search

samsledje / D-SCRIPT

A structure-aware interpretable deep learning model for sequence-based prediction of protein-protein interactions
http://dscript.csail.mit.edu
MIT License
87 stars 18 forks source link

What is the purpose of the argument dumb-embed-switch? #28

Closed redsphinx closed 2 years ago

redsphinx commented 2 years ago

As I understand now, if set to False, it allows the interaction model in train.py to train using embedding.FullyConnectedEmbed to extract embeddings inside the run_training_from_args function. However, I'm confused because in predict.py you don't use this model for the embeddings. What am I missing? How can there be 2 different embedding models?

samsledje commented 2 years ago

Hi redsphinx,

This option is completely a holdover from development and testing and should be removed in the current version of D-SCRIPT if you install via pip. If you're cloning the latest commit from the repository, it may exist as I've been doing some bug fixing, but should be removed in the next release.

redsphinx commented 2 years ago

I see, but for now, should I just set it to True? Does train.py train independently from the embedding model?

samsledje commented 2 years ago

You can safely ignore the flag altogether, it is false by default (as it should be) it will only have an effect if set. dscript train requires that embeddings have been generated using dscript embed or methods from the lm_embed module.