From TensorFlow to PyTorch with some help from Rust (Hebrew)

Length: 40
Presented at In person Rust June 2025 at AWS in Tel Aviv on 2025.06.08

When moving a machine learning project from TensorFlow to PyTorch, we wanted to keep using our existing training data files that were in the TFRecord format. Unfortunately, there was no good method of reading this format from PyTorch without adding TensorFlow as a dependency, and even that caused performance problems.

To address this, I wrote rustfrecord, a Python package written in Rust with the help of PyO3, and published to PyPI using Maturin.

In the talk I'll discuss the problem, the solution, and what can be learned from this for other projects.