def logLik(self, x, z):
'''
model specific likelihood
:return: log-P(x|z)
'''
logPxz = mathT.gaussInit(x, self.varn)
z1z2 = T.prod(z, axis=1, keepdims=True)
return logPxz(T.concatenate([z1z2, z1z2], axis=1))
def __init__(self):
'''
:return:
'''
super(banana, self).__init__()
self.priormus = utils.sharedf(np.zeros(2))
self.priorvar = utils.sharedf(np.eye(2))
self.stdn = utils.sharedf([.5, .5])
self.varn = nlinalg.diag(T.sqr(self.stdn))
self.logPz = mathT.gaussInit(self.priormus, self.priorvar)
def initLogPrior(self):
priorvar = np.eye(self.dim)
self.logPrior = mathT.gaussInit(np.zeros(self.dim), priorvar)
_, self.priorGen = mathZ.gaussInit(np.zeros(self.dim),priorvar)
import theano.tensor as T
import numpy as np
import nodes
import nodes.layers as layers
import nodes.iaf as iaf
import utils.mathT as mathT
import utils.mathZ as mathZ
import utils.plotZ as plotZ
f32 = theano.config.floatX
DIM = 2
SAMPLING_E = int(1e+5)
LR = 0.1
# defining target distribution
logPZ = mathT.gaussInit([0., 0.], np.eye(2) / 2., mean=True)
# defining normalising flow model
iafmodel = iaf.IafStack('IAF_simple', dim=DIM)
layer1 = iaf.IafLinear({
'name': 'iaf-1',
'dim': 2,
'bias': True,
'initmethod': 'identity',
'lr': LR
})
layer1.setWeight(weights=np.array([[0.3, 0.2], [0, 0.2]]), bias=[0, 0])
layer2 = iaf.IafTanh({'name': 'tanh'})
layer3 = iaf.IafPermute({'name': 'per', 'dim': 2})
def __init__(self,name):
self.name=name
self.layers = list()
self.logPrior = mathT.gaussInit(np.zeros(self.dim), np.e)
def initLogPrior(self):
noisevar = N.sharedf(np.eye(self.dim))*T.sqr(self.priorstd)
noisemu = N.sharedf(np.zeros(self.dim))
self.logPrior = mathT.gaussInit(noisemu, noisevar)