:lollipop: Epigenetics Dataloader for BigWig files

Fast batched dataloading of BigWig files containing epigentic track data and corresponding sequences powered by GPU for deep learning applications.

⚠️ BREAKING CHANGE (v0.3.0+): The output matrix dimensionality has changed from (n_tracks, batch_size, sequence_length) to (batch_size, sequence_length, n_tracks). This change was long overdue and eliminates the need for (potentially memory expensive) transpose operations downstream. If you're upgrading from an earlier version, please update your code accordingly (probaby you need to delete one transpose in your code).

✨ NEW FEATURE (v0.3.0+): Full bfloat16 support! You can now specify dtype="bfloat16" to get output tensors in bfloat16 format, reducing memory usage by 50%.

⚠️ Cupy and bfloat16 support Because cupy does not support bfloat16 yet, the cupy array is typed as uint64, but the actual data behind it is in bfloat16. So when converting the array to a tensor in a framework that DOES support bfloat16 like pytorch, tensorflow or JAX should be followed by a "view" method that just changes how the underlying bytes are interpreted (and not actually casting to bfloat16, which would change the underlaying data). In the bigwig_loader.pytorch.PytorchBigWigDataset this has already been done for you (when you set dtype="bfloat16").

Quickstart

Installation with Pixi

Using pixi to install bigwig-loader is highly recommended. Please take a look at this example pixi.toml:

[workspace]
channels = ["rapidsai", "conda-forge", "nvidia", "bioconda", "dataloading"]
name = "bigwig-loader"
platforms = ["linux-64"]
version = "0.1.0"

[tasks]
download-example-data = { cmd = "python -m bigwig_loader.download_example_data"}

[feature.bigwig-loader.system-requirements]
cuda = "12"

[dependencies]
python = "==3.11"
pip = "*"

[feature.bigwig-loader.dependencies]
cuda-version = "12.8.*"
pytorch-gpu = ">=2.6"
cuda-nvcc = "*"
kvikio = "<=25.08.00"
bigwig-loader = "*"
numpy = "*"
pandas = "*"

[pypi-dependencies]
python-dotenv = "*"
pydantic = "*"
pydantic-settings = "*"
universal-pathlib = "*"
fsspec = { version = "*" }
s3fs = "*"
pyfaidx = "*"
numcodecs ="*"

[environments]
default = {features = ["bigwig-loader"]}

If you just want to use bigwig-loader, just copy that into a pixi.toml file and add the other libraries you need. (you don't need to clone this repo, pixi will download bigwig-loader from the conda "dataloading" channel):

• Install pixi, if not installed:

  curl -fsSL https://pixi.sh/install.sh | sh
  

• change directory to wherever you put the pixi.toml, and:

  pixi run <my_training_command>
  

The pixi.toml I included in this repository works for both the released version and for development of bigwig-loader, but assumes you cloned this repo.

Installation with conda/mamba

Alternatively, bigwig-loader can be installed using conda/mamba. To create a new environment with bigwig-loader installed:

mamba create -n my-env -c rapidsai -c conda-forge -c bioconda -c dataloading bigwig-loader

Or add this to you environment.yml file:

name: my-env
channels:
  - rapidsai
  - conda-forge
  - bioconda
  - dataloading
dependencies:
    - bigwig-loader

and update:

mamba env update -f environment.yml

Installation with pip

Bigwig-loader can also be installed using pip in an environment which has the rapidsai kvikio library and cupy installed already:

pip install bigwig-loader

PyTorch Example

We wrapped the BigWigDataset in a PyTorch iterable dataset that you can directly use:

# examples/pytorch_example.py
import pandas as pd
import torch
from torch.utils.data import DataLoader
from bigwig_loader import config
from bigwig_loader.pytorch import PytorchBigWigDataset
from bigwig_loader.download_example_data import download_example_data

# Download example data to play with
download_example_data()
example_bigwigs_directory = config.bigwig_dir
reference_genome_file = config.reference_genome

train_regions = pd.DataFrame({"chrom": ["chr1", "chr2"], "start": [0, 0], "end": [1000000, 1000000]})

dataset = PytorchBigWigDataset(
    regions_of_interest=train_regions,
    collection=example_bigwigs_directory,
    reference_genome_path=reference_genome_file,
    sequence_length=1000,000,
    center_bin_to_predict=500,000,
    window_size=1,
    batch_size=1,
    super_batch_size=4,
    batches_per_epoch=100,
    maximum_unknown_bases_fraction=0.1,
    sequence_encoder="onehot",
    n_threads=4,
    return_batch_objects=True,
    dtype="bfloat16"
)

# Don't use num_workers > 0 in DataLoader. The heavy
# lifting/parallelism is done on cuda streams on the GPU.
dataloader = DataLoader(dataset, num_workers=0, batch_size=None)


class MyTerribleModel(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = torch.nn.Linear(4, 2)

    def forward(self, batch):
        return self.linear(batch)


model = MyTerribleModel()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)

def poisson_loss(pred, target):
    return (pred - target * torch.log(pred.clamp(min=1e-8))).mean()

for batch in dataloader:
    # batch.sequences.shape = n_batch x sequence_length x onehot encoding (4)
    pred = model(batch.sequences)
    # batch.values.shape = n_batch x center_bin_to_predict x n_tracks
    loss = poisson_loss(pred[:, 250000:750000, :], batch.values)
    print(loss)
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

Other frameworks

A framework agnostic Dataset object can be imported from bigwig_loader.dataset. This dataset object returns cupy tensors. Cupy tensors adhere to the cuda array interface and can be zero-copy transformed to JAX or tensorflow tensors.

from bigwig_loader.dataset import BigWigDataset

dataset = BigWigDataset(
    regions_of_interest=train_regions,
    collection=example_bigwigs_directory,
    reference_genome_path=reference_genome_file,
    sequence_length=1000,
    center_bin_to_predict=500,
    window_size=1,
    batch_size=32,
    super_batch_size=1024,
    batches_per_epoch=20,
    maximum_unknown_bases_fraction=0.1,
    sequence_encoder="onehot",
)

See the examples directory for more examples.

Background

This library is meant for loading batches of data with the same dimensionality, which allows for some assumptions that can speed up the loading process. As can be seen from the plot below, when loading a small amount of data, pyBigWig is very fast, but does not exploit the batched nature of data loading for machine learning.

In the benchmark below we also created PyTorch dataloaders (with set_start_method('spawn')) using pyBigWig to compare to the realistic scenario where multiple CPUs would be used per GPU. We see that the throughput of the CPU dataloader does not go up linearly with the number of CPUs, and therefore it becomes hard to get the needed throughput to keep the GPU, training the neural network,saturated during the learning steps.

This is the problem bigwig-loader solves. This is an example of how to use bigwig-loader:

Installation

1. git clone git@github.com:pfizer-opensource/bigwig-loader
2. cd bigwig-loader
3. create the conda environment" conda env create -f environment.yml

In this environment you should be able to run pytest -v and see the tests succeed. NOTE: you need a GPU to use bigwig-loader!

Development

This section guides you through the steps needed to add new functionality. If anything is unclear, please open an issue.

Environment

The pixi.toml includes a dev environment that has bigwig-loader installed as an editable pypi dependency.

1. git clone git@github.com:pfizer-opensource/bigwig-loader
2. cd bigwig-loader
3. optional: pixi install -e dev
4. run pre-commit install to install the pre-commit hooks

Run Tests

Tests are in the tests directory. One of the most important tests is test_against_pybigwig which makes sure that if there is a mistake in pyBigWIg, it is also in bigwig-loader.

In order to run these tests you need gpu.

pixi run -e dev test

When github runners with GPU's will become available we would also like to run these tests in the CI. But for now, you can run them locally.

Citing

If you use this library, consider citing:

Retel, Joren Sebastian, Andreas Poehlmann, Josh Chiou, Andreas Steffen, and Djork-Arné Clevert. “A Fast Machine Learning Dataloader for Epigenetic Tracks from BigWig Files.” Bioinformatics 40, no. 1 (January 1, 2024): btad767. https://doi.org/10.1093/bioinformatics/btad767.

@article{
    retel_fast_2024,
    title = {A fast machine learning dataloader for epigenetic tracks from {BigWig} files},
    volume = {40},
    issn = {1367-4811},
    url = {https://doi.org/10.1093/bioinformatics/btad767},
    doi = {10.1093/bioinformatics/btad767},
    abstract = {We created bigwig-loader, a data-loader for epigenetic profiles from BigWig files that decompresses and processes information for multiple intervals from multiple BigWig files in parallel. This is an access pattern needed to create training batches for typical machine learning models on epigenetics data. Using a new codec, the decompression can be done on a graphical processing unit (GPU) making it fast enough to create the training batches during training, mitigating the need for saving preprocessed training examples to disk.The bigwig-loader installation instructions and source code can be accessed at https://github.com/pfizer-opensource/bigwig-loader},
    number = {1},
    urldate = {2024-02-02},
    journal = {Bioinformatics},
    author = {Retel, Joren Sebastian and Poehlmann, Andreas and Chiou, Josh and Steffen, Andreas and Clevert, Djork-Arné},
    month = jan,
    year = {2024},
    pages = {btad767},
}