tf.set_random_seed(1)
# input and output placeholder
x_ph_bin = tf.placeholder(tf.float32, [M, len(binfeats)], name='x_bin') # binary inputs
x_ph_cont = tf.placeholder(tf.float32, [M, len(contfeats) + pnoise_size], name='x_cont') # continuous inputs
t_ph = tf.placeholder(tf.float32, [M, 1])
y_ph = tf.placeholder(tf.float32, [M, 1])
x_ph = tf.concat([x_ph_bin, x_ph_cont], 1)
activation = tf.nn.elu
"""Model Building"""
# CEVAE model (decoder)
# p(z)
mu_z = tf.zeros([tf.shape(x_ph)[0], d])
sigma_z = tf.ones([tf.shape(x_ph)[0], d])
z = Normal(loc=mu_z, scale=sigma_z)
# p(x|z)
hx = fc_net(z, (nh - 1) * [h], [], 'px_z_shared', lamba=lamba, activation=activation)
logits = fc_net(hx, [h], [[len(binfeats), None]], 'px_z_bin'.format(i + 1), lamba=lamba, activation=activation)
x1 = Bernoulli(logits=logits, dtype=tf.float32, name='bernoulli_px_z')
mu, sigma = fc_net(hx, [h], [[len(contfeats) + pnoise_size, None], [len(contfeats) + pnoise_size, tf.nn.softplus]], 'px_z_cont', lamba=lamba,
activation=activation)
x2 = Normal(loc=mu, scale=sigma, name='gaussian_px_z')
# p(t|z)
logits = fc_net(z, [h], [[1, None]], 'pt_z', lamba=lamba, activation=activation)
t = Bernoulli(logits=logits, dtype=tf.float32)
# p(y|t,z)
mu2_t0 = fc_net(z, nh * [h], [[1, None]], 'py_t0z', lamba=lamba, activation=activation)
mu2_t1 = fc_net(z, nh * [h], [[1, None]], 'py_t1z', lamba=lamba, activation=activation)
if task != 'ihdp':
y = BernoulliWithSigmoidProbs(logits=t * mu2_t1 + (1. - t) * mu2_t0, dtype=tf.float32)
else:
y = Normal(loc=t * mu2_t1 + (1. - t) * mu2_t0, scale=tf.ones_like(mu2_t0))
xj_ph_cont = tf.placeholder(tf.float32, [M, len(contfeats)], name='xj_cont') # continuous inputs
tj_ph = tf.placeholder(tf.float32, [M, 1])
yj_ph = tf.placeholder(tf.float32, [M, 1])
xi_ph = tf.concat([xi_ph_bin, xi_ph_cont], 1)
xj_ph = tf.concat([xj_ph_bin, xj_ph_cont], 1)
activation = tf.nn.elu
# CEVAE model (decoder)
# p(z)
zi = Normal(loc=tf.zeros([tf.shape(xi_ph)[0], d]), scale=tf.ones([tf.shape(xi_ph)[0], d]))
zj = Normal(loc=tf.zeros([tf.shape(xj_ph)[0], d]), scale=tf.ones([tf.shape(xj_ph)[0], d]))
# p(x|z)
hxi = fc_net(zi, (nh - 1) * [h], [], 'px_z_shared', lamba=lamba, activation=activation)
logitsxi = fc_net(hxi, [h], [[len(binfeats), None]], 'px_z_bin'.format(it + 1), lamba=lamba, activation=activation)
xi1 = Bernoulli(logits=logitsxi, dtype=tf.float32, name='bernoulli_pxi_zi')
mui, sigmai = fc_net(hxi, [h], [[len(contfeats), None], [len(contfeats), tf.nn.softplus]], 'px_z_cont', lamba=lamba,
activation=activation)
xi2 = Normal(loc=mui, scale=sigmai, name='gaussian_pxi_zi')
hxj = fc_net(zj, (nh - 1) * [h], [], 'px_z_shared', lamba=lamba, activation=activation, reuse=True)
logitsxj = fc_net(hxj, [h], [[len(binfeats), None]], 'px_z_bin'.format(it + 1), lamba=lamba, activation=activation, reuse=True)
xj1 = Bernoulli(logits=logitsxj, dtype=tf.float32, name='bernoulli_pxj_zj')
muj, sigmaj = fc_net(hxj, [h], [[len(contfeats), None], [len(contfeats), tf.nn.softplus]], 'px_z_cont', lamba=lamba,
activation=activation, reuse=True)
xj2 = Normal(loc=muj, scale=sigmaj, name='gaussian_pxj_zj')
