Uploaded image for project: 'Data Management'
  1. Data Management
  2. DM-10612

Implement strict monotonicity

    Details

    • Type: Story
    • Status: Done
    • Resolution: Done
    • Fix Version/s: None
    • Component/s: meas_deblender
    • Labels:
      None

      Description

      The current deblender uses an update scheme similar to ADMM, where constraints are not strictly enforced in any given step but instead the solution converges to meet the constraint. It should be possible to make monotonicity a proximal operator applied to the likelihood gradient, as opposed to a linear constraint, which will force a strict monotonic solution.

      This ticket will be an attempt to add the option of using strict monotonicity in the deblender.

        Attachments

          Activity

          fred3m Fred Moolekamp created issue -
          fred3m Fred Moolekamp made changes -
          Field Original Value New Value
          Epic Link DM-10353 [ 32070 ]
          fred3m Fred Moolekamp made changes -
          Epic Link DM-10353 [ 32070 ] DM-8140 [ 27595 ]
          fred3m Fred Moolekamp made changes -
          Sprint DRP F17-1 [ 614 ] DRP S17-6 [ 365 ]
          fred3m Fred Moolekamp made changes -
          Status To Do [ 10001 ] In Progress [ 3 ]
          Hide
          fred3m Fred Moolekamp added a comment - - edited

          Strict monotonicity has been tested using simulated data and HSC data.

          We see that it gives much more accurate results (without even invoking a sparsity penalty) and in the case of the HSC data does a very good job matching the colors of the sources

          The overlapping bright objects still need to see significant improvement. Much of this is likely due to imprecise positions (which we know to cause issues due to an inexact symmetry operator) and undetected sources that cause other footprints to grow in order to contain the flux. Tickets DM-10611 and DM-10310 (planned for the next sprint) are likely to help.

          Show
          fred3m Fred Moolekamp added a comment - - edited Strict monotonicity has been tested using simulated data and HSC data . We see that it gives much more accurate results (without even invoking a sparsity penalty) and in the case of the HSC data does a very good job matching the colors of the sources The overlapping bright objects still need to see significant improvement. Much of this is likely due to imprecise positions (which we know to cause issues due to an inexact symmetry operator) and undetected sources that cause other footprints to grow in order to contain the flux. Tickets DM-10611 and DM-10310 (planned for the next sprint) are likely to help.
          fred3m Fred Moolekamp made changes -
          Reviewers Peter Melchior [ pmelchior ]
          Status In Progress [ 3 ] In Review [ 10004 ]
          Hide
          pmelchior Peter Melchior added a comment -

          Could you please update the link for the simulation notebook: they both currently point to the HSC example notebook.

          Show
          pmelchior Peter Melchior added a comment - Could you please update the link for the simulation notebook: they both currently point to the HSC example notebook.
          Hide
          fred3m Fred Moolekamp added a comment -

          Sorry about that, I updated my previous comment with the correct link to the simulated notebook.

          Show
          fred3m Fred Moolekamp added a comment - Sorry about that, I updated my previous comment with the correct link to the simulated notebook.
          Hide
          pmelchior Peter Melchior added a comment -

          strict monotonicity does indeed work as we originally wanted it. remaining issues that can be seen in both notebooks are related to undetected objects that get picked up with spike-like patterns. However, because of the new M constraint, the influence of those is already noticeably reduced.

          Show
          pmelchior Peter Melchior added a comment - strict monotonicity does indeed work as we originally wanted it. remaining issues that can be seen in both notebooks are related to undetected objects that get picked up with spike-like patterns. However, because of the new M constraint, the influence of those is already noticeably reduced.
          pmelchior Peter Melchior made changes -
          Status In Review [ 10004 ] Reviewed [ 10101 ]
          fred3m Fred Moolekamp made changes -
          Resolution Done [ 10000 ]
          Status Reviewed [ 10101 ] Done [ 10002 ]

            People

            • Assignee:
              fred3m Fred Moolekamp
              Reporter:
              fred3m Fred Moolekamp
              Reviewers:
              Peter Melchior
              Watchers:
              Fred Moolekamp, Peter Melchior
            • Votes:
              0 Vote for this issue
              Watchers:
              2 Start watching this issue

              Dates

              • Created:
                Updated:
                Resolved:

                Summary Panel