Multi-labels

[lucamarini22]

Hello! First of all, thanks for this awesome library.

Is there a way to pass to a loss function a matrix of 0 and 1 (or False and True) of every possible pair instead of passing the class labels?

I'm asking this to manage the case where a sample belongs to multiple classes.

[KevinMusgrave]

For contrastive losses (ContrastiveLoss, NTXentLoss etc.) you can omit the labels and pass in indices_tuple:

loss_fn = ContrastiveLoss()
loss = loss_fn(embeddings, indices_tuple=indices_tuple)

Here's an example of how you can specify positive and negative pairs using indices_tuple: #435 (comment)

The discussion in the link is about the output of a miner, but you don't have to use a miner. You can manually create the indices_tuple yourself.

[lucamarini22]

Great, thanks!

[lucamarini22]

For people that might be interested: for the full list of losses that can use indices_tuples, see here under "Losses for which you can pass in indices_tuple without labels"

[lucamarini22]

Hi! Is there a way to pass the indices_tuples directly to the miner as well?

[KevinMusgrave]

No unfortunately there isn't. What behavior would you like when passing indices_tuple to a miner?

[lucamarini22]

I'd like to select the hard_pairs based on the indices_tuple as in my case I am also using indices_tuple when computing the loss, as I don't have classes that are categories. Does this clarify the expected behaviour? Please let me know in case you have other questions

[lucamarini22]

Actually I think I found a way to make it work, I will open a PR when I will have some time to clean the code!