How can I use OTK for multiple instance learning with varying bag size?

[HHenryD]

Hi, thanks for your great work!

I'm working on a multiple instance leanring problem, where I have some bags of instances, and each bag can have different number of instances. In my multiple instance learning framework, I used attention pooling / mean pooling to obtain the bag-level representation, followed by a MLP for downstream classification. I used batch size of 1 because the bag size varies. Each bag is represented by NxD, where N is the bag size, D is the feature dimension.

Given the superior performance of OTK, I'm interested in replacing the attention pooling / mean pooling by OTK layer in both supervised and unsupervised way. I'm wondering if your code supports batch size 1 or bags of varying sizes in multiple instance learning? If so, what should I do to implement the idea?

Thank you!