Problem when shifting labels to zero in binary classification problems with negative labels

[DominguesPH]

Hello!
I would like to report a possible bug.

Code: v2s_main.py
Lines: 43-45

When we have multi-class problems, such as ECG 5000 where the original labels are [1,2,3,4,5], the mod function applied in lines 43-45 shifts the labels to zero correctly, so we obtain the values [0,1,2,3,4] as labels.

However, when we have binary classification with negative labels, such as ECG 200, where the original labels are [-1,1], the mod function used yields 1 as remainder for both cases, and the label vector becomes an array of ones.

Do you agree?

[huckiyang]

Hello @DominguesPH, thank you for asking.
I think in the L30 of v2s_main.py https://github.com/huckiyang/Voice2Series-Reprogramming/blob/main/v2s_main.py#L30

    y_train[y_train == -1] = 0
    y_test[y_test == -1] = 0 

We pre-assigned negative labels back to positive values.
If you want to use the other dataset containing negative values, please consider this value check.
Hope this clarifies your question.

[DominguesPH]

Hello @DominguesPH, thank you for asking. I think in the L30 of v2s_main.py https://github.com/huckiyang/Voice2Series-Reprogramming/blob/main/v2s_main.py#L30

    y_train[y_train == -1] = 0
    y_test[y_test == -1] = 0 

We pre-assigned negative labels back to positive values. If you want to use the other dataset containing negative values, please consider this value check. Hope this clarifies your question.

Yes, Huckiyang, I agree with you!
Thank you for the answer!
I only mention this because the ECG200 is not the only dataset with negative labels (the dataset 0 - Ford-A also considers negative labels) and the L30-L32 code is conditioned according to dataset 2 (ECG200).
I don't know if other datasets you've used for testing also have these labels.

[huckiyang]

Hello @DominguesPH,
Yes, in our experimental setup, we did some data cleanup. I just made some modifications to this public code for all setups.
https://github.com/huckiyang/Voice2Series-Reprogramming/blob/main/v2s_main.py#L30

y_train = [np.uint32(i) for i in y_train]
y_test = [np.uint32(i) for i in y_test]

And I got ~94% for 20 eps by running python v2s_main.py --dataset 0 --eps 20 --mod 2 --seg 18 without tuning the dropout. Thank you again for this discussion.