search

Doragd / Chinese-Chatbot-PyTorch-Implementation

:four_leaf_clover: Another Chinese chatbot implemented in PyTorch, which is the sub-module of intelligent work order processing robot. 👩‍🔧
Apache License 2.0
879 stars 192 forks source link

关于训练时出现的NaN问题 #5

Closed super233 closed 4 years ago

super233 commented 4 years ago

大佬您好，我再使用您的代码训练qingyun数据时，出现Encoder的输出全部为nan，进而导致loss为nan，请问您当时遇到过这种情况吗？

zhifliao commented 4 years ago

遇到相同的问题，发现只要训练到2000次以上就会出现，请问大佬当时是如何解决的？

super233 commented 4 years ago

我当时把 decoder_learning_ratio设为1.0 后，训练起来就还不错

image-20191223091056754

image-20191223091220405

zhifliao commented 4 years ago

收到，非常感谢，我修改运行试试。

------------------ 原始邮件 ------------------

发件人: Super233 notifications@github.com

发送时间: 2020-04-22 14:28:51

收件人:Doragd/Chinese-Chatbot-PyTorch-Implementation chinese-chatbot-pytorch-implementation@noreply.github.com

抄送:廖志芳 13902220306@139.com,Comment comment@noreply.github.com

主题: Re: [Doragd/Chinese-Chatbot-PyTorch-Implementation] 关于训练时出现的NaN问题 (#5)

我当时把 decoder_learning_ratio设为1.0 后，训练起来就还不错

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.