where is the ScatterWrapper?

[shenjun1994]

I can't find the ScatterWrapper define in all the files, it used in the main.py,compute_validation_loss function, could you tell me where is the define about the ScatterWrapper?

[artes14]

I'm having same issue here, there's 1 problem just not getting out of project errors:

Unresolved reference 'ScatterWrapper'

[jolly12138]

I'm also having same issue here, there's 1 problem just not getting out of project errors:

Unresolved reference 'ScatterWrapper'

[GT84]

Has anyone found a solution for this? I'm facing the same problem!

[GT84]

Ok, searching pypi, conda etc. results in, no basic package providing this class. But, inspecting repositories history lead to commit "Optimized training on multiple GPUs."

The class has been removed/replaced with this commit.

My current solution is adding the content of the "old" code into my current train.py file.

class ScatterWrapper:
    """ Input is any number of lists. This will preserve them through a dataparallel scatter. """

    def __init__(self, *args):
        for arg in args:
            if not isinstance(arg, list):
                print('Warning: ScatterWrapper got input of non-list type.')
        self.args = args
        self.batch_size = len(args[0])

    def make_mask(self):
        out = torch.Tensor(list(range(self.batch_size))).long()
        if args.cuda:
            return out.cuda()
        else:
            return out

    def get_args(self, mask):
        device = mask.device
        mask = [int(x) for x in mask]
        out_args = [[] for _ in self.args]

        for out, arg in zip(out_args, self.args):
            for idx in mask:
                x = arg[idx]
                if isinstance(x, torch.Tensor):
                    x = x.to(device)
                out.append(x)

        return out_args

I didn't check the impacts very detailed yet, but basic training does work without any problems.