Grant McConachie
Grant McConachie
Hi @vroulet! Yes I think this is possible! I will work on it and let you know.
Hello again @vroulet, I tried the following ``` norm_emb = jnp.linalg.norm(embeddings, axis=1, keepdims=True) norm_emb = jnp.where(norm_emb == 0.0, 1.0, norm_emb) embeddings = embeddings / norm_emb xcs = jnp.matmul(embeddings, embeddings.T) /...
Hi @vroulet! You are definitely right. Adding this epsilon term to the cosine similarity only fixed a 0 vector embedding issue. For the issue where all labels are the same,...
Hi @vroulet! Just wanted to let you know I added a doctest! Let me know what you think.
Hi @fabianp! Yes I can add this in.
Ah I see. Is there a way to get the "POM" representation from the existing codebase, or would one need to train the ensemble model, and then generate the POM...
Any updates on this? I am running into the same issues.
Hi @gjhuizing, nothing yet! luckly I saved some featurizations so I've been working with those. I saw your other issue too and I think I agree it must be a...