Code for the paper "Physically Parameterized Differentiable MUSIC for DoA Estimation with Uncalibrated Arrays" available at the following link "https://arxiv.org/pdf/2411.15144"
In order to configurate the simulation environment, a conda-env setup file is provided: requirements.txt. Simply run:
conda create --name venv --file requirements.txt
to setup the environment.
The project is separated into two folders:
- core contains the source code: training routines, model architectures...
- paper experiments contains the paper code:
- data_generation contains the data generation routine as well as the data folder
- nn_training contains the training scripts for each of the paper's experiments
- post_training_viz contains the trained models as well as the plot functions
Before running the experiments, one should create the datasets. This done by running sample_generation.py in paper_experiments/data_generation for the following values:
- N = 16, M = 5
-
eta = 1/2, sigma2_g = 0.36, snr = 30dB
- T = 10, 20, 30, 50, 100
-
eta = 1/2, sigma2_g = 0.36, T = 100
- snr = 0, 5, 10, 20, 30
-
When the models are re-trained, the saved models should be moved from the folder .pt_dir in nn_training to .post_training in post_training_viz.