search

databio / gtars

Performance-critical tools to manipulate, analyze, and process genomic interval data. Primarily focused on building tools for geniml - our genomic machine learning python package.
2 stars 1 forks source link

Hierarchical tokenizers #27

Open nleroy917 opened 1 month ago

nleroy917 commented 1 month ago

This PR introduces a few new things:

1. A new "tokenizer-config" spec

It is now preferred that we instantiate tokenizers with a tokenizers.toml file. This was added to support hierarchical tokenizers which might take many BED files to instantiate.

Documentation is needed for the tokenizer.toml config files

2. A new hierarchical tokenizer

This directly addresses discussion over in #25, https://github.com/databio/geniml_dev/issues/85, and https://github.com/databio/geniml_dev/issues/79.

A hierarchical tokenizer can take many universes as input, establishing a priority of tokenization. The primary goal here is to significantly reduce the number of UNK token hits when tokenizing datasets

3. A new MetaTokenizer

Another extension of the tokenizers, the MetaTokenizer implements the "meta-tokenization" idea we had. In brief, clusters of highly similar regions (regions that cluster super close in embedding space), are all mapped to a singular "meta token", with the primary goal of drastically reducing vocab sizes for our models to improve training and inference speed and RAM requirements.

And of course, python bindings for all of it are implemented. I've also removed a lot of code that was antiquated and unused. While doing so, I spent considerable time revamping the documentation and tests.

codecov-commenter commented 1 month ago

Welcome to Codecov :tada:

Once you merge this PR into your default branch, you're all set! Codecov will compare coverage reports and display results in all future pull requests.

Thanks for integrating Codecov - We've got you covered :open_umbrella: