Skip to content

Latest commit

 

History

History
199 lines (138 loc) · 10.3 KB

README.md

File metadata and controls

199 lines (138 loc) · 10.3 KB
Raw

FedCorr: Multi-Stage Federated Learning for Label Noise Correction

This is the official PyTorch code for the following CVPR 2022 paper:

Title: FedCorr: Multi-Stage Federated Learning for Label Noise Correction. (Paper Link)

Authors: Jingyi Xu, Zihan Chen (equal contribution), Tony Q.S. Quek, and Kai Fong Ernest Chong

Abstract: Federated learning (FL) is a privacy-preserving distributed learning paradigm that enables clients to jointly train a global model. In real-world FL implementations, client data could have label noise, and different clients could have vastly different label noise levels. Although there exist methods in centralized learning for tackling label noise, such methods do not perform well on heterogeneous label noise in FL settings, due to the typically smaller sizes of client datasets and data privacy requirements in FL. In this paper, we propose FedCorr, a general multi-stage framework to tackle heterogeneous label noise in FL, without making any assumptions on the noise models of local clients, while still maintaining client data privacy. In particular, (1) FedCorr dynamically identifies noisy clients by exploiting the dimensionalities of the model prediction subspaces independently measured on all clients, and then identifies incorrect labels on noisy clients based on per-sample losses. To deal with data heterogeneity and to increase training stability, we propose an adaptive local proximal regularization term that is based on estimated local noise levels. (2) We further finetune the global model on identified clean clients and correct the noisy labels for the remaining noisy clients after finetuning. (3) Finally, we apply the usual training on all clients to make full use of all local data. Experiments conducted on CIFAR-10/100 with federated synthetic label noise, and on a real-world noisy dataset, Clothing1M, demonstrate that FedCorr is robust to label noise and substantially outperforms the state-of-the-art methods at multiple noise levels.

Citation

@inproceedings{xu2022fedcorr,
    author = {Xu, Jingyi and Chen, Zihan and Quek, Tony Q.S. and Chong, Kai Fong Ernest},
    title = {FedCorr: Multi-Stage Federated Learning for Label Noise Correction},
    booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year = {2022}
}

​

Errata [Updated on November 1, 2022]

  1. Supplementary material, Section 3.1, line 7 (arXiv version: page 13, line 1): Typo. The noise setting used for CIFAR-100 with non-IID data partition should be ( $\rho$, $\tau$ )=(0.4,0).

0. Illustration

1. Parameters

1.1. Descriptions

parameters description
iteration1 Number of iterations in pre-processing stage
rounds1 Number of rounds in finetuning stage
rounds2 Number of rounds in finetuning stage
frac1 Fraction for client selection in pre-processing stage
frac2 Fraction for client selection in other two stage
num_users Number of clients
local_bs Batch size for local training
beta Coefficient for local proximal term, default: 5
LID_k K for number of neighboring samples to estimate LID score, default: 20
level_n_system Ratio of noisy clients ($\rho$ in the paper)
level_n_lowerb Lower bound of noise level ($\tau$ in the paper)
relabel_ratio Ratio of relabeled samples among selected noisy samples
confidence_thres Threshold of model's confidence on each sample
clean_set_thres Threshold to filter 'clean' set used in finetuning stage, default: 0.1
fine_tuning Action, whether to include fine-tuning stage, default: store_false
correction Action, whether to correct noisy labels, default: store_false
model neural network model
dataset Dataset, options:cifar10,cifar100 and clothing1m
iid Action IID or non-IID data partition, default: store_true
non_iid_prob_class Non-IID sampling probability for class ($p$ in the paper)
alpha_dirichlet Parameter for Dirichlet distribution ($\alpha_{DIR}$ in the paper)
pretrained Action, whether to use pre-trained model, default: store_true
mixup Action, whether to use Mixup, default: store_true
alpha Parameters for Mixup, default: 1

1.2. Values

parameters CIFAR10 CIFAR100 CLOTHING1M
lr 0.03 0.01 0.001
iteration1 5 10 2
rounds1 500 450 50
rounds2 450 450 50
frac1 0.01 0.02 0.002
frac2 0.1 0.1 0.02
num_users 100 50 500
local_bs 10 10 16
relabel_ratio 0.5 0.5 0.8
confidence_thres 0.5 0.5 0.9
model resnet18 resnet34 resnet50

2. Heterogeneous Label Noise Model

In this paper, we propose a general framework for easy generation of heterogeneous label noise, which considers both local data statistics (i.e. IID or non-IID data partition) and local label quality (diverse noise levels for all clients).

2.1. Noisy datasets (CIFAR-10/100) used in this paper

For data partition, you can control the variability in both class distribution and the sizes of local datasets by varying args.non_iid_prob_class ($p$ in the paper) and args.alpha_dirichlet ($\alpha_{DIR}$) in util/options.py.

When adding noise, the diverse noise levels on all clients are controlled by args.level_n_system ($\rho$) and args.level_n_lowerb ($\tau$).

from util.util import add_noise
from util.dataset import get_dataset

dataset_train, dataset_test, dict_users = get_dataset(args)

# ------------------- Add Noise -----------------------
y_train = np.array(dataset_train.targets)
y_train_noisy, gamma_s, real_noise_level = add_noise(args, y_train, dict_users)
dataset_train.targets = y_train_noisy

2.2 Other datasets

If you want to apply this framework to your own datasets:

from util.util import add_noise
from util.sampling import iid_sampling, non_iid_dirichlet_sampling

dataset_train, dataset_test = get_your_own_dataset()
y_train = np.array(dataset_train.targets)

# ------------------- Data Partition -----------------------
if args.iid:
    dict_users = iid_sampling(n_train, args.num_users, args.seed)
else:
    dict_users = non_iid_dirichlet_sampling(y_train, args.num_classes, args.non_iid_prob_class, args.num_users, args.seed, args.alpha_dirichlet)

# ------------------- Add Noise -----------------------
y_train_noisy, gamma_s, real_noise_level = add_noise(args, y_train, dict_users)
dataset_train.targets = y_train_noisy

2.3 Illustration of local data partition

Three examples of different non-IID data partition (after sorting) on CIFAR-10 among 100 clients. Details of the figure and the its code can be found in plot.ipynb.

2.4 Illustration of noise model

An example of noisy label distribution on the first five clients, conducted on CIFAR-10 with IID data partition and noise setting $(\rho,\tau)=(0.6,0.5) $ among 100 clients. For each client, we give the confusion matrix of local dataset before training.

3. Combating Heterogeneous Label Noise using FedCorr

  • To train on CIFAR-10 with IID data partition and noise setting $(\rho,\tau)=(0.6,0.5)$, over 100 clients:
python main.py --dataset cifar10 --model resnet18 --iid --level_n_system 0.6 --level_n_lowerb 0.5 --iteration1 5 --rounds1 500 --rounds2 450 --seed 1 --mixup --lr 0.03 --beta 5
  • To train on CIFAR-10 with non-IID data partition with $(p,\alpha_{Dir})=(0.7,10)$ and noise setting $(\rho,\tau)=(0.6,0.5)$, over 100 clients:
python main.py --dataset cifar10 --model resnet18 --non_iid_prob_class 0.7 --alpha_dirichlet 10 --level_n_system 0.6 --level_n_lowerb 0.5 --iteration1 5 --rounds1 500 --rounds2 450 --seed 1 --mixup --lr 0.03 --beta 5
  • To train on CIFAR-100 with IID data partition and noise setting $(\rho,\tau)=(0.6,0.5)$, over 50 clients:
python main.py --dataset cifar100 --model resnet34 --num_users 50 --frac1 0.02 --iid --level_n_system 0.6 --level_n_lowerb 0.5 --iteration1 10 --rounds1 450 --rounds2 450 --seed 1 --mixup --lr 0.01 --beta 5
  • To train on Clothing1M with non-IID data partition over 500 clients:
python main.py --dataset clothing1m --model resnet50 --pretrained --num_users 500 --frac1 0.002 --frac2 0.02 --level_n_system 0 --level_n_lowerb 0 --iteration1 2 --rounds1 50 --rounds2 50 --local_bs 16 --beta 5 --confidence_thres 0.9 --relabel_ratio 0.8 --lr 0.001 --seed 1 --mixup

Please find more details of Clothing1M at https://github.com/Cysu/noisy_label.

The directory structure should be

data/
	├── clothing1m/
        ├── category_names_chn.txt
        ├── category_names_eng.txt
        ├── clean_label_kv.txt
        ├── clean_test_key_list.txt
        ├── clean_train_key_list.txt
        ├── clean_val_key_list.txt
        ├── images
        │   ├── 0
        │   ├── ⋮
        │   └── 9
        ├── noisy_label_kv.txt
        └── noisy_train_key_list.txt
    ├── cifar10/
    └── cifar100/
FedCorr/
	├── model/
	├── util/
	└── main.py