Towards Real World Federated Learning

Machine Learning and Deep Learning 2023

Politecnico di Torino

Code for the Federated Learning project.

Setup

Environment

The project is designed to run on google colab, to run it be sure to follow this steps:

1. Install required packages:

   pip install wandb
   pip install overrides
   

2. Add Idda dataset in data/idda
3. Add GTA dataset in data/gta
4. If wandb must be used to save metrics and models, login using the command wandb login (otherwise add the argument --not_use_wandb in run.py)

Datasets

The repository supports experiments on the following datasets:

1. Reduced Federated IDDA from FedDrive [1]
   • Task: semantic segmentation for autonomous driving
   • 24 users
2. Reduced GTA
   • Task: semantic segmentation for autonomous driving
   • 20 classes

How to run

The run.py orchestrates everything. All arguments need to be specified through the args parameter (options can be found in config/args.py). Example of experiment:

python run.py
  --project exam-project 
  --exp_name big-round 
  --seed 0 
  --training_ds idda
  --test_ds idda
  --model deeplabv3_mobilenetv2 
  --num_epochs 5 
  --bs 4 
  --optimizer SGD 
  --lr 0.1 
  --weight_decay 0 
  --momentum 0.9 
  --lr_policy poly 
  --lr_power 0.9 
  --lr_decay_step 15 
  --lr_decay_factor 0.1 
  --rrc_transform 
  --min_scale 0.5 
  --max_scale 2.0 
  --h_resize 756 
  --w_resize 1344 
  --norm eros_norm 
  --jitter 
  --phase all
  --not_use_wandb 
  federated 
  --num_rounds 500 
  --clients_per_round 5

How to test best ClAvBN and SiloBN model

• ClAvBN:
  1. follow step 2 and 3 from the Environment section
  2. import the file clavbn.sh in Google Colab
  3. run the command !chmod +x clavbn.sh in a code cell
  4. run the command !./clavbn.sh
• SiloBN:
  1. follow step 2 and 3 from the Environment section
  2. import the file silobn.sh in Google Colab
  3. run the command !chmod +x silobn.sh in a code cell
  4. run the command !./silobn.sh

References

[1] Fantauzzo, Lidia, et al. "FedDrive: generalizing federated learning to semantic segmentation in autonomous driving." 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022.