def createObjectiveFunction(self):
'''
@escription: initialize objective function and minimization function
@X,y data matrix/vector
@u random noise for simulator
@v standard normal for reparametrization trick
'''
X,u = T.dmatrices("X","u")
f, y, v = T.dcols("f", "y", "v")
mu = self.params[0]
logSigma = self.params[1]
logLambda = sharedX(np.log(self.sigma_e),name='logLambda')
#logLambda = self.params[2]
negKL = 0.5*self.dimTheta+0.5*T.sum(2*logSigma - mu ** 2 - T.exp(logSigma) ** 2)
f = self.regression_simulator(X,u,v,mu,logSigma)
logLike = -self.m*(0.5 * np.log(2 * np.pi) + logLambda)-0.5*T.sum((y-f)**2)/(T.exp(logLambda)**2)/self.Lu
elbo = (negKL + logLike)
obj = -elbo
self.lowerboundfunction = th.function([X, y, u, v], obj, on_unused_input='ignore')
derivatives = T.grad(obj,self.params)
self.gradientfunction = th.function([X,y,u,v], derivatives, on_unused_input='ignore')
def createGradientFunctions(self):
#create
X = T.dmatrices("X")
mu, logSigma, u, v, f, R = T.dcols("mu", "logSigma", "u", "v", "f", "R")
mu = sharedX( np.random.normal(10, 10, (self.dimTheta, 1)), name='mu')
logSigma = sharedX(np.random.uniform(0, 4, (self.dimTheta, 1)), name='logSigma')
logLambd = sharedX(np.matrix(np.random.uniform(0, 10)),name='logLambd')
logLambd = T.patternbroadcast(T.dmatrix("logLambd"),[1,1])
negKL = 0.5 * T.sum(1 + 2*logSigma - mu ** 2 - T.exp(logSigma) ** 2)
theta = mu+T.exp(logSigma)*v
W=theta
y=X[:,0]
X_sim=X[:,1:]
f = (T.dot(X_sim,W)+u).flatten()
gradvariables = [mu, logSigma, logLambd]
logLike = T.sum(-(0.5 * np.log(2 * np.pi) + logLambd) - 0.5 * ((y-f)/(T.exp(logLambd)))**2)
logp = (negKL + logLike)/self.m
optimizer = -logp
self.negKL = th.function([mu, logSigma], negKL, on_unused_input='ignore')
self.f = th.function(gradvariables + [X,u,v], f, on_unused_input='ignore')
self.logLike = th.function(gradvariables + [X, u, v], logLike,on_unused_input='ignore')
derivatives = T.grad(logp,gradvariables)
derivatives.append(logp)
self.gradientfunction = th.function(gradvariables + [X, u, v], derivatives, on_unused_input='ignore')
self.lowerboundfunction = th.function(gradvariables + [X, u, v], logp, on_unused_input='ignore')
self.optimizer = BatchGradientDescent(objective=optimizer, params=gradvariables,inputs = [X,u,v],conjugate=True,max_iter=1)
def createObjectiveFunction(self):
'''
@escription: initialize objective function and minimization function
@X,y data matrix/vector
@u random noise for simulator
@v standard normal for reparametrization trick
'''
X, u = T.dmatrices("X", "u")
f, y, v = T.dcols("f", "y", "v")
mu = self.params[0]
logSigma = self.params[1]
logLambda = sharedX(np.log(self.sigma_e), name='logLambda')
#logLambda = self.params[2]
negKL = 0.5 * self.dimTheta + 0.5 * T.sum(2 * logSigma - mu**2 -
T.exp(logSigma)**2)
f = self.regression_simulator(X, u, v, mu, logSigma)
logLike = -self.m * (0.5 * np.log(2 * np.pi) +
logLambda) - 0.5 * T.sum(
(y - f)**2) / (T.exp(logLambda)**2) / self.Lu
elbo = (negKL + logLike)
obj = -elbo
self.lowerboundfunction = th.function([X, y, u, v],
obj,
on_unused_input='ignore')
derivatives = T.grad(obj, self.params)
self.gradientfunction = th.function([X, y, u, v],
derivatives,
on_unused_input='ignore')
def createGradientFunctions(self):
#Create the Theano variables
W1,W2,W3,W4,W5,W6,x,eps = T.dmatrices("W1","W2","W3","W4","W5","W6","x","eps")
#Create biases as cols so they can be broadcasted for minibatches
b1,b2,b3,b4,b5,b6 = T.dcols("b1","b2","b3","b4","b5","b6")
z1 = T.col("z1")
if self.continuous:
#convolve x
# no_filters = 100, stride = 4, filter_size = 50
h_encoder = T.tanh(T.dot(W1,x) + b1)
#h_encoder = T.dot(W1,x) + b1
else:
h_encoder = T.tanh(T.dot(W1,x) + b1)
mu_encoder = T.dot(W2,h_encoder) + b2
log_sigma_encoder = 0.5*(T.dot(W3,h_encoder) + b3)
mu_encoder = T.dot(W2,h_encoder) + b2
log_sigma_encoder = 0.5*(T.dot(W3,h_encoder) + b3)
#Find the hidden variable z
z = mu_encoder + T.exp(log_sigma_encoder)*eps
prior = 0.5* T.sum(1 + 2*log_sigma_encoder - mu_encoder**2 - T.exp(2*log_sigma_encoder))
#Set up decoding layer
if self.continuous:
h_decoder = T.nnet.softplus(T.dot(W4,z) + b4)
h_dec = T.nnet.softplus(T.dot(W4,z1) + b4)
#h_decoder = T.dot(W4,z) + b4
#h_dec = T.dot(W4,z1) + b4
mu_decoder = T.tanh(T.dot(W5,h_decoder) + b5)
mu_dec = T.tanh(T.dot(W5,h_dec) + b5)
log_sigma_decoder = 0.5*(T.dot(W6,h_decoder) + b6)
logpxz = T.sum(-(0.5 * np.log(2 * np.pi) + log_sigma_decoder) - 0.5 * ((x - mu_decoder) / T.exp(log_sigma_decoder))**2)
gradvariables = [W1,W2,W3,W4,W5,W6,b1,b2,b3,b4,b5,b6]
else:
h_decoder = T.tanh(T.dot(W4,z) + b4)
y = T.nnet.sigmoid(T.dot(W5,h_decoder) + b5)
logpxz = -T.nnet.binary_crossentropy(y,x).sum()
gradvariables = [W1,W2,W3,W4,W5,b1,b2,b3,b4,b5]
logp = logpxz + prior
#Compute all the gradients
derivatives = T.grad(logp,gradvariables)
#Add the lowerbound so we can keep track of results
derivatives.append(logp)
self.get_z = th.function(gradvariables+[x,eps],z,on_unused_input='ignore')
self.generate = th.function(gradvariables+[z1,x,eps],mu_dec,on_unused_input='ignore')
self.predict = th.function(gradvariables+[x,eps],mu_decoder,on_unused_input='ignore')
self.gradientfunction = th.function(gradvariables + [x,eps], derivatives, on_unused_input='ignore')
self.lowerboundfunction = th.function(gradvariables + [x,eps], logp, on_unused_input='ignore')
def createGradientFunctions(self):
#Create the Theano variables
W1, W2, W3, W4, W5, W6, x, eps = T.dmatrices("W1", "W2", "W3", "W4",
"W5", "W6", "x", "eps")
#Create biases as cols so they can be broadcasted for minibatches
b1, b2, b3, b4, b5, b6 = T.dcols("b1", "b2", "b3", "b4", "b5", "b6")
if self.continuous:
h_encoder = T.nnet.softplus(T.dot(W1, x) + b1)
else:
h_encoder = T.tanh(T.dot(W1, x) + b1)
mu_encoder = T.dot(W2, h_encoder) + b2
log_sigma_encoder = 0.5 * (T.dot(W3, h_encoder) + b3)
#Find the hidden variable z
z = mu_encoder + T.exp(log_sigma_encoder) * eps
prior = 0.5 * T.sum(1 + 2 * log_sigma_encoder - mu_encoder**2 -
T.exp(2 * log_sigma_encoder))
#Set up decoding layer
if self.continuous:
h_decoder = T.nnet.softplus(T.dot(W4, z) + b4)
mu_decoder = T.nnet.sigmoid(T.dot(W5, h_decoder) + b5)
log_sigma_decoder = 0.5 * (T.dot(W6, h_decoder) + b6)
logpxz = T.sum(-(0.5 * np.log(2 * np.pi) + log_sigma_decoder) -
0.5 *
((x - mu_decoder) / T.exp(log_sigma_decoder))**2)
gradvariables = [W1, W2, W3, W4, W5, W6, b1, b2, b3, b4, b5, b6]
else:
h_decoder = T.tanh(T.dot(W4, z) + b4)
y = T.nnet.sigmoid(T.dot(W5, h_decoder) + b5)
logpxz = -T.nnet.binary_crossentropy(y, x).sum()
gradvariables = [W1, W2, W3, W4, W5, b1, b2, b3, b4, b5]
logp = logpxz + prior
#Compute all the gradients
derivatives = T.grad(logp, gradvariables)
#Add the lowerbound so we can keep track of results
derivatives.append(logp)
self.gradientfunction = th.function(gradvariables + [x, eps],
derivatives,
on_unused_input='ignore')
self.lowerboundfunction = th.function(gradvariables + [x, eps],
logp,
on_unused_input='ignore')
self.zfunction = th.function(gradvariables + [x, eps],
z,
on_unused_input='ignore')
def createGradientFunctions(self):
#create
X = T.dmatrices("X")
mu, logSigma, u, v, f, R = T.dcols("mu", "logSigma", "u", "v", "f",
"R")
mu = sharedX(np.random.normal(10, 10, (self.dimTheta, 1)), name='mu')
logSigma = sharedX(np.random.uniform(0, 4, (self.dimTheta, 1)),
name='logSigma')
logLambd = sharedX(np.matrix(np.random.uniform(0, 10)),
name='logLambd')
logLambd = T.patternbroadcast(T.dmatrix("logLambd"), [1, 1])
negKL = 0.5 * T.sum(1 + 2 * logSigma - mu**2 - T.exp(logSigma)**2)
theta = mu + T.exp(logSigma) * v
W = theta
y = X[:, 0]
X_sim = X[:, 1:]
f = (T.dot(X_sim, W) + u).flatten()
gradvariables = [mu, logSigma, logLambd]
logLike = T.sum(-(0.5 * np.log(2 * np.pi) + logLambd) - 0.5 *
((y - f) / (T.exp(logLambd)))**2)
logp = (negKL + logLike) / self.m
optimizer = -logp
self.negKL = th.function([mu, logSigma],
negKL,
on_unused_input='ignore')
self.f = th.function(gradvariables + [X, u, v],
f,
on_unused_input='ignore')
self.logLike = th.function(gradvariables + [X, u, v],
logLike,
on_unused_input='ignore')
derivatives = T.grad(logp, gradvariables)
derivatives.append(logp)
self.gradientfunction = th.function(gradvariables + [X, u, v],
derivatives,
on_unused_input='ignore')
self.lowerboundfunction = th.function(gradvariables + [X, u, v],
logp,
on_unused_input='ignore')
self.optimizer = BatchGradientDescent(objective=optimizer,
params=gradvariables,
inputs=[X, u, v],
conjugate=True,
max_iter=1)
def createGradientFunctions(self):
#Create the Theano variables
W1,W2,W3,W4,W5,W6,x,eps = T.dmatrices("W1","W2","W3","W4","W5","W6","x","eps")
#Create biases as cols so they can be broadcasted for minibatches
b1,b2,b3,b4,b5,b6,pi = T.dcols("b1","b2","b3","b4","b5","b6","pi")
if self.continuous:
h_encoder = T.nnet.softplus(T.dot(W1,x) + b1)
else:
h_encoder = T.tanh(T.dot(W1,x) + b1)
print type(pi)
rng = T.shared_randomstreams.RandomStreams(seed=124)
i = rng.choice(size=(1,), a=self.num_model, p=T.nnet.softmax(pi.T).T.flatten())
mu_encoder = T.dot(W2[i[0]*self.dimZ:(1+i[0])*self.dimZ],h_encoder) + b2[i[0]*self.dimZ:(1+i[0])*self.dimZ]
log_sigma_encoder = (0.5*(T.dot(W3[i[0]*self.dimZ:(1+i[0])*self.dimZ],h_encoder)))+ b3[i[0]*self.dimZ:(1+i[0])*self.dimZ]
z = mu_encoder + T.exp(log_sigma_encoder)*eps
prior = 0
for i in range(self.num_model):
prior += T.exp(pi[i][0])*0.5* T.sum(1 + 2*log_sigma_encoder[int(i)*self.dimZ:(1+int(i))*self.dimZ] - mu_encoder[int(i)*self.dimZ:(1+int(i))*self.dimZ]**2 - T.exp(2*log_sigma_encoder[int(i)*self.dimZ:(1+int(i))*self.dimZ]))
prior /= T.sum(T.exp(pi))
#Set up decoding layer
if self.continuous:
h_decoder = T.nnet.softplus(T.dot(W4,z) + b4)
mu_decoder = T.nnet.sigmoid(T.dot(W5,h_decoder) + b5)
log_sigma_decoder = 0.5*(T.dot(W6,h_decoder) + b6)
logpxz = T.sum(-(0.5 * np.log(2 * np.pi) + log_sigma_decoder) - 0.5 * ((x - mu_decoder) / T.exp(log_sigma_decoder))**2)
gradvariables = [W1,W2,W3,W4,W5,W6,b1,b2,b3,b4,b5,b6,pi]
else:
h_decoder = T.tanh(T.dot(W4,z) + b4)
y = T.nnet.sigmoid(T.dot(W5,h_decoder) + b5)
logpxz = -T.nnet.binary_crossentropy(y,x).sum()
gradvariables = [W1,W2,W3,W4,W5,b1,b2,b3,b4,b5,pi]
logp = logpxz + prior
#Compute all the gradients
derivatives = T.grad(logp,gradvariables)
#Add the lowerbound so we can keep track of results
derivatives.append(logpxz)
self.gradientfunction = th.function(gradvariables + [x,eps], derivatives, on_unused_input='ignore')
self.lowerboundfunction = th.function(gradvariables + [x,eps], logp, on_unused_input='ignore')
self.hiddenstatefunction = th.function(gradvariables + [x,eps], z, on_unused_input='ignore')
def createGradientFunctions(self):
#Create the Theano variables
W1, W2, W3, W4, W5, W7, x, eps = T.dmatrices("W1", "W2", "W3", "W4",
"W5", "W7", "x", "eps")
#Create biases as cols so they can be broadcasted for minibatches
b1, b2, b3, b4, b5, b7 = T.dcols("b1", "b2", "b3", "b4", "b5", "b7")
if self.continuous_data:
h_encoder = T.nnet.softplus(T.dot(W1, x) + b1)
else:
h_encoder = T.tanh(T.dot(W1, x) + b1)
mu_encoder = T.dot(W2, h_encoder) + b2
log_sigma_encoder = 0.5 * (T.dot(W3, h_encoder) + b3)
L_u = T.tril(log_L_u - T.diag(T.diag(log_L_u)) +
T.diag(T.exp(T.diag(log_L_u))))
# To do: Better ways of paramterising the covariance (see: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.31.494&rep=rep1&type=pdf)
#Compute GP objects
K_ff = self.ker.RBF(sf2, ell, X)
K_uu = self.ker.RBF(sf2, ell, X_u)
K_uu_inv = nlinalg.matrix_inverse(K_uu)
L_f = slinalg.cholesky(K_ff - T.dot(K_fu, T.dot(K_uu_inv, K_fu.T)))
# f_i make up the columns of f, simiarly for m_u_i
u = m_u + T.dot(L_u, eps_u) #n_induce iid pseudo inducing sets
f = T.dot(K_fu, T.dot(K_uu_inv, u)) + T.dot(L_f, X)
#Find the hidden variable z
# log_sigma_lhood = 0.5*(T.dot(W9,f) + b9) # the var GP maps to both mean *and* covariance
sigma_var_lhood = sigma_z**2 * T.eye(self.dimZ)
L_z = slinalg.cholesky(sigma_var_lhood)
z = f + T.dot(L_z, eps_z)
# z = mu_encoder + T.exp(log_sigma_encoder)*eps
prior = 0.5 * T.sum(1 + 2 * log_sigma_encoder - mu_encoder**2 -
T.exp(2 * log_sigma_encoder))
#Set up decoding layer
if self.continuous_data:
h_decoder = T.nnet.softplus(T.dot(W4, z) + b4)
mu_decoder = T.nnet.sigmoid(T.dot(W5, h_decoder) + b5)
log_sigma_decoder = 0.5 * (T.dot(W7, h_decoder) + b7)
logpxz = T.sum(-(0.5 * np.log(2 * np.pi) + log_sigma_decoder) -
0.5 *
((x - mu_decoder) / T.exp(log_sigma_decoder))**2)
gradvariables = [
W1, W2, W3, W4, W5, W7, b1, b2, b3, b4, b5, b7, sf2, ell, X_u,
m_u, L_u
]
else:
h_decoder = T.tanh(T.dot(W4, z) + b4)
y = T.nnet.sigmoid(T.dot(W5, h_decoder) + b5)
logpxz = -T.nnet.binary_crossentropy(y, x).sum()
gradvariables = [
W1, W2, W3, W4, W5, b1, b2, b3, b4, b5, sf2, ell, X_u, m_u, L_u
]
#Set up auxiliary layer
if self.continuous_data:
h_auxiliary = T.nnet.softplus(T.dot(W6, [x, z]) + b6)
mu_auxiliary = T.nnet.sigmoid(T.dot(W7, h_auxiliary) + b7)
log_sigma_auxiliary = 0.5 * (T.dot(W8, h_auxiliary) + b8)
else:
pass #to do
logp = logpxz + prior
#Compute KL terms
# KL_qp = -0.5*T.sum(1.0 + 2*log_sigma_lhood - f**2 - T.exp(2*log_sigma_lhood))
KL_qp = 0.5 * (T.dot(f.T, f) +
T.trace(sigma_var_lhood + T.log(T.eye(self.dimZ)) -
T.log(sigma_var_lhood)) - self.dimZ)
KL_qr = 0.5 * (T.dot(
(mu_auxiliary - mu_encoder).T,
T.dot(T.diag(1.0 / T.exp(log_sigma_auxiliary)),
mu_auxiliary - mu_decoder)) + T.trace(
T.dot(T.diag(1.0 / T.exp(log_sigma_auxiliary)),
T.dot(L_u, L_u.T)) + log_sigma_auxiliary -
log_sigma_encoder) - self.dimXf - self.dimf)
#Compute bound and all the gradients
stoch_bound = logpxz - KL_qp - KL_qr
derivatives = T.grad(stoch_bound, gradvariables)
#Add the lowerbound so we can keep track of results
derivatives.append(stoch_bound)
self.gradientfunction = th.function(gradvariables +
[x, eps_u, eps_z, X],
derivatives,
on_unused_input='ignore')
self.lowerboundfunction = th.function(gradvariables +
[x, eps_u, eps_z, X],
stoch_bound,
on_unused_input='ignore')
self.zfunction = th.function(gradvariables + [x, eps_u, eps_z, X],
z,
on_unused_input='ignore')
def createGradientFunctions(self):
#Create the Theano variables
W1,W2,W3,W4,W5,W6,x,eps = T.dmatrices("W1","W2","W3","W4","W5","W6","x","eps")
#Create biases as cols so they can be broadcasted for minibatches
b1,b2,b3,b4,b5,b6 = T.dcols("b1","b2","b3","b4","b5","b6")
if self.continuous:
h_encoder = T.nnet.softplus(T.dot(W1,x) + b1)
else:
# h_encoder = T.tanh(T.dot(W1,x) + b1)
def relu(x):
return 0.5 * (x + abs(x))
h_encoder = relu((T.dot(W1,x) + b1))
mu_encoder = T.dot(W2,h_encoder) + b2
log_sigma_encoder = 0.5*(T.dot(W3,h_encoder) + b3)
#Find the hidden variable z
z = mu_encoder + T.exp(log_sigma_encoder)*eps
KL_divergence = 0.5* T.sum(1 + 2*log_sigma_encoder - mu_encoder**2 - T.exp(2*log_sigma_encoder))
#Set up decoding layer
if self.continuous:
h_decoder = T.nnet.softplus(T.dot(W4,z) + b4)
mu_decoder = T.nnet.sigmoid(T.dot(W5,h_decoder) + b5)
log_sigma_decoder = 0.5*(T.dot(W6,h_decoder) + b6)
logpxz = T.sum(-(0.5 * np.log(2 * np.pi) + log_sigma_decoder) - 0.5 * ((x - mu_decoder) / T.exp(log_sigma_decoder))**2)
logpxz_individual = T.sum(-(0.5 * np.log(2 * np.pi) + log_sigma_decoder) - 0.5 * ((x - mu_decoder) / T.exp(log_sigma_decoder))**2, axis=0)
gradvariables = [W1,W2,W3,W4,W5,W6,b1,b2,b3,b4,b5,b6]
else:
# h_decoder = T.tanh(T.dot(W4,z) + b4)
def relu(x):
return 0.5 * (x + abs(x))
h_decoder = relu(T.dot(W4,z) + b4)
y = T.nnet.sigmoid(T.dot(W5,h_decoder) + b5)
logpxz_individual = -T.nnet.binary_crossentropy(y,x).sum(axis=0)
logpxz = -T.nnet.binary_crossentropy(y,x).sum()
gradvariables = [W1,W2,W3,W4,W5,b1,b2,b3,b4,b5]
logp = logpxz + KL_divergence
#Compute all the gradients
derivatives = T.grad(logp,gradvariables)
#Add the lowerbound so we can keep track of results
# derivatives.append(logp)
## This is more confusing
self.gradientfunction = th.function(gradvariables + [x,eps], derivatives, on_unused_input='ignore')
self.lowerboundfunction = th.function(gradvariables + [x,eps], logp, on_unused_input='ignore')
self.latentvarfunction = th.function(gradvariables + [x], [mu_encoder.T, log_sigma_encoder.T], on_unused_input='ignore')
# if self.continuous:
# self.reconstructionfunction = th.function(gradvariables + [x, eps], [mu_decoder.T, log_sigma_decoder.T], on_unused_input='ignore')
# else:
# self.reconstructionfunction = th.function(gradvariables + [x, eps], logpxz_individual.T, on_unused_input='ignore')
if not self.continuous:
self.recon_mnist = th.function(gradvariables + [x, eps], y, on_unused_input='ignore')
self.reconstructionfunction = th.function(gradvariables + [x, eps], logpxz_individual.T, on_unused_input='ignore')