def gumbel_loss(x, x_hat):
q_y = K.reshape(logits_y, (-1, N, M))
q_y = softmax(q_y)
log_q_y = K.log(q_y + 1e-20)
kl_tmp = q_y * (log_q_y - K.log(1.0/M))
KL = K.sum(kl_tmp, axis=(1, 2))
elbo = latent_dim * bce(x, x_hat) - KL
# elbo = latent_dim * mse(x, x_hat) - KL
return elbo
def vae_loss(x, x_mean):
xent_loss = img_dim * bce(x_s, x_mean)
kl_z_loss = mvn_kl(z_mean_en, z_logvar_en)
kl_a_loss = mvn_kl(a_mean_en, a_logvar_en, a_mean_de, a_logvar_de)
kl_loss = kl_z_loss + kl_a_loss
# return kl_loss + xent_loss
return kl_weight * kl_loss + xent_loss
def vae_label_loss(x, x_prob_de): x_loss = img_dim * bce(x, x_prob_de) y_loss = - K.log(nb_classes) kl_a_loss = mvn_kl(a_mean_en, a_logvar_en, a_mean_de, a_logvar_de) kl_z_loss = mvn_kl(z_mean_en, z_logvar_en) rec_loss = x_loss + y_loss kl_loss = kl_a_loss + kl_z_loss total_loss = rec_loss + kl_weight * kl_loss return total_loss
def vae_unlabel_loss(x_u, x_u_prob_de): for i in xrange(nb_classes): y_i_prob = y_u_probs_en[:, i] x_u_prob_de_y = x_u_prob_de[i] a_u_mean_de_y = a_u_mean_de_y[i] a_u_logvar_de_y = a_u_logvar_de_y[i] z_u_mean_en_y = z_u_mean_en_y[i] z_u_logvar_en_y = z_u_logvar_en_y[i] x_u_loss = img_dim * bce(x_u, x_u_prob_de_y) kl_y_loss = K.log(nb_classes * y_i_prob) kl_a_u_loss = mvn_kl(a_u_mean_en, a_u_logvar_en, a_u_mean_de_y, a_u_logvar_de_y) kl_z_u_loss = mvn_kl(z_u_mean_en_y, z_u_logvar_en_y) if i==0: rec_loss = y_i_prob * x_u_loss kl_loss = y_i_prob * (kl_y_loss + kl_a_u_loss + kl_z_u_loss) else: rec_loss += y_i_prob * x_u_loss kl_loss += y_i_prob * (kl_y_loss + kl_a_u_loss + kl_z_u_loss) total_loss = rec_loss + kl_weight * kl_loss return total_loss
def bcce(true, pred):
if K.ndim(true) == 3:
return cce(K.clip(true, 1e-8, 1-1e-8), K.clip(pred, 1e-8, 1-1e-8))
else:
return bce(K.clip(true, 1e-8, 1-1e-8), K.clip(pred, 1e-8, 1-1e-8))
def BCE(x, y):
return bce(K.batch_flatten(x),
K.batch_flatten(y))