libdnn
libdnn copied to clipboard
poweic
It seems like the label is being fed into the network while training
I was wondering why you added 1 to the AffineTransform layer geometry in https://github.com/botonchou/libdnn/blob/master/src/nnet.cpp#L233
After some nosing around I found you feed the label into the network as being one of its' inputs! (I printed the data at https://github.com/botonchou/libdnn/blob/master/src/nn-train.cpp#L164). I don't think this is desired behaviour. Can you please explain this?
In rand(out + 1, in + 1), the first +1 is for biases in this layer, and second +1 is reserved for +1 in the next layer. ( [x, 1] )
No, label won't be fed into the network. I'm pretty sure about that. Maybe you can tell me the command line arguments so that I can reproduce your situation.
I printed the data being fed into the network, it included my label. It is relatively easy to see when testing the XOR example I showed in issue #18.
I just printed. It didn't include your label. The last row are always 1.
About the problem you encountered, I think it's because I count the number of \n to see how many lines in the file. Since the data you provided (xor.train.dat and xor.test.dat) have no trailing \n at line 4, it might cause undefined behaviors.