gongzir1/yolo_fl

Federated Learning: Object detection with YOLOv5

Repository of the conceptual work around Federated Learning to train the object detection model YOLOv5.

Clone repository

git clone -b develop https://github.com/gongzir1/yolo_fl.git fedlearn
cd fedlearn
pip3 install -r requirements.txt

Repository folder structure:

• .vscode: vscode configuration.
• data: datasets and configuration files about the datasets.
• models: implemented models and configuration files about the models.
• src: source code.
  • client: implements the logic of the client in FL
  • server: implements the logic of the server in FL
  • strategy: implements the strategies that guide the server and the parameters aggregation
  • utils: custom helper functions
• data_splitter.py: Splits a folder of images that follow YOLOv5 guidelines into partitions to try a federated approach to train a model
• yolo_train_solo.py: trains yolo in a centralized way
• runner_client.py: sets up a client to train the net
• runner_server.py: sets up the server to handle the clients
• utils: yolo helper functions.
• detect.py: used to try the model in the webcam or images
• train.py: implements the train pipeline of the model. Adapted to be used through the clients
• val.py: implements the validation of the model. Adapted to be used through the server.

SetUp

Set up a Federated Learning process!

- server_address: ip of the server
- rounds: number of federated rounds
- loads_params: controls if the server provides the initial parameters or it gets them from a random client (`--no_load_params`)

Default useful parameters:

- weights: initial weights path. `default=ROOT / "yolov5s.pt"`
- cfg: path to the specifications about the structure of the model to be trained. default="models/yolov5s.yaml"
- hyp: path to the specifications about the hyperparameters of the model. default=ROOT / "data/camacuc2.yaml"
- data: path to the specifications about the data to be trained. default=ROOT / "data/hyps/hyp.scratch.yaml"

Control the number of clients that participate in a Federated training

- min_num_clients: Minimum number of available clients required for sampling (default: 2)
- min_sample_size: Minimum number of clients used for fit/evaluate (default: 2)
- sample_fraction: Fraction of available clients used for fit/evaluate (default: 1.0

- server_address: ip of the server. if executed in the same machine keep default. If not, check ip address of the server: XXX.XX.X.XX:8080
- cid: client id
- epochs: number of epochs per round

Default useful parameters:

- weights: initial weights path. `default=ROOT / "yolov5s.pt"`
- cfg: path to the specifications about the structure of the model to be trained. default="models/yolov5s.yaml"
- hyp: path to the specifications about the hyperparameters of the model. default=ROOT / "data/camacuc2.yaml"
- data: path to the specifications about the data to be trained. default=ROOT / "data/hyps/hyp.scratch.yaml"

Memory Handling:

In order to reduce the amount of memory in RAM and specially for embedded devices, lowering these parameters may help

- img: train, val image size (pixels)
- batch-size: number of images to pass through each batch of an epoch

Federated Object Detection on the Edge

This repository is able to run on edge devices such as Raspberry Pi 4 or Pi 3B.

In order to flash the raspberries follow these tutorials:

• Install 64 bit OS (Tested Bullseye): https://qengineering.eu/install-raspberry-64-os.html
• Install Pytorch and Pytorchvision wheels: https://qengineering.eu/install-pytorch-on-raspberry-pi-4.html

Then clone the repository and install requirements with pip3.

It is very important to handle RAM memory.

Run YOLOv5 centralized

You can run YOLOv5 as the original repository:

$ python yolo_train_solo.py --data data/camacuc2.yaml --cfg yolov5s.yaml --weights '' --batch-size 64
                                                            yolov5m                                40
                                                                                                   24
                                                                                                   16

Running code:

Start the server first:

$ python3 runner_server.py --server_address [::]:8080 --rounds 30 --load_params --min_sample_size 2 --min_num_clients 2 --device cuda:7 --data data/voc1.yaml

Clients: There are two ways to start the client, first is to start clients one by one:

python3 runner_client.py --server_address [::]:8080 --cid 0 --epochs 10 --data data/meat.yaml

You can also determine the number of clients first and run 'sh file' to start them all automatically:

sh ./run_client.sh 10

How to divide the dataset:

1. rename images and labels name based on their categorizes: e.g. rename_images(topside, topside, "Topside_cap_off", "jpg")
2. move all the images and labels into one dir e.g. move_files_to_directory(data/og_datasets,data/combine-data)
3. split train and val dataset: e.g. split_dataset_val_train(path='E:\DP\yolov5FL1\images\', img_ext="jpg", train_split=0.8)

Reference

This code is based on the original work from the GitHub repository: https://github.com/JQ-MC/object-detection-fedlearn.git