def build_loss(self, *args, **kargs):
if self.config.loss == "softmax_loss":
self.loss, _ = point_wise_loss.softmax_loss(self.logits, self.gold_label,
*args, **kargs)
elif self.config.loss == "sparse_amsoftmax_loss":
self.loss, _ = point_wise_loss.sparse_amsoftmax_loss(self.logits, self.gold_label,
self.config, *args, **kargs)
elif self.config.loss == "focal_loss_multi_v1":
self.loss, _ = point_wise_loss.focal_loss_multi_v1(self.logits, self.gold_label,
self.config, *args, **kargs)
if self.config.with_center_loss:
self.center_loss, _ = point_wise_loss.center_loss_v2(self.sent_repres,
self.gold_label, centers=self.memory,
config=self.config,
*args, **kargs)
self.loss = self.loss + self.config.center_gamma * self.center_loss
if self.config.get("mode", "train") == "train":
if self.config.with_label_regularization:
print("===with class regularization===")
self.class_loss, _ = point_wise_loss.focal_loss_multi_v1(
self.class_logits, self.gold_label,
self.config, *args, **kargs)
self.loss += self.config.class_penalty * self.class_loss
trainable_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, self.scope)
print("List of Variables:")
for v in trainable_vars:
print(v.name)
def build_loss(self, *args, **kargs):
if self.config.loss == "softmax_loss":
self.loss, _ = point_wise_loss.softmax_loss(
self.logits, self.gold_label, *args, **kargs)
elif self.config.loss == "sparse_amsoftmax_loss":
self.loss, _ = point_wise_loss.sparse_amsoftmax_loss(
self.logits, self.gold_label, self.config, *args, **kargs)
elif self.config.loss == "focal_loss_binary_v2":
self.loss, _ = point_wise_loss.focal_loss_binary_v2(
self.logits, self.gold_label, self.config, *args, **kargs)
elif self.config.loss == "contrastive_loss":
if self.config.metric in ["Hyperbolic"]:
self.loss = pair_wise_loss.hyper_contrastive_loss(
self.dist, self.gold_label, self.config, is_quardic=True)
elif self.config.metric in ["Euclidean", "Arccosine", "Cosine"]:
self.loss = pair_wise_loss.contrastive_loss(self.dist,
self.gold_label,
self.config,
is_quardic=True)
if self.config.get("weight_decay", None):
model_vars = tf.trainable_variables()
lossL2 = tf.add_n([tf.nn.l2_loss(v) for v in model_vars \
if 'bias' not in v.name ])
lossL2 *= self.config.weight_decay
self.loss += lossL2
if self.config.rnn == "universal_transformer":
print("====encoder type====", self.config.rnn)
self.loss += (self.act_loss / 2.0)
def build_loss(self, *args, **kargs):
if self.config.loss == "softmax_loss":
self.loss, _ = point_wise_loss.softmax_loss(self.logits, self.gold_label,
*args, **kargs)
elif self.config.loss == "sparse_amsoftmax_loss":
self.loss, _ = point_wise_loss.sparse_amsoftmax_loss(self.logits, self.gold_label,
self.config, *args, **kargs)
elif self.config.loss == "focal_loss_binary_v2":
self.loss, _ = point_wise_loss.focal_loss_binary_v2(self.logits, self.gold_label,
self.config, *args, **kargs)
def _get_loss(self, *args, **kargs):
if self.config.loss == "softmax_loss":
self.loss, _ = point_wise_loss.softmax_loss(self.logits, self.labels,
*args, **kargs)
elif self.config.loss == "sparse_amsoftmax_loss":
self.loss, _ = point_wise_loss.sparse_amsoftmax_loss(self.logits, self.labels,
self.config, *args, **kargs)
elif self.config.loss == "focal_loss_binary_v2":
self.loss, _ = point_wise_loss.focal_loss_binary_v2(self.logits, self.labels,
self.config, *args, **kargs)
elif self.config.loss == "focal_loss_multi_v1":
self.loss, _ = point_wise_loss.focal_loss_multi_v1(self.logits, self.labels,
self.config, *args, **kargs)
def build_loss(self, *args, **kargs):
if self.config.loss == "softmax_loss":
self.loss, _ = point_wise_loss.softmax_loss(
self.logits, self.gold_label, *args, **kargs)
elif self.config.loss == "sparse_amsoftmax_loss":
self.loss, _ = point_wise_loss.sparse_amsoftmax_loss(
self.logits, self.gold_label, self.config, *args, **kargs)
elif self.config.loss == "focal_loss_multi_v1":
self.loss, _ = point_wise_loss.focal_loss_multi_v1(
self.logits, self.gold_label, self.config, *args, **kargs)
if self.config.with_center_loss:
self.center_loss, _ = point_wise_loss.center_loss_v2(
self.sent_repres, self.gold_label, self.config, *args, **kargs)
self.loss = self.loss + self.config.center_gamma * self.center_loss
def build_loss(self, *args, **kargs):
with tf.device('/device:GPU:%s' % gpu_id):
if self.args.loss == "softmax_loss":
soft_loss, _ = point_wise_loss.softmax_loss(
self.estimation, self.target, *args, **kargs)
self.loss = soft_loss
elif self.args.loss == "sparse_amsoftmax_loss":
soft_loss, _ = point_wise_loss.sparse_amsoftmax_loss(
self.estimation, self.target, *args, **kargs)
self.loss = soft_loss
elif self.args.loss == "focal_loss_binary_v2":
soft_loss, _ = point_wise_loss.focal_loss_binary_v2(
self.estimation, self.target, *args, **kargs)
self.loss = soft_loss
def build_loss(self, *args, **kargs):
if self.config.loss == "softmax_loss":
self.loss, _ = point_wise_loss.softmax_loss(
self.logits, self.gold_label, *args, **kargs)
elif self.config.loss == "sparse_amsoftmax_loss":
self.loss, _ = point_wise_loss.sparse_amsoftmax_loss(
self.logits, self.gold_label, self.config, *args, **kargs)
elif self.config.loss == "focal_loss_binary_v2":
self.loss, _ = point_wise_loss.focal_loss_binary_v2(
self.logits, self.gold_label, self.config, *args, **kargs)
if self.config.get("weight_decay", None):
for var in set(tf.get_collection('reg_vars', self.scope)):
weight_decay = tf.multiply(tf.nn.l2_loss(var),
self.config.weight_decay,
name="{}-wd".format('-'.join(
str(var.op.name).split('/'))))
self.loss += weight_decay
def build_loss(self, *args, **kargs):
if self.config.loss == "softmax_loss":
self.loss, _ = point_wise_loss.softmax_loss(self.logits, self.gold_label,
*args, **kargs)
elif self.config.loss == "sparse_amsoftmax_loss":
self.loss, _ = point_wise_loss.sparse_amsoftmax_loss(self.logits, self.gold_label,
self.config, *args, **kargs)
elif self.config.loss == "focal_loss_binary_v2":
self.loss, _ = point_wise_loss.focal_loss_binary_v2(self.logits, self.gold_label,
self.config, *args, **kargs)
if self.config.l2_loss:
if self.config.sigmoid_growing_l2loss:
weights_added = tf.add_n([tf.nn.l2_loss(tensor) for tensor in tf.trainable_variables() if tensor.name.endswith("weights:0") and not tensor.name.endswith("weighted_sum/weights:0") or tensor.name.endswith('kernel:0')])
full_l2_step = tf.constant(self.config.weight_l2loss_step_full_reg , dtype=tf.int32, shape=[], name='full_l2reg_step')
full_l2_ratio = tf.constant(self.config.l2_regularization_ratio , dtype=tf.float32, shape=[], name='l2_regularization_ratio')
gs_flt = tf.cast(self.global_step , tf.float32)
half_l2_step_flt = tf.cast(full_l2_step / 2 ,tf.float32)
l2loss_ratio = tf.sigmoid( ((gs_flt - half_l2_step_flt) * 8) / half_l2_step_flt) * full_l2_ratio
else:
l2loss = tf.add_n([tf.nn.l2_loss(tensor) for tensor in tf.trainable_variables() if tensor.name.endswith("weights:0") or tensor.name.endswith('kernel:0')]) * tf.constant(config.l2_regularization_ratio , dtype='float', shape=[], name='l2_regularization_ratio')
self.loss += l2loss
