| | 8 | We've been working on getting ready for the demo month. I built new flat-field correction frames for the new data set, using the images from the previous flat-correction round. We started running the first set of the magic verification data on Monday, but quickly discovered that there were some mis-masked portions of the array. After a couple of iterations, we were able to run some of the verification data through to stack, diff, and magic. We were not happy with the stack and diff results in the first pass, so Paul and I took another look at possible sources of errors in the code. We discovered two important algorithm issues: 1) the normalization term, which was solved for as part of the kernel analysis, is coupled to the solutions. sometimes the analysis can get the right answer, but sometimes the normalization is driven to the wrong value and the resulting kernels are very poor. we solved this by carefully measuring and applying aperture magnitudes to determine the relative normalizations of the two images. 2) the penalty function in the dual kernel analysis (used only for diffs, not stacks) had a mismatched scale, so that the penalties (used to prevent the kernels from going to large values) were 10^8^ smaller than the terms they were penalizing (thus they had no real impact). By adjusting the choice of scaling factor, we solved this problem with the result that dual diffs now do not drive to large kernels. Paul also found an additional error that resulted in an error in the variance for those difference images with partially illuminated skycells. Our second pass on the magic verification suite turned up another (one line) bug that caused the masking to grow enormously. We fixed this error and are running our third attempt. Other minor points: I generated a new y-band fringe correction frame so that the y-band data can be processed for the magic verification test, and I generated a our first flat-corrected flat-field image for w-band (there were some minor configuration issues), so the w-band data may also be processed. |
