Implement clustering accuracy

[moetayuko]

🚀 Feature

Motivation

Clustering accuracy is a popular metric. In addition to classification accuracy, it employs the Hungarian algorithm to align the predicted pseudo labels and the ground truth labels.

Current implementations of clustering accuracy use either scipy.optimize.linear_sum_assignment or the munkres package for Hungarian. I'm not sure if this is allowed for torchmetrics, and a custom implementation needs to be added if not.

Pitch

Implement clustering accuracy in torchmetrics.clustering

[github-actions]

Hi! thanks for your contribution!, great first issue!

[SkafteNicki]

@moetayuko thanks for opening this issue. Do you have a reference to a source (possible research paper) where they describe the metric in details?

[moetayuko]

@moetayuko thanks for opening this issue. Do you have a reference to a source (possible research paper) where they describe the metric in details?

Sec. 6.2.1 of https://arxiv.org/abs/2206.07579

FYI here are some random implementations I found:
https://github.com/bdy9527/SDCN/blob/da6bb007b7d07362ac04db3e146b4944e2acb883/evaluation.py#L9
https://github.com/google-research/google-research/blob/cc787a6a513cb6d2042cd6e286e2cbe42c41a863/stacked_capsule_autoencoders/capsules/eval.py#L29
https://github.com/aeon-toolkit/aeon/blob/01e8424df1bee9e614a25a25b5268383a8fd9893/aeon/performance_metrics/clustering.py#L12

[SkafteNicki]

@moetayuko thanks for the references, it really helped understanding how the metric is intended to work.
Hopefully, I have time to fully implement the metric in the next couple of days. I have already the logic figured out using https://github.com/ivan-chai/torch-linear-assignment for solving the linear sum assignment problem:

from torchmetrics.functional.classification import multiclass_confusion_matrix
import torch
# pip install git+https://github.com/ivan-chai/torch-linear-assignment.git@main
from torch_linear_assignment import batch_linear_assignment

preds = torch.tensor([0, 0, 1, 1])
target = torch.tensor([1, 1, 0, 0])

confmat = multiclass_confusion_matrix(preds, target, num_classes=5)
print(confmat)

confmat = confmat[None]

assignment = batch_linear_assignment(confmat.max() - confmat)
print(assignment)

confmat = confmat[0]

tps = confmat[torch.arange(confmat.size(0)), assignment.flatten()]

acc = tps.sum() / len(preds)
print(acc)