def testReparameterizeDiag(self):
S, N, D = 4, 3, 2
mean = np.random.randn(S, N, D)
var = np.random.randn(S, N, D)**2
z = np.random.randn(S, N, D)
f = mean + z * (var + 1e-6)**0.5
with tf.Session() as sess:
assert_allclose(
f, sess.run(reparameterize(tf.identity(mean), var, z)))
def sample_from_conditional(self, X, z=None, full_cov=False):
"""
Calculates self.conditional and also draws a sample
If z=None then the tensorflow random_normal function is used to generate the
N(0, 1) samples, otherwise z are used for the whitened sample points
:param X: Input locations (S,N,D_in)
:param full_cov: Whether to compute correlations between outputs
:param z: None, or the sampled points in whitened representation
:return: mean (S,N,D), var (S,N,N,D or S,N,D), samples (S,N,D)
"""
mean, var = self.conditional(X, full_cov=full_cov)
if z is None:
z = tf.random_normal(tf.shape(mean), dtype=settings.float_type)
samples = reparameterize(mean, var, z, full_cov=full_cov)
return samples, mean, var
def sample_from_conditional(self, X, z=None, full_cov=False):
"""
Calculates self.conditional and also draws a sample, adding input propagation if necessary
If z=None then the tensorflow random_normal function is used to generate the
N(0, 1) samples, otherwise z are used for the whitened sample points
:param X: Input locations (S,N,D_in)
:param full_cov: Whether to compute correlations between outputs
:param z: None, or the sampled points in whitened representation
:return: mean (S,N,D), var (S,N,N,D or S,N,D), samples (S,N,D)
"""
mean, var = self.conditional_SND(X, full_cov=full_cov)
# set shapes
S = tf.shape(X)[0]
N = tf.shape(X)[1]
D = self.num_outputs
mean = tf.reshape(mean, (S, N, D))
if full_cov:
var = tf.reshape(var, (S, N, N, D))
else:
var = tf.reshape(var, (S, N, D))
if z is None:
z = tf.random_normal(tf.shape(mean), dtype=settings.float_type)
samples = reparameterize(mean, var, z, full_cov=full_cov)
if self.input_prop_dim:
shape = [tf.shape(X)[0], tf.shape(X)[1], self.input_prop_dim]
X_prop = tf.reshape(X[:, :, :self.input_prop_dim], shape)
samples = tf.concat([X_prop, samples], 2)
mean = tf.concat([X_prop, mean], 2)
if full_cov:
shape = (tf.shape(X)[0], tf.shape(X)[1], tf.shape(X)[1],
tf.shape(var)[3])
zeros = tf.zeros(shape, dtype=settings.float_type)
var = tf.concat([zeros, var], 3)
else:
var = tf.concat([tf.zeros_like(X_prop), var], 2)
return samples, mean, var
def testReparameterizeFullCov(self):
S, N, D = 4, 3, 2
mean = np.random.randn(S, N, D)
U = np.random.randn(S, N, N, D)
var = np.einsum('SnNd,SmNd->Snmd', U,
U) + np.eye(N)[None, :, :, None] * 1e-6
var_flat = np.reshape(np.transpose(var, [0, 3, 1, 2]), [S * D, N, N])
L_flat = np.linalg.cholesky(var_flat + np.eye(N)[None, :, :] * 1e-6)
L = np.transpose(np.reshape(L_flat, [S, D, N, N]), [0, 2, 3, 1])
z = np.random.randn(S, N, D)
f = mean + np.einsum('SnNd,SNd->Snd', L, z)
with tf.Session() as sess:
assert_allclose(
f,
sess.run(
reparameterize(tf.identity(mean), var, z, full_cov=True)))