training.py 文件源码

def __get_batch(self, inputs, targets):
        data_size = targets.size(0)

        inds = torch.floor(torch.rand(self.batch_size) * data_size).long().cuda()
        # bug: floor(rand()) sometimes gives 1
        inds[inds >= data_size] = data_size - 1

        if type(inputs) == tuple:
            inp = tuple([Variable( i.index_select(0, inds).cuda() ) for i in inputs])
        else:
            inp = Variable( inputs.index_select(0, inds).cuda() )

        targ = Variable( targets.index_select(0, inds).cuda() )
        return inp, targ