search

google-deepmind / alphafold

Open source code for AlphaFold.
Apache License 2.0
12.23k stars 2.18k forks source link

ETL (msa+templates) should be separated from inference #542

Open andreadisimone opened 2 years ago

andreadisimone commented 2 years ago

Dear all,

for our application, we need the possibility to run ETL (in this case MSA+templates) separate from inference.

At the moment this is not supported, but having a look at the code it should be a relatively simple thing to add a flag to stop the pipeline after the ETL is done. And if I understand correctly, the pipeline already resumes from the right point when use_precomputed_msa is used (although for some reason, the templates are always recomputed, but that does not take too long).

Is such a feature already implemented? Maybe it's planned? If not, are you accepting PRs? I would be glad to contribute some code, rather than manually hacking my private version.

Andrea.

zjq1011 commented 2 years ago

Hello, I am not sure if we want exactly the same function, though. The problem I met is the recomputation of the monomer pipeline in multimer predictions.

tcoates5 commented 1 year ago

I would also like to be able to run the CPU intensive operations (MSA and templates) separately from the GPU intensive (inference) steps. As Augustin-Zidek added the feature request label, I imagine they are open to PRs