RUPQ: Relative Update-Preserving Quantizer

This is the official pytorch implementation of the paper "RUPQ: Improving low-bit quantization by equalizing relative updates of quantization parameters" (link), British Machine Vision Conference, 2023. This repository contains the reproduction of the results presented in the paper for image classification (ResNet-18 and MobileNet-v2), super-resolution (SRResNet and EDSR) and object detection (YOLO-v3).

Reference

If you find our work useful, please cite:

@inproceedings{Buchnev_2023_BMVC,
author    = {Valentin Buchnev and Jiao He and Fengyu Sun and Ivan Koryakovskiy},
title     = {RUPQ: Improving low-bit quantization by equalizing relative updates of quantization parameters},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year      = {2023},
url       = {https://papers.bmvc2023.org/0307.pdf}
}

Quick Start

Prerequisites

It is recommended to use python 3.7.13.

Run this command to install all the necessary libraries:

pip install -r rupq/requirements.txt

Add the following line to the end of your ~/.bashrc file:

export PYTHONPATH=/path/to/this/repository:$PYTHONPATH

Dataset

Change the rupq/dataloaders/datasets_info.txt and enter the actual path to the required dataset.

• The DIV2K dataset can be found here.
• The Set14 dataset can be found here.
• The COCO dataset can be downloaded by running this script.

Models

• EDSR model implementation is taken from here.
• YOLO-v3 implementation is taken from here.

Training the full-precision (FP) model:

Run this command to start training FP model:

python rupq/main.py --config rupq/examples/float/classification/resnet18/fp.yaml --logdir fp_resnet18_logdir --gpu=0,1

Quantizing the model:

To start the w2a2 quantization of ResNet-18, run this command:

python rupq/main.py --config rupq/examples/quantized/classification/resnet18/w2a2_rupq.yaml --logdir w2a2_resnet18_logdir --load fp_resnet18_logdir/version_0/model.ckpt --gpu=0,1

Run the tensorboard to check the additional information about quantization process:

tensorboard --logdir w2a2_resnet18_logdir

For quantizing custom model on custom dataset, the following things are required:

1. Provide the code for dataloader in `rupq/dataloaders`.

2. Provide model definition in `rupq/models`.

3. Customize training loop in `rupq/tools/training`, if needed.

4. Create fp config, similar to the ones in `rupq/examples/float`. Train the fp model.

5. Create quantization config, similarly to the ones in `rupq/examples/quantization`. Train the quantized model.

Note: Applying standardization to input activations requires additional hyperparameters tuning, therefore we recommend to try first the setting without this feature (remove transformation from input_quantizer in the config) for your custom model. Our experience tells that input activations standardization brings a very slight improvement for quantized model quality.