def __init__(self, x, n_in, n_hid, n_out, nlayers=1, y=None, eps=None):
super(GaussianMLP, self).__init__(x, n_in, n_hid, nlayers=nlayers, prefix='GaussianMLP_hidden')
self.mu_layer = HiddenLayer(
input=self.hidden_layers[-1].output,
n_in=self.hidden_layers[-1].n_out,
n_out=n_out,
activation=None,
prefix='GaussianMLP_mu'
)
# log(sigma^2)
self.logvar_layer = HiddenLayer(
input=self.hidden_layers[-1].output,
n_in=self.hidden_layers[-1].n_out,
n_out=n_out,
activation=None,
prefix='GaussianMLP_logvar'
)
self.mu = self.mu_layer.output
self.var = T.exp(self.logvar_layer.output)
self.sigma = T.sqrt(self.var)
self.params = self.params + self.mu_layer.params +\
self.logvar_layer.params
# for use as encoder
if eps:
assert(y is None)
# XXX separate reparametrization
self.out = self.mu + self.sigma * eps
# for use as decoder
if y:
assert(eps is None)
# XXX specific to [0, 1] outputs
self.out = T.nnet.sigmoid(self.mu)
self.cost = -T.sum(log_diag_mvn(self.out, self.var)(y))
def __init__(self, x, n_in, n_hid, n_out, nlayers=1, y=None, eps=None):
super(GaussianMLP, self).__init__(x,
n_in,
n_hid,
nlayers=nlayers,
prefix='GaussianMLP_hidden')
self.mu_layer = HiddenLayer(input=self.hidden_layers[-1].output,
n_in=self.hidden_layers[-1].n_out,
n_out=n_out,
activation=None,
prefix='GaussianMLP_mu')
# log(sigma^2)
self.logvar_layer = HiddenLayer(input=self.hidden_layers[-1].output,
n_in=self.hidden_layers[-1].n_out,
n_out=n_out,
activation=None,
prefix='GaussianMLP_logvar')
self.mu = self.mu_layer.output
self.var = T.exp(self.logvar_layer.output)
self.sigma = T.sqrt(self.var)
self.params = self.params + self.mu_layer.params +\
self.logvar_layer.params
# for use as encoder
if eps:
assert (y is None)
# XXX separate reparametrization
self.out = self.mu + self.sigma * eps
# for use as decoder
if y:
assert (eps is None)
# XXX specific to [0, 1] outputs
self.out = T.nnet.sigmoid(self.mu)
self.cost = -T.sum(log_diag_mvn(self.out, self.var)(y))
def __init__(self,
x,
n_in,
n_hid,
n_out,
nlayers=1,
activation=None,
y=None,
eps=None,
COV=False): ##Lksi=None,
# if Lksi:
# if eps and (y is None): #encoder!
# super(GaussianMLP, self).__init__(x, n_in, n_hid, nlayers=nlayers, prefix='GaussianMLP_hidden', Lksi)
# elif (eps is None) and y: #decoder!
# super(GaussianMLP, self).__init__(x, n_in, n_hid, nlayers=nlayers, prefix='GaussianMLP_hidden', Lksi, self.G)
# else:
super(GaussianMLP, self).__init__(x,
n_in,
n_hid,
nlayers=nlayers,
prefix='GaussianMLP_hidden')
##mu&logvar are affine from h when encode
self.mu_layer = HiddenLayer(
input=self.hidden_layers[-1].output,
n_in=self.hidden_layers[-1].n_out,
n_out=n_out,
activation=
activation, ##None T.nnet.softplus, not much diff. if logvar>0 freyfaces
prefix='GaussianMLP_mu')
# log(sigma^2) ##h generate logvar, not sigma! logvar=2logsigma
self.logvar_layer = HiddenLayer(
input=self.hidden_layers[-1].output,
n_in=self.hidden_layers[-1].n_out,
n_out=n_out,
activation=
activation, ##None, ReLU|sigmoid, keep logvar>0 for freyfaces
prefix='GaussianMLP_logvar')
self.mu = self.mu_layer.output
self.var = T.exp(self.logvar_layer.output)
self.params = self.params + self.mu_layer.params + self.logvar_layer.params
def SampleKsi(d, u, mu, eps): # icml14SBP(20)
dn = 1.0 / d
uDnu = T.sum(u * u * dn)
coeff = (1 - 1.0 / T.sqrt(1.0 + uDnu)) / (uDnu + SMALLNUM)
u = u.reshape((u.shape[0], 1))
R = T.diag(T.sqrt(dn)) - coeff * T.dot(
T.dot(T.diag(dn), T.dot(u, u.T)), T.diag(T.sqrt(dn)))
return mu + T.dot(R, eps)
if COV == False:
self.sigma = T.sqrt(self.var)
if eps: # for use as encoder
assert (y is None)
self.out = self.mu + self.sigma * eps
if y: # for use as decoder
assert (eps is None)
self.out = T.nnet.sigmoid(
self.mu) ##the grey degree of each pixel
#Gaussian-LL of data y under (z, params)
self.cost = -T.sum(log_diag_mvn(
self.out, self.var)(y)) ##(self.out, self.var)
else:
self.cov_u_layer = HiddenLayer(input=self.hidden_layers[-1].output,
n_in=self.hidden_layers[-1].n_out,
n_out=n_out,
activation=activation,
prefix='GaussianMLP_cov_u')
self.u = self.cov_u_layer.output
self.params = self.params + self.cov_u_layer.params
if eps: ##icml14(21)
assert (y is None)
self.out, _ = scan(SampleKsi,
sequences=[self.var, self.u, self.mu, eps])
if y: # for use as decoder
assert (eps is None)
self.out = T.nnet.sigmoid(
self.mu) ##the grey degree of each pixel
self.cost = -T.sum(
log_nondiag_mvn(self.mu, self.var, self.u)(y))
def __init__(self, x, n_in, n_hid, n_out, nlayers=1, activation=None, y=None, eps=None, COV=False): ##Lksi=None,
# if Lksi:
# if eps and (y is None): #encoder!
# super(GaussianMLP, self).__init__(x, n_in, n_hid, nlayers=nlayers, prefix='GaussianMLP_hidden', Lksi)
# elif (eps is None) and y: #decoder!
# super(GaussianMLP, self).__init__(x, n_in, n_hid, nlayers=nlayers, prefix='GaussianMLP_hidden', Lksi, self.G)
# else:
super(GaussianMLP, self).__init__(x, n_in, n_hid, nlayers=nlayers, prefix='GaussianMLP_hidden')
##mu&logvar are affine from h when encode
self.mu_layer = HiddenLayer(
input=self.hidden_layers[-1].output,
n_in=self.hidden_layers[-1].n_out,
n_out=n_out,
activation=activation, ##None T.nnet.softplus, not much diff. if logvar>0 freyfaces
prefix='GaussianMLP_mu'
)
# log(sigma^2) ##h generate logvar, not sigma! logvar=2logsigma
self.logvar_layer = HiddenLayer(
input=self.hidden_layers[-1].output,
n_in=self.hidden_layers[-1].n_out,
n_out=n_out,
activation=activation, ##None, ReLU|sigmoid, keep logvar>0 for freyfaces
prefix='GaussianMLP_logvar'
)
self.mu = self.mu_layer.output
self.var = T.exp(self.logvar_layer.output)
self.params = self.params + self.mu_layer.params + self.logvar_layer.params
def SampleKsi(d, u, mu, eps): # icml14SBP(20)
dn = 1.0/d
uDnu = T.sum(u*u*dn)
coeff = ( 1-1.0/T.sqrt(1.0+uDnu) ) / (uDnu+SMALLNUM)
u = u.reshape((u.shape[0],1))
R = T.diag(T.sqrt(dn)) - coeff*T.dot( T.dot(T.diag(dn),T.dot(u,u.T)), T.diag(T.sqrt(dn)) )
return mu + T.dot(R,eps)
if COV == False:
self.sigma = T.sqrt(self.var)
if eps: # for use as encoder
assert(y is None)
self.out = self.mu + self.sigma * eps
if y: # for use as decoder
assert(eps is None)
self.out = T.nnet.sigmoid(self.mu) ##the grey degree of each pixel
#Gaussian-LL of data y under (z, params)
self.cost = -T.sum(log_diag_mvn(self.out, self.var)(y)) ##(self.out, self.var)
else:
self.cov_u_layer = HiddenLayer(
input=self.hidden_layers[-1].output,
n_in=self.hidden_layers[-1].n_out,
n_out=n_out,
activation=activation,
prefix='GaussianMLP_cov_u'
)
self.u = self.cov_u_layer.output
self.params = self.params + self.cov_u_layer.params
if eps: ##icml14(21)
assert(y is None)
self.out, _ = scan(SampleKsi, sequences=[self.var, self.u, self.mu, eps])
if y: # for use as decoder
assert(eps is None)
self.out = T.nnet.sigmoid(self.mu) ##the grey degree of each pixel
self.cost = -T.sum(log_nondiag_mvn(self.mu, self.var, self.u)(y))
