/FedNoisy

Benchmark for federated noisy label learning

Primary LanguagePythonApache License 2.0Apache-2.0

FedNoisy: Federated Noisy Label Learning Benchmark

Benchmark for federated noisy label learning.

[toc]

License arXiv

Federated learning has gained popularity for distributed learning without aggregating sensitive data from clients. But meanwhile, the distributed and isolated nature of data isolation may be complicated by data quality, making it more vulnerable to noisy labels. Many efforts exist to defend against the negative impacts of noisy labels in centralized or federated settings. However, there is a lack of a benchmark that comprehensively considers the impact of noisy labels in a wide variety of typical FL settings.

In this work, we serve the first standardized benchmark that can help researchers fully explore potential federated noisy settings. Also, we conduct comprehensive experiments to explore the characteristics of these data settings and unravel challenging scenarios on the federated noisy label learning, which may guide method development in the future. We highlight the 20 basic settings for more than 5 datasets proposed in our benchmark and standardized simulation pipeline for federated noisy label learning. We hope this benchmark can facilitate idea verification in federated learning with noisy labels.

Environment requirements

Pytorch installation via conda:

$ conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c nvidia -y

Extra dependencies:

$ pip install -r requirements.txt

Federated Noisy Labe Settings

Current supported datasets:

Noise Scene Dataset #Train #Validation #Test #Class ImageSize Noise Ratio (%)

Globalized

&

Localized

 MNIST 60K N/A 10K 10 28×28 N/A
SVHN 73K N/A 26K 10 32×32×3 N/A
CIFAR-10 50K N/A 10K 10 32×32×3 N/A
CIFAR-100 50K N/A 10K 100 32×32×3 N/A
Real-world Clothing1M 1M 14K 10K 14 224×224×3 ≈39.46
WebVision 2.4M 50K 50K 1000 256×256×3 ≈20.00

Federated noise scenes provided in $\texttt{FedNoisy}$ is shown as below. (Left: globalized noise; middle: localized noise; right: real-word noise. Globalized noise follows a constant noise ratio $\varepsilon_{global}$ in label corruption, while localized noise draws localized noise ratio $\varepsilon_{k}$ from a uniform distribution on each client $k$.)

Prepare

Raw dataset should be downloaded in to local folder before data-build process. The folder structure can be:

.
├── FedNoisy
│   ├── LICENSE
│   ├── README.md
│   ├── build_dataset_fed.py
│   ├── fednoisy
│   ├── imgs
│   ├── requirements.txt
│   └── scripts
├── rawdata
│   ├── cifar10
│   │   ├── cifar-10-batches-py
│   │   └── cifar-10-python.tar.gz
│   ├── cifar100
│   │   ├── cifar-100-python
│   │   └── cifar-100-python.tar.gz
│   ├── mnist
│   │   └── MNIST
│   │       └── raw
│   │           ├── t10k-images-idx3-ubyte
│   │           ...
│   │           └── train-labels-idx1-ubyte.gz
│   ├── 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
│   │   │   ├── 1
│   │   │   ...
│   │   │   └── 9
│   │   ├── noisy_label_kv.txt
│   │   ├── noisy_train_key_list.txt
│   │   ├── README.md
│   │   └── venn.png
│   └── svhn
│       ├── extra_32x32.mat
│       ├── test_32x32.mat
│       └── train_32x32.mat
├── fedNLLdata  # to store fedNLL data settings 
└── Fed-Noisy-checkpoint  # to store algorithm logging output and final models

  • Use torchvision.dataset to download of CIFAR-10/CIFAR-100/SVHN/MNIST directly:

    import torchvision

mnist_train = torchvision.datasets.MNIST(root="rawdata/mnist", train=True, download=True)
mnist_test = torchvision.datasets.MNIST(root="rawdata/mnist", train=False, download=True)

cifar10_train = torchvision.datasets.CIFAR10(root="rawdata/cifar10", train=True, download=True)
cifar10_test = torchvision.datasets.CIFAR10(root="rawdata/cifar10", train=False, download=True)

cifar100_train = torchvision.datasets.CIFAR100(root="rawdata/cifar100", train=True, download=True)
cifar100_test = torchvision.datasets.CIFAR100(root="rawdata/cifar100", train=False, download=True)

svhn_train = torchvision.datasets.SVHN(root="rawdata/svhn", split="train", download=True)
svhn_test = torchvision.datasets.SVHN(root="rawdata/svhn", split="test", download=True)

  • To download Clothing1M, please contact tong.xiao.work[at]gmail[dot]com to get the download link. Untar the images and unzip the annotations under rawdata/clothing1M.

Build dataset with noisy label

The basic command usage is

$ cd FedNoisy 

# under dir FedNoisy/
$ python build_dataset_fed.py --dataset cifar10 \
     --partition iid \
     --num_clients 10 \
     --globalize \
     --noise_mode clean \
     --raw_data_dir ../rawdata/cifar10 \
     --data_dir ../fedNLLdata/cifar10 \
     --seed 1

Noise related setting

  • Clean: --globalize --noise_mode clean for data setting without noise

  • Globalized noise

    • --globalize --noise_ratio 0.4 --noise_mode sym for globalized symmetric noise $\varepsilon_{global}=0.4$
    • --globalize --noise_ratio 0.4 --noise_mode asym for globalized asymmetric noise $\varepsilon_{global}=0.4$

  • Localized noise

    • --min_noise_ratio 0.3 --max_noise_ratio 0.5 --noise_mode sym for localized symmetric noise $\varepsilon_k \sim \mathcal{U}(0.3, 0.5)$
    • --min_noise_ratio 0.3 --max_noise_ratio 0.5 --noise_mode asym for localized asymmetric noise $\varepsilon_k \sim \mathcal{U}(0.3, 0.5)$

  • Real noise (only works for Clothing1M): --dataset clothing1m --globalize --noise_mode real --num_sampels 64000

    • --num_samples is for specifying number of training samples used for Clothing1M, the default is 64000

Dataset partition setting

  • MNIST: --dataset mnist

    • IID: --partition iid --num_clients 10
    • Non-IID quantity skew: --partition noniid-quantity --num_clients 10 --dir_alpha 0.1
    • Non-IID Dirichlet-based label skew: --partition noniid-labeldir --dir_alpha 0.1 --num_clients 10
    • Non-IID quantity-based label skew: --partition noniid-#label --major_classes_num 3 --num_clients 10

  • SVHN: --dataset svhn

  • CIFAR-10: --dataset cifar10

  • CIFAR-100: --dataset cifar100

  • Clothing1M: --dataset cloting1m

    • IID: --partition iid --num_clients 10
    • Non-IID quantity skew: --partition noniid-quantity --num_clients 10 --dir_alpha 0.1
    • Non-IID Dirichlet-based label skew: --partition noniid-labeldir --dir_alpha 0.1 --num_clients 10
    • Non-IID quantity-based label skew: --partition noniid-#label --major_classes_num 5 --num_clients 10

Algorithms

Federated Algorithm Noisy Label Algorithm Paper
Category Method
FedAvg Robust Regularizaiton Mixup [2018 ICLR] Mixup: Beyond empirical risk minimization
RobustLoss Function SCE [2019 ICCV] Symmetric Cross Entropy for Robust Learning with Noisy Labels
GCE [2018 NeurIPS] Generalized Cross Entropy Loss for Training Deep Neural Networks with Noisy Labels
MAE [2017 AAAI] Robust Loss Functions under Label Noise for Deep Neural Networks
Loss Adjustment M-DYR-H [2019 ICML] Unsupervised Label Noise Modeling and Loss Correction
M-DYR-S [2019 ICML] Unsupervised Label Noise Modeling and Loss Correction
DM-DYR-SH [2019 ICML] Unsupervised Label Noise Modeling and Loss Correction
Sample Selection Co-teaching [2018 NeurIPS] Co-teaching: Robust Training of Deep Neural Networks with Extremely Noisy Labels

  • FedAvg

    • Vanilla FedAvg

      # under dir FedNoisy/
$ python python fednoisy/algorithms/fedavg/main.py --dataset mnist \
    --model SimpleCNN \
    --partition iid \
    --num_clients 10 \
    --globalize \
    --noise_mode sym \
    --noise_ratio 0.4 \
    --seed 1 \
    --sample_ratio 1.0 \
    --com_round 500 \
    --epochs 5 \
    --momentum 0.9 \
    --lr 0.01 \
    --weight_decay 0.0005 \
    --data_dir ../fedNLLdata/mnist \
    --out_dir ../Fed-Noisy-checkpoint/mnist/

    • FedAvg + Symmetric Cross Entropy

      # under dir FedNoisy/
$ python fednoisy/algorithms/fedavg/main.py --dataset mnist \
    --model SimpleCNN \
    --partition iid \
    --num_clients 10 \
    --globalize \
    --noise_mode sym \
    --noise_ratio 0.4 \
    --data_dir ../fedNLLdata/mnist \
    --out_dir ../Fed-Noisy-checkpoint/mnist/ \
    --com_round 500 \
    --epochs 5 \
    --sample_ratio 1.0 \
    --lr 0.01 \
    --momentum 0.9 \
    --weight_decay 0.0005 \
    --criterion sce \
    --sce_alpha 0.01 \
    --sce_beta 1.0 \
    --seed 1

    • FedAvg + DM-DYR-SH

      # under dir FedNoisy/
$ python fednoisy/algorithms/fedavg/main.py --dataset mnist \
    --model SimpleCNN \
    --partition iid \
    --num_clients 10 \
    --globalize \
    --noise_mode sym \
    --noise_ratio 0.4 \
    --data_dir ../fedNLLdata/mnist \
    --out_dir ../Fed-Noisy-checkpoint/mnist/ \
    --com_round 500 \
    --epochs 5 \
    --sample_ratio 1.0 \
    --lr 0.01 \
    --momentum 0.9 \
    --weight_decay 1e-4 \
    --dynboot \
    --dynboot_alpha 32 \
    --dynboot_mixup dynamic \
    --dynboot_reg 1.0 \
    --seed 1

    • FedAvg + Mixup

      # under dir FedNoisy/
$ python fednoisy/algorithms/fedavg/main.py --dataset mnist \
    --model SimpleCNN \
    --partition iid \
    --num_clients 10 \
    --globalize \
    --noise_mode sym \
    --noise_ratio 0.4 \
    --data_dir ../fedNLLdata/mnist \
    --out_dir ../Fed-Noisy-checkpoint/mnist/ \
    --com_round 500 \
    --epochs 5 \
    --sample_ratio 1.0 \
    --lr 0.01 \
    --momentum 0.9 \
    --weight_decay 5e-4 \
    --mixup \
    --mixup_alpha 1.0 \
    --seed 1

    • FedAvg + Co-teaching

      # under dir FedNoisy/
$ python fednoisy/algorithms/fedavg/main.py --dataset mnist \
    --model SimpleCNN \
    --partition iid \
    --num_clients 10 \
    --globalize \
    --noise_mode sym \
    --noise_ratio 0.4 \
    --data_dir ../fedNLLdata/mnist  \
    --out_dir ../Fed-Noisy-checkpoint/mnist/ \
    --com_round 500 \
    --epochs 5 \
    --sample_ratio 1.0 \
    --lr 0.01 \
    --momentum 0.9 \
    --weight_decay 5e-4 \
    --coteaching \
    --coteaching_forget_rate 0.4 \  # depends on noise ratio, possible value can be noise ratio it self 
    --coteaching_num_gradual 25 \  # default setting is 10 for 200 epochs, here we set similar ratio with number of global round
    --seed 1

For more scripts, please check scripts folder.

References

[1] LeCun, Y. (1998). The MNIST database of handwritten digits. http://yann.lecun.com/exdb/mnist/.

[2] Krizhevsky, A., & Hinton, G. (2009). Learning multiple layers of features from tiny images.

[3] Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., & Ng, A. Y. (2011). Reading digits in natural images with unsupervised feature learning.

[4] Xiao, T., Xia, T., Yang, Y., Huang, C., & Wang, X. (2015). Learning from massive noisy labeled data for image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2691-2699).

[5] Li, W., Wang, L., Li, W., Agustsson, E., & Van Gool, L. (2017). Webvision database: Visual learning and understanding from web data. arXiv preprint arXiv:1708.02862.