使用Multi-gpu时成员变量的Pytorch错误值

Question

Here is a simple class for running in multi-gpu environment. The member variable self.firstIter should be False after the first iteration.

Class TestNetwork(nn.Module):

    def __init__(self):
        super(TestNetwork, self).__init__()
        self.firstIter = True #indicates whether it's the first iteration

    def forward(self, input):
        print 'is firstIter: ', self.firstIter #always True!!
        if self.firstIter is True:
            self.firstIter = False
        # do otherthings

The code works as expected when using only one gpu.

However when using multi-gpu (ie nn.DataParallel ), the value of self.firstIter is always printed as True .

Why does this happen? What is wrong with the code?

Using PyTorch version 0.3.1.

Answer 1

Basically, DataParallel operates on model replicas, and changes made to replicas (during forward) are not visible outside forward/backward calls if number of divices is large than 1.

Plz refer to https://discuss.pytorch.org/t/nonetype-attribute-when-using-dataparallel/11566 for details.