The July Offline Leads ‘meeting’ was handled via email and a google document. We heard from over half the experiments, not bad for a July. Next meeting is live.
LArSoft Report – Erica Snider
Floating Point Exceptions policy decision. After a discussion at the 6/27/17 LArSoft Coordination meeting led by Herb Greenlee, we concluded we don’t want to ignore FPEs. Thomas Junk has been tracking down where exceptions occur under the umbrella milestone of https://cdcvs.fnal.gov/redmine/issues/17047. Summary of the discussion at the June 27 Coordination Meeting:
- Experiments with support from LArSoft should mount a campaign to eliminate uninitialized variables and FPEs from the code with the goal of enabling FPEs for
- Divide by zero
- NAN where a number is expected
- Over and underflows are not necessarily bugs, so education is needed to handle these properly when these results are possible. Enabling FPEs by default for these was not recommended (??)
Questions to the offline leads: What is the decision from your experiment regarding this plan? Will you participate in trying to eliminate uninitialized variables and FPEs? If so, what should the plan be moving forward? (We assume there would be a period of finding problems followed by attempts to enable exceptions.)
As usual, our current priorities are in the Steering Group document available at: https://indico.fnal.gov/conferenceDisplay.py?confId=14198 If there are other areas that experiments would like to see LArSoft focus on, please let us know.
DUNE – Thomas Junk, Andrew John Norman
As mentioned in the issues discussion, the generative adversarial neural network has been pursued by Dan Smith of Boston University; he gave a talk on its application to LArIAT data at the DUNE Collaboration Meeting in May: http://indico.fnal.gov/getFile.py/access?contribId=49&sessionId=15&resId=0&materialId=slides&confId=12345
Some work has been done to track down Floating-Point Errors in the past two weeks, starting with the collaboration meeting tutorial example with FPE signals turned on.
Jason Stock has been improving the radiological model for the DUNE FD.
ICARUS – Daniele Gibin
LArIAT – Jonathan Asaadi
MicroBooNE – Herbert Greenlee, Wesley Robert Ketchum
ProtoDUNE – Robert Sulej
Continuing the topic raised in DUNE section: Best results of EM/track separation on real data are still obtained without GAN filter. See also Daniel’s presentation on the upcoming LArIAT CM.
What concerns running CNN in the production mode: there is a possibility to ask for GPU equipped nodes on OSG: https://indico.fnal.gov/getFile.py/access?contribId=5&resId=0&materialId=slides&confId=14818 This may be an option to investigate once Tensorflow API is available in UPS.
Refactoring LArG4 (issue 14454) is required for making significant progress on dual-phase input to DUNE TDR.
SBND – Roxanne Guenette, Andrzej Szelc
We are still finalizing the new Geometry idea – expert moving. Once this is done, we have a student starting up that could follow up on whether the bugs are gone in the newest version of ROOT. In the meantime Gianluca with Vito’s help is putting sbndcode in the continuous integration framework – this will be very useful.
We are hoping to launch a larger scale data production chain at which point we will be able to test a larger set of pieces of code.
Open issues from previous meetings:
- From 3/22/17 meeting: Since SBND has been trying to include files inside GDML, have run into problems. Gianluca has been helping debug this. New version of ROOT may be a solution.
- 4/18/17 update: The version of root used by LArSoft has changed a few times over the last few weeks. Andrzej Szelc said they haven’t looked into this because they were focusing on getting the basic geometry in. Once that is done, will look at this and hope the new ROOT version may have fixed it. So, no ticket has been written yet (to ROOT or LArSoft) about this as waiting on whether it is still an issue.
- 5/25/17 update – still waiting.
- 6/14/17 update – Once SBND gets the latest version of ROOT in, they’ll see if this fixes the issue.
- 7/27/17 update – SBND is still finalizing the new Geometry idea – expert moving. Once this is done, we have a student starting up that could follow up on whether the bugs are gone in the newest version of ROOT. In the meantime Gianluca with Vito’s help is putting sbndcode in the continuous integration framework – this will be very useful.
2. From May: Feedback at a recent SBN analysis meeting on proposed restructuring of G4 simulation step was that these were potentially very important changes. Is it possible to see all of this in place this summer for ICARUS large-scale processing and potentially MicroBooNE processing campaign? Along with the work itself, how one updates/does the backtracking remains an unanswered question: not technically difficult, but could imply significant changes in downstream code based on how it is done. This may take another or a few more people working on it to really see it through.
- The project and schedule are being tracked in issue 14454, https://cdcvs.fnal.gov/redmine/issues/14454
3. From 6/14/17 meeting: MicroBooNE asked about how to give credit to authors, experiments and institutions for LArSoft code written.
- 7/27/17 update –
- Have updated the recommendation on documenting code at http://larsoft.org/important-concepts-in-larsoft/design/.
- We committed to presenting a proposed documentation template for header files at a LArSoft Coordination Meeting. (This was done on 6/27/17).
- LArSoft has a place to give author credit for algorithms and services on larsoft.org – as well as in the code itself. This is not being done enough, so suggest a multi-prong education campaign to encourage people to sign their name on the code they write and to contribute to the http://larsoft.org/doc-algorithms/ page. This will be highlighted in a LArSoft Coordinate Meeting with follow up but this needs to be done by all experiments, not just LArSoft.
- Presented proposal at June 27 Coordination Meeting. The action items from that discussion
– Add institution and experiment to the template
– Publish the template on a web page
- What are the plans from the experiments to push this practice out to the code?
- Recommend adding template documentation whenever code is updated?
- Make it the policy of the experiments
4. CLOSED From 4/19/17 meeting: Presentation from student on CNN and adversarial network? According to Robert Sulej, it is possible to think of making such machine-learning-based filter for the detector/E-field response simulation, but it is a future work. They are working now on the idea rather to provide data-driven training set for the CNN model preparation. Need time to understand the results to tell what is the limitation of the tool and what isn’t.
- 5/25/17 update – This has been on the schedule for a coordination meeting, and has been postponed at the request of the authors
- 6/14/17 update – No update
- 7/27/17 update – The generative adversarial neural network has been pursued by Dan Smith of Boston University; he gave a talk on its application to LArIAT data at the DUNE Collaboration Meeting in May: http://indico.fnal.gov/getFile.py/access?contribId=49&sessionId=15&resId=0&materialId=slides&confId=12345