def sample_v_given_hs(self, h_sample, s_sample, rng=None, size=None):
"""
Generates sample from p(v | h, s)
"""
v_mean = self.v_given_hs(h_sample, s_sample)
rng = self.theano_rng if rng is None else rng
size = size if size else self.batch_size
if self.flags['truncate_v']:
v_sample = truncated.truncated_normal(
size=(size, self.n_v),
avg = v_mean,
std = T.sqrt(1./self.lambd_prec),
lbound = -self.truncation_bound['v'],
ubound = self.truncation_bound['v'],
theano_rng = rng,
dtype=floatX)
else:
v_sample = rng.normal(
size=(size, self.n_v),
avg = v_mean,
std = T.sqrt(1./self.lambd_prec),
dtype=floatX)
return v_sample
def sample_s_given_ghv(self, g_sample, h_sample, v_sample, rng=None, size=None):
"""
Generates sample from p(s | g, h, v)
"""
s_mean = self.s_given_ghv(g_sample, h_sample, v_sample)
rng = self.theano_rng if rng is None else rng
size = size if size else self.batch_size
if self.flags['truncate_s']:
s_sample = truncated.truncated_normal(
size=(size, self.n_s),
avg = s_mean,
std = T.sqrt(1./self.alpha_prec),
lbound = self.truncation_bound['s'],
ubound = self.truncation_bound['s'],
theano_rng = rng,
dtype=floatX)
else:
s_sample = rng.normal(
size=(size, self.n_s),
avg = s_mean,
std = T.sqrt(1./self.alpha_prec),
dtype=floatX)
return s_sample
def sample_v_given_gsht(self, g_sample, s_sample, h_sample, t_sample, rng=None):
v_mean = self.v_given_gsht(g_sample, s_sample, h_sample, t_sample)
rng = self.theano_rng if rng is None else rng
v_sample = truncated.truncated_normal(
size=(self.batch_size, self.n_v),
avg = v_mean,
std = T.sqrt(1./self.lambd_prec),
lbound = -self.vbound,
ubound = self.vbound,
theano_rng = rng,
dtype=floatX)
return v_sample
def sample_v_given_hs(self, h_sample, s_sample, rng=None):
v_mean = self.v_given_hs(h_sample, s_sample)
rng = self.theano_rng if rng is None else rng
if self.flags.get('truncated_normal', False):
v_sample = truncated.truncated_normal(
size=(self.batch_size, self.n_v),
avg=v_mean,
std=T.sqrt(1. / self.beta_prec),
lbound=-self.vbound,
ubound=self.vbound,
theano_rng=rng,
dtype=floatX)
else:
v_sample = rng.normal(size=(self.batch_size, self.n_v),
avg=v_mean,
std=T.sqrt(1. / self.beta_prec),
dtype=floatX)
return v_sample
def sample_v_given_hs(self, h_sample, s_sample, rng=None):
v_mean = self.v_given_hs(h_sample, s_sample)
rng = self.theano_rng if rng is None else rng
if self.flags.get('truncated_normal', False):
v_sample = truncated.truncated_normal(
size=(self.batch_size, self.n_v),
avg = v_mean,
std = T.sqrt(1./self.beta_prec),
lbound = -self.vbound,
ubound = self.vbound,
theano_rng = rng,
dtype=floatX)
else:
v_sample = rng.normal(
size=(self.batch_size, self.n_v),
avg = v_mean,
std = T.sqrt(1./self.beta_prec), dtype=floatX)
return v_sample
def sample_t_given_gshv(self, g_sample, s_sample, h_sample, v_sample, rng=None):
t_mean = self.t_given_gshv(g_sample, s_sample, h_sample, v_sample)
rng = self.theano_rng if rng is None else rng
if self.flags['truncate_s']:
t_sample = truncated.truncated_normal(
size=(self.batch_size, self.n_h),
avg = t_mean,
std = T.sqrt(1./self.beta_prec),
lbound = self.eta - self.truncation_bounds['t'],
ubound = self.eta + self.truncation_bounds['t'],
theano_rng = rng,
dtype=floatX)
else:
t_sample = rng.normal(
size=(self.batch_size, self.n_h),
avg = t_mean,
std = T.sqrt(1./self.beta_prec), dtype=floatX)
return t_sample
def sample_s_given_ghtv(self, g_sample, h_sample, t_sample, v_sample, rng=None):
s_mean = self.s_given_ghtv(g_sample, h_sample, t_sample, v_sample)
rng = self.theano_rng if rng is None else rng
if self.flags['truncate_s']:
s_sample = truncated.truncated_normal(
size=(self.batch_size, self.n_g),
avg = s_mean,
std = T.sqrt(1./self.alpha_prec),
lbound = self.mu - self.truncation_bounds['s'],
ubound = self.mu + self.truncation_bounds['s'],
theano_rng = rng,
dtype=floatX)
else:
s_sample = rng.normal(
size=(self.batch_size, self.n_g),
avg = s_mean,
std = T.sqrt(1./self.alpha_prec), dtype=floatX)
return s_sample
def sample_v_given_hs(self, h_sample, s_sample, rng=None, size=None):
"""
Generates sample from p(v | h, s)
"""
v_mean = self.v_given_hs(h_sample, s_sample)
rng = self.theano_rng if rng is None else rng
size = size if size else self.batch_size
if self.flags['truncate_v']:
v_sample = truncated.truncated_normal(
size=(size, self.n_v),
avg=v_mean,
std=T.sqrt(1. / self.lambd_prec),
lbound=-self.truncation_bound['v'],
ubound=self.truncation_bound['v'],
theano_rng=rng,
dtype=floatX)
else:
v_sample = rng.normal(size=(size, self.n_v),
avg=v_mean,
std=T.sqrt(1. / self.lambd_prec),
dtype=floatX)
return v_sample
