Skip to content

nyuolab/WEM-Cross-Batch-Attention

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

WEM-Cross-Batch-Attention

WEM (Whole Exome Model) is a model designed to handle the entire exome in context at once. It does this through a cross batch attention mechanism, which will pass information in between the "batches" in a run. While not batches in the sense that they represent completely independent input to the model, the name comes from the idea that with a simple transpose of the input matrix we can run self attention on a dimension that will pass information between multiple documents in the same forward pass.

We are applying this to whole exomes by treating the Genes as the batch dimension and sequeneces for those genes as the sequence dimension. We hope that this architecture will allow us to perform tasks with very large context lengths

Setup

Enroot container

Enroot containers should be able to be set up automatically using the EnrootSetup.sh bash script. If this does not work, you can also build a sqsh file using the dockerfile and then create an enroot container from the sqsh file.

Running the model

WEM models can be run from the scripts in the run folder.

WEM_model_training_local.sh is a standalone script that will run on a single node using torch run

WEM_model_training_slurm.sh is an sbatch script that can run on multiple nodes.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published