search

macournoyer / neuralconvo

Neural conversational model in Torch
777 stars 346 forks source link

the result is not like what u said #58

Closed csrgxtu closed 7 years ago

csrgxtu commented 8 years ago

i use this command train:

th train.lua --dataset 50000 --hiddenSize 1000

though it is CPU, but after weeks, It trained, and test like this: result

CodePothunter commented 7 years ago

yep... You should know this author is not that of the paper and the data set is also different... Actually, I doubted that the samples given by Google could be reproduced, because it looked so "intelligent"!

macournoyer commented 7 years ago

Sadly, neither can I reproduce the results anymore :/ I added a warning about this in the README.

The cause seems to be a change in one of dependencies. I don't know which one. But I'm currently working on a new implementation based on harvardnlp/seq2seq-attn.

csrgxtu commented 7 years ago

this always confuse me, cause people say that a algorithm can achieve 97% accuracy, but in reality in my situation, it is only 70%.

nabihach commented 7 years ago

@macournoyer, any news about your new implementation? Eagerly looking forward to it.

ghost commented 7 years ago

@macournoyer I got the same issue as @CodePothunter mentioned when using harvardnlp/seq2seq-attn. Can anyone get the reasonable responses using the existing open source implementations?