fetch is Fast Extragalactic Transient Candidate Hunter. It has been detailed in the paper Towards deeper neural networks for Fast Radio Burst detection
We suggest using anaconda for using FETCH.
First we need to install cudatoolkit matching the installed cuda version.
For cuda 8.0 conda install -c anaconda cudatoolkit==8.0 tensorflow-gpu==1.4.1
For cuda 9.0 conda install -c anaconda cudatoolkit==9.0 tensorflow-gpu==1.12.0
For cuda 9.2 conda install -c anaconda cudatoolkit==9.2 tensorflow-gpu==1.12.0
For cuda 10. conda install -c anaconda cudatoolkit==10.0.130 tensorflow-gpu==1.13.1
Note: tensorflow installation from conda-forge channel does not work with GPUs.
You would also require pysigproc to create the candidate files which can be found here.
Now we can install fetch like this:
conda install -c anaconda keras scikit-learn pandas scipy numpy matplotlib scikit-image tqdm numba pyyaml=3.13
git clone https://github.com/devanshkv/fetch.git
cd fetch
python setup.py install
The installation will put predict.py,candmaker.py and train.py in your PYTHONPATH.
First create a candidate file (cands.csv) of the following format:
/path/to/filterbank/myfilterbank.fil,S/N,start_time,dm,boxcar_width,label,path_to_kill_mask
here boxcar_width is in units of int(log2(number of samples)). path_to_kill_mask is a numpy readable file with channel numbers to kill. If not required, this field can be left empty.
Next, to generate the candidate files containing DM-time and Frequency-time arrays for classification use candmaker.py. Saving candidate h5s with their parameters in cands.csv to a directory /data/canddidates/ and rebinning the time and frequency axis to 256 bins using decimation can be done by:
candmaker.py --frequency_size 256 --time_size 256 --cand_param_file cands.csv --plot --fout /data/candidates/
To predict a these candidate h5 files living in the directory /data/candidates/ use predict.py for model a as follows:
predict.py --data_dir /data/candidates/ --model a
To fine-tune the model a, with a bunch of candidates, put them in a pandas readable csv, candidate.csv with headers 'h5' and 'label'. Use
train.py --data_csv candidates.csv --model a --output_path ./
This would train the model a and save the training log, and model weights in the output path.
Test filterbank data can be downloaded from here. The folder contains three filterbanks: 28.fil 29.fil 34.fil. Heimdall results for each of the files are as follows:
for 28.fil
16.8128 1602 2.02888 1 127 475.284 22 1601 1604
for 29.fil
18.6647 1602 2.02888 1 127 475.284 16 1601 1604
for 34.fil
13.9271 1602 2.02888 1 127 475.284 12 1602 1604
The cand.csv would look like the following:
28.fil,16.8128,2.02888,475.284,1
29.fil,18.6647,2.02888,475.284,1
34.fil,13.9271,2.02888,475.284,1
Running candmaker.py will create three files:
cand_tstart_58682.620316710374_tcand_2.0288800_dm_475.28400_snr_13.92710.h5
cand_tstart_58682.620316710374_tcand_2.0288800_dm_475.28400_snr_16.81280.h5
cand_tstart_58682.620316710374_tcand_2.0288800_dm_475.28400_snr_18.66470.h5
Running predict.py with model a will give results_a.csv:
,candidate,probability,label
0,cand_tstart_58682.620316710374_tcand_2.0288800_dm_475.28400_snr_18.66470.h5,1.0,1.0
1,cand_tstart_58682.620316710374_tcand_2.0288800_dm_475.28400_snr_16.81280.h5,1.0,1.0
2,cand_tstart_58682.620316710374_tcand_2.0288800_dm_475.28400_snr_13.92710.h5,1.0,1.0