MXNet-LEAF

The federated learning paradigm was recently introduced and is now gaining traction as a means of training across many users' devices. Prior to this repository, there were limited resources to build and test federated learning models in the MXNet environment. MXNet-LEAF changes that.

Instructions for Use

Requirements for Talwalkar Lab's LEAF Dataset

cd MXNet-LEAF
pip install -r requirements.txt
pip3 install -r requirements.txt # install requirements for both Python 2 and 3, just in case
git clone https://github.com/TalwalkarLab/leaf.git

Note, we are not yet done setting up LEAF. Each LEAF dataset has its own instructions, and I recommend visiting https://github.com/TalwalkarLab/leaf to figure out how to set them up. For example, if you are interesting in using the Femnist dataset, just like the instructions ask, go to the Femnist folder and run ./preprocess.sh -s niid --sf 0.05 -k 0 -t sample.

Additionally worth noting, the setup for Reddit dataset is different from the rest--be mindful!

Other Requirements

pip3 install mxnet
pip3 install gluonnlp

Workflow

1. Install requirements

2. Clone the LEAF repository

3. Install the dataset of choice using LEAF instructions

4. If using dataset Sent140, run /embeddings/get_embs.sh to install GloVe embedding

5. Run main.py. Examples included below

   python3 main.py --dataset FEMNIST --lr 0.1 --nbyz 0 --p 0 --server_pc 0 --aggregation simple --batch_size 3
   python3 main.py --dataset FEMNIST --lr 0.005 --nbyz 0 --p 0 --server_pc 0 --aggregation trim
   
   python3 main.py --dataset CELEBA --lr 0.001 --nbyz 0 --p 0 --server_pc 0 --aggregation simple --batch_size 3
   python3 main.py --dataset CELEBA --lr 0.001 --nbyz 0 --p 0 --server_pc 0 --aggregation median --batch_size 3
   
   python3 main.py --dataset SHAKESPEARE --lr 0.1 --nbyz 0 --p 0 --server_pc 0 --aggregation simple
   
   python3 main.py --dataset SENT140 --lr 0.1 --nbyz 0 --p 0 --server_pc 0 --aggregation simple --batch_size 2
   
   python3 main.py --dataset REDDIT --lr 7 --nbyz 0 --p 0 --server_pc 0 --aggregation simple --batch_size 3
   

Note, LEAF dataset names are to be completely capitalized to avoid confusion. The LEAF dataset names are as follows:

FEMNIST
CELEBA
SENT140
SHAKESPEARE
REDDIT
SYNTHETIC # no plans to implement

Extra

mxnet_from_source.md contains instructions for installing MXNet from source without root/sudo permission