Moving_tensors_to_GPU

data = 2 * data #This is performed on CPU
data_gpu = 2 * data.to(device='cuda') #This is performed on gpu

data_gpu tensor is not brought back to the CPU once the result has been computed.
Execution of the last line results in:

1. Copying of data tensor to gpu.
2. Allocation of a new tensor on gpu to store the results of multiplication
3. A handle for this new tensor is created 

In order to transfer data to cpu,
data_cpu = data_gpu.to(device='cpu')

# Shorthands:
data_gpu = data.cuda()
data_gpu = data.cuda(0) #0 refers to the index of gpu, similar to cuda:0