torch seems to have a memleak in this case, even though you'd expect it to not, with the collectgarbage()

torchmemtest.lua

require 'image'
local memTestClass = torch.class('image.memTestClass')
function memTestClass:__newindex__(key, value)
   if type(key) == 'number' then 
      return 1 -- do this so that rawset isn't called anyways, stupid of me to miss this
   else rawset(self, key, value) end
end

function testMemory()
   n = 50000
   data=image.memTestClass()
   for i=1,n do
      local img = image.lena()
      data[i] = img
      collectgarbage()
   end
end
testMemory()