This release contains two new implementations of Perceiver:
1. one supports multiple modalities of data in the same batch (lucidrains' implementation accepted only a single data modality at the time this repo was forked).
2. the second implements an experimental approach where text embeddings are chunked and one chunk given to one layer of the perceiver during cross-attention. This could be beneficial when you need to learn from text with other modalities, use with caution, do your own evaluations of performance.