Note: this package is not 100% compatible with the CBOR specification. See the Not implemented section for more details.
TL;DR ApET introduces an attention-free, approximation-error guided token compression framework for VLMs that maximally preserves visual information by pruning tokens ...