How to find labels as clustered results (ex: clustered images....)?

[edwardcho]

Hello, Sir,

According your examples (I used cifar10 example).
I did training (pretext -> clustering -> selflabeling...)
Then I did evaling....

I think that I should find labels as final results.
How to find labels for clustered images??

Thanks,
Edward Cho.

[edwardcho]

Hello Sir,

Still, I couldn't find clustered images as final outputs.

As I mentioned,
I want to find separated images as clustered results.
How to get this results ??

Thanks,
Edward Cho.

[wvangansbeke]

Hi,

I'm not sure I understand your questions correctly, but the outputs after the self-labeling step are the cluster ids. For example, let's say we know CIFAR10 has 10 ground truth classes. Now, we can set the output dimension of our model to 10 clusters (ids: 0 - 9). After training, the Hungarian matching algorithm can map each cluster id to a ground truth class for evaluation purposes.

[edwardcho]

Hello Sir,

Thank you for your reply quickly.
But still I can't understand completely.

Using CIFAR-10, I got self-labeled model.

Then using eval.py, I tried to get the outputs for clustered ids of CIFAR-10 test-data.
In hungarian_evaluate function of utils/evaluate_utils.py, the count of targets and predictions are 10000.
(As you know, the count of test-data of CIFAR-10 is 10000)
But, the count of match (results of _hungrian_match function) is 10.
I think that this values should do equal.

If I'm wrong, what's wrong?

Thanks,
Edward Cho.

P.S.
In your e-mail,
"the outputs after the self-labeling step are the cluster ids"
=> What is the value in that function (hungarian_evaluate)?

"reordered_preds"... Is this a self-labeling result?