search

cjiang2 / VDCNN

Implementation of Very Deep Convolutional Neural Network for Text Classification
171 stars 41 forks source link

Training fails #15

Closed zjfheart closed 3 years ago

zjfheart commented 5 years ago

Hi,

I run your code on ag_news, but every time encounter a training failure. (test acc is 25%).

Look like the way you recovered has some issues. any ideas about this?

WindHunter-Yang commented 5 years ago

me, too How to solve this problem

Valenser commented 5 years ago

have the same problem. any help? :)

zjfheart commented 5 years ago

Pls, refer to this paper https://arxiv.org/abs/1902.10887. Make your h smaller. This can helps with training.

zhongzhh8 commented 5 years ago

have the same problem. any help? :)

how did you solve this problem finally?I need your help

zjfheart commented 5 years ago

Screen Shot 2019-05-09 at 11 51 27 For AG-NEWS dataset, this is my results. 25 acc signify training failure.

zhongzhh8 commented 5 years ago

Pls, refer to this paper https://arxiv.org/abs/1902.10887. Make your h smaller. This can helps with training. I have the error of a constantly 25 acc in AF_NEWS and do not know how to solve this problem. Do you mean that you change the parameter lr=0.1 in "optimizer=SGD(lr=0.01, momentum=0.9)"? I try it but it does not work. Please tell me how you solved this problem in detail. I need your help.

zhongzhh8 commented 5 years ago

@zjfheart thank you very much. In AG_NEWS and deep 9, I try to set lr=0.001,and it works! After 1 epoch it has raised to 80 acc . Your answer reminds me of the training step. Thank you .