Torch7, how to calculate the number of parameters in a convNet

Question

I am looking for a way to calculate the number of parameters in a convolutional Neural Network. In particular, I am using the Resnet model in https://github.com/facebook/fb.resnet.torch. Do you know if there is any function that could calculate the total number of parameters? Do you have other suggestion for doing it? Thanks in advance.

Answer 1

You basically have to go through each layer of your network and count the number of parameters in that layer. Here is a sample function that does that:

-- example model to be fed to the function
model = nn.Sequential()
model:add(nn.SpatialConvolution(3,12,1,1))
model:add(nn.Linear(2,3))
model:add(nn.ReLU())

function countParameters(model)
local n_parameters = 0
for i=1, model:size() do
   local params = model:get(i):parameters()
   if params then
     local weights = params[1]
     local biases  = params[2]
     n_parameters  = n_parameters + weights:nElement() + biases:nElement()
   end
end
return n_parameters
end