def get_loss(self, image):
recon_mean, latent_mean, latent_log_std, latent_samples = \
self.forward(image)
# kl term
kl_q = modeling_lib.get_kl_q_standard_normal(latent_mean, latent_log_std)
# bernoulli likelihood
loglik = modeling_lib.get_bernoulli_loglik(recon_mean, image)
return -loglik + kl_q
def get_loss(self, image):
recon_mean, latent_mean, latent_log_std, latent_samples = \
self.forward(image)
# kl term
kl_q = modeling_lib.get_kl_q_standard_normal(latent_mean,
latent_log_std)
# bernoulli likelihood
loglik = modeling_lib.get_bernoulli_loglik(recon_mean, image)
return -loglik + kl_q
def get_loss_cond_pixel_1d(self, image, pixel_1d):
# forward
recon_mean, latent_mean, latent_log_std, latent_samples, pixel_2d = \
self.forward_cond_pixel_1d(image, pixel_1d)
# kl term
kl_latent = \
modeling_lib.get_kl_q_standard_normal(latent_mean, latent_log_std)
# bernoulli likelihood
loglik = modeling_lib.get_bernoulli_loglik(recon_mean, image)
return -loglik + kl_latent
def get_loss_cond_pixel_1d(self, image, pixel_1d):
# forward
recon_mean, latent_mean, latent_log_std, latent_samples, pixel_2d = \
self.forward_cond_pixel_1d(image, pixel_1d)
# kl term
kl_latent = \
modeling_lib.get_kl_q_standard_normal(latent_mean, latent_log_std)
# bernoulli likelihood
loglik = modeling_lib.get_bernoulli_loglik(recon_mean, image)
return -loglik + kl_latent
def get_loss_cond_pixel_1d(self, one_hot_pixel, image, \
use_cached_image = False):
# forward
if use_cached_image:
image_ = None
else:
image_ = image
recon_mean, latent_mean, latent_log_std, latent_samples = \
self.forward_cond_pixel_1d(one_hot_pixel, image_)
# kl term
kl_latent = \
modeling_lib.get_kl_q_standard_normal(latent_mean, latent_log_std)
# bernoulli likelihood
loglik = modeling_lib.get_bernoulli_loglik(recon_mean, image)
return -loglik + kl_latent