This repo guide you trhought the process of replicating the results from the Ray-Tracing MIMO Channel Dataset for Machine Learning Applied to V2V Communication, avaliable here: https://ieeexplore.ieee.org/abstract/document/10000783/
First, you'll need to download or generate the dataset to work with, you can either download the baseline dataset from the Raymobtime site or download the raw data and preprocess it. If you wish to use the baseline data, just go to Step 2, but if you wish to preprocess your data, and change it, just keep reading.
First, download the raw data from Raymobtime site, that will be used to preprocess your data. Second, you'll need to generate your input data, for that, you'll need the CoordVehiclesRxPerScene.csv file and run the following code:
python3 preprocessing/process_coord_matrix_input.py
Feel free to change variables in the code to reach different outcomes, such as: analysis_area or analysis_area_resolution. By the end of the code you'll have an coord_matrix_input.npz that it'll be your input dataset. Then, you'll need to generate the output dataset, for that, download the ray_tracing_data_v002_carrier60GHz.zip and run the following code:
python3 preprocessing/process_beam_output.py -e 2500 -p path/to/ray_tracing_data_v002_carrier60GHz.zip -c cfg/training_pipeline.json
With the flag -e being the number of episodes, and -p being the path to the ray_tracing_data_v002_carrier60GHz.zip folder (ex: /home/user/data/ray_tracing_data_v002_carrier60GHz). Feel free to change parameters if you want acheive different results, such as nTx, nRx (but remember to change in both json files). By the end of it, you'll have your output dataset beam_output.npz.
With the dataset, you can run the Deep Neural Network, but first, you'll need onw first step to preprocess the data for the DLL, so first, run the following code:
python3 src/datastructure.py -c cfg/data_structure.json
Remember to change input_data and output_data to indicate the path of your dataset. Feel free to play around and change parameters from the json, such as: nTx, nRx that change the number of the antennas in each veicle, splits that are the percentages of split for train, validation and test datasets.This will generate the structure of folders for the dataset. Then run the following code:
python3 src/datagenerator.py -c cfg/training_pipeline.json
With that, the preprocessing is finished and you can run the DNN. To train simply run:
python3 training_pipeline.py -c cfg/training_pipeline.json
Feel free to change the parameters of the Optmizers from the training_pipeline.json. Then you can test your DNN, run:
python3 testing_pipeline.py -c cfg/training_pipeline.json