def lowerBound(self, x, y, z, z_m, z_lv): """ Compute densities and lower bound given all inputs (mc_samps X n_obs X n_dim) """ l_px = self.compute_logpx(x,y,z) l_py = dgm.multinoulliUniformLogDensity(y) l_pz = dgm.standardNormalLogDensity(z) l_qz = dgm.gaussianLogDensity(z, z_m, z_lv) return tf.reduce_mean(l_px + l_py + l_pz - l_qz, axis=0)
def qy_loss(self, x, y=None, a=None, expand_y=True):
if a is None:
_, _, a = self.sample_a(x)
qy_in = tf.reshape(
tf.concat([
tf.tile(tf.expand_dims(x, 0), [self.mc_samples, 1, 1]), a
],
axis=-1), [-1, self.n_x + self.n_a])
else:
qy_in = tf.reshape(tf.concat([x, a], axis=-1),
[-1, self.n_x + self.n_a])
y_ = tf.reshape(self.q_y_ax_model(qy_in),
[self.mc_samples, -1, self.n_y])
if y is not None and expand_y == True:
y = tf.tile(tf.expand_dims(y, 0), [self.mc_samples, 1, 1])
if y is None:
return dgm.multinoulliUniformLogDensity(y_)
else:
return dgm.multinoulliLogDensity(y, y_)
def lowerBound(self, x, y, z, z_m, z_lv, a, qa_m, qa_lv):
""" Helper function for loss computations. Assumes each input is a rank(3) tensor """
pa_in = tf.reshape(tf.concat([y, z], axis=-1),
[-1, self.n_y + self.n_z])
pa_m, pa_lv = dgm.forwardPassGauss(pa_in,
self.pa_yz,
self.n_hid,
self.nonlinearity,
self.bn,
scope='pa_yz')
pa_m, pa_lv = tf.reshape(pa_m,
[self.mc_samples, -1, self.n_a]), tf.reshape(
pa_lv, [self.mc_samples, -1, self.n_a])
l_px = self.compute_logpx(x, y, z, a)
l_py = dgm.multinoulliUniformLogDensity(y)
l_pz = dgm.standardNormalLogDensity(z)
l_pa = dgm.gaussianLogDensity(a, pa_m, pa_lv)
l_qz = dgm.gaussianLogDensity(z, z_m, z_lv)
l_qa = dgm.gaussianLogDensity(a, qa_m, qa_lv)
return tf.reduce_mean(l_px + l_py + l_pz + l_pa - l_qz - l_qa, axis=0)
def qy_loss(self, x, y=None):
y_ = self.q_y_x_model(x)
if y is None:
return dgm.multinoulliUniformLogDensity(y_)
else:
return dgm.multinoulliLogDensity(y, y_)
