def postX_from_chol(tc1, tc2, lm):
"""
postX = (Lambda1 + S)^{-1}.(Lambda1_ij.*Mu_j + X^T_k.*S_kj;i.*X_j)
"""
return blk_chol_inv(tc1,
tc2,
blk_chol_inv(tc1, tc2, lm),
lower=False,
transpose=True)
def symbsample_X(self, Y=None):
Y = self.Y if Y is None else Y
Nsamps, Tbins = Y.shape[0], Y.shape[1]
TheChol = theano.clone(self.TheChol, replace={self.Y : Y})
postX = theano.clone(self.postX, replace={self.Y : Y})
normSamps = srng.normal([Nsamps, Tbins, self.xDim])
noise, _ = theano.scan(lambda tc1, tc2, ns : blk_chol_inv(tc1, tc2, ns, lower=False, transpose=True),
sequences=(TheChol[0], TheChol[1], normSamps))
return postX + noise
def symbsample_Xalt(self, Y=None, X=None):
if Y is None: Y = self.Y
if X is None: X = self.lat_ev_model.get_X()
Xgen = self.lat_ev_model.get_X()
Nsamps, Tbins = Y.shape[0], Y.shape[1]
TheChol = theano.clone(self.TheChol, replace={self.Y : Y, Xgen : X})
postX = theano.clone(self.postX, replace={self.Y : Y, Xgen : X})
normSamps = srng.normal([Nsamps, Tbins, self.xDim])
noise, _ = theano.scan(lambda tc1, tc2, ns :
blk_chol_inv(tc1, tc2, ns, lower=False, transpose=True),
sequences=(TheChol[0], TheChol[1], normSamps))
return postX, noise
def sample_noise(self, Ydata, Xdata, Nsamps=1):
Xgen = self.lat_ev_model.get_X()
N, T = Xdata.shape[0], Xdata.shape[1]
normSamps = srng.normal([N, T, self.xDim])
TheChol = self.TheChol
noise, _ = theano.scan(lambda tc1, tc2, ns :
blk_chol_inv(tc1, tc2, ns, lower=False, transpose=True),
sequences=(TheChol[0], TheChol[1], normSamps))
noise_vals = np.zeros((Nsamps, N, T, self.xDim))
for samp in range(Nsamps):
noise_vals[samp] = noise.eval({self.Y : Ydata, Xgen : Xdata})
return noise_vals
def sample_postX(self):
"""
"""
Nsamps, NTbins, xDim = self.Nsamps, self.NTbins, self.xDim
prenoise_NxTxd = tf.random_normal([Nsamps, NTbins, xDim], dtype=DTYPE)
aux_fn = lambda _, seqs: blk_chol_inv(
seqs[0], seqs[1], seqs[2], lower=False, transpose=True)
noise = tf.scan(fn=aux_fn,
elems=[
self.TheChol_2xxNxTxdxd[0],
self.TheChol_2xxNxTxdxd[1], prenoise_NxTxd
],
initializer=tf.zeros_like(prenoise_NxTxd[0],
dtype=DTYPE))
noisy_postX_ng = tf.add(self.postX_ng_NxTxd, noise, name='noisy_postX')
noisy_postX = tf.add(self.postX_NxTxd, noise, name='noisy_postX')
return noisy_postX, noisy_postX_ng
def postX_from_chol(tc1, tc2, lm):
"""
postX = (Lambda1 + Lambda2)^{-1}.Lambda1.Mu
"""
return blk_chol_inv(tc1, tc2, blk_chol_inv(tc1, tc2, lm), lower=False, transpose=True)
def postX_from_chol(tc1, tc2, lm):
return blk_chol_inv(tc1, tc2, blk_chol_inv(tc1, tc2, lm))
