get_iid_Pi

[milhidaka]

Hi,
I have found a problem that may be a minor bug and would like to report it.

FedUL/utils/data_utils.py

Line 141 in c900bd0

iid_Pi = iid_Pi / torch.sum(iid_Pi, dim=0)

The result iid_Pi is used in get_U_sets_Multiclass function to split bag_sizes[i] into classnum classes (thetas=iid_Pi) .

FedUL/utils/data_utils.py

Line 105 in c900bd0

n_this = int(bag_sizes[i] * thetas[i][cls])

However, torch.sum(thetas[i]) is not 1. I think it is unnatural.

iid_Pi = iid_Pi / torch.sum(iid_Pi, dim=1, keepdim=True)

may be correct?

In practice, this problem is minor because the deviation from 1 is small when given the arguments used in the experiment.

torch.sum(get_iid_Pi(5, 10, 10), dim=1)
tensor([0.9961, 0.9798, 1.0240, 0.9745, 0.9325, 1.0388, 1.0151, 1.0014, 0.9772,
        1.0606], dtype=torch.float64)