Utilizing multiple gpus with different memory to full use

[suhaspillai]

I am running my network on 2 gpus using DataParallelTable .Following is the snippet

function loadNetworkParallel(net,nGPU)

  if nGPU>1 then
    require 'cudnn'
    require 'cutorch'
    require 'cunn'
    gpus = torch.range(1,nGPU):totable() 
    net_parallel = nn.DataParallelTable(1):add(net,{gpus })
    return net_parallel:cuda()
  elseif nGPU==1 then
    require 'cudnn'
    require 'cutorch'
    require 'cunn'
    return net:cuda()
  end
end

1. GTX 980 - RAM: 4035MB
2. GTX 1080 - RAM: 8114MB

I can run batch size of 15, where approx 4000MB is used by GTX980 and 4000MB by GTX 1080 . There is still (4000MB) left on GTX 1080 but I think DataParallelTable allocates input equally and so I am unable to run batch size greater than 15. Any idea on how can I make it work, or do I need to allocate inputs manually to each gpu by checking how much memory is still left.