今天在翻espnet源码的时候,发现了这个东西RNNP,注明是带projection layer的,这玩意好像在NLP那边经常用,但是语音识别这边确实不多。 LSTM的公式和图如下[1]: LSTMP的公式和图如下: 可以看出,关键的区别在于cell-memory之后有一个projection。 按照官方的描述,如果增加cell_memory会同时以二次增加
a湿陷性土层层底深度7.00～12.25m Falls wet soil layer level bottom depth 7.00~12.25m [translate] aGalapagos 加拉帕戈斯 [translate] aA new day has come . A new month has come .The last month of 2011 has come and is leaving .What I have harvested and lost is not significant .The most ...