Meaning of -1.000 in AP and AR

[AnhMinhTran]

As I understand, the AP amd AR calculation are as of follow:
Precision = truepositive/ (truepositive + falsepositive)
Recall = truepositive/ (truepositive + falsenegative)

Which is what I found in evaluation dir

precision = cum_true_positives.astype(float) / (cum_true_positives + cum_false_positives)
    recall = cum_true_positives.astype(float) / num_gt

So if my AR and AP is consistently equal -1.000, the model doesn't learn and uodate any weights throughout the epochs?

[rwightman]

@AnhMinhTran I don't think I've ever seen -1 and I've run a lot of models using this code and elsewhere through the coco evaluator, there must be something wrong with the predictions... it's generally 0. if it's not learning.

[AnhMinhTran]

I have figured out the issue, its due to labeling categories in the JSON file. The JSON I got the annotation starts with 0 index, which caused this issue