def compute_logpx(self, x, y, z):
px_in = tf.reshape(tf.concat([y,z], axis=-1), [-1, self.n_y+self.n_z])
if self.x_dist == 'Gaussian':
mean, log_var = dgm.forwardPassGauss(px_in, self.px_yz, self.n_hid, self.nonlinearity, self.bn, scope='px_yz')
mean, log_var = tf.reshape(mean, [self.mc_samples, -1, self.n_x]), tf.reshape(log_var, [self.mc_samples, -1, self.n_x])
return dgm.gaussianLogDensity(x, mean, log_var)
elif self.x_dist == 'Bernoulli':
logits = dgm.forwardPassCatLogits(px_in, self.px_yz, self.n_hid, self.nonlinearity, self.bn, scope='px_yz')
logits = tf.reshape(logits, [self.mc_samples, -1, self.n_x])
return dgm.bernoulliLogDensity(x, logits)
def compute_logpx(self, x, z):
px_in = tf.reshape(z, [-1, self.n_z])
if self.x_dist == 'Gaussian':
mean, log_var = self.p_x_z_mean(px_in), self.p_x_z_log_var(px_in)
mean = tf.reshape(mean, [self.mc_samples, -1, self.n_x])
log_var = tf.reshape(log_var, [self.mc_samples, -1, self.n_x])
return dgm.gaussianLogDensity(x, mean, log_var)
elif self.x_dist == 'Bernoulli':
logits = self.p_x_z_mean(px_in)
logits = tf.reshape(logits, [self.mc_samples, -1, self.n_x])
return dgm.bernoulliLogDensity(x, logits)
def compute_logpx(self, x, z):
""" compute the log density of x under p(x|z) """
px_in = tf.reshape(z, [-1, self.n_z])
if self.x_dist == 'Gaussian':
mean, log_var = dgm.forwardPassGauss(px_in, self.px_z, self.n_hid, self.nonlinearity, self.bn, scope='px_z')
mean, log_var = tf.reshape(mean, [self.mc_samples, -1, self.n_x]), tf.reshape(log_var, [self.mc_samples, -1, self.n_x])
return dgm.gaussianLogDensity(x, mean, log_var)
elif self.x_dist == 'Bernoulli':
logits = dgm.forwardPassCatLogits(px_in, self.px_z, self.n_hid, self.nonlinearity, self.bn, scope='px_z')
logits = tf.reshape(logits, [self.mc_samples, -1, self.n_x])
return dgm.bernoulliLogDensity(x, logits)
def compute_logpx(self, x, y, z):
px_in = tf.reshape(tf.concat([y, z], axis=-1),
[-1, self.n_y + self.n_z])
if self.x_dist == 'Gaussian':
mean, log_var = self.p_x_yz_mean(px_in), self.p_x_yz_log_var(px_in)
mean, log_var = tf.reshape(
mean, [self.mc_samples, -1, self.n_x]), tf.reshape(
log_var, [self.mc_samples, -1, self.n_x])
return dgm.gaussianLogDensity(x, mean, log_var)
elif self.x_dist == 'Bernoulli':
logits = self.p_x_yz_mean_model(px_in)
logits = tf.reshape(logits, [self.mc_samples, -1, self.n_x])
return dgm.bernoulliLogDensity(x, logits)
def compute_logpx(self, x, y, z, a):
""" compute the log density of x under p(x|y,z,a) """
px_in = tf.reshape(tf.concat([y, z, a], axis=-1),
[-1, self.n_y + self.n_z + self.n_a])
if self.x_dist == 'Gaussian':
mean, logVar = dgm.forwardPassGauss(px_in,
self.px_yza,
self.n_hid,
self.nonlinearity,
self.bn,
scope='px_yza')
mean, logVar = tf.reshape(
mean, [self.mc_samples, -1, self.n_x]), tf.reshape(
logVar, [self.mc_samples, -1, self.n_x])
return dgm.gaussianLogDensity(x, mean, logVar)
elif self.x_dist == 'Bernoulli':
logits = dgm.forwardPassCatLogits(px_in,
self.px_yza,
self.n_hid,
self.nonlinearity,
self.bn,
scope='px_yza')
logits = tf.reshape(logits, [self.mc_samples, -1, self.n_x])
return dgm.bernoulliLogDensity(x, logits)