This blog post was originally published at Qualcomm’s website. It is reprinted here with the permission of Qualcomm.
AIMET now supports AdaRound and RNN quantization, plus a new GitHub open source project features code from prominent Qualcomm AI Research papers
Over the past couple of years, we’ve been expanding our collaboration efforts by starting open-source GitHub projects to share state-of-the-art techniques from Qualcomm AI Research. In May 2020, Qualcomm Innovation Center (QuIC) open sourced the AI Model Efficiency Toolkit (AIMET) on GitHub to provide a simple library plugin for AI developers to utilize for state-of-the-art quantization and compression techniques. In January 2021, QuIC contributed a collection of popular pre-trained models optimized for 8-bit inference to GitHub in the form of AIMET Model Zoo. Together with the models, AIMET Model Zoo provides the recipe for quantizing popular 32-bit floating point (FP32) models to 8-bit integer (INT8) models with little loss in accuracy.
New quantization techniques added to AIMET
We are continuing to actively contribute to AIMET, and are happy to share the news that support has been added for Adaptive Rounding (AdaRound) to achieve 4-bit quantization without sacrificing much accuracy, as well as support for the quantization of recurrent neural networks (RNNs), broadening AIMET’s ability to target networks that typically address temporal dynamic behavior, like speech recognition.
New open source project with source code from prominent papers
Today, I’m also excited to announce that we are increasing our open collaboration by introducing the new Qualcomm AI Research GitHub page. At academic AI conferences, such as NeurIPS, ICLR, and CVPR, novel papers are a primary way to contribute innovative and impactful AI research to the rest of the community. It is by sharing these new discoveries with AI researchers and engineers that we can collaborate, build on others work, and push the AI industry forward. Now, this new GitHub page will expand on these efforts by including source code associated with some key papers from Qualcomm AI Research. We hope that providing the code will allow other researchers and developers to easily build on top of it, advancing our research and leading to new innovations. Our initial contribution will include code from 4 papers, including three accepted papers from CVPR 2021:
- FrameExit: Conditional Early Exiting for Efficient Video Recognition (CVPR 2021 oral)
- InverseForm: A Loss Function for Structured Boundary-Aware Segmentation (CVPR 2021 oral)
- Skip-Convolutions for Efficient Video Processing (CVPR 2021)
- Probabilistic Numeric Convolutional Neural Networks (ICLR 2021)
This is just the beginning for the Qualcomm AI Research GitHub page, and we will continue publishing code from future AI papers on a regular basis. We also plan to make more of our datasets available to the AI community through this GitHub, as we have done in the past with Qualcomm Keyword Speech Dataset and QAST: A Dataset of Tensor Programs Execution Times (which were previously released through the Qualcomm Developer Network). I hope that our research sparks your interest, and I look forward to seeing how the AI community builds upon our work.
Dr. Joseph Soriaga
Senior Director of Technology, Qualcomm Technologies
