Sklearn through precision_ recall_ Curve get F1 maximum

20220718 -

In order to check the performance of classifiers , Make use of precision_recall_curve To get accuracy and recall , And calculate F1 The maximum of , Problem of passing threshold , Check the maximum performance of the classifier , The code is as follows , And then calculate

precision, recall, thresholds = precision_recall_curve(y_test, y_test_predicted_probas)
f1_scores = 2*recall*precision/(recall+precision)
print('Best threshold: ', thresholds[np.argmax(f1_scores)])
print('Best F1-Score: ', np.max(f1_scores))

But in the process , Will meet RuntimeWarning: invalid value encountered in true_divide, This warning , Then output the highest F1 when , The result is nan.

Before , Because it doesn't happen every time , So I didn't care about it . Now it's time to do experiments on a large scale , I often report mistakes , So Google checked , In essence, the error reporting is generally clear , That is, some values do not meet the division conditions , But I tried to add a very small number , It didn't work .

Then I searched and found the corresponding solution .

precision, recall, thresholds = precision_recall_curve(y_test, y_test_predicted_probas)
numerator = 2 * recall * precision
denom = recall + precision
f1_scores = np.divide(numerator, denom, out=np.zeros_like(denom), where=(denom!=0))
max_f1 = np.max(f1_scores)
max_f1_thresh = thresholds[np.argmax(f1_scores)]

Essentially , Let division avoid these values .