search

UPennEoR / ml_rfi

Machine learning applied to RFI flagging
BSD 2-Clause "Simplified" License
8 stars 4 forks source link
machine-learning radio-astronomy

readme

ml_rfi

A Deep Fully Convolutional Neural Net implementation in Tensorflow as applied to RFI flagging in waterfall visibilities.

Dependencies:

Dependent software versions specified here work with the current iteration of ml_rfi, older or newer versions of these software packages may work but cannot be guaranteed.

numpy 1.14

scipy 1.0

h5py 2.7

tensorflow 1.8

sklearn 0.19

pyuvdata 1.2

DEEP ML RFI Scripts

run_analysis.sh - Bash script for running runDFCN.py

runDFCN.py - Neural net training and evaluation script which takes input model type and runs through batch training/evaluation.

pipelineDFCN.py - Script for strictly running for prediction in some type of pipeline, functionality is based on pyuvdata datasets but has some support for other types.

AmpModel.py - Tensorflow model for amplitude input DFCN.

AmpPhsModel.py - Tensorflow model for amplitude-phase input DFCN.

helper_functions.py - Functions used for loading datasets and defining tensorflow objects used in the models.

Training Datasets

Simulated Data

SimVisRFI_15_120_v3.h5 - Contains 1000 simulated waterfall visibilities over baseline lengths of 15m to 120m. RFI includes stations and more random events (timescales >= 30). Deprecated!!!

SimVisRFI_v2.h5 - Contains 1000 simulated waterfall visibilities for one baseline type. RFI includes stations and speckled RFI. Deprecated!!!

SimVisRFI_15_120_NoStations.h5 - Contains 300 simulated waterfall visibilities over baseline lengths of 15m to 120m. No RFI stations (e.g. ORBCOM), all RFI is randomly placed across all times & frequencies. Addittionaly includes randomized burst events where RFI is placed at all frequencies for a time sample. Deprecated!!!

SimVis_3000_v13.h5 - Amalgamation of several previous simulated training datasets, and the most recently used for training.

Real Data

Ground truth is NOT 100% certain for this dataset. It's based entirely on what XRFI perceives as RFI and is prone to it's biases.

RealVisRFI_v3.h5 - Contains 3584 real HERA (37-ish) waterfall visibilities all flagged using XRFI. !!! The majority of visibilities up to 900 look decent enough to train/evaluate on, however everything after this looks like it's been very poorly flagged by XRFI, i.e. missing ORBCOM !!!

IDR21TrainingData_Raw_vX.h5 - Small hand flagged dataset which should have a very well identified ground truth.

Training - Evaluation Strategies

Training is performed using entirely simulated data from HERAsim across a wide range of baseline types, mock sky scenarios, and RFI instances.

Evaluation is performed in two ways:

  1. Evaluate on 20% of unused training data to check for overfitting.
  2. Evaluate on real data. (e.g. IDR21TrainingData_Raw_vX.h5)

Current strategy:

Training dataset:

SimVis_3000_v13.h5

Evaluation dataset:

IDR21TrainingData_Raw_vX.h5