Update metrics.py

[unai-gurbindo]

Possible error in the ap_per_class function when calculating the precision-recall curve

Hello,

I have identified a possible error in the ap_per_class function when calculating the metrics for the precision-recall curve. Below, I provide an example to explain it more clearly:

Let’s assume the unique_classes list is [0, 1, 2], and the model never predicted class 1. In this case, on line 583 of the utils/metrics.py file, the loop will skip to the next class (class 2) because n_p == 0 for class 1.

However, before moving on to the next class, a zero-filled array corresponding to class 1 should be added to the prec_values list. Otherwise, the precision-recall curve for class 2 will incorrectly appear as if it belongs to class 1, and the legend will also fail to show the correct label for class 2.

To fix this behavior, before the continue statement, the code should be modified as follows:

if n_p == 0 or n_l == 0:
    prec_values.append(np.zeros(1000))
    continue

I hope my explanation is clear, and I have correctly identified the issue. While this is a rare case where the model never predicts one of the classes in the dataset, it is still something to consider to avoid errors in the precision-recall curve visualizations.

Best regards.