brightmart
commented
6 years ago general speaking, people load pretrain word embedding and continue to train word embedding on domain related task. the reason is we need a general representation of words, but we also need the representation fit well on domain related task.
we apply a small learning rate during fine-tuning stage to make sure the adjustment is small and effective.
如果用了Tencent_AILab_ChineseEmbedding_100w.txt来做pre-trained word embedding，tf.assign之后不需要tf.stop_gradient吗？不然训练的时候embedding就被改了，难道只是起一下更好的初始化的作用？