PFIO

PFIO is an IO abstraction library developed by PFN, optimized for deep learning training with batteries included. It supports

• Filesystem API abstraction with unified error semantics,
• Explicit user-land caching system,
• IO performance tracing and metrics stats, and
• Fileset container utilities to save metadata.

Dependency

• HDFS client and libhdfs for HDFS access
• CPython >= 3.9

Installation and Document build

Installation

$ git clone https://github.com/pfnet/pfio.git
$ cd pfio
$ pip install .

Documentation

$ cd pfio/docs
$ make html
$ open build/html/index.html

How to use

Please refer to the official document for more information about the usage.

Release

Check the official document for latest release procedure.

Run tests locally:

$ pip install tox
$ tox

Bump version numbers in pfio/version.py .

Push and open a pull request to invoke CI. Once CI passed and the pull request merged, tag a release:

$ git tag -s X.Y.Z
$ git push --tags

Build:

$ rm -rf dist
$ pip3 install --user build
$ python3 -m build

Release to PyPI:

$ python3 -m pip install --user --upgrade twine
$ python3 -m twine upload --repository testpypi dist/*