Parameter and configuration of Neural Network (gl.neuralnet_classifier.create)

User 1747 | 4/10/2015, 6:41:01 PM

I have started using <i>gl.neuralnet_classifier.create()</i> for creating neural network and training, I am using a <i> net = gl.deeplearning.load('network.conf') </i> file to configure neural network, where network.conf is somewhat like

<i> netconfig = start layer[0->1] = fullc initrandom = gaussian padding = 2 stride = 4 initbias = 0 initsigma = 0.01 numchannels = 93 numgroups = 1 kernelsize = 5 numhiddenunits = 100 layer[1->2] = relu layer[2->3] = fullc initrandom = gaussian padding = 2 stride = 4 initbias = 0 initsigma = 0.01 numchannels = 100 numgroups = 1 kernelsize = 3 # -- '' -- #Many such layers #-- '' -- numhiddenunits = 9 layer[16->17] = softmax netconfig = end

network parameters

initrandom = xavier learningrate = 0.1 inputshape = 1,1,93 batchsize = 256 biaslearningrate = 0.2 learningrategamma = 0.1 initsigma = 0.01 biasl2regularization = 0.001 learningratestep = 1000 learningrateschedule = exponentialdecay l2_regularization = 0.005 momentum = 0.9

end network parameters

</i>

I am unable to visualize the figure of neural network. neither I am getting the meaning of some parameter mentioned in file, just got this file in some example and using to by changing parameters. I tried and search to find some tutorial where this can be explained but did not find any. I have studied Neural Network and know the terminology of that, but getting hard to make sense of each parameters here.

Can anybody help me to understand the architecture of NN made by these type of files and tutorial for parameters of each layers. Please give me some links or tutorial for NN in <i>graphlab </i>

Comments

User 1190 | 4/10/2015, 8:20:16 PM

Hi @NikeshBajaj,

We have documentation of each layers with paper reference in the API docs https://dato.com/products/create/docs/graphlab.toolkits.deeplearning.html

This book by Yoshua Bengio is also a good reference http://www.iro.umontreal.ca/~bengioy/dlbook/

If you think any of our documentation is unclear, please let us know, and we are more than happy to fix that.

Thanks, -jay