def _print_verbose1(self,
epoch,
fd,
sess,
acc_train,
acc_test,
avg_var=None):
am_test, alv_test, a_test = self._sample_a(self.x_test, 1)
am_train, alv_train, a_train = self._sample_a(self.x_train, 1)
zm_test, zlv_test, z_test = self._sample_Z(self.x_test, self.y_test,
a_test, 1)
zm_train, zlv_train, z_train = self._sample_Z(self.x_train,
self.y_train, a_train, 1)
lpx_test, lpx_train, klz_test, klz_train = sess.run([
self._compute_logpx(self.x_test, z_test, self.y_test),
self._compute_logpx(self.x_train, z_train, self.y_train),
dgm._gauss_kl(zm_test, tf.exp(zlv_test)),
dgm._gauss_kl(zm_train, tf.exp(zlv_train))
],
feed_dict=fd)
print(
'Epoch: {}, logpx: {:5.3f}, klz: {:5.3f}, Train: {:5.3f}, Test: {:5.3f}'
.format(epoch, np.mean(lpx_train), np.mean(klz_train), acc_train,
acc_test))
def _compute_ELBO(self, x):
z_mean, z_log_var, z = self._sample_Z(x)
KLz = dgm._gauss_kl(z_mean, z_log_var)
l_qz = dgm._gauss_logp(z, z_mean, z_log_var)
l_pz = dgm._gauss_logp(z, tf.zeros_like(z), tf.ones_like(z))
l_px = self._compute_logpx(x, z)
total_elbo = l_px - self.beta * (KLz)
return tf.reduce_mean(total_elbo), tf.reduce_mean(l_px), tf.reduce_mean(KLz)
def _labeled_loss(self, x, y):
""" Compute necessary terms for labeled loss (per data point) """
q_mean, q_log_var, z = self._sample_Z(x, y, self.Z_SAMPLES)
l_px = self._compute_logpx(x, z)
l_py = self._compute_logpy(y, x, z)
l_pz = dgm._gauss_logp(z, tf.zeros_like(z), tf.log(tf.ones_like(z)))
l_qz = dgm._gauss_logp(z, q_mean, q_log_var)
klz = dgm._gauss_kl(q_mean, q_log_var)
return l_px + l_py + self.beta * (l_pz - l_qz)
def _print_verbose1(self, epoch, fd, sess):
self.phase = False
zm_test, zlv_test, z_test = self._sample_Z(self.x_test, self.y_test, 1)
zm_train, zlv_train, z_train = self._sample_Z(self.x_train,
self.y_train, 1)
lpx_test, lpx_train, klz_test, klz_train, acc_train, acc_test = sess.run(
[
self._compute_logpx(self.x_test, z_test, self.y_test),
self._compute_logpx(self.x_train, z_train, self.y_train),
dgm._gauss_kl(zm_test, tf.exp(zlv_test)),
dgm._gauss_kl(zm_train, tf.exp(zlv_train)), self.train_acc,
self.test_acc
],
feed_dict=fd)
print(
'Epoch: {}, logpx: {:5.3f}, klz: {:5.3f}, Train: {:5.3f}, Test: {:5.3f}'
.format(epoch, np.mean(lpx_train), np.mean(klz_train), acc_train,
acc_test))
def _kl_W(self):
kl = 0
for i in range(len(self.NUM_HIDDEN)):
mean, logvar = self.Pzx_y['W' + str(i) +
'_mean'], self.Pzx_y['W' + str(i) +
'_logvar']
kl += tf.reduce_sum(dgm._gauss_kl(mean, logvar))
mean, logvar = self.Pzx_y['b' + str(i) +
'_mean'], self.Pzx_y['b' + str(i) +
'_logvar']
kl += tf.reduce_sum(
dgm._gauss_kl(tf.expand_dims(mean, 1),
tf.expand_dims(logvar, 1)))
mean, logvar = self.Pzx_y['Wout_mean'], self.Pzx_y['Wout_logvar']
kl += tf.reduce_sum(dgm._gauss_kl(mean, logvar))
mean, logvar = self.Pzx_y['bout_mean'], self.Pzx_y['bout_logvar']
kl += tf.reduce_sum(
dgm._gauss_kl(tf.expand_dims(mean, 1), tf.expand_dims(logvar, 1)))
return kl
def _labeled_loss_W(self, x, y):
""" Compute necessary terms for labeled loss (per data point) """
d = tf.cast(self.TRAINING_SIZE, tf.float32)
q_mean, q_logvar, z = self._sample_Z(x, y, self.Z_SAMPLES)
l_px = self._compute_logpx(x, z)
l_py = self._compute_logpy(y, x, z)
l_pz = dgm._gauss_logp(z, tf.zeros_like(z), tf.log(tf.ones_like(z)))
l_qz = dgm._gauss_logp(z, q_mean, q_logvar)
klz = dgm._gauss_kl(q_mean, q_logvar)
klw = self._kl_W() / d
return l_px + l_py + self.beta * (l_pz - l_qz) - klw