what's the difference between train mode and evalutation mode

[Mata Fu]

As the net mentioned https://github.com/amitshaked/resmatch/blob/master/src/networks/network.lua convert the fc layer to conv layer in test net, after changing the net, it sets

evaluate()

in the end, what's this for ? 

function network.getTestNetwork(model)
   -- Replace the model with fully-convolutional network
   -- with the same weights, and pad it to maintain resolution


   local testModel = model:clone('weight', 'bias')


   -- replace linear with 1X1 conv
   local nodes, containers = testModel:findModules('nn.Linear')
   for i = 1, #nodes do
      for j = 1, #(containers[i].modules) do
         if containers[i].modules[j] == nodes[i] then


            local w = nodes[i].weight
            local b = nodes[i].bias
            local conv = nn.SpatialConvolution1_fw(w:size(2), w:size(1)):cuda()
            conv.weight:copy(w)
            conv.bias:copy(b)
            -- Replace with a new instance
            containers[i].modules[j] = conv
         end
      end
   end


   -- replace reshape with concatenation
   nodes, containers = testModel:findModules('nn.Reshape')
   for i = 1, #nodes do
      for j = 1, #(containers[i].modules) do
         if containers[i].modules[j] == nodes[i] then
            -- Replace with a new instance
            containers[i].modules[j] = nn.Concatenation():cuda()
         end
      end
   end


   -- pad convolutions
   padConvs(testModel)


   -- switch to evalutation mode
   testModel:evaluate()




   return testModel
end

[Vislab]

Layers like nn.Dropout() or nn.BatchNormalization() have different behaviours during train/test (nn.Dropout() during test is essentially a nn.Identity() module). Then most practical way to switch all those layers at once in a sequence of layers is to have a training and evaluation methods to trigger the right behaviour.