zhaoyingjun / chatbot

ChatGPT带火了聊天机器人,主流的趋势都调整到了GPT类模式,本项目也与时俱进,会在近期更新GPT类版本。基于本项目和自己的语料可以训练出自己想要的聊天机器人,用于智能客服、在线问答、闲聊等场景。
3.53k stars 1.02k forks source link

seqGAN模型中discrimination是RNN的问题 #64

Open onlyhyl opened 4 years ago

onlyhyl commented 4 years ago

作者您好,很感谢您开源的代码! 我想请教一下,seqGAN中判别器是用到了RNN吗?为什么不是CNN?

zhaoyingjun commented 4 years ago

用cnn需要解决时序位置的关系,如果你想要使用cnn 可以参考使用cnn来做的自然语言处理项目

onlyhyl commented 4 years ago

了解!那在该模型中,我在预训练生成器后,利用预训练好的生成器为判别器生成伪数据的时候,该生成的伪数据基本上都是_UNK,只有零星几句是包含有一两个生活文字,请问这是正常的吗?而且在训练2000步的时候,生成的伪数据仍然如此。

train_answer_sentence: [3, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] b'_UNK' b'_UNK' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' lable: 1 train_answer_sentence: [434, 667, 179, 9, 11, 285, 19, 19, 3, 0, 0, 0, 0, 0, 0] b'\xe5\xa5\xb3\xe7\x8e\x8b' b'\xe9\x99\x9b\xe4\xb8\x8b' b'\xe6\x88\x91\xe9\x94\x99' b'\xe4\xba\x86' b'\xef\xbc\x81' b'\xe5\xaf\xb9\xe4\xb8\x8d\xe8\xb5\xb7' b'\xe2\x80\xa6' b'\xe2\x80\xa6' b'_UNK' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' lable: 1 train_answer_sentence: [58, 28, 7, 3, 3, 3, 3, 3, 17, 0, 0, 0, 0, 0, 0] b'\xe5\xb0\x8f\xe9\x80\x9a' b'\xe5\x90\x83' b'\xe7\x9a\x84' b'_UNK' b'_UNK' b'_UNK' b'_UNK' b'_UNK' b'~' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' lable: 1 train_answer_sentence: [5, 10, 3, 7, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] b'\xe6\x88\x91' b'\xe6\x98\xaf' b'_UNK' b'\xe7\x9a\x84' b'_UNK' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' lable: 0 train_answer_sentence: [3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3] b'_UNK' b'_UNK' b'_UNK' b'_UNK' b'_UNK' b'_UNK' b'_UNK' b'_UNK' b'_UNK' b'_UNK' b'_UNK' b'_UNK' b'_UNK' b'_UNK' b'_UNK' lable: 0 train_answer_sentence: [3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] b'_UNK' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' b'_PAD' lable: 0

麻烦您了,感谢!