1

I have my own dataset. I want to do a classification task. But I built same symbol network in Mxnet and Keras. Even the optimizer rules are same. But the results are different.

my Mxnet code is here: enter image description here

and results: it looks like ramdon? enter image description here

But my keras code are same network: enter image description here

enter image description here

but the result is much better. In training set, i can be 100% enter image description here

I still cannot figure out why there are same network architectures and data are same. However, the classification result between two frames is large.

Hope someone could give some suggestion. Thx.

Luv
  • 263
  • 3
  • 9
  • What data are you using to train? What data iterator are you using? Try adjusting the learning rate and the batch size. If the data is publicly available, I can help adjusting parameters to get your code working. – Indhu Bharathi Mar 27 '17 at 19:52
  • My own dataset. But I am confused, because in Keras and Mxnet, I used the same training data and validation. And learning rate are all 0.1,batch size is 32, the same. Why they have different results. About the data iterator, in mxnet I use mx.io.NDArrayIter – Luv Mar 28 '17 at 05:04
  • Would you be able to provide the code in text form instead of image @Luv? Also could you confirm the Keras backend you're using? See `$HOME/.keras/keras.json` file. – Thom Lane Feb 23 '18 at 01:38

1 Answers1

0

This question appears to have been abandoned by the original poster.

Simon Corston-Oliver
  • 226
  • 1
  • 6