def _gradient_penalty(self, X, X_hat, L, G, Z, Z_hat, EPSILON):
ax1 = [i for i in range(1, len(X.get_shape()))]
ax2 = [i for i in range(1, len(Z.get_shape()))]
with tf.name_scope('grad_penalty'):
with tf.name_scope('x'):
x_inter = ops.interpolate(X, X_hat, EPSILON, name='x_inter')
d_inter_x = self._discriminator(x_inter,
L,
is_training=True,
reuse=True)
x_grads, z_grads = tf.gradients([d_inter_x], [x_inter, L],
name="grads_x")
x_slopes = tf.sqrt(tf.reduce_sum(x_grads**2, axis=ax1))
tf.summary.scalar('x_slopes', tf.reduce_mean(x_slopes))
delta_x = X - X_hat
x_unit = delta_x / ops.norm(delta_x, ax1, True)
x_alpha = tf.reduce_sum(x_grads * x_unit, ax1) / x_slopes
tf.summary.scalar('x_alpha', tf.reduce_mean(x_alpha))
tf.summary.scalar('x_unit_grads',
tf.reduce_mean(ops.norm(x_unit, ax1)))
x_pen = tf.sqrt(tf.reduce_sum((x_grads - x_unit)**2, axis=ax1))
gp_loss_x = tf.reduce_mean(x_pen, name='gp_loss_x')
tf.add_to_collection('losses', gp_loss_x)
tf.summary.scalar('gp_loss_x', gp_loss_x)
with tf.name_scope('z'):
z_inter = ops.interpolate(Z, Z_hat, EPSILON, name='z_inter')
d_inter_z = -self._discriminator(
G, z_inter, is_training=True, reuse=True)
_, z_grads = tf.gradients([d_inter_z], [G, z_inter],
name="grads_z")
delta_z = Z - Z_hat
z_unit = delta_z / ops.norm(delta_z, ax2, True)
z_pen = tf.sqrt(tf.reduce_sum((z_grads - z_unit)**2, axis=ax2))
gp_loss_z = tf.reduce_mean(z_pen, name='gp_loss_z')
tf.add_to_collection('losses', gp_loss_z)
tf.summary.scalar('gp_loss_z', gp_loss_z)
gp_loss = gp_loss_x
# gp_loss = gp_loss_x + gp_loss_z
tf.add_to_collection('losses', gp_loss)
tf.summary.scalar('gp_loss', gp_loss)
return gp_loss
def _gradient_penalty(self, X, G, EPSILON, D_fn):
with tf.name_scope('grad_penalty'):
x_inter = ops.interpolate(X, G, EPSILON, name='x_inter')
d_inter = D_fn(x_inter, is_training=True, reuse=True)
grad = tf.gradients(d_inter, x_inter, name="grads")[0]
n_dim = len(grad.get_shape())
slopes = tf.sqrt(
tf.reduce_sum(tf.square(grad),
reduction_indices=[i for i in range(1, n_dim)]))
tf.summary.histogram('slopes', slopes)
gp_loss = tf.reduce_mean((slopes - 1.)**2, name='gp_loss')
tf.add_to_collection('losses', gp_loss)
tf.summary.scalar('gp_loss', gp_loss)
return gp_loss