Metadata-Version: 2.1
Name: aqtp
Version: 0.0.1
Summary: AQT: Accurate Quantized Training
Home-page: https://github.com/google/aqt
Author: Cerebra Catalyst team
Author-email: noreply@google.com
License: UNKNOWN
Platform: UNKNOWN
Classifier: Development Status :: 3 - Alpha
Classifier: Intended Audience :: Developers
Classifier: Intended Audience :: Science/Research
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Programming Language :: Python :: 3.7
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Description-Content-Type: text/markdown
Requires-Dist: numpy (>=1.16.4)
Requires-Dist: jax (>=0.2.7)
Requires-Dist: tensorflow (>=2.3.1)

# Accurate Quantized Training

This directory contains libraries for running and analyzing
neural network quantization experiments in JAX and flax.

Summary about this work is presented at paper [Pareto-Optimal Quantized ResNet Is Mostly 4-bit](https://arxiv.org/abs/2105.03536).
Please cite the paper in your publications if you find the source code useful for your research.

Contributors: Shivani Agrawal, Lisa Wang, Jonathan Malmaud, Lukasz Lew,
Pouya Dormiani, Phoenix Meadowlark, Oleg Rybakov.

## AQT Quantization Library

`Jax` and `Flax` quantization libraries provides `what you serve is what you train`
quantization for convolution and matmul. See [this README.md](./jax/README.md).

## Reporting Tool

After a training run has completed, the reporting tool in
`report_utils.py` allows to generate a concise experiment report with aggregated
metrics and metadata. See [this README.md](./utils/README.md).


