search

pochih / RL-Chatbot

🤖 Deep Reinforcement Learning Chatbot
MIT License
418 stars 140 forks source link

您好，我重新训练SEQ2SEQ模型后用 text.py 但是得到以下结果: #6

Closed chaiyixuan closed 6 years ago

chaiyixuan commented 6 years ago

您好，我重新训练SEQ2SEQ模型后用 text.py 但是得到以下结果: Concentration concentration concentration concentration planning priest. Concentration concentration concentration concentration planning concentration planning priest. Concentration concentration concentration concentration planning priest. Concentration concentration concentration concentration planning concentration planning priest. Concentration concentration concentration concentration planning priest. Concentration concentration concentration concentration planning concentration planning priest. 您知道是什么原因么？

pochih commented 6 years ago

很抱歉， 我自己執行時沒有這個錯誤 不知道您在執行時版本是否正確呢？

chaiyixuan commented 6 years ago

之前Adam由于版本原因无法使用，我改用GradientDecent训练，导致无法收敛。后来做代码兼容性调整，还是用Adam训练，模型loss收敛到7.8左右。（我用batch80 和epoch50）但是test得出的结果好像不太好，您知道是什么原因么？

附结果： question => Have you heard about 'machine learning and having it deep and structured'? generated_sentence => President's poem. question => How are you? generated_sentence => Street cemetery doe. question => What's your name? generated_sentence => Scum baltimore result casino. Casino casino doyle. question => Hello generated_sentence => Scum baltimore result landed. Street couldn't zira ruin ha. question => Thank God! If I had to hear one more story about your coiffure generated_sentence => End. question => You never wanted to go out with me, did you?

pochih commented 6 years ago

可以試試看 RMSprop GradientDecent 收斂太慢。