SoftMaxTree:sharedClone bug

[akhti]

Hi,

I've noticed the following bug. When I create a shared copy SoftMaxTree node of size X, I end up with 1.5 * X memory instead of X.

Minimal working example for BillionWord hierarchy from dp package:

    require 'torch'; require 'nn'; require 'nnx';
    require 'cutorch'; require 'cunn'; require 'cutorch'
    require 'dp'

    function report_gpu_usage(comment)
      local props = cutorch.getDeviceProperties(cutorch.getDevice())
      local usage = props["totalGlobalMem"] - props["freeGlobalMem"]
      print(string.format("GPU usage (%s): %.1fMB", comment, usage / 1024 ^ 2))
    end

    dataset = dp.BillionWords()

    report_gpu_usage("initial")
    sm = nn.SoftMaxTree(100, dataset:hierarchy() , dataset:rootId()):cuda()
    report_gpu_usage("created softmax")
    sm_sh_clone = sm:sharedClone()
    report_gpu_usage("shared clone")
    sm_clone = sm:clone()
    report_gpu_usage("full clone")
    collectgarbage()
    report_gpu_usage("collectgarbage")

Output:

GPU usage (initial): 80.4MB
GPU usage (created softmax): 772.7MB
GPU usage (shared clone): 1112.1MB
GPU usage (full clone): 1804.4MB
GPU usage (collectgarbage): 1804.4MB

As you can see, shared copy requires extra 1112 - 772 = 340MB. That's a number

[nicholas-leonard]

This is the culprit : https://github.com/clementfarabet/lua---nnx/blob/master/SoftMaxTree.lua#L329 .

Can you send a Pull Request?