torch_finetuning.lua

    if(opt.traintype == 'finetuning' and epoch == 1) then
        -- find the number of elements for which the learning rates to be reduced
        totalParams = 0;
        isAfterCrossNeighbor = 0;
        dontFreeze = 0;

        for index, node in ipairs(model.modules) do

            --currParams = node:getParameters()
            
            -- since the weights of initial conv layers are shared, half the count after CrossInputNeighborhood layer
            --if(torch.typename(node) == 'nn.CrossInputNeighborhood') then
            --    totalParams = totalParams/2;
                
            --    finalLearningRates[{{1, totalParams}}] = 0.01;
            --end
            
            if(torch.typename(node) == 'nn.Linear') then
                print(torch.typename(node) .. ' found')
                
                --if(lastLinearLayer == nil) then
                --    finalWeightDecays[{{1, totalParams}}] = 0;
                --    node.accGradParameters = function(self,i,o) end
                --print(torch.typename(node) .. ' gradient not accumulated')
                --end
                
                dontFreeze = 1;
                lastLinearLayer = node;
                node:reset()
            end
            
            --freeze the parameters of conv layer
            if(dontFreeze == 0) then
                --node.updateGradInput = function(self,i,o) end -- for the gradInput
                node.accGradParameters = function(self,i,o) end -- for freezing the parameters
            end

            
            -- check if the layer has parameters
            --if(#currParams:size() ~= 0) then
            --    currSize = #currParams

                -- if the layer is not fully connected, set the learning rate as 0.01
            --   if(torch.typename(node) == 'nn.Linear') then
            --        linearLayerParams = linearLayerParams + currSize[1];
            --        finalWeightDecays[{{totalParams + 1, totalParams + currSize[1]}}] = 5e-4;
            --        node:reset()
            --    end

            --    totalParams = totalParams + currSize[1];
                -- print(currSize[1] .. ', total: ', totalParams)
            --end
        end
        --print(parameters:size())
        --print('total params: '.. totalParams .. ', linear layer params: ' .. linearLayerParams)
        --print(finalLearningRates)
        --total = torch.sum(finalWeightDecays:eq(5e-4))
        --print('total decays ' .. total)
        --total = torch.sum(finalLearningRates:eq(1))
        --print('total learning rates ' .. total)
    
        --reset last linear layer
        --lastLinearLayer:reset();
        --total = torch.sum(parameters:eq())
        --print('total 0s ' .. total)
        --io.read()        
    end