0

I am new to python and try to use lasagne and theano to enhance performance of machine learning by computing it on GPU (installed numpy+mlk). I call NeuralNet.fit(X, y) which sometimes proceeds successfully, but sometimes (and more often) crashes with open an alert window saying following:

"Python.exe has stopped working"

I can explore "problem details", where I can find "Fault Module Name: m16938b3c266c80e639557073d4dbcc9b.pyd"

Also, sometimes the trn loss and val loss are very huge (100 digits long number) and in some epoch, they turn to be "nan".

I use MNIST as the demo and for speed pick only the first 50 digits. Here the neural network structure:

## Layer information

  #  name     size
---  -------  --------
  0  input    1x28x28
  1  conv1    16x26x26
  2  conv2    16x24x24
  3  hidden4  500
  4  output   10

any ideas or suggestions? Thank you very much

fairtrax
  • 416
  • 2
  • 8
  • It seems that theano has memory issues when the BLAS library is not dynamic. Depending on your knowledge about BLAS and stuff like that, and your operating system, you can try to provide a proper dynamic BLAS for it. Or, if possible, simply change to tensorflow. (Tensorflow solved this problem for me) – Daniel Möller Sep 15 '17 at 16:35
  • Try clearing theano cache. `theano-cache clear` – Sentient07 Sep 23 '17 at 10:21
  • @Sentient07 : which tool consumes this command? – fairtrax Sep 25 '17 at 11:34
  • there is no tool, this is inbuilt in theano. these are the caches created by theano – Sentient07 Sep 25 '17 at 19:03
  • @Sentient07 ahaa, I found that, thanks. I will try it. But now I figured out, the problem occurs only when I use CNN, if there is no ConvolutionalLayer used, python doesn't crash. – fairtrax Sep 25 '17 at 20:41

0 Answers0