PyTorch implementation for Partially View-aligned Representation Learning with Noise-robust Contrastive Loss (CVPR 2021)

Last update: Nov 01, 2022

Overview

2021-CVPR-MvCLN

This repo contains the code and data of the following paper accepted by CVPR 2021

Partially View-aligned Representation Learning with Noise-robust Contrastive Loss

Requirements

pytorch==1.5.0

numpy>=1.18.2

scikit-learn>=0.22.2

munkres>=1.1.2

logging>=0.5.1.2

Configuration

The hyper-parameters, the training options (including the ratiao of positive to negative, aligned proportions and switch time) are defined in the args. part in run.py.

Datasets

The Scene-15 and Reuters-dim10 datasets are placed in "datasets" folder. The NoisyMNIST and Caltech101 datasets could be downloaded from Google cloud or Baidu cloud with password "rqv4".

Usage

After setting the configuration and downloading datasets from the cloud desk, one could run the following code to verify our method on NoisyMNIST-30000 dataset for clustering task.

python run.py --data 3

The expected outputs are as follows:

******** Training begin, use RobustLoss: 1.0*m, use gpu 0, batch_size = 1024, unaligned_prop = 0.5, NetSeed = 64, DivSeed = 249 ********
=======> Train epoch: 0/80
margin = 5
distance: pos. = 2.5, neg. = 2.5, true neg. = 2.5, false neg. = 2.49
loss = 3.41, epoch_time = 12.07 s
******** testing ********
CAR=0.1012, kmeans: acc=0.1791, nmi=0.0435, ari=0.021
******* neg_dist_mean >= 1.0 * margin, start using fine loss at epoch: 3 *******
=======> Train epoch: 10/80
distance: pos. = 0.76, neg. = 5.38, true neg. = 5.83, false neg. = 1.34
loss = 0.09, epoch_time = 15.17 s
******** testing ********
CAR=0.8712, kmeans: acc=0.9462, nmi=0.8705, ari=0.8862
......
=======> Train epoch: 80/80
distance: pos. = 0.25, neg. = 5.34, true neg. = 5.8, false neg. = 1.17
loss = 0.03, epoch_time = 14.18 s
******** testing ********
CAR=0.8753, kmeans: acc=0.9459, nmi=0.8744, ari=0.8859
******** End, training time = 1276.29 s ********

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{yang2021MvCLN,
   title={Partially View-aligned Representation Learning with Noise-robust Contrastive Loss},
   author={Mouxing Yang, Yunfan Li, Zhenyu Huang, Zitao Liu, Peng Hu, Xi Peng},
   booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
   month={June},
   year={2021}
}

PyTorch implementation for Partially View-aligned Representation Learning with Noise-robust Contrastive Loss (CVPR 2021)

Related tags

Overview

2021-CVPR-MvCLN

Requirements

Configuration

Datasets

Usage

Citation

Owner

XLearning Group

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

An attempt at the implementation of GLOM, Geoffrey Hinton's paper for emergent part-whole hierarchies from data

:hot_pepper: R²SQL: "Dynamic Hybrid Relation Network for Cross-Domain Context-Dependent Semantic Parsing." (AAAI 2021)

CS50x-AI - Artificial Intelligence with Python from Harvard University

Detecting Human-Object Interactions with Object-Guided Cross-Modal Calibrated Semantics

This repository will be a summary and outlook on all our open, medical, AI advancements.

Scalable machine learning based time series forecasting

Perform Linear Classification with Multi-way Data

Data augmentation for NLP, accepted at EMNLP 2021 Findings

PyTorch implementation of "Contrast to Divide: self-supervised pre-training for learning with noisy labels"

Self-Supervised Contrastive Learning of Music Spectrograms

Repository for "Space-Time Correspondence as a Contrastive Random Walk" (NeurIPS 2020)

Run Effective Large Batch Contrastive Learning on Limited Memory GPU

IEEE-CIS Technical Challenge on Predict+Optimize for Renewable Energy Scheduling

Optimized Gillespie algorithm for simulating Stochastic sPAtial models of Cancer Evolution (OG-SPACE)

The official implementation of the IEEE S&P`22 paper "SoK: How Robust is Deep Neural Network Image Classification Watermarking".

FAST Aiming at the problems of cumbersome steps and slow download speed of GNSS data

PyTorch code of "SLAPS: Self-Supervision Improves Structure Learning for Graph Neural Networks"

YKKDetector For Python

TensorFlow Tutorials with YouTube Videos