This repository contains Dockerfiles for running algorithms for the Fast Radio Burst detection benchmark on the Eyra Benchmark Platform and the code for the evaluation of these algorithms.
The algorithms are provided as Docker image and implement a specific interface.
Each Docker image will read data from the file /data/input/test_data (a filterbank file, containing injected FRBs - for details and examples, see the benchmark details), and should write detected FRB candidates to a the file /data/output. One line per candidate, with the following fields (separated with a space character): <DM> <S/N> <TIME> <DOWNSAMPLE> <FREQ_REF>.
These output files are then evaluated by the evaluation algorithm, which expects the algorithms output at /data/input/implementation_output, and the actual injected FRBs at /data/input/ground_truth. It will then output a JSON file containing totals (detected/missed/false positives) to /data/output.
Most of the algorithms in this repository require a GPU to run. This requires some additional Docker configuration. The algorithms have been tested on an AWS p3.2xlarge (with a single NVIDIA Tesla V100 GPU) machine, running Ubuntu 18.04 and prepared using the setup.sh file.
To build a the Docker image for an algorithm, clone this repository and run the following command for example:
docker build -t frb-heimdall:1 algorithms/heimdall.- To make the data available to the Docker container it's best to prepare a data directory (e.g.
/data), which will be 'mounted' into the container.mkdir /data && mkdir /data/input && cp <filterbank_file> /data/input/test_data. - Then run the container, while mounting the data folder:
docker run --rm -it --gpus all -v /data:/data frb-heimdall:1. After completion, output should be at/data/output. - Next you could run the evaluation:
docker build -t frb-evaluation:1 evaluation
mv /data/output /data/implementation_output
cp <ground_truth_file> /data/ground_truth
docker run --rm -it -v /data:/data frb-evaluation:1
Again, output should end up in /data/output.