0

I am doing small project using Convolution Neural Network. I code is base on dennybritz. Here my Convect code and here is two function load data and CNN architect. I stuck in the evaluation process, my evaluation is always around 0.42 (The red star lines)training results.

The result should be above 0.6. I tried change parameter of the Network or change the size of data but nothing change. Could anyone suggest what is wrong with my code and how optimize to get a better results.

ngoduyvu
  • 241
  • 4
  • 16
  • what are the blue points? – miraculixx Sep 26 '16 at 04:50
  • If you going to the link on Convect you can see that on the training process It will generate accuracy for each step it is the blue point. The red dots are the accuracy when estimate. – ngoduyvu Sep 26 '16 at 05:04
  • to increase the chances you get a good answer, please provide a more succint description of what you are trying to achieve and what you have tried so far – miraculixx Sep 26 '16 at 05:09
  • Oh sorry my bad. So I am trying to do text classification using Convolution Neural Network on text document like dennybritz did. However, my data has 45 labels instead of 2 labels. I success run his model on data and use it to classify 45 label. But in the evaluate step, the model run on test data giving really low accuracy. I am tried to figure out what I am doing wrong and how to fix it. – ngoduyvu Sep 26 '16 at 05:20
  • Did not look at your code. But your accuracy on training is much better than test dataset, which signals a potential issue of over-fitting. Could you try to either 1) decrease the keep prob on your dropout layer 2) to decrease the dimension of your embedding or the number of filter banks (to decrease the number of model parameters). – greeness Sep 28 '16 at 22:19
  • Thanks greeness, I changed all of the parameter one by one to see how it affect the outcome. Here is the [result](https://www.dropbox.com/home?preview=Data+result.pdf). It is little long but I made a summary at the final section. – ngoduyvu Sep 28 '16 at 22:57
  • [file outcome](https://www.dropbox.com/s/qhg92pgyua3zt5u/Data%20result.pdf?dl=0) – ngoduyvu Sep 28 '16 at 23:05
  • Looks like you did get better results. That's nice. BTW, the keep prob of dropout should be between 0 and 1; it does not make sense to make it beyond 1.0. – greeness Sep 29 '16 at 01:49

0 Answers0