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

Related tags

Deep LearningTWIST
Overview

TWIST: Self-Supervised Learning by Estimating Twin Class Distributions

Architecture

Codes and pretrained models for TWIST:

@article{wang2021self,
  title={Self-Supervised Learning by Estimating Twin Class Distributions},
  author={Wang, Feng and Kong, Tao and Zhang, Rufeng and Liu, Huaping and Li, Hang},
  journal={arXiv preprint arXiv:2110.07402},
  year={2021}
}

TWIST is a novel self-supervised representation learning method by classifying large-scale unlabeled datasets in an end-to-end way. We employ a siamese network terminated by a softmax operation to produce twin class distributions of two augmented images. Without supervision, we enforce the class distributions of different augmentations to be consistent. In the meantime, we regularize the class distributions to make them sharp and diverse. TWIST can naturally avoid the trivial solutions without specific designs such as asymmetric network, stop-gradient operation, or momentum encoder.

formula

Models and Results

Main Models for Representation Learning

arch params epochs linear download
Model with multi-crop and self-labeling
ResNet-50 24M 850 75.5% backbone only full ckpt args log eval logs
ResNet-50w2 94M 250 77.7% backbone only full ckpt args log eval logs
DeiT-S 21M 300 75.6% backbone only full ckpt args log eval logs
ViT-B 86M 300 77.3% backbone only full ckpt args log eval logs
Model without multi-crop and self-labeling
ResNet-50 24M 800 72.6% backbone only full ckpt args log eval logs

Model for unsupervised classification

arch params epochs NMI AMI ARI ACC download
ResNet-50 24M 800 74.4 57.7 30.1 40.5 backbone only full ckpt args log
Top-3 predictions for unsupervised classification

Top-3

Semi-Supervised Results

arch 1% labels 10% labels 100% labels
resnet-50 61.5% 71.7% 78.4%
resnet-50w2 67.2% 75.3% 80.3%

Detection Results

Task AP all AP 50 AP 75
VOC07+12 detection 58.1 84.2 65.4
COCO detection 41.9 62.6 45.7
COCO instance segmentation 37.9 59.7 40.6

Single-node Training

ResNet-50 (requires 8 GPUs, Top-1 Linear 72.6%)

python3 -m torch.distributed.launch --nproc_per_node=8 --use_env train.py \
  --data-path ${DATAPATH} \
  --output_dir ${OUTPUT} \
  --aug barlow \
  --batch-size 256 \
  --dim 32768 \
  --epochs 800

Multi-node Training

ResNet-50 (requires 16 GPUs spliting over 2 nodes for multi-crop training, Top-1 Linear 75.5%)

python3 -m torch.distributed.launch --nproc_per_node=8 --use_env \
  --nnodes=${WORKER_NUM} \
  --node_rank=${MACHINE_ID} \
  --master_addr=${HOST} \
  --master_port=${PORT} train.py \
  --data-path ${DATAPATH} \
  --output_dir ${OUTPUT}

ResNet-50w2 (requires 32 GPUs spliting over 4 nodes for multi-crop training, Top-1 Linear 77.7%)

python3 -m torch.distributed.launch --nproc_per_node=8 --use_env \
  --nnodes=${WORKER_NUM} \
  --node_rank=${MACHINE_ID} \
  --master_addr=${HOST} \
  --master_port=${PORT} train.py \
  --data-path ${DATAPATH} \
  --output_dir ${OUTPUT} \
  --backbone 'resnet50w2' \
  --batch-size 60 \
  --bunch-size 240 \
  --epochs 250 \
  --mme_epochs 200

DeiT-S (requires 16 GPUs spliting over 2 nodes for multi-crop training, Top-1 Linear 75.6%)

python3 -m torch.distributed.launch --nproc_per_node=8 --use_env \
  --nnodes=${WORKER_NUM} \
  --node_rank=${MACHINE_ID} \
  --master_addr=${HOST} \
  --master_port=${PORT} train.py \
  --data-path ${DATAPATH} \
  --output_dir ${OUTPUT} \
  --backbone 'vit_s' \
  --batch-size 128 \
  --bunch-size 256 \
  --clip_norm 3.0 \
  --epochs 300 \
  --mme_epochs 300 \
  --lam1 -0.6 \
  --lam2 1.0 \
  --local_crops_number 6 \
  --lr 0.0005 \
  --momentum_start 0.996 \
  --momentum_end 1.0 \
  --optim admw \
  --use_momentum_encoder 1 \
  --weight_decay 0.06 \
  --weight_decay_end 0.06

ViT-B (requires 32 GPUs spliting over 4 nodes for multi-crop training, Top-1 Linear 77.3%)

python3 -m torch.distributed.launch --nproc_per_node=8 --use_env \
  --nnodes=${WORKER_NUM} \
  --node_rank=${MACHINE_ID} \
  --master_addr=${HOST} \
  --master_port=${PORT} train.py \
  --data-path ${DATAPATH} \
  --output_dir ${OUTPUT} \
  --backbone 'vit_b' \
  --batch-size 64 \
  --bunch-size 256 \
  --clip_norm 3.0 \
  --epochs 300 \
  --mme_epochs 300 \
  --lam1 -0.6 \
  --lam2 1.0 \
  --local_crops_number 6 \
  --lr 0.00075 \
  --momentum_start 0.996 \
  --momentum_end 1.0 \
  --optim admw \
  --use_momentum_encoder 1 \
  --weight_decay 0.06 \
  --weight_decay_end 0.06

Linear Classification

For ResNet-50

python3 evaluate.py \
  ${DATAPATH} \
  ${OUTPUT}/checkpoint.pth \
  --weight-decay 0 \
  --checkpoint-dir ${OUTPUT}/linear_multihead/ \
  --batch-size 1024 \
  --val_epoch 1 \
  --lr-classifier 0.2

For DeiT-S

python3 -m torch.distributed.launch --nproc_per_node=8 evaluate_vitlinear.py \
  --arch vit_s \
  --pretrained_weights ${OUTPUT}/checkpoint.pth \
  --lr 0.02 \
  --data_path ${DATAPATH} \
  --output_dir ${OUTPUT} \

For ViT-B

python3 -m torch.distributed.launch --nproc_per_node=8 evaluate_vitlinear.py \
  --arch vit_b \
  --pretrained_weights ${OUTPUT}/checkpoint.pth \
  --lr 0.0015 \
  --data_path ${DATAPATH} \
  --output_dir ${OUTPUT} \

Semi-supervised Learning

Command for training semi-supervised classification

1% Percent (61.5%)

python3 evaluate.py ${DATAPATH} ${MODELPATH} \
  --weights finetune \
  --lr-backbone 0.04 \
  --lr-classifier 0.2 \
  --train-percent 1 \
  --weight-decay 0 \
  --epochs 20 \
  --backbone 'resnet50'

10% Percent (71.7%)

python3 evaluate.py ${DATAPATH} ${MODELPATH} \
  --weights finetune \
  --lr-backbone 0.02 \
  --lr-classifier 0.2 \
  --train-percent 10 \
  --weight-decay 0 \
  --epochs 20 \
  --backbone 'resnet50'

100% Percent (78.4%)

python3 evaluate.py ${DATAPATH} ${MODELPATH} \
  --weights finetune \
  --lr-backbone 0.01 \
  --lr-classifier 0.2 \
  --train-percent 100 \
  --weight-decay 0 \
  --epochs 30 \
  --backbone 'resnet50'

Detection

Instruction

  1. Install detectron2.

  2. Convert a pre-trained MoCo model to detectron2's format:

    python3 detection/convert-pretrain-to-detectron2.py ${MODELPATH} ${OUTPUTPKLPATH}

  3. Put dataset under "detection/datasets" directory, following the directory structure requried by detectron2.

  4. Training: VOC

    cd detection/
python3 train_net.py \
  --config-file voc_fpn_1fc/pascal_voc_R_50_FPN_24k_infomin.yaml \
  --num-gpus 8 \
  MODEL.WEIGHTS ../${OUTPUTPKLPATH}

    COCO

    python3 train_net.py \
  --config-file infomin_configs/R_50_FPN_1x_infomin.yaml \
  --num-gpus 8 \
  MODEL.WEIGHTS ../${OUTPUTPKLPATH}
Owner
Bytedance Inc.
Bytedance Inc.
GitHub Repository
Code for the paper One Thing One Click: A Self-Training Approach for Weakly Supervised 3D Semantic Segmentation, CVPR 2021.

One Thing One Click One Thing One Click: A Self-Training Approach for Weakly Supervised 3D Semantic Segmentation (CVPR2021) Code for the paper One Thi

44 Dec 12, 2022
Official code for our ICCV paper: "From Continuity to Editability: Inverting GANs with Consecutive Images"

GANInversion_with_ConsecutiveImgs Official code for our ICCV paper: "From Continuity to Editability: Inverting GANs with Consecutive Images" https://a

QingyangXu 38 Dec 07, 2022
The Official Implementation of Neural View Synthesis and Matching for Semi-Supervised Few-Shot Learning of 3D Pose [NIPS 2021].

Neural View Synthesis and Matching for Semi-Supervised Few-Shot Learning of 3D Pose Release Notes The offical PyTorch implementation of Neural View Sy

Angtian Wang 20 Oct 09, 2022
Talk covering the features of skorch

Skorch Talk Skorch - A Union of Scikit-learn and PyTorch Presentation The slides can be downloaded at: download link. Google Colab Part One - MNIST Pa

Thomas J. Fan 3 Oct 20, 2020
PyTorch implementation of saliency map-aided GAN for Auto-demosaic+denosing

Saiency Map-aided GAN for RAW2RGB Mapping The PyTorch implementations and guideline for Saiency Map-aided GAN for RAW2RGB Mapping. 1 Implementations B

Yuzhi ZHAO 20 Oct 24, 2022
(Personalized) Page-Rank computation using PyTorch

torch-ppr This package allows calculating page-rank and personalized page-rank via power iteration with PyTorch, which also supports calculation on GP

Max Berrendorf 69 Dec 03, 2022
Multi-Glimpse Network With Python

Multi-Glimpse Network Multi-Glimpse Network: A Robust and Efficient Classification Architecture based on Recurrent Downsampled Attention arXiv Require

9 May 10, 2022
Official implementation of "Watermarking Images in Self-Supervised Latent-Spaces"

🔍 Watermarking Images in Self-Supervised Latent-Spaces PyTorch implementation and pretrained models for the paper. For details, see Watermarking Imag

Meta Research 32 Dec 13, 2022
ThunderSVM: A Fast SVM Library on GPUs and CPUs

What's new We have recently released ThunderGBM, a fast GBDT and Random Forest library on GPUs. add scikit-learn interface, see here Overview The miss

Xtra Computing Group 1.4k Dec 22, 2022
This project helps to colorize grayscale images using multiple exemplars.

Multiple Exemplar-based Deep Colorization (Pytorch Implementation) Pretrained Model [Jitendra Chautharia](IIT Jodhpur)1,3, Prerequisites Python 3.6+ N

jitendra chautharia 3 Aug 05, 2022
PyTorch implementation code for the paper MixCo: Mix-up Contrastive Learning for Visual Representation

How to Reproduce our Results This repository contains PyTorch implementation code for the paper MixCo: Mix-up Contrastive Learning for Visual Represen

opcrisis 46 Dec 15, 2022
LSTC: Boosting Atomic Action Detection with Long-Short-Term Context

LSTC: Boosting Atomic Action Detection with Long-Short-Term Context This Repository contains the code on AVA of our ACM MM 2021 paper: LSTC: Boosting

Tencent YouTu Research 9 Oct 11, 2022
The ARCA23K baseline system

ARCA23K Baseline System This is the source code for the baseline system associated with the ARCA23K dataset. Details about ARCA23K and the baseline sy

4 Jul 02, 2022
Implementation of the paper: "SinGAN: Learning a Generative Model from a Single Natural Image"

SinGAN This is an unofficial implementation of SinGAN from someone who's been sitting right next to SinGAN's creator for almost five years. Please ref

35 Nov 10, 2022
A curated list of programmatic weak supervision papers and resources

A curated list of programmatic weak supervision papers and resources

Jieyu Zhang 118 Jan 02, 2023
CoMoGAN: continuous model-guided image-to-image translation. CVPR 2021 oral.

CoMoGAN: Continuous Model-guided Image-to-Image Translation Official repository. Paper CoMoGAN: continuous model-guided image-to-image translation [ar

166 Dec 31, 2022
This is a Pytorch implementation of paper: DropEdge: Towards Deep Graph Convolutional Networks on Node Classification

DropEdge: Towards Deep Graph Convolutional Networks on Node Classification This is a Pytorch implementation of paper: DropEdge: Towards Deep Graph Con

401 Dec 16, 2022
This is the code for the paper "Contrastive Clustering" (AAAI 2021)

Contrastive Clustering (CC) This is the code for the paper "Contrastive Clustering" (AAAI 2021) Dependency python=3.7 pytorch=1.6.0 torchvision=0.8

Yunfan Li 210 Dec 30, 2022
PyTorch implementation of Memory-based semantic segmentation for off-road unstructured natural environments.

MemSeg: Memory-based semantic segmentation for off-road unstructured natural environments Introduction This repository is a PyTorch implementation of

11 Nov 28, 2022
A simple Neural Network that predicts the label for a series of handwritten digits

Neural_Network A simple Neural Network that predicts the label for a series of handwritten numbers This program tries to predict the label (1,2,3 etc.

Ty 1 Dec 18, 2021