1

I have a very simple model that is the data input to a hidden layer with 2 units that then outputs to an activation layer and then to the softmax layer.

mxnet will report the computed bias and weight with something like

fullyconnected0_bias -> [ 1.8431 -1.84309]
fullyconnected0_weight -> 
[
 [-1.22873 -1.22873]
 [ 1.22872  1.22872]
]

given the geometric intuition that this represents a separating hyperplane, how would one go about plotting this? For some reason I can't find any examples of anyone doing this in my searching? What do the weights and bias represent in terms of plot-able inputs?

sail0r
  • 445
  • 1
  • 5
  • 18

1 Answers1

1

One thing you may want to check out is the tensorboard-lite project which works with MXNet, with which you can easily visualize the network weights/outputs. https://github.com/dmlc/tensorboard

eric-haibin-lin
  • 377
  • 2
  • 9