Free tensor memory in torch lua after using it

[shab]

Hello, I'm new to torch and cuda and am getting confused now. 

I define a sample big cudaTensor as follows:

a = torch.CudaTensor(1,224,1040,960):fill(124.15)

This uses about 1.1 G of GPU memory, and the change can be seen both by nvidia-smi and with command cutorch.getMemoryUsage(1) (I have one GPU)

Now if I resize the storage and the tensor as follows:

a:storage():resize(1)

a:resize(0)

cutorch.getMemoryUsage(1) shows that the memory is being freed, but it's not the same in nvidia_smi. (collectgarbage() does not make any changes either). 

Can anyone help me to figure out what the problem is? and what should I do to free memory on torch Lua after not needing a cudaTensor? replacing it by nil and calling garbage collector is not working. 

Thanks