codefish1990

Question：multihead_attention()’s output and ff()‘s output need dropout？

1、v0.8.0的dien.py中，auxiliary_net()中的tf.layers.batch_normalization()没有设置training，此bn是否还能生效？ 2、从auxiliary_net换成了DNN背后的原因是？

inputs.py文件，embedding_lookup()函数中，if fc.use_hash = False，那mask_feat_list不起作用。

about the "Center-Wise Clustering for Source Clustering" loss

1

in ./ESAM/test/center_loss.py，the function def get_center_loss(features, labels, num_classes): ... return (loss_0 + loss_1)/(2*10+2*2) why divide by (2*10+2*2) ？ why is (2*10+2*2)？