pengwwang / leaf

Leaf: A Benchmark for Federated Settings

LEAF: A Benchmark for Federated Settings

Resources

• Homepage: leaf.cmu.edu
• Paper: "LEAF: A Benchmark for Federated Settings"

Datasets

1. FEMNIST

• Overview: Image Dataset
• Details: 62 different classes (10 digits, 26 lowercase, 26 uppercase), images are 28 by 28 pixels (with option to make them all 128 by 128 pixels), 3500 users
• Task: Image Classification

2. Sentiment140

• Overview: Text Dataset of Tweets
• Details 660120 users
• Task: Sentiment Analysis

3. Shakespeare

• Overview: Text Dataset of Shakespeare Dialogues
• Details: 1129 users
• Task: Next-Character Prediction

4. Celeba

• Overview: Image Dataset based on the Large-scale CelebFaces Attributes Dataset
• Details: 9343 users (we exclude celebrities with less than 5 images)
• Task: Image Classification (Smiling vs. Not smiling)

5. Synthetic Dataset

• Overview: We propose a process to generate synthetic, challenging federated datasets. The high-level goal is to create devices whose true models are device-dependant. To see a description of the whole generative process, please refer to the paper
• Details: The user can customize the number of devices, the number of classes and the number of dimensions, among others
• Task: Classification

Notes

• Install the libraries listed in requirements.txt
  • I.e. with pip: run pip3 install -r requirements.txt
• Go to directory of respective dataset for instructions on generating data
• models directory contains instructions on running baseline reference implementations