How to Compute mAP about rml-cnn

[bbenligiray]

The labels for your second image is for 11 classes instead of 8, I'm guessing that's a mistake. I deleted the last three labels:

labels = np.array([1,0,1,0,0,0,0,1,0,1,0,0,0,0,0,0]).reshape(2,8)
preds = np.array([0.6,0.2,0.7,0.3,0.2,0.3,0.4,0.3,0.2,0.8,0.1,0.2,0.3,0.4,0.7,0.3]).reshape(2,8)
print(calculate_mAP(labels, preds))

Prints 37.5

[mymuli]

Thank you very much,First !

rml-cnn/metrics.py

Line 55 in d3f9f28

ground_truth = labels[:, ind_class]

ground_truth = labels[:, ind_class] gets the values for each column，
np.sum(ground_truth) is the sum of each column，
labels = np.array([1,0,1,0,0,0,0,1,0,1,0,0,0,0,0,0]).reshape(2,8)
——> [[1,0,1,0,0,0,0,1],
[0,1,0,0,0,0,0,0]],
当 ground_truth=[0，0], np.sum(ground_truth)=0, rec = tp / np.sum(ground_truth) is NAN
This is where I am confused. Could you explain it?

[bbenligiray]

That function is the rewrite of the function below in Python, so I would consider asking its main author:
https://github.com/zhufengx/SRN_multilabel/blob/master/tools/AP_VOC.m

However, I think the main problem is that you are trying to calculate AP for classes that you have no positive example of. For example the label for your 4th class is 0 for both examples, which doesn't make sense. Why would you use a class if none of your examples have it?

[mymuli]

I understand it. Thank you very much.