search

devanshkv / fetch

A set of deep learning models for FRB/RFI binary classification.
GNU General Public License v3.0
42 stars 32 forks source link

How to make the candidates file? #5

Closed parulj3795 closed 5 years ago

parulj3795 commented 5 years ago

Hi. I am following your paper 'Towards deeper neural networks for Fast Radio Burst detection'. Could you please direct on how to make the candidate file? The information in README.md is not very clear to me.

devanshkv commented 5 years ago

To create a candidate.csv file, you would need

  1. The path filterbank file
  2. Signal to noise ratio of the candidate
  3. Start time (all single pulse search pipeline lines like heimdall or singlepulsesearch.py Give the start time of the candidate in the filterbank file). In case of heimdall its the third column.
  4. DM of the candidate
  5. Box car width, this is samples as a power of two. So if you boxcar width is 8 samples, this number is int(log2(8)) = 3.
  6. Label is not required for prediction and can be ignored. If you are planning to train the classifiers then the label is 0 for RFI and 1 for FRBs.
  7. Kill mask. This is a simple text file with channel numbers in the filterbank which you would like to mask. This can also be ignored if there are no bad channels.

Stitch this into a .csv file which looks like: /path/to/filterbank/myfilterbank.fil,S/N,start_time,dm,boxcar_width

Hope this helps, let me know if you have more questions. If you have a specific candidate file from a pipeline for which you want help, please go ahead and paste it below.

parulj3795 commented 5 years ago

In order to train my network, I need to have a training data set. Can I create fake FRB candidates using candmaker.py ?

vandana55 commented 5 years ago

Hi Devansh Could you please elaborate how many total features you have considered while generating candidate file? following are few features

  1. Signal to noise ratio of the candidate
  2. DM of the candidate
  3. Box car width is there any other feature considered while designing candidate file.
vandana55 commented 5 years ago

With respect to above query.. I think these are the Parameters / features used... List of candidate parameters (fil_name, snr, width, dm, label, tcand(s))

devanshkv commented 5 years ago

@parulj3795 candmaker.py cannot make fake FRBs. You can try using burst_search or simpulse.

fetch is reasonably generic and you may not have to retrain it. Do you see it often fail in general or for some particular kind of examples?

devanshkv commented 5 years ago

@vandana55 do you mean the candidate.csv for training FETCH? For that see Table 2 and Figure 4 in our paper.