Closed
Description
While adding examples to docstrings, two models (three classes) show an odd
behavior, i.e. they give different results under different circumstances and they
are not random, since the random seed and rand_state are fixed.
The results are deterministic under each setting, but change from setup to
setup.
For instance (observed in PR #12124), the EllipticEnvelope, has the following
issue (this is a failiur on travis).
073 >>> import numpy as np
074 >>> from sklearn.covariance import EllipticEnvelope
075 >>> real_cov = np.array([[.8, .3],
076 ... [.3, .4]])
077 >>> np.random.seed(0)
078 >>> X = np.random.multivariate_normal(mean=[0, 0],
079 ... cov=real_cov,
080 ... size=300)
081 >>> cov = EllipticEnvelope(random_state=0).fit(X)
082 >>> cov.covariance_ # doctest: +ELLIPSIS
Expected:
array([[0.7411..., 0.2535...],
[0.2535..., 0.3053...]])
Got:
array([[0.81478325, 0.28653659],
[0.28653659, 0.30913504]])
The same issue was observed in PR #11732, for GraphicalLasso and
GraphicalLassoCV.
Please note that the results are deterministic, i.e. changing the values to
what's reported by travis, would make the test pass, as I've done for the
PR #11732 .
The corresponding code resulting in the issue, is the following:
import numpy as np
from scipy import linalg
from sklearn.datasets import make_sparse_spd_matrix
from sklearn.covariance import GraphicalLasso, log_likelihood
n_samples = 60
n_features = 20
prng = np.random.RandomState(1)
prec = make_sparse_spd_matrix(n_features, alpha=.98,
smallest_coef=.4,
largest_coef=.7,
random_state=prng)
cov = linalg.inv(prec)
d = np.sqrt(np.diag(cov))
cov /= d
cov /= d[:, np.newaxis]
X = prng.multivariate_normal(np.zeros(n_features), cov, size=n_samples)
emp_cov = np.dot(X.T, X) / n_samples
model = GraphicalLasso()
loglik_est = -model.fit(X).score(X)
loglik_real = -log_likelihood(emp_cov, prec)
print("estimated negative log likelihood: %g" % loglik_est)
[here the difference between systems is: 26.1847 vs 26.1927]
print("real negative log likelihood: %g" % loglik_real)
[here the difference between systems is: 28.1526 vs 28.1067]Metadata
Metadata
Assignees
Labels
No labels