PMNet with SionnaRT

Overview

• This repository provides detailed instructions on the training of PMNet—an NN tailored for path loss map prediction.

• The PMNet are trained with site-specific channel measurement data obtained through Sionna RT.

• This training includes two main steps:

  1. Generating path loss map dataset by utilizing Sionna RT;
  2. Training PMNet with the pathloss map dataset to predict path loss map.

• Example: Ground Truth (SionnaRT measurement) vs. Predicted (PMNet prediction)

  
  

Dataset: SionnaRT-based Pathloss Map (USC Campus Area)

• A 3D map of the USC campus, created with Blender OSM, was utilized. These models were then exported to create scenes in Sionna RT.

   

Pre-Processing

1. Configuration (TX/RX/Channel/etc.):
   • Details will be updated...
2. Map Generation:
   • For each scene at a specific TX location, three types of maps are generated:
     1. Pathloss Maps: These are grayscale images that visualize pathloss (or pathgain) across regions of interest (RoI).
        • Gray conversion: $-255 \sim 0$ [dBm] pathgain $\rightarrow$ $55 \sim 255$ grayscale
     2. City Maps: These are grayscale images showing RoI and buildings.
        • Grayscale mapping: $0 \sim 54$ [m] height $\rightarrow$ $0 \sim 54$ grayscale and $255$ (White) gray value represent building and ROI area, respectively.
     3. TX Maps: These are grayscale images indicating the TX locations, which is highlited with $255$ (White) gray value.
3. Cropping:
   • Images cropped into 256x256 pixels, ensuring inclusion of TX point and are further augmented.
   • A total of $9164$ cropped images are produced for the USC campus map dataset.

"How to Prepare (Pre-Process) Pathloss Map Data?"

• To pre-process the pathloss map data, simply run the following script. Please replace [START] and [END] with the TX points you want to start and end data mining with. A bigger range will require a lot of memory. A good estimate to have is a range of 5. In order to mine data for all $158$ TX, you can run the file updating the [START] and [END] arguments. Additionally, the total number of TX points are $175$, out of which $17$ give NaN values in their coverage map which are automatically discarded.

  python data/preprocess.py [START] [END]
  

Model: PMNet

• To train the PMNet model, we use stacked cropped City and TX maps from the data/cropped folder as input to predict the Pathloss map as the output.

"How to Train?"

• To train PMNet, simply run the train.py.

  python train.py
  

"How to Evaluate?"

• To evaluate a PMNet, refer to the following commands. Please update the path to model for evaluation. Similarly make sure the data is already present in the data/cropped folder else follow the above section to prepare the data.

  python eval.py \
      --model_to_eval '[PATH_TO_MODEL]' 
  

Download: Dataset & Checkpoint

• Dataset:

  • Full Dataset: .zip
  • Uncropped: images
  • Cropped: images

• Checkpoint (Pre-trained PMNet):

  • PMNet (RMSE: 0.00098): ckpt