[tensorflow 연어] tf.clip_by_global_norm

다음 두 블로그는 이 문제에 대해 매우 상세하게 말한다.https://blog.csdn.net/u010814042/article/details/76154391https://blog.csdn.net/u013713117/article/details/56281715아래의 코드는 쌍방향 lstm+crf로 손실을 계산하는 부분이지만 사고방식은 통용된다. 

with tf.variable_scope('crf'):
    self.log_likelihood, self.transition_params = crf.crf_log_likelihood(inputs=self.logits,tag_indices=self.targets,sequence_lengths=self.sequence_length)

with tf.name_scope('loss'):
    #      self.log_likelihood  ，      -self.log_likelihood    
    self.log_loss = tf.reduce_mean(-self.log_likelihood)
    #  
    self.l2_loss = tf.contrib.layers.apply_regularization(
        regularizer=tf.contrib.layers.l2_regularizer(0.0001),
        weights_list=tf.trainable_variables())
    self.loss = self.log_loss+self.l2_loss

self.train_summary = tf.summary.scalar('loss',self.loss)
self.vali_summary = tf.summary.scalar('loss',self.loss)
# self.merged = tf.summary.merge_all()
optimizer = tf.train.AdamOptimizer(self.learning_rate)
# optimizer = tf.train.GradientDescentOptimizer(self.learning_rate)
#          
tvars = tf.trainable_variables()
#gradients     ，   loss   trainable   。
#clip_by_global_norm          trainable     ，   trainable  ，     clip    ，   globalnorm
grads, _ = tf.clip_by_global_norm(tf.gradients(self.loss, tvars), self.clip_grad)
#apply_gradients tf.train.Optimizer.minimize           ，minizie         ，       tf.train.Optimizer.compute_gradients        
#tf.train.Optimizer.apply_gradients，             ，              ，        zip(    ，    )
#tvars           ：https://blog.csdn.net/u014595019/article/details/52805444
self.train_op = optimizer.apply_gradients(zip(grads, tvars))