HazyResearch / safari

Convolutions for Sequence Modeling
Apache License 2.0
848 stars 70 forks source link

Encoder decoder #31

Open him4318 opened 12 months ago

him4318 commented 12 months ago

Hi Can we make the hyena model work like a full vanilla transformer? where we can pass encoder last hidden state as memory to the decoder. I was trying to build OCR with the hyena model so I tried prepending the image embeddings with text embeddings but it seems to not learn anything.

Thanks