PerSOTA: A Robust-To-Noise Personalized Over The Air Federated Learning for Human Activity Recognition
This is the official code for the submitted paper PerSOTA FL: A ROBUST-TO-NOISE PERSONALIZED OVER THE AIR FEDERATED LEARNING FOR HUMAN ACTIVITY RECOGNITION. For now, the code only works on HARBAX dataset. we will keep updating the code in the upcoming days.
Update: the proposed PerSOTA FL can now be tested on both in-distribution test data (saved as local test accuracy) and out-of-distribution test data (saved as general test accuracy).
Please install the following packages before running main.py .
!pip install tensorflow==2.10
First, download the HARBOX dataset from Here. Click on the dropbox list and download Large_Scale_HARBox.zip.
Second, go to \utils\dataset.py and find the function load_HARboX_data_v2() and modify read_path = r'YOUR_PATH\large_scale_HARBox/'+str(client_id)+'/'+class_id+'.txt' to address the dataset properly.
Finally, run the following code that specifies the report path, number of local updates, number of rounds, the SNR, The number of receivers and the personalization parameter.
python main.py --save_path 'report' --E 3 --R 120 --SNR 5 --N_r 100 --alpha 0.5
Or, PerSOTA_FL_notebook in google colab.
The results are with the following setting python main.py --save_path 'report' --E 3 --R 120 --SNR 5 --N_r 100 --alpha 0.8
Pending