Realistic FL evaluation requires careful dataset selection and environment setup. As such, FedScale provides two categories of datasets:
- Workload datasets that represent diverse FL tasks; and
- Environment datasets that reflect settings FL will be deployed.
When running in the benchmark evaluation mode, FedScale runtime can leverage the latter to evaluate the former in realistic settings.
You can run fedscale dataset
(or use download.sh
) to download or remove individual datasets.
# Run `fedscale dataset help` for more details
`fedscale dataset download [dataset_name]` # Or `bash download.sh download [dataset_name] `
We are continuously adding more datasets! Please contribute.
We provide real-world datasets for the federated learning community, and plan to release much more soon! Each one is associated with its training, validation and testing dataset. A summary of statistics for training datasets can be found in the table below, and you can refer to each folder for more details.
You can use this example code to explore any of the FedScale datasets.
Dataset | Data Type | # of Clients | # of Samples | Example Task |
---|---|---|---|---|
iNature | Image | 2,295 | 193K | Classification |
FMNIST | Image | 3,400 | 640K | Classification |
OpenImage | Image | 13,771 | 1.3M | Classification, Object detection |
Google Landmark | Image | 43,484 | 3.6M | Classification |
Charades | Video | 266 | 10K | Action recognition |
VLOG | Video | 4,900 | 9.6k | Video classification, Object detection |
Waymo Motion | Video | 496,358 | 32.5M | Motion prediction |
Dataset | Data Type | # of Clients | # of Samples | Example Task |
---|---|---|---|---|
Europarl | Text | 27,835 | 1.2M | Text translation |
Blog Corpus | Text | 19,320 | 137M | Word prediction |
Stackoverflow | Text | 342,477 | 135M | Word prediction, classification |
Text | 1,660,820 | 351M | Word prediction | |
Amazon Review | Text | 1,822,925 | 166M | Classification, Word prediction |
CoQA | Text | 7,685 | 116K | Question Answering |
LibriTTS | Text | 2,456 | 37K | Text to speech |
Google Speech | Audio | 2,618 | 105K | Speech recognition |
Common Voice | Audio | 12,976 | 1.1M | Speech recognition |
Dataset | Data Type | # of Clients | # of Samples | Example Task |
---|---|---|---|---|
Taxi Trajectory | Text | 442 | 1.7M | Sequence Prediction |
Puffer | Text | 121,551 | 15.4M | Sequence Prediction |
Taobao | Text | 182,806 | 20.9M | Recommendation |
Go dataset | Text | 150,333 | 4.9M | Reinforcement learning |
Note that no details were kept of any of the participants age, gender, or location, and random ids were assigned to each individual. In using these datasets, we will strictly obey to their licenses, and these datasets provided in this repo should be used for research purpose only.
A speech recognition dataset with over ten thousand clips of one-second-long duration. Each clip contains one of the 35 common words (e.g., digits zero to nine, "Yes", "No", "Up", "Down") spoken by thousands of different people.
OpenImage is a vision dataset collected from Flickr, an image and video hosting service. It contains a total of 16M bounding boxes for 600 object classes (e.g., Microwave oven). We clean up the dataset according to the provided indices of clients.
Reddit (StackOverflow) consists of comments from the Reddit (StackOverflow) website. It has been widely used for language modeling tasks, and we consider each user as a client. In our benchmark, we restrict to the 30k most frequently used words, and represent each sentence as a sequence of indices corresponding to these 30k frequently used words. We use Transformers to tokenize these sequences with a block size 64.
.
|---- data # Dictionary of each datasets
|---- donwload.sh # Download tool of each dataset
One of the key challenges in FL is reproducing the environment where FL will likely be deployed. To this end, we provide environmental datasets to reproduce heterogeneous device performance and device availability traces.
We use the AIBench dataset and MobiPerf dataset. AIBench dataset provides the computation capacity of different models across a wide range of devices. As specified in real FL deployments, we focus on the capability of mobile devices that have > 2GB RAM in this benchmark. To understand the network capacity of these devices, we clean up the MobiPerf dataset, and provide the available bandwidth when they are connected with WiFi, which is preferred in FL as well.
We use a large-scale real-world user behavior dataset from FLASH. It comes from a popular input method app (IMA) that can be downloaded from Google Play, and covers 136k users and spans one week from January 31st to February 6th in 2020. This dataset includes 180 million trace items (e.g., battery charge or screen lock) and we consider user devices that are in charging to be available, as specified in real FL deployments.
Please read and/or cite as appropriate to use FedScale code or data or learn more about FedScale.
@inproceedings{fedscale-icml22,
title={{FedScale}: Benchmarking Model and System Performance of Federated Learning at Scale},
author={Fan Lai and Yinwei Dai and Sanjay S. Singapuram and Jiachen Liu and Xiangfeng Zhu and Harsha V. Madhyastha and Mosharaf Chowdhury},
booktitle={International Conference on Machine Learning (ICML)},
year={2022}
}
and
@inproceedings{oort-osdi21,
title={Oort: Efficient Federated Learning via Guided Participant Selection},
author={Fan Lai and Xiangfeng Zhu and Harsha V. Madhyastha and Mosharaf Chowdhury},
booktitle={USENIX Symposium on Operating Systems Design and Implementation (OSDI)},
year={2021}
}