Does average_precision work for multilabel classification?

[KevinMusgrave]

By multilabel, I mean that a sample can have a target vector like [1, 0, 0, 1].

I noticed that the score doesn't match sklearn.metrics.average_precision_score when average="macro".

Here's my code:

import torch
from torchmetrics.functional import average_precision
from sklearn.metrics import average_precision_score

y_true = torch.randint(0,2,size=(100,10))
y_scores = torch.sigmoid(torch.randn(100, 10))

for average in ["macro", "micro"]:
    score = average_precision_score(y_true.numpy(), y_scores.numpy(), average=average)
    score2 = average_precision(y_scores, y_true, average=average)
    print("average:",average)
    print("sklearn:", score)
    print("torchmetrics:", score2.item())
    print("\n\n")

Here's the output:

average: macro
sklearn: 0.4853133981804489
torchmetrics: 0.47312888503074646



average: micro
sklearn: 0.4731288604445855
torchmetrics: 0.47312888503074646

Edit: It seems like passing in num_classes=10 gets rid of the discrepancy. Is that the correct usage?