Deep learning 4J

KNIME Extensions Deep Learning
#1

I've read in the forum that non-integer division should be handled automatically by DL4J. So why do I get the following errors (many like the one below):

ERROR DL4J Feedforward Learner (Classification) 0:14:1     Execute failed: Invalid input configuration (layer name = "null") for height: inHeight=50, kernelH=3, padH=0, strideH=2; (50-3+2*0)/2 is not an integer

And where can I define padding? I haven't found it in the configuration, but I see above padH=0.

Thanks,

Francesco

 

 

 

#2

Hi fcivardi,

currently the padding can't be adjusted. It will be always zero. This error indicates that the dimension of your images you want to use does not match the created architecture. You could try to reduce the kernel size or use a stride of 1, but that strongly depends on the number of layers, their configuration, and the size of the used images.

Kind regards

David