def gradient_penalty(inter_data, theta):
_, inter = mlp(inter_data, theta)
gradients = tf.gradients([inter], [inter_data])[0]
slopes = tf.sqrt(
tf.reduce_mean(tf.square(gradients), reduction_indices=[1]))
gradient_penalty = tf.reduce_mean((slopes - 1)**2)
return gradient_penalty
var] = normalize_toydata(toydata, testcase, var)
# ------------------ Network architecture ------------------------------
X = tf.placeholder(tf.float32, shape=[None, X_dim])
z = tf.placeholder(tf.float32, shape=[None, z_dim])
""" Generator: G(z) """
theta_G = theta_def([z_dim, 128, 128, 128, X_dim])
""" Encoder: E(X) """
theta_E = theta_def([X_dim, 128, 128, 128, z_dim])
""" Discriminator: D(x) """
theta_D = theta_def([X_dim, 128, 128, 128, 1])
# ------------------ Setup criteria/loss functions ---------------------
# Encode and decode data
_, _ze = mlp(X, theta_E)
_Xr_prob, _Xr_logit = mlp(_ze, theta_G)
# Latent interpolation
alpha = 0.5
zi = (1. - 1. * alpha) * z + 1. * alpha * _ze
# Sample from random z
_X_prob, _X_logit = mlp(z, theta_G)
_Xi_prob, _Xi_logit = mlp(zi, theta_G)
D_real, D_real_logit, f_real = mlp_feat(X, theta_D)
D_recon, D_recon_logit, f_recon = mlp_feat(_Xr_prob, theta_D)
D_fake, D_fake_logit, f_fake = mlp_feat(_X_prob, theta_D)
D_inter, D_inter_logit, f_inter = mlp_feat(_Xi_prob, theta_D)
def gradient_penalty(inter_data, theta):
_, inter = mlp(inter_data, theta)
gradients = tf.gradients([inter], [inter_data])[0]
slopes = tf.sqrt(tf.reduce_mean(tf.square(gradients), reduction_indices=[1]))
gradient_penalty = tf.reduce_mean((slopes - 1) ** 2)
return gradient_penalty