DapperFL source + Final project code for Distributed Deep Learning Course 25

Scripts we have implmented for final project:

To keep our final project implementation code separate from the original source code, we have implemented all our logic in separate python scripts.

Main program - fedml_experiments/standalone/domain_generalization/main_text.py
Dataset - sorted_data_acl
Backbone - fedml_api/standalone/domain_generalization/backbone/LSTM.py
Dataset Handler - fedml_api/standalone/domain_generalization/datasets/amazon_product_reviews.py
Federated Model - fedml_api/standalone/domain_generalization/models/utils/federated_model_text.py
Dapper FL Text - fedml_api/standalone/domain_generalization/models/dapperfl_text.py
Training - fedml_api/standalone/domain_generalization/utils/training_text.py

To install requirements:

pip install -r requirements.txt

UBELIX specific instructions:

module load OpenMPI/4.1.1-GCC-11.2.0

export OMPI_MCA_pml=^ucx

export PYTHONPATH=/path/to/the/repo/Spring25-DDL-DapperFL:$PYTHONPATH

We use wandb to keep a log of our experiments. If you don't have a wandb account, just install it and use it as offline mode.

pip install wandb
wandb off

To train the model(s) in the paper, run this command:

python main.py --model dapperfl --dataset fl_officecaltech --backbone resnet18

To train our text model version, run this command:

python main_text.py --model dapperfl_text --dataset fl_amazonreviews --backbone lstm

You can modify the arguments to run DapperFL on other settings. The arguments are described as follows:

Arguments	Description
`prefix`	A prefix for logging.
`communication_epoch`	Total communication rounds of Federated Learning.
`local_epoch`	Local epochs for local model updating.
`parti_num`	Number of participants.
`model`	Name of FL framework.
`dataset`	Datasets used in the experiment. Options: `fl_officecaltech`, `fl_digits`.
`pr_strategy`	Pruning ratio used to prune local models. Options: `0` (without pruning), `0.1` ~ `0.9`, `AD` (adaptive pruning).
`backbone`	Backbone global model. Options: `resnet10`, `resnet18`.
`alpha`	Coefficient alpha in co-pruning. Default: `0.9`.
`alpha_min`	Coefficient alpha_min in co-pruning. Default: `0.1`.
`epsilon`	Coefficient epsilon in co-pruning. Default: `0.2`.
`reg_coeff`	Coefficient for L2 regularization. Default: `0.01`.
`seed`	Random seed.