Calling cuda() with async results in SyntaxError

问题

I'm trying to run this PyTorch code:

for i, (input, target) in enumerate(train_loader):

    input = input.float().cuda(async=True)
    target = target.cuda(async=True)
    input_var = torch.autograd.Variable(input)
    target_var = torch.autograd.Variable(target)

    output = model(input_var)

But when I try I am getting this error message:

input = input.float().cuda(async=True)
                               ^
SyntaxError: invalid syntax
Process finished with exit code 1

What am I doing wrong? I already installed cuda.

回答1:

Your code does not work because:

• async is a reserved keyword in python which cannot be used in that way, that is why you get the SyntaxError

• cuda() no longer has an argument async. The constructor looks like this:

cuda(device=None, non_blocking=False) → Tensor

• Previously there was an argument async but this replaced by non_blocking as async became a reserved keyword in Python 3.7.
  https://github.com/pluskid/fitting-random-labels/pull/5

Use instead non_blocking:

The argument non_blocking has the same effect as async previously had:

• non_blocking (bool):
  If True and the source is in pinned memory, the copy will be asynchronous with respect to the host. Otherwise, the argument has no effect. Default: False.
  
  https://pytorch.org/docs/stable/tensors.html#torch.Tensor.cuda

---

As an add-on: If you are interested in what async is actually used for you can take a look here: https://www.python.org/dev/peps/pep-0492/#new-syntax

回答2:

There was an async parameter but it was deprecated now, because async became a reserved word in Python 3.7. Details are included in this issue rename .cuda(async=..) parameters. You can use non_blocking as an alternative.

来源：https://stackoverflow.com/questions/53201534/calling-cuda-with-async-results-in-syntaxerror