How to extract class labels from GenericCLScenario class

I created a custom dataset by using the AvalancheDataset class and the dataset_benchmark function, hence obtaining a GenericCLScenario instance. Everything works fine, the dataset is correct and I can use it without any problems.

Now, my aim is to extract, for each stream and experience in the GenericCLScenario, the class labels contained in such experience.
I am aware that the class has the property classes_in_experience, however that returns a set of tensors (containing the labels), which is not optimal to quickly extract the unique class labels.

Is there a better way to obtain the class labels from a GenericCLScenario?

Thanks for your attention.

Hi @DWarez, using the sintax self.classes_in_experience[stream_name][experience_id] would give you access to a set of unique class ids for a specific experience. Is this what you want, right?

You are using that classes_in_experience contains tensors and their values are not unique? If so, this may be a bug. If you post an issue on github with code to reproduce it we can fix it.

Hello, thanks a lot for the replies.

As @AntonioCarta said, it seems like that classes_in_experience returns a set containing non-unique values (I have no idea how this is possible since I recall that in Python sets cannot contain the same item more than one).

This is the output I get while debugging. Note that self.data is an attribute that contains the return value of the function dataset_benchmark

As you can see the set contains the collection of labels, which implies that the same class-label is repeated several times.

I cannot share the code since it’s under a NDA, however I can try to replicate the bug and eventually open a GitHub issue, if this isn’t the expected behavior.

yes, this is unexpected. The dataset constructor should convert the tensor to a list internally (otherwise the equality doesn’t work), but in your case it isn’t.

Thank you Antonio, I opened an issue here, even though I haven’t managed to write a MWE yet. I hope this can still help in fixing the bug.

1 Like