shuuki4 / batch-normalization Goto Github PK

Hi,
I doubt about if your code really work on dealing with training and testing differently.
Once the theano function are built, the set_runmode() method , 0 for training , 1 for testing, will not
affect the calculating when running. The self.run_mode will always be 0.

I fix all the random seed firstly. Then, I removed the all the script to set_runmode. It get totally same result as not.

I think your code to decay learning rate and dealing with dropout suffer from the same problem as above.