Deep Learning RFI mitigation code. This repository contains the deep learning method used in the paper Deep learning for RFI mitigation of Nançay data and its impact on pulsar timing and the trained weights.
The network architecture is a U-net using the mobileNetV2 as features extractor. The data used for trains, validations and tests are pulsar observations from the Nançay radio telescope in France.
Download/close this repository, you can find in the table below a small description of what you can find in any file.
| filename | functions |
|---|---|
| callbacks | callback to print the epoch at the end of each one |
| generate_rfi | create dynamic spectrum with rfi structures |
| metrics | modified tensorflow metric for mean IOU |
| psrlog | data generator to feed the network |
| segnet | build the network and return model |
| training | training script |
| utils | clean observations, prepare data, plots results |
You can load the already trained model by using this line :
model = keras.models.load_model("./saved_net", custom_objects={'ConfusionMatrixMetric':ConfusionMatrixMetric(2)})Disclaimer : this model has been trained with Nançay observations and may give you different results. It is recommended to train again the network with your own data. To do so, please refer to the data processing section.
To use your own data with the network, you will need to either use psrchive with python3 or prepreprocess your observations. The network use dynamic spectrum with 3 channels : median(bins), median_absolute_deviation(bins), peak_to_peak(bins). These transformations are made by the datagerenator but it will require that you have observations saved in a numpy array format with shape (nsub,nchan,nbins).