def model_loss(self, data_ph, model_params):
with tf.variable_scope("loss"):
label = data_ph.get_label()
mask = data_ph.get_mask()
l2_loss_list = list()
for i, deconv in enumerate(self.predict_list):
deconv = self._filter_mask(deconv, mask)
label = self._filter_mask(label, mask)
l2_loss = mf.image_l2_loss(deconv, label, "l2_loss_%d" % i)
l2_loss_list.append(l2_loss)
tf.add_to_collection("losses", l2_loss)
l1_loss = mf.image_l1_loss(deconv, label, "l1_loss_%d" % i)
self.l1_loss = l1_loss
self.l2_loss = tf.add_n(l2_loss_list)
# Add domain loss
if model_params['use_da']:
pred = tf.reshape(self.da_cls, [-1, 2])
da_label = data_ph.get_da_label()
total_da_loss = tf.nn.softmax_cross_entropy_with_logits(
pred, da_label
)
weight_da_loss = data_ph.get_da_weight() * total_da_loss
self.da_loss = tf.reduce_mean(weight_da_loss)
tf.add_to_collection("losses", self.da_loss)
self.loss = tf.add_n(tf.get_collection('losses'), name='total_loss')
def model_loss(self, data_ph, model_params):
with tf.variable_scope("loss"):
label = data_ph.get_label()
mask = data_ph.get_mask()
l1_loss_list = list()
l2_loss_list = list()
for i, deconv in enumerate(self.predict_list):
deconv = self._filter_mask(deconv, mask)
label = self._filter_mask(label, mask)
l2_loss = mf.image_l2_loss(deconv, label, "l2_loss_%d" % i)
l2_loss_list.append(l2_loss)
tf.add_to_collection("losses", l2_loss)
l1_loss = mf.image_l1_loss(deconv, label, "l1_loss_%d" % i)
tf.add_to_collection("losses", l1_loss)
l1_loss_list.append(l1_loss)
self.l1_loss = l1_loss_list
self.l2_loss = tf.add_n(l2_loss_list)
self.loss = tf.add_n(tf.get_collection('losses'),
name='total_loss')
def _image_l2_loss(self, label, mask, predict_list, index, model_params):
"""
Args:
label: [b, h, w, c]
mask: [b, h, w, c]
predict_list: list of [b, h, w, c]
"""
l2_loss_list = list()
for i, deconv in enumerate(predict_list):
deconv = self._filter_mask(deconv, mask)
label = self._filter_mask(label, mask)
l2_loss = mf.image_l2_loss(deconv, label,
"image_loss_%d_%d" % (index, i))
l2_loss_list.append(l2_loss)
count_diff = mf.count_diff(deconv, label,
"count_diff_%d_%d" % (index, i))
tf.summary.scalar("count_diff_%d_%d" % (index, i), count_diff)
#tf.add_to_collection("losses", l2_loss)
l2_loss = tf.add_n(l2_loss_list)
return l2_loss
