alfiuman / whydahgally-lstm-mlp Goto Github PK

Hello All ! Using the default project I see the CUDA utilization at only 15% using a 1050 Ti (GPU-Z).
Do you know how change the parameters to push this up to fully utilize the GPU ?

I recompiled your code for win7 under VS2012. the LSTM algo does not seem to converge with a number of numCell other than 20 with the example you provide.
Playing with the other params like alpha or lossFunct does not seem to help. I have incorporated your latest change in matrix.h, did not fit it.
Chosing LOSSFUNCTLOGPOW3 or LOSSFUNCTPOW3 create NaN. Can you help ?

The problem is in computeLoss(). The 1st arg1 to lossFunctLogPow3(arg1,arg2) and the others should never be <= 0. because of the log() funciton they call.
The fix is to replace any value <= 0 by a tiny very tiny pos number (+1.0e-20f for example)
This fix eliminates the NaN and the convergence problem for any type of loss function and/or numCell