Skip to content

gtars-tokenizers

Genomic region tokenizers for machine learning applications.

Features

  • Multiple tokenization strategies configurable via TOML
  • HuggingFace integration support
  • Configurable tokenization schemes
  • Optimized for ML pipelines

Python Usage

from gtars.tokenizers import Tokenizer

# load tokenizer from a BED file (creates vocabulary from regions)
tokenizer = Tokenizer.from_bed("path/to/bedfile.bed")

# load from configuration file
tokenizer = Tokenizer.from_config("tokenizer_config.toml")

# load pretrained tokenizer
tokenizer = Tokenizer.from_pretrained("path/to/model")

# tokenize regions
tokens = tokenizer.tokenize(["chr1:1000-2000", "chr2:3000-4000"])

Rust Usage

use gtars_tokenizers::Tokenizer;

// load tokenizer
let tokenizer = Tokenizer::from_bed("regions.bed")?;

// tokenize regions
let tokens = tokenizer.tokenize(&regions)?;

Integration with HuggingFace Transformers

The tokenizers were designed to be as compatible as possible with HuggingFace Transformers. You can easily integrate them into your ML pipelines.

from gtars.tokenizers import Tokenizer

tokenizer = Tokenizer.from_pretrained("databio/atacformer-base-hg38")

vocab_size = tokenizer.vocab_size
special_tokens_map = tokenizer.special_tokens_map