can not use lbc about cgl_fairness HOT 5 CLOSED

Hello, please use Please use the reweighting option for using LBC. You can check the details here:
#1 (comment)

from cgl_fairness.

Hi, I'm sorry for a typo. It is iteration, not iter. And the iter argument is used in trainer/rewieghting.py. It represents the number of iterations for outer update of eta in a minimax optimization.

We modified our README as well. Thanks.

from cgl_fairness.

@sangwon79 @SanghyukChun Thanks! However there still exists another error:

Traceback (most recent call last):
  File "/data1/cgl_fairness/main.py", line 124, in <module>
    main()
  File "/data1/cgl_fairness/main.py", line 98, in main
    trainer_.train(train_loader, test_loader, args.epochs, writer=writer)
  File "/data1/cgl_fairness/trainer/reweighting.py", line 46, in train
    weight_set = self.debias_weights(Y_train, S_train, extended_multipliers, num_groups, num_classes)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/data1/cgl_fairness/trainer/reweighting.py", line 207, in debias_weights
    weights = w_matrix[sen_attrs, label]
              ~~~~~~~~^^^^^^^^^^^^^^^^^^
RuntimeError: indices should be either on cpu or on the same device as the indexed tensor (cpu)

so I use this in reweighting.py:

sen_attrs = sen_attrs.to(w_matrix.device)
label = label.to(w_matrix.device)
weights = w_matrix[sen_attrs, label]

It makes it run, but the score keeps in a low degree:

from cgl_fairness.

Thanks again.
The first issue may be due to the difference in the version of Torch package. It works well in our setting.
We left out a line of code in the process of publishing our code.
The first and second issues have been solved in the modified code. Please git pull again.

from cgl_fairness.

Thanks a lot! This has been solved.

from cgl_fairness.