Intel Low Precision Inference Tool (iLiT)

Intel Low Precision Inference Tool (iLiT) is an open-source python library which is intended to deliver a unified low-precision inference interface cross multiple Intel optimized DL frameworks on both CPU and GPU. It supports automatic accuracy-driven tuning strategies, along with additional objectives like performance, model size, or memory footprint. It also provides the easy extension capability for new backends, tuning strategies, metrics and objectives.

WARNING

GPU support is under development.

Currently supported Intel optimized DL frameworks are:

• Tensorflow*
• PyTorch*
• Apache* MXNet

Currently supported tuning strategies are:

• Basic
• Bayesian
• MSE

Documentation

• Introduction explains iLiT infrastructure, design philosophy, supported functionality, details of tuning strategy implementations and tuning result on popular models.
• Tutorial provides comprehensive step-by-step instructions of how to enable iLiT on sample models.

Install from source

git clone https://github.com/intel/lp-inference-kit.git
cd lp-inference-kit
python setup.py install

Install from binary

# install from pip
pip install ilit

# install from conda
conda config --add channels intel
conda install ilit

System Requirements

Hardware

iLiT supports systems based on Intel 64 architecture or compatible processors.

Software

iLiT requires to install Intel optimized framework version for TensorFlow, PyTorch, and MXNet.

Tuning Zoo

The followings are the examples integrated with iLiT for auto tuning.

Known Issues

1. KL Divergence Algorithm is very slow at TensorFlow

   Due to TensorFlow not supporting tensor dump naturally, current solution of dumping the tensor content is adding print op and dumpping the value to stdout. So if the model to tune is a TensorFlow model, please restrict calibration.algorithm.activation and calibration.algorithm.weight in user yaml config file to minmax.

2. MSE tuning strategy doesn't work with PyTorch adaptor layer

   MSE tuning strategy requires to compare FP32 tensor and INT8 tensor to decide which op has impact on final quantization accuracy. PyTorch adaptor layer doesn't implement this inspect tensor interface. So if the model to tune is a PyTorch model, please not choose MSE tuning strategy.

Support

Please submit your questions, feature requests, and bug reports on the GitHub issues page. You may also reach out to ilit.maintainers@intel.com.

Contributing

We welcome community contributions to iLiT. If you have an idea on how to improve the library:

• For changes impacting the public API, submit an RFC pull request.
• Ensure that the changes are consistent with the code contribution guidelines and coding style.
• Ensure that you can run all the examples with your patch.
• Submit a pull request.

For additional details, see contribution guidelines.

This project is intended to be a safe, welcoming space for collaboration, and contributors are expected to adhere to the Contributor Covenant code of conduct.

License

iLiT is licensed under Apache License Version 2.0. This software includes components with separate copyright notices and license terms. Your use of the source code for these components is subject to the terms and conditions of the following licenses.

Apache License Version 2.0:

• Intel TensorFlow Quantization Tool

MIT License:

• bayesian-optimization

See accompanying LICENSE file for full license text and copyright notices.

Legal Information

Citing

If you use iLiT in your research or wish to refer to the tuning results published in the Tuning Zoo, please use the following BibTeX entry.

@misc{iLiT,
  author =       {Feng Tian, Chuanqi Wang, Guoming Zhang, Penghui Cheng, Pengxin Yuan, Haihao Shen, and Jiong Gong},
  title =        {Intel Low Precision Inference Tool},
  howpublished = {\url{https://github.com/intel/lp-inference-kit}},
  year =         {2020}
}