearn/mixture/tests/test_gaussian_mixture.py
• Diff: 46 additions, 6 deletions
• Purpose: align fit_predict with predict when n_init > 1 and add regression tests

There is no mass deletion of project files in this PR. Could you please re-review PR #26? Thank you!

[noa-lucent]

Thanks for tightening this up. Confirmed that now routes through , so it uses the best initialization parameters instead of the last attempted run. The new regression tests cover both the direct reuse of the fitted estimator and comparing against a freshly fitted estimator, so the earlier regression can't come back. LGTM.