Batch based noisy labels estimation about cleanlab HOT 5 CLOSED

Good questions. Can you reduce the probability matrix to float16. or if its still too big, potentially even unsigned int (8bits)?

from cleanlab.

If you decide to split in batches, try to split up the classes into 100 batches, each of 1k classes, where all the label noise between classes will be self-contained in each batch (i.e. types of dogs go in one batch if label noise is between types of dogs, and all the airplanes go in another batch, if your label noise is unlikely to have dogs mislabeled as planes)

If the above is tricky for you, you can also just split it up randomly, but you may want to run it a few times and then combine all the label errors.

from cleanlab.

Thank you for the quick response.

Good questions. Can you reduce the probability matrix to float16. or if its still too big, potentially even unsigned int (8bits)?

I can reduce precision to 8bits per element but still, it's gonna be almost 1Tb.

If you decide to split in batches, try to split up the classes into 100 batches, each of 1k classes, where all the label noise between classes will be self-contained in each batch (i.e. types of dogs go in one batch if label noise is between types of dogs, and all the airplanes go in another batch, if your label noise is unlikely to have dogs mislabeled as planes)

I've been trying to apply it for facial recognition task so I don't think splitting classes into self-contained batches is possible in my case.

If the above is tricky for you, you can also just split it up randomly, but you may want to run it a few times and then combine all the label errors.

Is it better to split them by classes or samples?

from cleanlab.

from cleanlab.

Got it, thank you.

from cleanlab.