glassroom/heinsen_tree
Makes hierarchical classification easy, and also more efficient, enabling greater scale. See here for an example of hierarchical classification over a large semantic tree. To get you started, here's a toy example:
ClassTree is a PyTorch module implementing the methods we propose in our paper. These methods are algebraically expressible as tensor transformations that common software frameworks for machine learning, like PyTorch, execute efficiently, particularly in hardware accelerators like GPUs and TPUs. Our methods enable efficient hierarchical classification in parallel. For details, see our paper.
When training a model, filter out padding values to flatten mapped scores into a matrix and mapped labels into a vector, for computing a classification loss (e.g., cross-entropy) at every applicable level of depth in parallel:
We recommend that you restrict the space of allowed predictions to only paths that exist in the tree, which ClassTree stores in a buffer named paths (corresponding to matrix P in the paper).
