LOFAR Labelling

[mesarcik]

LOFAR Labelling

[mesarcik]

Description

• I have labelled up to sample 160. I think this is sufficient for testing.
• The purpose of this dataset is to determine whether training models on AOFlagger will result in "good" performance.

To do

• Re-read the hand-labelled examples and make a plot comparing my flags with the original data.
• Create datafile containing the test data that I have labelled
• Determine optimal AOFlagger thresholding strategy based on the our now labelled ground truth
• Train the UNET and AE using the AOflagged data and evaluate it on the ground truth.

Dataset structure

def generate_lofar(...):
    # loop through both the default and defaultanot directory and read in the masks as np arrays concatenating them on each iteration 
    # select the remaining 370 images as training data, but do not create any masks 
    # when we read in the dataset we need to generate the masks using the selected AOFlagger strategy 

[mesarcik]

Dataset problems

• tldr i need to start labelling from scratch.
• As albert-jan pointed out, I stupidly uploaded the dataset with compression, in effect I blurred out much of the detail in my spectrogram see below:

Compressed:

Uncompressed

plt.imshow

[mesarcik]

Change of datasource

• In further inspection of the adder spectra i have found they all are version 1.4, i.e. the have some scaling factor saved separately to the visibilities (im not sure exactly what downsampling scheme is used)
• However it seems that when performing the compression there are a number of overflows (particularly in RFI infested regions), this is irrespective of bit-depth shown below (first is uint8, float32, float64)

• I there think it is a better idea to use the LTA data because from as far as i have seen these problems do not exist

Fixed labeling

• I changed from CVAT to lableme because CVAT doesnt allow 1x1px labels
  

[mesarcik]

Considerations

• it seems to me that there is both some under and over flagging from AOFLagger
• this may have unforseen cosequences for our models
  
  
  

[mesarcik]

Own labels comparison

• AUROC of 0.7788088786036889 between our masks and the AOflagger masks