implement cross-attention between image and language embeddings

[adityavipradas]

The current implementation in the repo performs a simple concatenation by prepending language embeddings with image embeddings. Performing cross-attention between the two embeddings will provide better token representation. I will be happy to implement this.

[andimarafioti]

We wanted to implement a self-attention architecture because it's simpler. Having the example with cross attention could be interesting, but the project changes to keep it simple aren't trivial. Honestly, it sounds like quite a hard first issue to open

[adityavipradas]

Understood. I was going for more embedding interactions but makes sense for the sake of simplicity. Do let me know if there is any interest in implementing this in the future.