Problem about implementation

[Stanwang218]

Hi,

I just read your paper about Noise-Robust CP. When I look at the implementation, I feel confused that why you resample the calibration set, which increase the number of calibration set to n_examples * n_classes. In the paper, the number of calibration set is still n(n_example). It is just to compute the weighted average score for all labels.

It would be grateful if you can help me understand.

[Stanwang218]

When you compute the q_level, why you use n_examples * n_classes as n rather than n_examples. Besides that, for the weights, why it is num_class - 1 in the denominator.