S2-BNN: Bridging the Gap Between Self-Supervised Real and 1-bit Neural Networks via Guided Distribution Calibration (CVPR 2021)

Zhiqiang Shen

Last update: Dec 24, 2022

Overview

S²-BNN (Self-supervised Binary Neural Networks Using Distillation Loss)

This is the official pytorch implementation of our paper:

"S2-BNN: Bridging the Gap Between Self-Supervised Real and 1-bit Neural Networks via Guided Distribution Calibration" (CVPR 2021)

by Zhiqiang Shen, Zechun Liu, Jie Qin, Lei Huang, Kwang-Ting Cheng and Marios Savvides.

In this paper, we introduce a simple yet effective self-supervised approach using distillation loss for learning efficient binary neural networks. Our proposed method can outperform the simple contrastive learning baseline (MoCo V2) by an absolute gain of 5.5∼15% on ImageNet.

The student models are not restricted to the binary neural networks, you can replace with any efficient/compact models.

Citation

If you find our code is helpful for your research, please cite:

@InProceedings{Shen_2021_CVPR,
	author    = {Shen, Zhiqiang and Liu, Zechun and Qin, Jie and Huang, Lei and Cheng, Kwang-Ting and Savvides, Marios},
	title     = {S2-BNN: Bridging the Gap Between Self-Supervised Real and 1-Bit Neural Networks via Guided Distribution Calibration},
	booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
	year      = {2021}

}

Preparation

1. Requirements:

Python
PyTorch
Torchvision

2. Data:

Download ImageNet dataset following https://github.com/pytorch/examples/tree/master/imagenet#requirements.

Training & Testing

To train a model, run the following scripts. All our models are trained with 8 GPUs.

1. Standard Two-Step Training:

Our enhanced MoCo V2:

Step 1:

cd Contrastive_only/step1
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders]  --mlp --moco-t 0.2 --aug-plus --cos -j 48

Step 2:

cd Contrastive_only/step2
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders]  --mlp --moco-t 0.2 --aug-plus --cos -j 48  --model-path ../step1/checkpoint_0199.pth.tar

Our MoCo V2 + Distillation Loss:

Download real-valued teacher network here. We use MoCo V2 800-epoch pretrained model, while you can choose other stronger self-supervised models as the teachers.

Step 1:

cd Contrastive+Distillation/step1
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders] --mlp --moco-t 0.2 --aug-plus --cos -j 48 --wd 0  --teacher-path ../../moco_v2_800ep_pretrain.pth.tar

Step 2:

cd Contrastive+Distillation/step2
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders] --mlp --moco-t 0.2 --aug-plus --cos -j 48 --wd 0  --teacher-path ../../moco_v2_800ep_pretrain.pth.tar --model-path ../step1/checkpoint_0199.pth.tar

Our Distillation Loss Only:

Step 1:

cd Distillation_only/step1
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders] --mlp --moco-t 0.2 --aug-plus --cos -j 48 --wd 0 --teacher-path ../../moco_v2_800ep_pretrain.pth.tar

Step 2:

cd Distillation_only/step2
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders] --mlp --moco-t 0.2 --aug-plus --cos -j 48 --wd 0 --teacher-path ../../moco_v2_800ep_pretrain.pth.tar --model-path ../step1/checkpoint_0199.pth.tar

2. Simple One-Step Training (Conventional):

Our enhanced MoCo V2:

cd Contrastive_only/step2
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders] --mlp --moco-t 0.2 --aug-plus --cos -j 48

Our MoCo V2 + Distillation Loss:

cd Contrastive+Distillation/step2
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders] --mlp --moco-t 0.2 --aug-plus --cos -j 48 --wd 0 --teacher-path ../../moco_v2_800ep_pretrain.pth.tar

Our Distillation Loss Only:

cd Distillation_only/step2
python main_moco.py --lr 0.0003 --batch-size 256 --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders] --mlp --moco-t 0.2 --aug-plus --cos -j 48 --wd 0 --teacher-path ../../moco_v2_800ep_pretrain.pth.tar

You can replace binary neural networks with any kinds of efficient/compact models on one-step training.

3. Testing:

To linearly evaluate a model, run the following script:

python main_lincls.py  --lr 0.1  -j 24  --batch-size 256  --pretrained  /home/szq/projects/s2bnn/checkpoint_0199.pth.tar --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 [imagenet-folder with train and val folders]

Results & Models

We provide pre-trained models with different training strategies, we report in the table #epochs, OPs, Top-1 accuracy on ImageNet validation set:

Models	#Epoch	FLOPs (x10⁸)	OPs (x10⁸)	Top-1 (%)	Trained models
MoCo V2 baseline	200	0.12	0.87	46.9	Download
Our enhanced MoCo V2	200	0.12	0.87	52.5	Download
Our MoCo V2 + Distillation Loss	200	0.12	0.87	56.0	Download
Our Distillation Loss Only	200	0.12	0.87	61.5	Download

Training Logs

Our linear evaluation logs are availabe at here.

Acknowledgement

MoCo V2 (Improved Baselines with Momentum Contrastive Learning)

ReActNet (ReActNet: Towards Precise Binary NeuralNetwork with Generalized Activation Functions)

MEAL V2 (MEAL V2: Boosting Vanilla ResNet-50 to 80%+ Top-1 Accuracy on ImageNet without Tricks)

Contact

Zhiqiang Shen, CMU (zhiqiangshen0214 at gmail.com)

Comments

CIFAR100 Results

Hi, I noticed that in Table 5, the accuracy of training from scratch is only 70.9% for full precision ReActNet-MobileNet and 37.2% for binary. May I ask why the accuracy for the binary version is so low? According to the paper of ReActNet, the ReActNet-MobileNet can achieve 69.4% Top-1 Acc on ImageNet datasets, why it does not perform well on CIFAR100?

opened by NicoNico6 3
Quick question about the difference of the test results with MoCo v2

Table 4 from S2-BNN Paper

Table 3 from MoCo v2

There's a big difference between the two papers, 46.9% (S2-BNN) vs. 60.6% (MoCo v2 - without any methods). Is there any difference in settings compare to the original paper?

opened by flrngel 2

[CVPR 2021] "The Lottery Tickets Hypothesis for Supervised and Self-supervised Pre-training in Computer Vision Models" Tianlong Chen, Jonathan Frankle, Shiyu Chang, Sijia Liu, Yang Zhang, Michael Carbin, Zhangyang Wang

The Lottery Tickets Hypothesis for Supervised and Self-supervised Pre-training in Computer Vision Models Codes for this paper The Lottery Tickets Hypo

59 Dec 28, 2022

[CVPR 2022] Official code for the paper: "A Stitch in Time Saves Nine: A Train-Time Regularizing Loss for Improved Neural Network Calibration"

MDCA Calibration This is the official PyTorch implementation for the paper: "A Stitch in Time Saves Nine: A Train-Time Regularizing Loss for Improved

21 Dec 22, 2022

QTool: A Low-bit Quantization Toolbox for Deep Neural Networks in Computer Vision

This project provides abundant choices of quantization strategies (such as the quantization algorithms, training schedules and empirical tricks) for quantizing the deep neural networks into low-bit counterparts.

51 Dec 10, 2022

CBREN: Convolutional Neural Networks for Constant Bit Rate Video Quality Enhancement

CBREN This is the Pytorch implementation for our IEEE TCSVT paper : CBREN: Convolutional Neural Networks for Constant Bit Rate Video Quality Enhanceme

3 Nov 4, 2022

The Self-Supervised Learner can be used to train a classifier with fewer labeled examples needed using self-supervised learning.

Published by SpaceML • About SpaceML • Quick Colab Example Self-Supervised Learner The Self-Supervised Learner can be used to train a classifier with

92 Nov 30, 2022

Official codes: Self-Supervised Learning by Estimating Twin Class Distribution

TWIST: Self-Supervised Learning by Estimating Twin Class Distributions Codes and pretrained models for TWIST: @article{wang2021self, title={Self-Sup

85 Dec 15, 2022

RODD: A Self-Supervised Approach for Robust Out-of-Distribution Detection

RODD Official Implementation of 2022 CVPRW Paper RODD: A Self-Supervised Approach for Robust Out-of-Distribution Detection Introduction: Recent studie

17 Oct 11, 2022

This is the unofficial code of Deep Dual-resolution Networks for Real-time and Accurate Semantic Segmentation of Road Scenes. which achieve state-of-the-art trade-off between accuracy and speed on cityscapes and camvid, without using inference acceleration and extra data

Deep Dual-resolution Networks for Real-time and Accurate Semantic Segmentation of Road Scenes Introduction This is the unofficial code of Deep Dual-re

113 Dec 23, 2022

Code for the ICML 2021 paper "Bridging Multi-Task Learning and Meta-Learning: Towards Efficient Training and Effective Adaptation", Haoxiang Wang, Han Zhao, Bo Li.

Bridging Multi-Task Learning and Meta-Learning Code for the ICML 2021 paper "Bridging Multi-Task Learning and Meta-Learning: Towards Efficient Trainin

57 Dec 15, 2022

S2-BNN: Bridging the Gap Between Self-Supervised Real and 1-bit Neural Networks via Guided Distribution Calibration (CVPR 2021)

Related tags

Overview

S2-BNN (Self-supervised Binary Neural Networks Using Distillation Loss)

Citation

Preparation

1. Requirements:

2. Data:

Training & Testing

1. Standard Two-Step Training:

Our enhanced MoCo V2:

Step 1:

Step 2:

Our MoCo V2 + Distillation Loss:

Step 1:

Step 2:

Our Distillation Loss Only:

Step 1:

Step 2:

2. Simple One-Step Training (Conventional):

Our enhanced MoCo V2:

Our MoCo V2 + Distillation Loss:

Our Distillation Loss Only:

3. Testing:

Results & Models

Training Logs

Acknowledgement

Contact

You might also like...

[CVPR 2021] "The Lottery Tickets Hypothesis for Supervised and Self-supervised Pre-training in Computer Vision Models" Tianlong Chen, Jonathan Frankle, Shiyu Chang, Sijia Liu, Yang Zhang, Michael Carbin, Zhangyang Wang

[CVPR 2022] Official code for the paper: "A Stitch in Time Saves Nine: A Train-Time Regularizing Loss for Improved Neural Network Calibration"

QTool: A Low-bit Quantization Toolbox for Deep Neural Networks in Computer Vision

CBREN: Convolutional Neural Networks for Constant Bit Rate Video Quality Enhancement

The Self-Supervised Learner can be used to train a classifier with fewer labeled examples needed using self-supervised learning.

Official codes: Self-Supervised Learning by Estimating Twin Class Distribution

RODD: A Self-Supervised Approach for Robust Out-of-Distribution Detection

This is the unofficial code of Deep Dual-resolution Networks for Real-time and Accurate Semantic Segmentation of Road Scenes. which achieve state-of-the-art trade-off between accuracy and speed on cityscapes and camvid, without using inference acceleration and extra data

Code for the ICML 2021 paper "Bridging Multi-Task Learning and Meta-Learning: Towards Efficient Training and Effective Adaptation", Haoxiang Wang, Han Zhao, Bo Li.

Comments

CIFAR100 Results

Quick question about the difference of the test results with MoCo v2

Table 4 from S2-BNN Paper

Table 3 from MoCo v2

Owner

Zhiqiang Shen

[MICCAI'20] AlignShift: Bridging the Gap of Imaging Thickness in 3D Anisotropic Volumes

LiDAR Distillation: Bridging the Beam-Induced Domain Gap for 3D Object Detection

This repo is the code release of EMNLP 2021 conference paper "Connect-the-Dots: Bridging Semantics between Words and Definitions via Aligning Word Sense Inventories".

Code for 'Self-Guided and Cross-Guided Learning for Few-shot segmentation. (CVPR' 2021)'

Source code of NeurIPS 2021 Paper ''Be Confident! Towards Trustworthy Graph Neural Networks via Confidence Calibration''

modelvshuman is a Python library to benchmark the gap between human and machine vision

PyTorch implementation of "Transparency by Design: Closing the Gap Between Performance and Interpretability in Visual Reasoning"

[CVPR 21] Vectorization and Rasterization: Self-Supervised Learning for Sketch and Handwriting, IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021.

Pytorch Implementation of Spiking Neural Networks Calibration, ICML 2021

Code used to generate the results appearing in "Train longer, generalize better: closing the generalization gap in large batch training of neural networks"

S²-BNN (Self-supervised Binary Neural Networks Using Distillation Loss)