def build_loss(self):
self.wd = self.config["weight_decay"]
self.loss = 0
for var in set(tf.get_collection('reg_vars', self.scope)):
weight_decay = tf.multiply(tf.nn.l2_loss(var),
self.wd,
name="{}-wd".format('-'.join(
str(var.op.name).split('/'))))
self.loss += weight_decay
with tf.name_scope("loss"):
if self.config["loss_type"] == "cross_entropy":
self.loss += tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=self.estimation, labels=self.gold_label))
elif self.config["loss_type"] == "focal_loss":
one_hot_target = tf.one_hot(self.gold_label, 2)
print(one_hot_target.get_shape())
self.loss += focal_loss(self.estimation, one_hot_target)
tf.add_to_collection('ema/scalar', self.loss)
print("List of Variables:")
for v in tf.trainable_variables():
print(v.name)
def build_loss(self):
with tf.name_scope("loss"):
if self.config["loss_type"] == "cross_entropy":
self.loss = tf.add(
tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=self.estimation, labels=self.gold_label)),
tf.reduce_sum(
tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)),
name="loss")
elif self.config["loss_type"] == "focal_loss":
one_hot_target = tf.one_hot(self.gold_label, 2)
print(one_hot_target.get_shape())
self.loss = focal_loss(self.estimation, one_hot_target)
elif self.config["loss_type"] == "contrastive_loss":
if self.config["distance_metric"] == "l1_similarity":
self.loss = tf.add(
tf.reduce_mean(
tf.losses.softmax_cross_entropy(
self.one_hot_label, self.pred_probs)),
tf.reduce_sum(
tf.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES)),
name="loss")
elif self.config["distance_metric"] == "cosine_constrastive":
self.loss = tf.add(
tf.reduce_mean(self.contrastive_distance),
tf.reduce_sum(
tf.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES)),
name="loss")
elif self.config["distance_metric"] == "l2_similarity":
self.loss = tf.add(
tf.reduce_mean(self.contrastive_distance),
tf.reduce_sum(
tf.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES)),
name="loss")
tf.add_to_collection('ema/scalar', self.loss)
print("List of Variables:")
for v in tf.trainable_variables():
print(v.name)
def build_loss(self):
self.wd = self.config.get("l2_reg", None)
with tf.name_scope("loss"):
if self.config["loss_type"] == "cross_entropy":
self.loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=self.estimation, labels=self.gold_label))
elif self.config["loss_type"] == "focal_loss":
one_hot_target = tf.one_hot(self.gold_label, 2)
print(one_hot_target.get_shape())
self.loss = focal_loss(self.estimation, one_hot_target)
elif self.config["loss_type"] == "contrastive_loss":
if self.config["distance_metric"] == "l1_similarity":
self.loss = tf.reduce_mean(
tf.losses.softmax_cross_entropy(
self.one_hot_label, self.pred_probs))
elif self.config["distance_metric"] == "cosine_constrastive":
self.loss = tf.reduce_mean(self.contrastive_distance)
elif self.config["distance_metric"] == "l2_similarity":
self.loss = tf.reduce_mean(self.contrastive_distance)
wd_loss = 0
print("------------wd--------------", self.wd)
if self.wd is not None:
for var in set(tf.get_collection('reg_vars', self.scope)):
weight_decay = tf.multiply(
tf.nn.l2_loss(var),
self.wd,
name="{}-wd".format('-'.join(
str(var.op.name).split('/'))))
wd_loss += weight_decay
print("---------using l2 regualarization------------")
self.loss += wd_loss
tf.add_to_collection('ema/scalar', self.loss)
print("List of Variables:")
for v in tf.trainable_variables():
print(v.name)