128 Embeddings

[duckroll]

Is there a way to use old models with 128 embeddings instead of 512?

[defi0x1]

I think:

• Low latency compute distance.
• Features extracted are condensed
• Reduce resource storage vector embed

[duckroll]

Thanks for answer. I meant to ask how to implement old models that use 128 embeddings, not the benefits of using 512 embeddings.

[lcc157]

Add an extra nn.Linear(512,128, bias=False)?