Maskformer
Implements the MaskFormer architecture in Salt. The maskformer utilises an encoder+decoder-based architecture and allows for the generation of truth objects with masks associating detector-level objects to the truth object. An example of this is vertexing, where the model can generate some N vertices. For each generated vertex embedding, a mask associating tracks to a given vertex is generated, as well as a classification of the vertex type, and regression of vertex properties.
- https://indico.cern.ch/event/1346358/#4-vertex-reconstruction-with-m
- https://arxiv.org/abs/2312.12272
This MR is an initial working version. There is one major difference to the original architecture - pooling to generate the global representation is done only on the encoder track embeddings, and does not utilise the decoder embeddings at this time. Onnx seems to throw a hissy fit when this happens, so we may see some sub-optimal performance in the flavour tagging tasks.