Forward mode

Benchmark result

-- Without the monkeypatching (standard nn)

rudy@rudy-Tower:~workspace/utc-caltech[master]
$ time luajit luatools/prediction.lua
data/res/divers-bignn/set09/V010.txt

real    1m1.162s
user    0m51.132s
sys     0m13.084s


-- With the monkeypatching
rudy@rudy-Tower:~workspace/utc-caltech[master]
$ time luajit luatools/prediction.lua
data/res/divers-bignn/set09/V010.txt

real    0m43.975s
user    0m39.715s
sys     0m7.190s

net.lua

-- the network that I was using
local get_detector = function()

   local cnn = nn.Sequential()
   cnn:add(nn.SpatialConvolution(3, 25, 5, 5))
   cnn:add(nn.PReLU())
   cnn:add(nn.SpatialDropout())


   cnn:add(nn.SpatialConvolution(25, 50, 5, 5))
   cnn:add(nn.PReLU())
   cnn:add(nn.SpatialDropout())

   cnn:add(nn.SpatialMaxPooling(2,2,2,2))

   cnn:add(nn.SpatialConvolution(50,75,5,5))
   cnn:add(nn.PReLU())
   cnn:add(nn.SpatialDropout())

   cnn:add(nn.SpatialConvolution(75,100,1,5))
   cnn:add(nn.PReLU())
   cnn:add(nn.SpatialDropout())

   cnn:add(nn.SpatialMaxPooling(2,2,2,2))

   cnn:add(nn.SpatialConvolution(100, 25*6*4, 4, 6, 1, 1))
   cnn:add(nn.PReLU())
   cnn:add(nn.Dropout())

   cnn:add(nn.SpatialConvolution(25*6*4 , 12*6*4, 1, 1))
   cnn:add(nn.PReLU())
   cnn:add(nn.Dropout())

   cnn:add(nn.SpatialConvolution(12*6*4, 2, 1, 1))

   cnn:cuda()

   return cnn
end



-- function to replace updateOutput for Sequential
local updateOutputCleaner= function(self, input)
   local currentOutput = input
   for i=1,#self.modules do
      currentOutput = self.modules[i]:updateOutput(currentOutput)
      self.modules[i].finput = nil
      self.modules[i].output = torch.CudaTensor()
      collectgarbage()
   end
   return currentOutput
end


-- where i monkeypatch
local switch_to_test_mode = function(cnn)

   cnn:evaluate()
   cnn:zeroGradParameters()
   cnn.updateOutput = updateOutputCleaner
   return cnn

end