How to get the Input embedding?

[wwzll123]

Thanks for the great work!

I wonder if it is possible to extract the input embeddings before they are fed into the diffusion model.

As highlighted in the red box below.

Thanks!

[arogozhnikov]

Hi @wwzll123

Clone project locally, use local install (pip install -e ), and modify necessary part (e.g by saving these tensors or making any other adjustments).

[wwzll123]

Hi @wwzll123

Clone project locally, use local install (pip install -e ), and modify necessary part (e.g by saving these tensors or making any other adjustments).

Thanks for your reply! I will try.

[navvye]

Just wanted to ask if these were different from the ESM embeddings that Chai-Discovery uses.

[wwzll123]

They are generated from the front trunk and should contain more complex structure information than the single ESM embedding.

…

---- Replied Message ---- | From | Navvye ***@***.***> | | Date | 12/16/2024 23:13 | | To | chaidiscovery/chai-lab ***@***.***> | | Cc | wwzll123 ***@***.***>, Mention ***@***.***> | | Subject | Re: [chaidiscovery/chai-lab] How to get the Input embedding? (Issue #224) | Just wanted to ask if these were different from the ESM embeddings that Chai-Discovery uses. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: ***@***.***>

[Garhorne0813]

Will Chai provide an interface similar to AlphaFold3? https://github.com/google-deepmind/alphafold3/commit/e3683222ce05c93479144890da65847406bad6b9

[arogozhnikov]

Will Chai provide an interface similar to AlphaFold3? https://github.com/google-deepmind/alphafold3/commit/e3683222ce05c93479144890da65847406bad6b9

it is output embeddings, not input embeddings

The code is straightforward, and all it takes is just one np.savez to store what you want - so I'm confused why this should be exposed.