Unexpected LeNet Results

[Arjun Kar]

Hi all,

I just recently started using Caffe, and was doing the LeNet tutorial.  Everything built and ran just fine, but my accuracy at the end was about 0.1, and the loss function stayed around 10^6 for the entire computation.  This is different than the results shown on the tutorial page here.  I should be getting about 0.99 accuracy and 0.03 loss function.  I thought it was a loss function divergence so I lowered the learning rate in lenet_solver.prototxt by a factor of 10, then 100, then 10^5, but all of those gave the same results.  What am I doing wrong here?  If it helps, this problem goes away when I use the CPU training mode, but that mode trains way too slow to finish.  Also, as an aside, is there a way to obtain the weights of the model for use in implementing neural network testing in a different language?  Or does Caffe only output the weights in protobuf binary and not in any sort of readable plaintext?

Thanks,

~Arjun