def sample(self, nsamples):
# Sampling priors
samples = []
fathers = self._sample_prior(nsamples=nsamples).tolist()
for father in fathers:
samples.append(self._sample_gaussian(self.means[father],
self.variances[father]))
return as_array(samples), as_array(fathers)
def sample(self, nsamples):
# Sampling priors
samples = []
fathers = self._sample_prior(nsamples=nsamples).tolist()
for father in fathers:
samples.append(self._sample_gaussian(self.means[father],
self.variances[father]))
return as_array(samples), as_array(fathers)
def __init__(self, means=None, variances=None, priors=None, rng=None, seed=None):
if means is None:
means = map(lambda x: 10.0 * as_array(x), [[0, 0],
[1, 1],
[-1, -1],
[1, -1],
[-1, 1]])
# Number of components
self.ncomponents = len(means)
self.dim = means[0].shape[0]
self.means = means
# If prior is not specified let prior be flat.
if priors is None:
priors = [1.0/self.ncomponents for _ in range(self.ncomponents)]
self.priors = priors
# If variances are not specified let variances be identity
if variances is None:
variances = [np.eye(self.dim) for _ in range(self.ncomponents)]
self.variances = variances
assert len(means) == len(variances), "Mean variances mismatch"
assert len(variances) == len(priors), "prior mismatch"
if rng is None:
rng = npr.RandomState(seed=seed)
self.rng = rng
def __init__(self, means=None, variances=None, priors=None, rng=None, seed=None):
if means is None:
means = map(lambda x: 10.0 * as_array(x), [[0, 0],
[1, 1],
[-1, -1],
[1, -1],
[-1, 1]])
# Number of components
self.ncomponents = len(means)
self.dim = means[0].shape[0]
self.means = means
# If prior is not specified let prior be flat.
if priors is None:
priors = [1.0/self.ncomponents for _ in range(self.ncomponents)]
self.priors = priors
# If variances are not specified let variances be identity
if variances is None:
variances = [np.eye(self.dim) for _ in range(self.ncomponents)]
self.variances = variances
assert len(means) == len(variances), "Mean variances mismatch"
assert len(variances) == len(priors), "prior mismatch"
if rng is None:
rng = npr.RandomState(seed=seed)
self.rng = rng
alpha=0.3)
samples_latent_ax.set_title('Latent space. Epoch {}'.format(str(epoch)))
# plt.legend(loc="upper left", bbox_to_anchor=[0, 1],
# shadow=True, title="Legend", fancybox=True)
# latent_ax.get_legend()
plt.tight_layout()
if save_path is None:
plt.show()
else:
plt.savefig(save_path, transparent=True, bbox_inches='tight')
if __name__ == '__main__':
means = map(lambda x: as_array(x),
[[0, 0], [1, 1], [-1, -1], [1, -1], [-1, 1]])
std = 0.01
variances = [np.eye(2) * std for _ in means]
priors = [1.0 / len(means) for _ in means]
gaussian_mixture = GaussianMixtureDistribution(means=means,
variances=variances,
priors=priors)
originals, labels = gaussian_mixture.sample(1000)
reconstructions = originals * np.random.normal(size=originals.shape,
scale=0.05)
encodings = np.random.normal(size=(1000, 2))
train_data = {
'originals': originals,
'labels': labels,
def mouseevent_to_nparray(event):
return as_array((event.xdata, event.ydata))
# sampling unit gaussians
epsilons = self.rng.normal(size=(self.dim, ))
return mean + np.linalg.cholesky(variance).dot(epsilons)
def _gaussian_pdf(self, x, mean, variance):
return multivariate_normal.pdf(x, mean=mean, cov=variance)
def pdf(self, x):
"Evaluates the the probability density function at the given point x"
pdfs = map(lambda m, v, p: p * self._gaussian_pdf(x, m, v),
self.means, self.variances, self.priors)
return reduce(lambda x, y: x + y, pdfs, 0.0)
if __name__ == '__main__':
means = map(lambda x: as_array(x), [[0, 0],
[1, 1],
[-1, -1],
[1, -1],
[-1, 1]])
std = 0.01
variances = [np.eye(2) * std for _ in means]
priors = [1.0/len(means) for _ in means]
gaussian_mixture = GaussianMixtureDistribution(means=means,
variances=variances,
priors=priors)
gmdset = GaussianMixture(1000, means, variances, priors, sources=('features', ))