The Dual Path Network (DPN) is a neural network architecture for blind symbol decoding and modulation classification. This repository includes the code that accompanies the paper [1]. It includes the code for DPN and the output post processing along with the code for data generation.
The training data was generated in realtime during training as described in [1]. The code for the generator is in this repo. A google drive link to the validation and test sets is available. The weights of the trained network used in the paper are also provided.
There is a known bug in the SNR values and a workaround is provided (See the IMPORTANT WARNING).
[1] S. Hanna, C. Dick, and D. Cabric, “Signal Processing Based Deep Learning for Blind Symbol Decoding and Modulation Classification,” arXiv:2106.10543 [cs, eess], Jun. 2021, Accessed: Jun. 21, 2021. [Online]. Available: http://arxiv.org/abs/2106.10543
-
The python packages used with this code are available in requirements.txt (exported from conda).
-
DPN code uses CuDNNGRU. Hence, an NVidia GPU is required to run the code.
-
The validation and test datasets provided in the link are about 10 GB each. A server with RAM >=32 GB is needed to load them into memory.
Note that these requirements are to replicate the authors' setup. The code might work with other versions of the packages.
If no GPU is available the GuDNNGRU layer can be replaced by a regular GRU in frm_nn_zoo_01.py (However, you might need some workaround to load the weights).
A smaller version of the validation and test dataset can be generated using dataset_creator.ipynb
Also note that data generation is run using multiprocessing with 10 workers. If you have fewer than 10 cores in your setup you might want to reduce this number in the fit_generator function in 001_d1_train.ipynb
There is a known bug in the code in the signal generation. A square root is missing in the generation of a noise. As a consequence, the signals generated have twice the required SNR in dB.
For example, when the input value of the SNR in the generator is 10dB, the true SNR of the generated signal is 20dB.
The workaround is to provide the input SNR value as half of the required value. For example, if you want a signal with a 10dB SNR, provide an input value of 5dB.
The bug is in line 30 in frm_dataset_creator.py. However, the datasets and results were generated before the bug was discovered and it was not fixed for backward compatibility.
001_d1_train.ipynb: DPN training
003_d1_demod_dsp.ipynb: Decode data using genie algorithm from [1]
004_d1_baseline_nets.ipynb: Train the SGRU network, which is used as a baseline
005_d1_demod_dpn.ipynb: Demodulate the signals using DPN output and store modulation classification (MC) output
006_d1_compare_demod.ipynb: Compare demodulation results between Genie and DPN
008_d1_compare_params.ipynb: Evaluate frequency and timing offsets
009_d1_pred_baseline.ipynb: Generate MC predictions for GRU
010_d1_mod_class.ipynb: Compare MC results for DPN and GRU
013_d1_demod_sample.ipynb: Plot a signal from the dataset
020_d1_confusion_matrix.ipynb: Plot the confusion matrix
dataset_creator.ipynb: Code to generate a dataset
frm_nn_zoo_01.py: The code for DPN
frm_dataset_creator.py: Code for data generation
frm_dataset_creator2.py: Optimized code for data generation
frm_modulations.py: Generating signals from different modulations
frm_modulations_fast.py: Optimized code for modulations
frm_train_generator.py: A keras generator for realtime sample generation
frm_dataset_loader.py: Code for reading the dataset from disk
frm_demod_utils.py: Functions used for demodulation
frm_eval_utils.py: Functions used in the evaulation
frm_nn_baseline.py: Neural network code for SGRU
frm_nn_functions.py: Keras functions used by DPN
conf_dataset_1.py: Configuration file for the datset used in [1]
datasets: dataset folder (contains google drive link)
models: The weights for trained models
outputs: Temporary outputs provided
html: HTML version of all jupyter notebooks for convenience
py: Python version of all jupyter notebooks for convenience
tmp: Temporary folder to store the weights
requirements.txt: List of python packages (with version numbers) used with this code. Exported from conda according to these instructions
Readme.md: This file