Comparison with standard model

[a7b23]

Any suggestions on how to compare the training time with a non-binary MLP of same complexity? i.e. maybe a set of opt flags that I need to turn off to convert the Main_BinaryNet_MNIST.py into a standard MLP model?
Also, I'm getting a test accuracy of only ~95%(~10 epochs) on MNIST. What maximum accuracy can I expect on full training? How many epochs will be needed?

[han-qiu]

Well, I have just run the code of mnist.py(The theano version, not this one). Here is the output of the last epoch:
LR: 3.02775865823e-07 training loss: 0.00297216895865 validation loss: 0.010455210574 validation error rate: 0.999999979511% best epoch: 980 best validation error rate: 0.939999978989% test loss: 0.00903553832935 test error rate: 0.929999979213%
It's about 25-30 seconds/epoch on a TITAN GPU.
Maybe I will compare the training time with other models