def image_loss(self, logits, labels):
"""Cross-entropy between the logits and labels."""
logits = tf.squeeze(logits, axis=-2)
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels,
logits=logits)
loss = tf.reduce_mean(loss, axis=0)
loss = base_utils.nats_to_bits(tf.reduce_sum(loss))
return loss / (64.0 * 64.0)
def loss(self, targets, logits, train_config, training, aux_output=None):
if training:
labels = targets['targets_slice']
else:
labels = targets['targets']
height, width, num_channels = labels.shape[1:]
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels,
logits=logits)
loss = tf.reduce_mean(loss, axis=0)
loss = base_utils.nats_to_bits(tf.reduce_sum(loss))
loss = loss / (height * width * num_channels)
return loss, {}
def loss(self, targets, logits, train_config, training, aux_output=None):
is_downsample = train_config.get('downsample', False)
downsample_res = train_config.get('downsample_res', 64)
if is_downsample and training:
labels = targets['targets_slice_%d' % downsample_res]
elif is_downsample:
labels = targets['targets_%d' % downsample_res]
elif training:
labels = targets['targets_slice']
else:
labels = targets['targets']
height, width, num_channels = labels.shape[1:4]
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels,
logits=logits)
loss = tf.reduce_mean(loss, axis=0)
loss = base_utils.nats_to_bits(tf.reduce_sum(loss))
loss = loss / (height * width * num_channels)
return loss, {}