Peng Foo comments

Results 4 comments of


                                            Peng Foo

haha its four months later and this bug fix is still not merged

Same for me.

原因应当是未shuffle，导致全唐诗后面的唐诗实际上是长短句，长短句相对于古诗loss会升高，导致epoch末尾输出的loss会高。解决方案： 1.从corpus中删除长短句 2.batch generation 那里每一epoch应当shuffle data。 gonna fix this

此外，输出sample级别的loss没有什么参考价值，至少应该输出mini batch的average loss gonna fix this