ashvardanian / tenpack

Fast Tensors Packaging library for text, image, video, and audio data compatible with PyTorch, TensorFlow, & NumPy 🖼️🎵🎥 ➡️ 🧠

TenPack

Three simple things this library does to you:

1. Guess the media type from raw bytes,
2. Parse its' dimensions, sizes, lengths, etc.,
3. Unpack data into regular preallocated Tensors.

Where do we use it? To connect the Data Storage layer of UKV to High-Performance Computing libraries like TensorFlow and PyTorch.

How it works?

Most common file-formats have "signatures" or "magic numbers" embedded into them. Often, as the prefix of the byte-stream.

• List of file signatures
• Magic numbers in programming

Libraries implementing the first step have been implemented for other languages:

• filetype for GoLang
• filetype.py for Python
• FileType for Elixir
• FileSignatures for C#

Alternatives for Tensor Exports

• Pillow and Pillow-SIMD for image formats.
• FFmpeg, for video formats.
• Nyquist, for audio formats.

In fact, TenPack is just a CMake-friendly generalization of those libraries with a C interface and focus on memory reusing.

tenpack's Issues

Pure CPython rewrite of the bindings

Add: Image resizing & channel reordering

Likely via OpenCV, as it supports multiple resizing policies.

• Docs
• Tutor

Splitting channels:

• Docs

Feature: efficient python test

Test will

• Take the raw data of decoded media
• Encode it in new media file
• Compare input output files contents