train on default setting on qa1 but accuracy stay about 20%

therne / dmn-tensorflow

Dynamic Memory Networks (https://arxiv.org/abs/1603.01417) in Tensorflow

240 stars 82 forks source link

train on default setting on qa1 but accuracy stay about 20% #12

Open tomsonsgs opened 7 years ago

tomsonsgs commented 7 years ago

default

michael-wzhu commented 6 years ago

yes, i also observe this problem. Do you have some thoughts about how to solve it?

michael-wzhu commented 6 years ago

I have switched the initialization method from "He et al " to Xavier, the accuracy immediately increased a lot.

And i found that the configs have some differences from the original paper, such as learning rate.

JaneShenYY commented 6 years ago

@michael-wzhu I have got the same problem. Would you please specify how did you modify the code? and what accuracy for task1 did you obtain after the modification please?