FedNLP: A Benchmarking Framework for Federated Learning in Natural Language Processing
FedNLP is a research-oriented benchmarking framework for advancing federated learning (FL) in natural language processing (NLP). It uses FedML repository as the git submodule. In other words, FedNLP only focuses on adavanced models and dataset, while FedML supports various federated optimizers (e.g., FedAvg) and platforms (Distributed Computing, IoT/Mobile, Standalone).
The figure below is the overall structure of FedNLP. 
Installation
After git clone-ing this repository, please run the following command to install our dependencies.
conda create -n fednlp python=3.7
conda activate fednlp
# conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.1 -c pytorch -n fednlp
pip install torch==1.6.0+cu101 torchvision==0.7.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt
cd FedML; git submodule init; git submodule update; cd ../;
Code Structure of FedNLP
-
FedML: a soft repository link generated usinggit submodule add https://github.com/FedML-AI/FedML. -
data: provide data downloading scripts and raw data loader to process original data and generate h5py files. Besides,data/advanced_partitionoffers some practical partition functions to split data for each client.
Note that in FedML/data, there also exists datasets for research, but these datasets are used for evaluating federated optimizers (e.g., FedAvg) and platforms. FedNLP supports more advanced datasets and models.
-
data_preprocessing: preprocessors, examples and utility functions for each task formulation. -
data_manager: data manager is responsible for loading dataset and partition data from h5py files and driving preprocessor to transform data to features. -
model: advanced NLP models. You can define your own models in this folder. -
trainer: please define your owntrainer.pyby inheriting the base class inFedML/fedml-core/trainer/fedavg_trainer.py. Some tasks can share the same trainer. -
experiments/distributed:experimentsis the entry point for training. It contains experiments in different platforms. We start fromdistributed.- Every experiment integrates FIVE building blocks
FedML(federated optimizers),data_manager,data_preprocessing,model,trainer. - To develop new experiments, please refer the code at
experiments/distributed/transformer_exps/fedavg_main_tc.py.
-
experiments/centralized:- This is used to get the reference model accuracy for FL.
Data Preparation
In order to set up correct data to support federated learning, we provide some processed data files and partition files. Users can download them for further training conveniently.
If users want to set up their own dataset, they can refer the scripts under data/raw_data_loader. We already offer a bunch of examples, just follow one of them to prepare your owned data!
download our processed files from Amazon S3.
Dwnload files for each dataset using these two scripts data/download_data.sh and data/download_partition.sh.
We provide two files for each dataset: data files are saved in data_files, and partition files are in directory partiton_files. You need to put the downloaded data_files and partition_files in the data folder here. Simply put, we will have data/data_files/*_data.h5 and data/partition_files/*_partition.h5 in the end.
Experiments for Centralized Learning (Sanity Check)
Transformer-based models
First, please use this command to test the dependencies.
# Test the environment for the fed_transformers
python -m model.fed_transformers.test
Run Text Classification model with distilbert:
DATA_NAME=20news
CUDA_VISIBLE_DEVICES=1 python -m experiments.centralized.transformer_exps.main_tc \
--dataset ${DATA_NAME} \
--data_file ~/fednlp_data/data_files/${DATA_NAME}_data.h5 \
--partition_file ~/fednlp_data/partition_files/${DATA_NAME}_partition.h5 \
--partition_method niid_label_clients=100.0_alpha=5.0 \
--model_type distilbert \
--model_name distilbert-base-uncased \
--do_lower_case True \
--train_batch_size 32 \
--eval_batch_size 8 \
--max_seq_length 256 \
--learning_rate 5e-5 \
--epochs 20 \
--evaluate_during_training_steps 500 \
--output_dir /tmp/${DATA_NAME}_fed/ \
--n_gpu 1
Experiments for Federated Learning
We already summarize some scripts for running federated learning experiments. Once you finished the environment settings, you can refer and run these scripts including run_text_classification.sh, run_seq_tagging.sh and run_span_extraction.sh under experiments/distributed/transformer_exps.
Citation
Please cite our FedNLP and FedML paper if it helps your research. You can describe us in your paper like this: "We develop our experiments based on FedNLP [1] and FedML [2]".
![[IMPORTANT]Client Sampling Frozen](https://avatars.githubusercontent.com/u/45058866?v=4)
431 Dec 19, 2022
125 Dec 20, 2022
3 May 25, 2022
12.3k Dec 31, 2022
8.4k Dec 30, 2022
2.9k Jan 2, 2023
2.3k Jan 7, 2023
1.5k Dec 26, 2022
148 Dec 26, 2022
9 Jul 17, 2021
0 Feb 21, 2022
4 Jul 1, 2022