git.sesse.net Git - voxel-flow/blob - utils/loss_utils.py

   1 """Implements various tensorflow loss layer.
   2 """
   3
   4 from __future__ import absolute_import
   5 from __future__ import division
   6 from __future__ import print_function
   7
   8 import tensorflow as tf
   9 import tensorflow.contrib.slim as slim
  10
  11 def l1_loss(predictions, targets):
  12   """Implements tensorflow l1 loss.
  13   Args:
  14   Returns:
  15   """
  16   total_elements = (tf.shape(targets)[0] * tf.shape(targets)[1] * tf.shape(targets)[2]
  17       * tf.shape(targets)[3])
  18   total_elements = tf.to_float(total_elements)
  19
  20   loss = tf.reduce_sum(tf.abs(predictions- targets))
  21   loss = tf.div(loss, total_elements)
  22   return loss
  23
  24 def l2_loss(predictions, targets):
  25   """Implements tensorflow l2 loss, normalized by number of elements.
  26   Args:
  27   Returns:
  28   """
  29   total_elements = (tf.shape(targets)[0] * tf.shape(targets)[1] * tf.shape(targets)[2]
  30       * tf.shape(targets)[3])
  31   total_elements = tf.to_float(total_elements)
  32
  33   loss = tf.reduce_sum(tf.square(predictions-targets))
  34   loss = tf.div(loss, total_elements)
  35   return loss
  36
  37 def tv_loss():
  38   #TODO
  39   pass
  40 def vae_loss(z_mean, z_logvar, prior_weight=1.0):
  41   """Implements the VAE reguarlization loss.
  42   """
  43   total_elements = (tf.shape(z_mean)[0] * tf.shape(z_mean)[1] * tf.shape(z_mean)[2]
  44       * tf.shape(z_mean)[3])
  45   total_elements = tf.to_float(total_elements)
  46
  47   vae_loss = -0.5 * tf.reduce_sum(1.0 + z_logvar - tf.square(z_mean) - tf.exp(z_logvar))
  48   vae_loss = tf.div(vae_loss, total_elements)
  49   return vae_loss
  50
  51 def bilateral_loss():
  52   #TODO
  53   pass