SRToANet

Codes for "Super-Resolution ToA estimation using Neural Networks", EUSIPCO 2020

Data Preperation

All the information of channels (time delays, complex attenuation) are stored in data/Pathset_train.mat, data/Pathset_test.mat, and data/Pathset_test_802.mat

Simply run data/CIR_Generation.m will generate the dataset for training and testing, stored in data/traindata and data/testdata respectively

Interpolation module

Require installation of Torchinterp1d. See https://github.com/aliutkus/torchinterp1d

Train the model

Run train.py to train the model. It will first train the super-resolution network and then the two regressors

usage: train.py [-h] [--name NAME] [--snr {high,low}] [--bandwidth BANDWIDTH]
            [--e_sr E_SR] [--lr_sr LR_SR] [--batch_sr BATCH_SR]
            [--lam LAM] [--up UP] [--e_reg E_REG] [--lr_reg LR_REG]
            [--batch_reg BATCH_REG] [--use_ori [USE_ORI]]
            [--window_len WINDOW_LEN] [--window_index WINDOW_INDEX]
            [--device {cpu,gpu}] [--print_interval PRINT_INTERVAL]

optional arguments:
  -h, --help            show this help message and exit
  --name NAME           name for the experiment folder
  --snr {high,low}      select high SNR scenario or low SNR scenario
  --bandwidth BANDWIDTH
                    define the system bandwidth
  --e_sr E_SR           number of epochs to train the super resolution network
  --lr_sr LR_SR         learning rate for the super resolution network
  --batch_sr BATCH_SR   batchsize for the super resolution network
  --lam LAM             weight for the time domain loss
  --up UP               upsamping rate for the super resolution net
  --e_reg E_REG         number of epochs to train the super resolution network
  --lr_reg LR_REG       learning rate for the super resolution network
  --batch_reg BATCH_REG
                    batchsize for the super resolution network
  --use_ori [USE_ORI]   whether to include the original observation for the
                    regressors
  --window_len WINDOW_LEN
                    window size for the second regressor
  --window_index WINDOW_INDEX
                    number of samples for the window for the second
                    regressor
  --device {cpu,gpu}    choose the device to run the code
  --print_interval PRINT_INTERVAL
                    number of iterations between each loss print

Test

Run test.py to test the model. You can test the customized channel model and the 802.15.4a channel model.

usage: test.py [-h] [--name NAME] [--snr SNR] [--bandwidth BANDWIDTH]
           [--up UP] [--use_ori [USE_ORI]] [--use_802 [USE_802]]
           [--window_len WINDOW_LEN] [--window_index WINDOW_INDEX]
           [--device {cpu,gpu}] [--num_test NUM_TEST]

optional arguments:
  -h, --help            show this help message and exit
  --name NAME           name for the experiment folder
  --snr SNR             select high SNR scenario or low SNR scenario
  --bandwidth BANDWIDTH
                    define the system bandwidth
  --up UP               upsamping rate for the super resolution net
  --use_ori [USE_ORI]   whether to include the original observation for the
                    regressors
  --use_802 [USE_802]   whether to test the 802.15.4a channel
  --window_len WINDOW_LEN
                    window size for the second regressor
  --window_index WINDOW_INDEX
                    number of samples for the window for the second
                    regressor
  --device {cpu,gpu}    choose the device to run the code
  --num_test NUM_TEST   number of test cirs

Example figures

CIR comparison

ToA estimation

Reference

Yao-shan Hsiao*, Mingyu Yang*, Hun-Seok Kim, "Super-Resolution Time-of-Arrival Estimation using Neural Networks", EUSIPCO 2020