[email protected]) Boosting Co-teaching with Compression Regularization for Label Noise | PythonRepo" /> [email protected]) Boosting Co-teaching with Compression Regularization for Label Noise | PythonRepo">

([email protected]) Boosting Co-teaching with Compression Regularization for Label Noise

Overview

Nested-Co-teaching

([email protected]) Pytorch implementation of paper "Boosting Co-teaching with Compression Regularization for Label Noise"

[PDF]

If our project is helpful for your research, please consider citing :

@inproceedings{chen2021boosting, 
	  title={Boosting Co-teaching with Compression Regularization for Label Noise}, 
	  author={Chen, Yingyi and Shen, Xi and Hu, Shell Xu and Suykens, Johan AK}, 
	  booktitle={CVPR Learning from Limited and Imperfect Data (L2ID) workshop}, 
	  year={2021} 
	}

Our model can be learnt in a single GPU GeForce GTX 1080Ti (12G), this code has been tested with Pytorch 1.7.1

Table of Content

1. Toy Results

The nested regularization allows us to learn ordered representation which would be useful to combat noisy label. In this toy example, we aim at learning a projection from X to Y with noisy pairs. By adding nested regularization, the most informative recontruction is stored in the first few channels.

Baseline, same MLP Nested200, 1st channel
gif gif
Nested200,first 10 channels Nested200, first 100 channels
gif gif

2. Results on Clothing1M and Animal

Clothing1M [Xiao et al., 2015]

  • We provide average accuracy as well as the standard deviation for three runs (%) on the test set of Clothing1M [Xiao et al., 2015]. Results with “*“ are either using a balanced subset or a balanced loss.
Methods [email protected] result_ref/download
CE 67.2 [Wei et al., 2020]
F-correction [Patrini et al., 2017] 68.9 [Wei et al., 2020]
Decoupling [Malach and Shalev-Shwartz, 2017] 68.5 [Wei et al., 2020]
Co-teaching [Han et al., 2018] 69.2 [Wei et al., 2020]
Co-teaching+ [Yu et al., 2019] 59.3 [Wei et al., 2020]
JoCoR [Wei et al., 2020] 70.3 --
JO [Tanaka et al., 2018] 72.2 --
Dropout* [Srivastava et al., 2014] 72.8 --
PENCIL* [Yi and Wu, 2019] 73.5 --
MLNT [Li et al., 2019] 73.5 --
PLC* [Zhang et al., 2021] 74.0 --
DivideMix* [Li et al., 2020] 74.8 --
Nested* (Ours) 73.1 ± 0.3 model
Nested + Co-teaching* (Ours) 74.9 ± 0.2 model

ANIMAL-10N [Song et al., 2019]

  • We provide test set accuracy (%) on ANIMAL-10N [Song et al., 2019]. We report average accuracy as well as the standard deviation for three runs.
Methods [email protected] result_ref/download
CE 79.4 ± 0.1 [Song et al., 2019]
Dropout [Srivastava et al., 2014] 81.3 ± 0.3 --
SELFIE [Song et al., 2019] 81.8 ± 0.1 --
PLC [Zhang et al., 2021] 83.4 ± 0.4 --
Nested (Ours) 81.3 ± 0.6 model
Nested + Co-teaching (Ours) 84.1 ± 0.1 model

3. Datasets

Clothing1M

To download Clothing1M dataset [Xiao et al., 2015], please refer to here. Once it is downloaded, put it into ./data/. The structure of the file should be:

./data/Clothing1M
├── noisy_train
├── clean_val
└── clean_test

Generate two random Clothing1M noisy subsets for training after unzipping :

cd data/
# generate two random subsets for training
python3 clothing1M_rand_subset.py --name noisy_rand_subtrain1 --data-dir ./Clothing1M/ --seed 123

python3 clothing1M_rand_subset.py --name noisy_rand_subtrain2 --data-dir ./Clothing1M/ --seed 321

Please refer to data/gen_data.sh for more details.

ANIMAL-10N

To download ANIMAL-10N dataset [Song et al., 2019], please refer to here. It includes one training and one test set. Once it is downloaded, put it into ./data/. The structure of the file should be:

./data/Animal10N/
├── train
└── test

4. Train

4.1. Stage One : Training Nested Dropout Networks

We first train two Nested Dropout networks separately to provide reliable base networks for the subsequent stage. You can run the training of this stage by :

  • For training networks on Clothing1M (ResNet-18). You can also train baseline/dropout networks for comparisons. More details are provided in nested/run_clothing1m.sh.
cd nested/ 
# train one Nested network
python3 train_resnet.py --train-dir ../data/Clothing1M/noisy_rand_subtrain1/ --val-dir ../data/Clothing1M/clean_val/ --dataset Clothing1M --arch resnet18 --lrSchedule 5 --lr 0.02 --nbEpoch 30 --batchsize 448 --nested 100 --pretrained --freeze-bn --out-dir ./checkpoints/Cloth1M_nested100_lr2e-2_bs448_freezeBN_imgnet_model1 --gpu 0
  • For training networks on ANIMAL-10N (VGG-19+BN). You can also train baseline/dropout networks for comparisons. More details are provided in nested/run_animal10n.sh.
cd nested/ 
python3 train_vgg.py --train-dir ../data/Animal10N/train/ --val-dir ../data/Animal10N/test/ --dataset Animal10N --arch vgg19-bn --lr-gamma 0.2 --batchsize 128 --warmUpIter 6000 --nested1 100 --nested2 100 --alter-train --out-dir ./checkpoints_animal10n/Animal10N_alter_nested100_100_vgg19bn_lr0.1_warm6000_bs128_model1 --gpu 0

4.2. Stage Two : Fine-tuning with Co-teaching

In this stage, the two trained networks are further fine-tuned with Co-teaching. You can run the training of this stage by :

  • For fine-tuning with Co-teaching on Clothing1M (ResNet-18) :
cd co_teaching_resnet/ 
python3 main.py --train-dir ../data/Clothing1M/noisy_rand_subtrain1/ --val-dir ../data/Clothing1M/clean_val/ --dataset Clothing1M --lrSchedule 5 --nGradual 0 --lr 0.002 --nbEpoch 30 --warmUpIter 0 --batchsize 448 --freeze-bn --forgetRate 0.3 --out-dir ./finetune_ckpt/Cloth1M_nested100_lr2e-3_bs448_freezeBN_fgr0.3_pre_nested100_100 --resumePthList ../nested/checkpoints/Cloth1M_nested100_lr2e-2_bs448_imgnet_freezeBN_model1_Acc0.735_K12 ../nested/checkpoints/Cloth1M_nested100_lr2e-2_bs448_imgnet_freezeBN_model2_Acc0.733_K15 --nested 100 --gpu 0

The two Nested ResNet-18 networks trained in stage one can be downloaded here: ckpt1, ckpt2. We also provide commands for training Co-teaching from scratch for comparisons in co_teaching_resnet/run_clothing1m.sh.

  • For fine-tuning with Co-teaching on ANIMAL-10N (VGG-19+BN) :
cd co_teaching_vgg/ 
python3 main.py --train-dir ../data/Animal10N/train/ --val-dir ../data/Animal10N/test/ --dataset Animal10N --arch vgg19-bn --lrSchedule 5 --nGradual 0 --lr 0.004 --nbEpoch 30 --warmUpIter 0 --batchsize 128 --freeze-bn --forgetRate 0.2 --out-dir ./finetune_ckpt/Animal10N_alter_nested100_lr4e-3_bs128_freezeBN_fgr0.2_pre_nested100_100_nested100_100 --resumePthList ../nested/checkpoints_animal10n/new_code_nested/Animal10N_alter_nested100_100_vgg19bn_lr0.1_warm6000_bs128_model1_Acc0.803_K14 ../nested/checkpoints_animal10n/new_code_nested/Animal10N_alter_nested100_100_vgg19bn_lr0.1_warm6000_bs128_model2_Acc0.811_K14 --nested1 100 --nested2 100 --alter-train --gpu 0

The two Nested VGG-19+BN networks trained in stage one can be downloaded here: ckpt1, ckpt2. We also provide commands for training Co-teaching from scratch for comparisons in co_teaching_vgg/run_animal10n.sh.

5. Evaluation

To evaluate models' ability of combating with label noise, we compute classification accuracy on a provided clean test set.

5.1. Stage One : Nested Dropout Networks

Evaluation of networks derived from stage one are provided here :

cd nested/ 
# for networks on 
python3 test.py --test-dir ../data/Clothing1M/clean_test/ --dataset Clothing1M --arch resnet18 --resumePthList ./checkpoints/Cloth1M_nested100_lr2e-2_bs448_imgnet_freezeBN_model1_Acc0.735_K12 --KList 12 --gpu 0

More details can be found in nested/run_test.sh. Note that "_K12" in the model's name denotes the index of the optimal K, and the optimal number of channels for the model is actually 13 (nb of optimal channels = index of channel + 1).

5.2. Stage Two : Fine-tuning Co-teaching Networks

Evaluation of networks derived from stage two are provided as follows.

  • Networks trained on Clothing1M:
cd co_teaching_resnet/ 
python3 test.py --test-dir ../data/Clothing1M/clean_test/ --dataset Clothing1M --arch resnet18 --resumePthList ./finetune_ckpt/Cloth1M_nested100_lr2e-3_bs448_freezeBN_fgr0.3_pre_nested100_100_model2_Acc0.749_K24 --KList 24 --gpu 0

More details can be found in co_teaching_resnet/run_test.sh.

  • Networks trained on ANIMAL-10N:
cd co_teaching_vgg/ 
python3 test.py --test-dir ../data/Animal10N/test/ --dataset Animal10N --resumePthList ./finetune_ckpt/Animal10N_nested100_lr4e-3_bs128_freezeBN_fgr0.2_pre_nested100_100_nested100_100_model1_Acc0.842_K12 --KList 12 --gpu 0

More details can be found in co_teaching_vgg/run_test.sh.

A program that takes in the hand gesture displayed by the user and translates ASL.

Interactive-ASL-Recognition Using the framework mediapipe made by google, OpenCV library and through self teaching, I was able to create a program tha

Riddhi Bajaj 3 Nov 22, 2021
An unofficial package help developers to implement ZATCA (Fatoora) QR code easily which required for e-invoicing

ZATCA (Fatoora) QR-Code Implementation An unofficial package help developers to implement ZATCA (Fatoora) QR code easily which required for e-invoicin

TheAwiteb 28 Nov 03, 2022
Source code of our TPAMI'21 paper Dual Encoding for Video Retrieval by Text and CVPR'19 paper Dual Encoding for Zero-Example Video Retrieval.

Dual Encoding for Video Retrieval by Text Source code of our TPAMI'21 paper Dual Encoding for Video Retrieval by Text and CVPR'19 paper Dual Encoding

81 Dec 01, 2022
Code for the paper: Fusformer: A Transformer-based Fusion Approach for Hyperspectral Image Super-resolution

Fusformer Code for the paper: "Fusformer: A Transformer-based Fusion Approach for Hyperspectral Image Super-resolution" Plateform Python 3.8.5 + Pytor

Jin-Fan Hu (胡锦帆) 11 Dec 12, 2022
Pure Javascript OCR for more than 100 Languages 📖🎉🖥

Version 2 is now available and under development in the master branch, read a story about v2: Why I refactor tesseract.js v2? Check the support/1.x br

Project Naptha 29.2k Jan 05, 2023
Code for the head detector (HeadHunter) proposed in our CVPR 2021 paper Tracking Pedestrian Heads in Dense Crowd.

Head Detector Code for the head detector (HeadHunter) proposed in our CVPR 2021 paper Tracking Pedestrian Heads in Dense Crowd. The head_detection mod

Ramana Subramanyam 76 Dec 06, 2022
This is a Computer vision package that makes its easy to run Image processing and AI functions. At the core it uses OpenCV and Mediapipe libraries.

CVZone This is a Computer vision package that makes its easy to run Image processing and AI functions. At the core it uses OpenCV and Mediapipe librar

CVZone 648 Dec 30, 2022
Implementation of our paper 'PixelLink: Detecting Scene Text via Instance Segmentation' in AAAI2018

Code for the AAAI18 paper PixelLink: Detecting Scene Text via Instance Segmentation, by Dan Deng, Haifeng Liu, Xuelong Li, and Deng Cai. Contributions

758 Dec 22, 2022
One Metrics Library to Rule Them All!

onemetric Installation Install onemetric from PyPI (recommended): pip install onemetric Install onemetric from the GitHub source: git clone https://gi

Piotr Skalski 49 Jan 03, 2023
[EMNLP 2021] Improving and Simplifying Pattern Exploiting Training

ADAPET This repository contains the official code for the paper: "Improving and Simplifying Pattern Exploiting Training". The model improves and simpl

Rakesh R Menon 138 Dec 26, 2022
computer vision, image processing and machine learning on the web browser or node.

Image processing and Machine learning labs   computer vision, image processing and machine learning on the web browser or node note Fast Fourier Trans

ryohei tanaka 487 Nov 11, 2022
A bot that plays TFT using OCR. Keeps track of bench, board, items, and plays the user defined team comp.

NOTES: To ensure best results, make sure you are running this on a computer that has decent specs. 1920x1080 fullscreen is required in League, game mu

francis 125 Dec 30, 2022
Detect and fix skew in images containing text

Alyn Skew detection and correction in images containing text Image with skew Image after deskew Install and use via pip! Recommended way(using virtual

Kakul 230 Dec 21, 2022
Detect textlines in document images

Textline Detection Detect textlines in document images Introduction This tool performs border, region and textline detection from document image data

QURATOR-SPK 70 Jun 30, 2022
A Vietnamese personal card OCR website built with Django.

Django VietCardOCR Installation Creation of virtual environments is done by executing the command venv: python -m venv venv That will create a new fol

Truong Hoang Thuan 4 Sep 04, 2021
CTPN + DenseNet + CTC based end-to-end Chinese OCR implemented using tensorflow and keras

简介 基于Tensorflow和Keras实现端到端的不定长中文字符检测和识别 文本检测:CTPN 文本识别:DenseNet + CTC 环境部署 sh setup.sh 注:CPU环境执行前需注释掉for gpu部分,并解开for cpu部分的注释 Demo 将测试图片放入test_images

Yang Chenguang 2.6k Dec 29, 2022
Official implementation of "An Image is Worth 16x16 Words, What is a Video Worth?" (2021 paper)

An Image is Worth 16x16 Words, What is a Video Worth? paper Official PyTorch Implementation Gilad Sharir, Asaf Noy, Lihi Zelnik-Manor DAMO Academy, Al

213 Nov 12, 2022
This is a tensorflow re-implementation of PSENet: Shape Robust Text Detection with Progressive Scale Expansion Network.My blog:

PSENet: Shape Robust Text Detection with Progressive Scale Expansion Network Introduction This is a tensorflow re-implementation of PSENet: Shape Robu

Michael liu 498 Dec 30, 2022
Kornia is a open source differentiable computer vision library for PyTorch.

Open Source Differentiable Computer Vision Library

kornia 7.6k Jan 06, 2023
Scene text recognition

AttentionOCR for Arbitrary-Shaped Scene Text Recognition Introduction This is the ranked No.1 tensorflow based scene text spotting algorithm on ICDAR2

777 Jan 09, 2023