This library is currently in the Beta stage and does not have a stable release. The API may change based on user feedback or performance. We are committed to bring this library to stable release, but future changes may not be completely backward compatible. If you have suggestions on the API or use cases you'd like to be covered, please open a GitHub issue. We'd love to hear thoughts and feedback.
TorchArrow is a machine learning preprocessing library over batch data, providing performant and Pandas-style easy-to-use API for model development. Currently it provides a Python DataFrame that allows extensible UDFs with Velox, with the following features:
- Seamless handoff with PyTorch or other model authoring, such as Tensor collation and easily plugging into PyTorch DataLoader and DataPipes
- Zero copy for external readers via Arrow in-memory columnar format
- Multiple execution runtimes support:
- High-performance C++ UDF support with vectorization
You will need Python 3.7 or later. Also, we highly recommend installing an Miniconda environment.
First, set up an environment. If you are using conda, create a conda environment:
conda create --name torcharrow python=3.7
conda activate torcharrow
The following is the corresponding torcharrow versions and supported Python versions.
torch |
torcharrow |
python |
|---|---|---|
main / nightly |
main / nightly |
>=3.7, <=3.10 |
1.13.0 |
0.2.0 |
>=3.7, <=3.10 |
Follow the instructions in this Colab notebook
Experimental nightly binary on macOS (requires macOS SDK >= 10.15) and Linux (requires glibc >= 2.17) for Python 3.7, 3.8, and 3.9 can be installed via pip wheels:
pip install --pre torcharrow -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html
If you are installing from source, you will need Python 3.7 or later and a C++17 compiler.
git clone --recursive https://github.com/pytorch/torcharrow
cd torcharrow
# if you are updating an existing checkout
git submodule sync --recursive
git submodule update --init --recursiveOn macOS
HomeBrew is required to install development tools on macOS.
# Install dependencies from Brew
brew install --formula ninja flex bison cmake ccache icu4c boost gflags glog libevent
# Build and install other dependencies
scripts/build_mac_dep.sh ranges_v3 fmt double_conversion folly re2On Ubuntu (20.04 or later)
# Install dependencies from APT
apt install -y g++ cmake ccache ninja-build checkinstall \
libssl-dev libboost-all-dev libdouble-conversion-dev libgoogle-glog-dev \
libgflags-dev libevent-dev libre2-dev libfl-dev libbison-dev
# Build and install folly and fmt
scripts/setup-ubuntu.shFor local development, you can build with debug mode:
DEBUG=1 python setup.py develop
And run unit tests with
python -m unittest -v
To build and install TorchArrow with release mode:
python setup.py install
You can find the API documentation here.
This 10 minutes tutorial provides a short introduction to TorchArrow, and you can also try it in this Colab.
You can find the example about integrating a TorchRec based training loop utilizing TorchArrow's on-the-fly preprocessing here. More examples are coming soon!
We welcome PRs! See the CONTRIBUTING file.
We'd love to hear from and work with early adopters to shape our design. Please reach out by raising an issue if you're interested in using this library for your project.
We hope to continue to expand the library, harden API, and gather feedback to enable future releases. Stay tuned!
TorchArrow is BSD licensed, as found in the LICENSE file.