Exploring Federated Learning for Semantic Segmentation in Autonomous Driving and Satellite Images Scenarios

(Developed for the Machine Learning and Deep Learning 2023 course)

Politecnico di Torino

Read the paper

About

Datasets

The repository supports experiments on the following datasets:

1. Reduced Federated IDDA from FedDrive [2]
   • Task: semantic segmentation for autonomous driving
   • 24 users
2. Reduced GTA5
   • Task: semantic segmentation for autonomous driving
   • downloadGta.py available to download it.
3. Reduced LoveDA
   • Task: semantic segmentation for satellite/aerial imagery.
   • downloadLoveda.py available to download it.

NOTE: to use the scripts and download the files run before (if needed): pip install pymegatools

How to run

The main.py setup the whole application. All arguments need to be specified through the args parameter (options can be found in utils/args.py). Example of experiments:

Centralized mode:

• IDDA

python main.py --dataset idda --centr --model deeplabv3_mobilenetv2 --num_rounds 200 --num_epochs 20 --clients_per_round 8 

Distributed mode:

• GTA5

python main.py --dataset idda --model deeplabv3_mobilenetv2 --num_rounds 200 --num_epochs 2 --clients_per_round 8 

FDA mode (pretraining + distributed mode):

• GTA as source and IDDA as target

python main.py --dataset idda --fda --model deeplabv3_mobilenetv2 --num_rounds 200 --num_epochs 2 --clients_per_round 8

Reproducing results:

Some checkpoints are already available in the repo. In particular they are located in models/checkpoints.

To reproduce the result you can use the args --load_from (e.g.):

• GTA as source and IDDA as target (FDA + Pseudo)

python main.py --dataset idda --fda --model deeplabv3_mobilenetv2 --load_from "models/checkpoints/gta5_fda.pth" --num_rounds 200 --num_epochs 2 --clients_per_round 8

• Loveda (Pseudo)

python main.py --dataset loveda --fda --model deeplabv3_mobilenetv2 --load_from "models/checkpoints/loveda_nofda.pth" --num_rounds 200 --num_epochs 2 --clients_per_round 8

• GTA5 CENTR TRANSFORMER

python main.py --dataset gta5 --model segformer --transformer_model b1 --load_from "models/checkpoints/gta5_nofda.pth" --num_rounds 1 --num_epochs 1 --clients_per_round 1

Checkpoints available:

gta5_fda

Load the best checkpoint of the model trained on gta5 dataset + FDA transformation.

gta5_nofda

Load the best checkpoint of the model trained on gta5 dataset w/o FDA.

loveda_nofda.pth

Load the best checkpoint of the model trained loveda.

All the checkpoints are saved on a run of 100 epochs.

For other checkpoints (e.g. Segformer checkpoints): Drive Checkpoints