Getting Started

Before diving into the specifics, ensure you have activated your preferred virtual environment.

Clone and Set Up the AutoFLIP Repository

git clone [AutoFLIP-repo-url] 
cd AutoFLIP 
git checkout [branch-name]

For the experimental code from this repository, see below.

Installation

⚠️ Note: AutoFLIP has been tested exclusively on Linux environments.

First, create a fresh conda environment and activate it:

conda env create -f environment.yml
conda activate autoflip

Repository and Code Structure

With AutoFLIP we dynamically prunes and compresses DL models within both the local clients and the global server. It leverages a federated loss exploration phase to investigate model gradient behavior across diverse datasets and losses, providing insights into parameter significance for build a efficent mask for the pruining. Our experiments showcase notable enhancements in scenarios with strong non-IID data, underscoring AutoFLIP capacity to tackle computational constraints and achieve superior global convergence.

Therefore we have to main components in our code: -AutoFlip_Exploration: Where we complete the exploration sesison for AutoFLIP. -AutoFLIP: Where we perform the Federated Leargnin traning with AutoFLIP.

Further, the repository structure is organized as follows: src: Contains the source code for algorithms, server and client implementations, and utilities. checkpoints: Stores pruned model states for both clients and the server. logs: Saves all logs from the experiments. plots: Contains the code for visualizations of the experimental results. results_paper: Holds the experimental outcomes referenced in the paper. AutoFLIP: Main directory for executing AutoFLIP experiments.

AutoFLIP/
├── src/
│   ├── algorithm/
│   │   ├── basealgorithm.py
│   │   ├── autoflip.py
│   │   ├── fedavg.ipynb
│   │   └── ...
│   ├── server/
│   │   ├── autoflipserver_exploration.py
│   │   ├── autoflipserver.py
│   │   └── ...
│   ├── client/
│   │   ├── autoflipclient_exploration.py
│   │   ├── autoflipclient.py
│   │   └── ...
│   ├── datasets/
│   ├── metrics/
│   └── models/
├── checkpoints/
│   ├── ClientPrunedGModel.py
│   └── InitPrunedGlobalModel.py
├── logs/
├── plots/
├── result_paper/
└── main.py

Example

Start your AutoFLIP journey with the following example, which reproduces the CIFAR10 experiment from the paper:

    python3 main.py\
        --exp_name "CIFAR10_${treeshold}_${n}" --device cuda --result_path ./result_paper/20K/RedDim/CIFAR10/AutoFLIP \
        --seed 1709566899 --dataset CIFAR10 \
        --split_type patho --mincls 2 --test_fraction 0.2 \
        --rawsmpl 1.0 --resize 24  --randhf 0.5  --randjit 0.5\
        --model_name TwoCNN --hidden_size 32 --dropout 0 --num_layers 4 --init_type xavier \
        --algorithm Autoflip --eval_type both --eval_every 10 --eval_metrics acc1 acc5 f1 precision recall\
        --K 20 --R 200 --E 5 --C 0.25 --B 350 --beta 0.9 \
        --optimizer Adam --lr  0.0003 --lr_decay 0.99 --lr_decay_step 1 --criterion CrossEntropyLoss\
        --Patience_mask 40 --epoochs_mask 150 --perc_clients_for_mask 1 --mask_pruining True --treeshold_pruining  0.30

Experiments

All runcommands for the ablation and experiments of AUtoFLIP are in commands/command_paper.sh.