Question about UNI-generated tile-level embeddings and CLAM-UNI’s extract_features_fp.py

[krudo-taco]

Hi there,

First of all, thank you for your work on integrating UNI with the CLAM pipeline! Before realized CLAM’s integration with UNI, I had already been using UNI to generate tile-level embeddings. It’s great to see that CLAM now supports this directly.

I do have one question, though: Because generating embeddings can be very time-consuming, I want to confirm whether the tile-level embeddings produced from input tiles -> UNI are essentially the same as those produced by running the CLAM extract_features_fp.py script with the UNI model. In other words, has there been any modification in the way these embeddings are computed within CLAM? If not, I’d prefer to use my existing embeddings directly, rather than re-generating them, to save time.

Thank you, and I appreciate any clarification you can provide!