CODEWITHSUNDEEP
pythongpupytorch
Kani
Kani

Reputation: 1424

How to get the device type of a pytorch module conveniently?

I have to stack some my own layers on different kinds of pytorch models with different devices.

E.g. A is a cuda model and B is a cpu model (but I don't know it before I get the device type). Then the new models are C and D respectively, where

class NewModule(torch.nn.Module):
    def __init__(self, base):
        super(NewModule, self).__init__()
        self.base = base
        self.extra = my_layer() # e.g. torch.nn.Linear()

    def forward(self,x):
        y = self.base(x)
        z = self.extra(y)
        return z

...

C = NewModule(A) # cuda
D = NewModule(B) # cpu

However I must move base and extra to the same device, i.e. base and extra of C are cuda models and D's are cpu models. So I tried this __inin__:

def __init__(self, base):
    super(NewModule, self).__init__()
    self.base = base
    self.extra = my_layer().to(base.device)

Unfortunately, there's no attribute device in torch.nn.Module(raise AttributeError).

What should I do to get the device type of base? Or any other method to make base and extra to be on the same device automaticly even the structure of base is unspecific?

Upvotes: 69

Views: 153564

Answers (3)

Marine Galantin
Marine Galantin

Reputation: 2279

@Duane's answer creates a parameter in the model (despite being a small tensor).

I think this answer is slightly more pythonic and elegant:

class Model(nn.Module):
    def __init__(self, *args, **kwargs):
        super().__init__()
        self.device = torch.device('cpu') # device parameter not defined by default for modules

    def _apply(self, fn):
        # https://stackoverflow.com/questions/54706146/moving-member-tensors-with-module-to-in-pytorch
        # override apply by moving the attribute device of the class object as well.
        # This allows to directly know where the class is when creating new attribute for the class object.
        super()._apply(fn)
        self.device = fn(self.device)
        return self

net.cuda(), net.float(), etc will all work as well, since those all call _apply rather than to (as can be seen in the source).

An alternative solution from the comment of @Kani (accepted answer) is also very elegant:

class Model(nn.Module):
    def __init__(self, *args, **kwargs):
        """
        Constructor for Neural Network.
        """
        super().__init__()

    @property
    def device(self):
        return next(self.parameters()).device

You access the device through model.device as for parameters. This solution does not work when you have no parameter inside the model.

Upvotes: 12

Duane
Duane

Reputation: 5140

My solution, works in 99% of cases.

class Net(nn.Module):
  def __init__()
    super().__init__()
    self.dummy_param = nn.Parameter(torch.empty(0))

  def forward(x):
    device = self.dummy_param.device
    ... etc

Thereafter, the dummy_param will always have the same device as the module Net, so you can get it anytime you want. eg:

net = Net()
net.dummy_param.device

'cpu'

net = net.to('cuda')
net.dummy_param.device

'cuda:0'

Upvotes: 43

Shagun Sodhani
Shagun Sodhani

Reputation: 3727

This question has been asked many times (1, 2). Quoting the reply from a PyTorch developer:

That’s not possible. Modules can hold parameters of different types on different devices, and so it’s not always possible to unambiguously determine the device.

The recommended workflow (as described on PyTorch blog) is to create the device object separately and use that everywhere. Copy-pasting the example from the blog here:

# at beginning of the script
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

...

# then whenever you get a new Tensor or Module
# this won't copy if they are already on the desired device
input = data.to(device)
model = MyModule(...).to(device)

Do note that there is nothing stopping you from adding a .device property to the models.

As mentioned by Kani (in the comments), if the all the parameters in the model are on the same device, one could use next(model.parameters()).device.

Upvotes: 91

Related Questions