brian14708 / tsar

Archive file format for storing large tensors efficiently

Tensor Archive

Archive file format for storing tensors, with optional lossy compression for better storage efficiency.

Features:

• floating-point compression (zfp)
• storing data in lower precision format (bfloat16, ...)
• compressing mantissa and exponents separately
• tools for building archives from ONNX format

Quick Start

Install tsar-py package from source (require Rust build environment):

pip install git+https://github.com/brian14708/tsar.git#subdirectory=tsar-py

ONNX format

# lossless compression
tsar-pack -e 0 "<ONNX-FILE>.onnx" output.tsar
# lossy compression (with maximum error 1e-6)
tsar-pack -e 1e-6 "<ONNX-FILE>.onnx" output.tsar

# extract file to model/ directory
tsar-unpack output.tsar model/

Python API

TODO

Results

Model	Compression	Size
ResNet-152	none	230.6 MiB
	gzip	215.4 MiB
	tsar (lossless)	197.4 MiB
	tsar (err=1e-6)	129.8 MiB
	tsar (err=1e-5)	108.7 MiB
	tsar (err=1e-4)	87.8 MiB
	tsar (err=1e-3)	60.6 MiB