Created by: vikmary
-
implemented almost all features of key-value networks (attention over knowledge base, encoder attention, embeddings instead of one-hots) -
released a seq2seq_gobot_v2.tar.gz model with BLEU=13.2 -
updated doc
Activity
101 101 # For batch norm it is necessary to update running averages 102 102 extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) 103 103 with tf.control_dependencies(extra_update_ops): 104 105 def clip_if_not_none(grad): 106 if grad is None: 107 return grad 108 return tf.clip_by_norm(grad, clip_norm) 109 104 110 opt = optimizer(learning_rate) 105 111 grads_and_vars = opt.compute_gradients(loss, var_list=variables_to_train) 106 112 if clip_norm is not None: 107 grads_and_vars = [(tf.clip_by_norm(grad, clip_norm), var) 108 for grad, var in grads_and_vars] # if grad is not None 113 grads_and_vars = [(clip_if_not_none(grad), var) 101 101 # For batch norm it is necessary to update running averages 102 102 extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) 103 103 with tf.control_dependencies(extra_update_ops): 104 105 def clip_if_not_none(grad):
Please register or sign in to reply