Description

GaussianMixture.sample(n) occasionally returns incorrectly-sized samples possibly due to rounding error.

Steps/Code to Reproduce

from sklearn.mixture import GaussianMixture
import numpy as np

np.random.seed(1011)

X = np.random.normal(size=(100,20))

mixture_model = GaussianMixture(n_components=10).fit(X)

Expected Results

I'd expect that if n_samples are requested, n_samples are provided, even when n_samples is not an even multiple of the number of mixture distributions.

Actual Results

Instead, it appears that my expectations are confounded in Line 388 of mixture/base.py, since

np.round(model.weights * n_samples).astype(int)

will not necessarily sum to n, depending on the weights and samples.

samples = mixture_model.sample(100)
print(samples[0].shape)
#prints: (100,20)
samples = mixture_model.sample(104)
print(samples[0].shape)
#prints: (103, 20)
samples = mixture_model.sample(1002)
print(samples[0].shape)
#prints: (1000,20)

Suggested Patch

I could be mistaken on this, but I think that the correct way to sample n_samples from a mixture distribution with frequency vector weights_ is to draw a count vector, np.random.multinomial(n_samples, weights_), and then draw the corresponding number of samples from each component distribution as comes out of the multinomial draw.

This would involve a one-line change to mixture/base.py replacing line 388 referenced above with:

rng.multinomial(n_samples, self.weights).astype(int)

Versions

>>> import platform; print(platform.platform())
Linux-4.7.6-1-ARCH-x86_64-with-arch
>>> import sys; print("Python", sys.version)
Python 3.5.2 |Continuum Analytics, Inc.| (default, Jul  2 2016, 17:53:06)
[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)]
>>> import numpy; print('NumPy', numpy.__version__)
NumPy 1.11.2
>>> import scipy; print('SciPy', scipy.__version__)
SciPy 0.18.1
>>> import sklearn; print('Scikit-Learn', sklearn.__version__)
Scikit-Learn 0.18