Neural-enhanced Demodulation

This module provides the codes for our neural-enhanced demodulation, including the preprocessing for symbol transformation and DNN-based demodulation.

Overview

• Neural-enhanced Demodulation
• Overview
• Getting Started
  • Dependencies
  • Load Dataset
• Run Experiments and Validate Results
  • From the Scratch
  • Direct Inference
  • Evaluation

Getting Started

Dependencies

• scipy==1.0.0
• numpy==1.16.4
• torchvision==0.3.0
• PyTorch==1.7.0

Load Dataset

1. Clone this repo: git clone --recursive https://github.com/hanqingguo/NELoRa-Sensys.git

2. Download the Raw Dataset from here

3. Pre-process data with matlab scripts. (take SF=7, bandwidth=125k) as example:

   • specify LORA_SF, LORA_BW in matlab/param_configs.m
   • run matlab/generation_dataset.m, specify raw data root and generated data path. (Long Process && Large File Warning!!)
     • This script adds random Gaussian noises to original collected chirp data.
     • The file name of the generated chirp symbol follows rules as below:
     • In default, the generated dataset will be saved in /data/Lora/sf7_125k as following name:

   {Code} _ {SNR} _ {SF} _ {BW} _ {batch_index} _ {Code Label}_ {packet_index}_ {symbol_index}.mat

Run Experiments and Validate Results

From the Scratch

To train a model from scratch, we recommand you have a NVIDIA video card, and support cuda acceleration.

1. Specify the root path, data path, and evaluation_dir in config.py

 ├ [DIR] root_path (.)
     ┬    
     ├ [DIR] evaluations_dir
        |-[DIR] [dir_comment]\_checkpoints
        |-[DIR] [dir_comment]\_samples
        |-[DIR] [dir_comment]\_testing
        |- e.g. : sf7_v1_checkpoints
 ├ [DIR] data_dir (/data/Lora/sf7_125k)
     |- This is same as the output of the matlab script

2. run the following command:

python main.py --dir_comment sf7_v1 --batch_size 16 --root_path . --data_dir /data/Lora/sf7_125k --groundtruth_code 35 --normalization --train_iter 100000 --ratio_bt_train_and_test 0.8 --network end2end

3. Check your loss with the std print. e.g.:

   • Iteration [ 1000/100000] | G_Y_loss: 5.5639| G_Image_loss: 2.6935| G_Class_loss: 2.8704
   • G_Y_loss: G_Image_loss + G_Class_loss
   • G_Image_loss: The loss between groundtruth chirp and your denoised chirp.
   • G_Class_loss: The loss of decoding correct symbol compares with the {Code Label}

4. Check your samples in [evaluations_dir]:

Example, a chirp code 24 under -23 dB noise.

Raw Chirp	GroundTruth	Denoised

5. Get a decode result with in your pytorch/. Named as: [dir_comment]_[sf]_[bw].mat (e.g., sf7_v1_7_125000.mat)

Direct Inference

1. Download the pretrained models from here
2. run the following command:

python main.py --dir_comment sf7_125k --batch_size 16 --root_path . --data_dir /data/Lora/sf7_125k --groundtruth_code 35 --normalization --train_iter 0 --ratio_bt_train_and_test 0.8 --network end2end --load yes --load_iters 100000

3. Get a decode result with in your pytorch/ directory.

Evaluation

1. Preparation:

• Get baseline performance: You can either download it from here, or run a baseline decode locally with our provided script: matlab/generate_baseline.m. Note that the result of abs accumulation decode method and phase compensation method reach similar decode accuracy. You can validate it with changing the abs_decode in our generate_baseline.m script.
• Get NELoRa performance:
• Copy your decode result from pytorch/*.mat to matlab/evaluation/

2. Run matlab/evaluation.m
3. You will get the Symbol Error Rate (SER)result: