Pytorch: Why is model.cuda() a class method instead of property ?

Created on 15 Apr 2017 · 3Comments · Source: pytorch/pytorch

import torch
from torch.autograd import Variable

a = Variable(torch.Tensor([1, 2]))
if torch.cuda.is_available():
    b = a.cuda()

b_data = b.data
b_grad = b.grad
b_grad_data = b.grad.data

If .data and .grad are class attributes, then .cuda() can be kept as property so that one a do b = a.cuda. I am making suggestion thinking that it would increase API consistency, as .cuda(), .data and .grad are torch.Tensor so cuda() can be changed to a property. By doing this, the GPU residing model can be accessed by model.cuda

Source

kdexd

Most helpful comment

cuda is made a method since it is a relatively expensive operation, which involves copying data from CPU to GPU. While grad and data are cheap operations (O(1)), which only retrieve references to the corresponding Tensor. I think making cuda a method is reasonable.

chenzhekl on 16 Apr 2017

👍10

All 3 comments

chenzhekl on 16 Apr 2017

👍10

.cuda() copies CPU data to GPU. You probably don't want to keep the data in GPU all the time. That means, you only store data in GPU when it's really necessary.

lg-zhang on 16 Apr 2017

as mentioned by both above comments, making .cuda a property is inappropriate. We return new objects when referencing .cuda() and a lot is happening.