The official implementation for "FQ-ViT: Fully Quantized Vision Transformer without Retraining".

Overview

FQ-ViT [arXiv]

This repo contains the official implementation of "FQ-ViT: Fully Quantized Vision Transformer without Retraining".

Table of Contents

Introduction

Transformer-based architectures have achieved competitive performance in various CV tasks. Compared to the CNNs, Transformers usually have more parameters and higher computational costs, presenting a challenge when deployed to resource-constrained hardware devices.

Most existing quantization approaches are designed and tested on CNNs and lack proper handling of Transformer-specific modules. Previous work found there would be significant accuracy degradation when quantizing LayerNorm and Softmax of Transformer-based architectures. As a result, they left LayerNorm and Softmax unquantized with floating-point numbers. We revisit these two exclusive modules of the Vision Transformers and discover the reasons for degradation. In this work, we propose the FQ-ViT, the first fully quantized Vision Transformer, which contains two specific modules: Powers-of-Two Scale (PTS) and Log-Int-Softmax (LIS).

Layernorm quantized with Powers-of-Two Scale (PTS)

These two figures below show that there exists serious inter-channel variation in Vision Transformers than CNNs, which leads to unacceptable quantization errors with layer-wise quantization.

Taking the advantages of both layer-wise and channel-wise quantization, we propose PTS for LayerNorm's quantization. The core idea of PTS is to equip different channels with different Powers-of-Two Scale factors, rather than different quantization scales.

Softmax quantized with Log-Int-Softmax (LIS)

The storage and computation of attention map is known as a bottleneck for transformer structures, so we want to quantize it to extreme lower bit-width (e.g. 4-bit). However, if directly implementing 4-bit uniform quantization, there will be severe accuracy degeneration. We observe a distribution centering at a fairly small value of the output of Softmax, while only few outliers have larger values close to 1. Based on the following visualization, Log2 preserves more quantization bins than uniform for the small value interval with dense distribution.

Combining Log2 quantization with i-exp, which is a polynomial approximation of exponential function presented by I-BERT, we propose LIS, an integer-only, faster, low consuming Softmax.

The whole process is visualized as follow.

Getting Started

Install

  • Clone this repo.
git clone https://github.com/linyang-zhh/FQ-ViT.git
cd FQ-ViT
  • Create a conda virtual environment and activate it.
conda create -n fq-vit python=3.7 -y
conda activate fq-vit
  • Install PyTorch and torchvision. e.g.,
conda install pytorch=1.7.1 torchvision cudatoolkit=10.1 -c pytorch

Data preparation

You should download the standard ImageNet Dataset.

├── imagenet
│   ├── train
|
│   ├── val

Run

Example: Evaluate quantized DeiT-S with MinMax quantizer and our proposed PTS and LIS

python test_quant.py deit_small <YOUR_DATA_DIR> --quant --pts --lis --quant-method minmax
  • deit_small: model architecture, which can be replaced by deit_tiny, deit_base, vit_base, vit_large, swin_tiny, swin_small and swin_base.

  • --quant: whether to quantize the model.

  • --pts: whether to use Power-of-Two Scale Integer Layernorm.

  • --lis: whether to use Log-Integer-Softmax.

  • --quant-method: quantization methods of activations, which can be chosen from minmax, ema, percentile and omse.

Results on ImageNet

This paper employs several current post-training quantization strategies together with our methods, including MinMax, EMA , Percentile and OMSE.

  • MinMax uses the minimum and maximum values of the total data as the clipping values;

  • EMA is based on MinMax and uses an average moving mechanism to smooth the minimum and maximum values of different mini-batch;

  • Percentile assumes that the distribution of values conforms to a normal distribution and uses the percentile to clip. In this paper, we use the 1e-5 percentile because the 1e-4 commonly used in CNNs has poor performance in Vision Transformers.

  • OMSE determines the clipping values by minimizing the quantization error.

The following results are evaluated on ImageNet.

Method W/A/Attn Bits ViT-B ViT-L DeiT-T DeiT-S DeiT-B Swin-T Swin-S Swin-B
Full Precision 32/32/32 84.53 85.81 72.21 79.85 81.85 81.35 83.20 83.60
MinMax 8/8/8 23.64 3.37 70.94 75.05 78.02 64.38 74.37 25.58
MinMax w/ PTS 8/8/8 83.31 85.03 71.61 79.17 81.20 80.51 82.71 82.97
MinMax w/ PTS, LIS 8/8/4 82.68 84.89 71.07 78.40 80.85 80.04 82.47 82.38
EMA 8/8/8 30.30 3.53 71.17 75.71 78.82 70.81 75.05 28.00
EMA w/ PTS 8/8/8 83.49 85.10 71.66 79.09 81.43 80.52 82.81 83.01
EMA w/ PTS, LIS 8/8/4 82.57 85.08 70.91 78.53 80.90 80.02 82.56 82.43
Percentile 8/8/8 46.69 5.85 71.47 76.57 78.37 78.78 78.12 40.93
Percentile w/ PTS 8/8/8 80.86 85.24 71.74 78.99 80.30 80.80 82.85 83.10
Percentile w/ PTS, LIS 8/8/4 80.22 85.17 71.23 78.30 80.02 80.46 82.67 82.79
OMSE 8/8/8 73.39 11.32 71.30 75.03 79.57 79.30 78.96 48.55
OMSE w/ PTS 8/8/8 82.73 85.27 71.64 78.96 81.25 80.64 82.87 83.07
OMSE w/ PTS, LIS 8/8/4 82.37 85.16 70.87 78.42 80.90 80.41 82.57 82.45

Citation

If you find this repo useful in your research, please consider citing the following paper:

@misc{
    lin2021fqvit,
    title={FQ-ViT: Fully Quantized Vision Transformer without Retraining}, 
    author={Yang Lin and Tianyu Zhang and Peiqin Sun and Zheng Li and Shuchang Zhou},
    year={2021},
    eprint={2111.13824},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}
DeiT: Data-efficient Image Transformers

DeiT: Data-efficient Image Transformers This repository contains PyTorch evaluation code, training code and pretrained models for DeiT (Data-Efficient

Facebook Research 3.2k Jan 06, 2023
This repository is for Contrastive Embedding Distribution Refinement and Entropy-Aware Attention Network (CEDR)

CEDR This repository is for Contrastive Embedding Distribution Refinement and Entropy-Aware Attention Network (CEDR) introduced in the following paper

phoenix 3 Feb 27, 2022
MAME is a multi-purpose emulation framework.

MAME's purpose is to preserve decades of software history. As electronic technology continues to rush forward, MAME prevents this important "vintage" software from being lost and forgotten.

Michael Murray 6 Oct 25, 2020
Breaking the Dilemma of Medical Image-to-image Translation

Breaking the Dilemma of Medical Image-to-image Translation Supervised Pix2Pix and unsupervised Cycle-consistency are two modes that dominate the field

Kid Liet 86 Dec 21, 2022
DeLiGAN - This project is an implementation of the Generative Adversarial Network

This project is an implementation of the Generative Adversarial Network proposed in our CVPR 2017 paper - DeLiGAN : Generative Adversarial Net

Video Analytics Lab -- IISc 110 Sep 13, 2022
MVSDF - Learning Signed Distance Field for Multi-view Surface Reconstruction

MVSDF - Learning Signed Distance Field for Multi-view Surface Reconstruction This is the official implementation for the ICCV 2021 paper Learning Sign

110 Dec 20, 2022
EMNLP 2021: Single-dataset Experts for Multi-dataset Question-Answering

MADE (Multi-Adapter Dataset Experts) This repository contains the implementation of MADE (Multi-adapter dataset experts), which is described in the pa

Princeton Natural Language Processing 68 Jul 18, 2022
Laplace Redux -- Effortless Bayesian Deep Learning

Laplace Redux - Effortless Bayesian Deep Learning This repository contains the code to run the experiments for the paper Laplace Redux - Effortless Ba

Runa Eschenhagen 28 Dec 07, 2022
PyTorch code of "SLAPS: Self-Supervision Improves Structure Learning for Graph Neural Networks"

SLAPS-GNN This repo contains the implementation of the model proposed in SLAPS: Self-Supervision Improves Structure Learning for Graph Neural Networks

60 Dec 22, 2022
Galaxy images labelled by morphology (shape). Aimed at ML development and teaching

Galaxy images labelled by morphology (shape). Aimed at ML debugging and teaching.

Mike Walmsley 14 Nov 28, 2022
SARS-Cov-2 Recombinant Finder for fasta sequences

Sc2rf - SARS-Cov-2 Recombinant Finder Pronounced: Scarf What's this? Sc2rf can search genome sequences of SARS-CoV-2 for potential recombinants - new

Lena Schimmel 41 Oct 03, 2022
[NeurIPS 2021] Garment4D: Garment Reconstruction from Point Cloud Sequences

Garment4D [PDF] | [OpenReview] | [Project Page] Overview This is the codebase for our NeurIPS 2021 paper Garment4D: Garment Reconstruction from Point

Fangzhou Hong 112 Dec 23, 2022
DeepMReye: magnetic resonance-based eye tracking using deep neural networks

DeepMReye: magnetic resonance-based eye tracking using deep neural networks

73 Dec 21, 2022
Submission to Twitter's algorithmic bias bounty challenge

Twitter Ethics Challenge: Pixel Perfect Submission to Twitter's algorithmic bias bounty challenge, by Travis Hoppe (@metasemantic). Abstract We build

Travis Hoppe 4 Aug 19, 2022
On-device speech-to-intent engine powered by deep learning

Rhino Made in Vancouver, Canada by Picovoice Rhino is Picovoice's Speech-to-Intent engine. It directly infers intent from spoken commands within a giv

Picovoice 510 Dec 30, 2022
Codes for realizing theories learned from Data Mining, Machine Learning, Deep Learning without using the present Python packages.

Codes-for-Algorithms Codes for realizing theories learned from Data Mining, Machine Learning, Deep Learning without using the present Python packages.

Tracy (Shengmin) Tao 1 Apr 12, 2022
Beancount-mercury - Beancount importer for Mercury Startup Checking

beancount-mercury beancount-mercury provides an Importer for converting CSV expo

Michael Lynch 4 Oct 31, 2022
Shape-aware Semi-supervised 3D Semantic Segmentation for Medical Images

SASSnet Code for paper: Shape-aware Semi-supervised 3D Semantic Segmentation for Medical Images(MICCAI 2020) Our code is origin from UA-MT You can fin

klein 125 Jan 03, 2023
Pytorch implementation of ICASSP 2022 paper Attention Probe: Vision Transformer Distillation in the Wild

Attention Probe: Vision Transformer Distillation in the Wild Jiahao Wang, Mingdeng Cao, Shuwei Shi, Baoyuan Wu, Yujiu Yang In ICASSP 2022 This code is

IIGROUP 6 Sep 21, 2022
A crash course in six episodes for software developers who want to become machine learning practitioners.

Featured code sample tensorflow-planespotting Code from the Google Cloud NEXT 2018 session "Tensorflow, deep learning and modern convnets, without a P

Google Cloud Platform 2.6k Jan 08, 2023