Convert a cudnn batchnorm module to nn backend (cycles all modules of a network).

convert_module.lua

local function ConvertBNcudnn2nn(net)
  local function ConvertModule(net)
    return net:replace(function(x)
        if torch.type(x) == 'cudnn.BatchNormalization' then
          return cudnn.convert(x, nn)
        else
          return x
        end
    end)
  end
  net:apply(function(x) return ConvertModule(x) end)
end

To be more comprehensive, it would be nice to include the spatial BN as well. The current one does not replace them.

if torch.type(x) == 'cudnn.BatchNormalization' or torch.type(x) == 'cudnn.SpatialBatchNormalization' then