Currently, we only support triplet tokenization, which takes a triple (code, value, time), generates vectors for each of the three and sums those vectors to produce a token. We should add support for
An arbitrary user defined encoder to tokenize a modality
Use of frozen embeddings. I.e. a user can define pre-cached tokens (of arbitrary shape) as inputs.
Currently, we only support triplet tokenization, which takes a triple (code, value, time), generates vectors for each of the three and sums those vectors to produce a token. We should add support for