This benchmark also integrates LEAF, and supports FEMNIST, CelebA. For these datasets, this benchmark does not partition them further.
Most of the datasets supported by this benchmark are integrated into torchvision.datasets, expect Tiny-ImageNet-200, Covid-19, Organ-S/A/CMNIST, DomainNet.
For those datasets, I prepare download scripts (in folder data/download) for you. 🤗
e.g.
cd data/download
sh tiny_imagenet.sh📢 All arguments have their default value.
| Arguments for general datasets | Description |
|---|---|
--dataset, -d |
The name of dataset. |
--iid |
Non-zero value for randomly partitioning data and disabling all other Non-IID partition methods. |
--client_num, -cn |
The amount of clients. |
--split, -sp |
Chooses from [sample, user]. user: partition clients into train-test groups; sample: partition each client's data samples into train-test groups. |
--val_ratio, -vr |
Propotion of valset data/clients. |
--test_ratio, -tr |
Propotion of testset data/clients. |
--plot_distribution, -pd |
Non-zero value for saving data distribution image. |
⭐ For CIFAR-100 specifically, this benchmark supports partitioning it into the superclass category (CIFAR-100's 100 classes can also be classified into 20 superclasses) by setting --super_class to non-zero.
Partition data evenly. Client data distributions are similar to each other. Note that this setting has the highest priority, means that activating this scheme will disable all others.
✨ IID partition can only process partial datasets and combines other Non-IID schemes.
--iid: Need to set in[0, 1].
python generate_data.py -d cifar10 --iid 1 -cn 20
# 50% data are partitioned IID, and the rest 50% are partitioned according to dirichlet parititon scheme: Dir(0.1)
python generate_data.py -d cifar10 --iid 0.5 --alpha 0.1 -cn 20
Refers to Measuring the Effects of Non-Identical Data Distribution for Federated Visual Classification (FedAvgM). Dataset would be splitted according to
--alpha, -a: The parameter for controlling intensity of label heterogeneity.--least_samples, -ls: The parameter for defining the minimum number of samples each client would be distributed. A small--least_samplesalong with small--alphaor big--client_nummight considerablely prolong the partition.
python generate_data.py -d cifar10 -a 0.1 -cn 20
Refers to Communication-Efficient Learning of Deep Networks from Decentralized Data (FedAvg). The whole dataset would be evenly splitted into many equal-size shards.
--shards, -s: Number of data shards that each client holds. The same partition method as in FedAvg.
python generate_data.py -d cifar10 -s 2 -cn 20
Each client would be allocated data that belongs to -c classes. And classes for each client are randomly choosed.
--classes, -c: Number of classes that each client's data belong to.
python generate_data.py -d cifar10 -c 2 -cn 20
Refers to What Do We Mean by Generalization in Federated Learning?. Each client's data are correspond to a gaussian distribution that generated by a gaussian mixture model. You can learn the whole process precedure in paper's Appendix D.
--semantic, -sm: Non-zero value for performing semantic partition.--efficient_net_type: The type of EfficientNet for computing the embeddings of data.--pca_components: The number of dimension for PCA decomposition.--gmm_max_iter: The maximum number of fitting iteration of the gaussian mixture model.--gmm_init_params: The way for initializing gaussian mixture model (kmeans/random).--use_cuda: Non-zero value for using CUDA to accelerate the computation.
python generate_data.py -d cifar10 -sm 1 -cn 20
Refers to Federated Optimization in Heterogeneous Networks
. The whole dataset are generated according to
-
(--gamma, --beta): The parameters$(\alpha, \beta)$ in paper respectively. -
--dimension: The dimension of synthetic data. -
--iid: Non-zero value for generating IID synthetic dataset.
python generate_data.py -d synthetic --beta 1 --gamma 1 -cn 20 
Argument details are in data/femnist/README.md and data/celeba/README.md
You should set all arguments well already when running preprocess.sh.
All generic arguments (except -d) in generate_data.py will be deactivated when processing LEAF datasets.
When processing LEAF datasets, generate_data.py only responsible for translating the output of preprocess.sh (json data files) to data.npy and targets.npy.
- Through
data/download/domain.shdownloading and decomporessing DomainNet raw data. cdtodata/domainand runpython preprocess.py(an interactive wizard).
Running python generate_data.py -d domain without additional arguments would build domain separation partition (each client has only has data from one domain).
- Note that
python generate_data.py -d domainis at the end ofdata/domain/preprocess.pyalready, so you don't need to run that command by yourself after runningpreprocess.py.
Combine argument of other schemes to build custom label heterogeneity.
Note that performing extra partitioning would make the number of clients reset by --client_num instead of the client for each domain in preprocess.py you set previously.
python generate_data.py -d domain -a 1.0 -cn 10
Set --ood_domains {domain_name...} to exclude domain(s) out of data partitioning and allocate corresponding data samples to test client(s).
Note that if --ood_domains is not empty, FL-bench will map the data labels from the class space to the domain space. So the data hetergeneity will be observed in the domain space instead of the class space.
One OOD domain for one test client.
python generate_data.py -d domain -a 1.0 -cn 10 --ood_domains sketch
python generate_data.py -d domain -a 1.0 -cn 10 --ood_domains sketch quickdraw
data/femnist, data/celeba, data/leaf_utils are copied from LEAF with subtle modifications to be integrated into this benchmark. data/femnist/README.md and data/celeba/README.md for full details.
FL-bench ignores the test set of Tiny-ImageNet-200 due to it is unlabeled.